0142T 011 dc22
Bioelectrical activity is associated with living excitable tissue. It has been known, owing to
efforts of numerous investigators, that bioelectrical activity is closely related to the mech
anisms and functions of excitable membranes in living organs such as the heart and the
brain. A better understanding of bioelectrical activity, therefore, will lead to a better under
standing of the functions of the heart and the brain as well as the mechanisms underlying
the bioelectric phenomena.
Bioelectrical activity can be better understood through two common approaches. The
first approach is to directly measure bioelectrical activity within the living tissue. A rep
resentative example is the direct measurement using microelectrodes or a microelectrode
array. In this direct measurement approach, important characteristics of bioelectrical activ
ity, such as transmembrane potentials and ionic currents, have been recorded to study the
bioelectricity of living tissue. Recently, direct measurement of bioelectrical activity has also
been made using optical techniques. These electrical and optical techniques have played
an important role in our investigations of the mechanisms of cellular dynamics in the heart
and the brain.
The second approach is to noninvasively study bioelectrical activity by means of mod
eling and imaging. Mathematical and computer models have offered a unique capability of
correlating vast experimental observations and exploring the mechanisms underlying ex
perimental data. Modeling also provides a virtual experimental setting, which enables well
controlled testing of hypothesis and theory. Based on the modeling of bioelectrical activity,
noninvasive imaging approaches have been developed to detect, localize, and image bio
electrical sources that generate clinical measurements such as electrocardiogram(ECG) and
electroencephalogram (EEG). Information obtained from imaging allows for elaboration
of the mechanisms and functions of organ systems such as the heart and the brain.
During the past few decades, significant progress has been made in modeling and
imaging of bioelectrical activity in the heart and the brain. Most literature, however, has
treated these research efforts in parallel. The similarity arises from the biophysical point of
view that membrane excitation in both cardiac cells and neurons can be treated as volume
current sources. The clinical observations of ECG and EEG are the results of volume con
duction of currents within a body volume conductor. The difference among bioelectrical
activity originating from different organ systems is primarily due to the different physio
logical mechanisms underlying the phenomena. From the methodological point of view,
vi Preface
therefore, modeling and imaging of bioelectrical activity can be treated within one theo
retical framework. Although this book focuses on bioelectric activity of the heart and the
brain, the theory, methodology, and stateoftheart research that are presented in this book
should also be applicable to a variety of applications.
The purpose of this book is to provide a stateoftheart coverage of basic principles,
theories , and methods of modeling and imaging of bioelectrical activity with applications to
cardiac and neural electrical activity. It is aimed at serving as a reference book for researchers
working in the field of modeling and imaging of bioelectrical activity, as an introduction
to investigators who are interested in entering the field or acquiring knowledge about the
current state of the field, and as a textbook for graduate students and seniors in a biomedical
engineering, bioengineering, or medical physics curriculum.
The first three chapters deal with the modeling of cellular activity, cell networks,
and whole organ for bioelectrical activity in the heart. Chapter I provides a systematic
review of onecell models and cell network models as applied to cardiac electrophysiology.
It illustrates how modeling can help elucidate the mechanisms of cardiac cells and cell
networks, and increase our understanding of cardiac pathology in threedimension and
whole heart models . Chapter 2 provides a thorough theoretical treatment of the forward
problem of bioelectricity, and in particular electrocardiography. Following a review of the
theoretical basis of equivalent dipole source models and stateoftheart numerical methods
of computing the electrical potential fields, Chapter 2 discusses the applications of forward
theory to whole heart modeling and defibrillation. Chapter 3 reviews important issues in
whole heart modeling and its implementation as well as various applications of whole
heart modeling and simulations of cardiac pathologies. Chapter 3 also illustrates important
clinical applications the modeling approach can offer.
The following two chapters review the theory and methods of inverse imaging with
applications to the heart . Chapter 4 provides a systematic treatment of the methods and
applications of heart surface inverse solutions . Many investigations have been made in
order to inversely estimate and reconstruct potential distribution over the epicardium, or
activation sequence, over the heart surface from body surface electrocardiograms. Progress
has also been made to estimate endocardial surface potentials and activation sequence from
catheter recordings. These approaches and activities are well reviewed in Chapter 4. Chapter
5 reviews the recent development in three dimensional electrocardiography tomographic
imaging. Recent research shows that, by incorporating a priori information into the inverse
solutions, it is possible to estimate threedimensional distributions of electrophysiological
characteristics such as activation time and transmembrane potentials, or equivalent current
dipole distribution. Inparticular, a wholeheartmodel based tomographic imaging approach
is introduced, which illustrates the close relationship between modeling and imaging and
the merits of modelbased imaging .
Chapter 6 deals with a noninvasive body surface mapping technology  surface Lapla
cian mapping. Compared with wellestablished body surface potential mapping, body sur
face Laplacian mapping has received relatively recent attention in its enhanced capability of
identifying and mapping spatially separated multiple activities . This chapter also illustrates
that a noninvasive mapping technique can be applied to imaging of bioelectrical activity
originated from different organ systems, such as the heart and the brain.
The subsequent two chapters treat inverse imaging of the brain from neuromagnetic
and neuroelectric measurements, as well as functional magnetic resonance imaging (fMRI).
Preface vii
Chapter 7 reviews the forward modeling of magnetoencephalogram(MEG), and neuromag
netic source imaging with a focus on spatial filtering approach. Chapter 8 provides a general
review of linear inverse solutions for EEG and MEG, and multimodal imaging inte
grating EEG, MEG and Along with Chapters 4 and 5, these four chapters are intended
to provide a solid foundation in inverse imaging methods as applied to imaging bioelectrical
activity.
Chapter 9 deals with tissue conductivity, an important parameter that is required in
bioelectric inverse solutions. The conductivity parameter is needed in establishing accurate
forward models of the body volume conductor and obtaining accurate inverse solutions
using modelbased inverse imaging. As most inverse solutions are derived from noninvasive
measurements with the assumption of known tissue conductivity distribution, the accuracy
of tissue conductivity is crucial in ensuring accurate and robust imaging of bioelectrical
activity. Chapter 9 systematically addresses this issue for various living tissues.
This book is a collective effort by researchers who specialize in the field of modeling
and imaging of bioelectrical activity. I am very grateful to them for their contributions
during their very busy schedules and their patience during this process. I am indebted to
Aaron Johnson Brian Halm, Shoshana Sternlicht, and Kevin Sequeira of Kluwer Academic
Publisher for their great support during this project. Financial support from the National
Science Foundation, through grants of NSF CAREER Award BES9875344, NSF BES
0218736 and NSF BES020l939, is also greatly appreciated.
We hope this book will provide an intellectual resource for your research and/or edu
cational purpose in the fascinating field of modeling and imaging of bioelectrical activity.
Bin He
Minneapolis
CONTENTS
1 FROM CELLULAR ELECTROPHYSIOLOGY TO
ELECTROCARDIOGRAPHy......................................................... 1
Nitish Thakor, VivekIyer; and Mahesh Shenai
1 Introduction 1
1.1 The Onecell Model 3
1.1.1 Voltage Gating Ion Channel Kinetics (HodgkinHuxleyFormalism) .. 3
1.1.2 Modeling the Cardiac Action Potential.......... .. 7
1.1.3 Modeling Pathologic Action Potentials 10
1.2 Network Models 17
1.2.1 Cellcell Coupling and Linear Cable Theory 17
1.2.2 Multidimensional Networks .. 18
1.2.3 Reconstructionof the Local Extracellular Electrogram(Forward Problem) 20
1.2.4 Modeling Pathology in Cellular Networks 23
1.3 Modeling Pathology in Threedimensional and Whole Heart Models 29
1.3.1 Myocardial Ischemia 31
1.3.2 Preexcitation Studies 31
1.3.3 Hypertrophic Cardiomyopathy 34
1.3.4 Drug Integration in Threedimensional Whole Heart Models 35
1.3.5 Genetic Integration in Threedimensional Whole Heart Models. 35
1.4 Discussion 36
References 38
2 THE FORWARDPROBLEM OF ELECTROCARDIOGRAPHY:
THEORETICAL UNDERPINNINGS AND APPLICATIONS................ 43
Ramesh M. Gulrajani
2.1 Introduction.................................................................................. 43
2.2 Dipole Source Representations 44
2.2.1 Fundamental Equations 44
2.2.2 The Bidomain Myocardium.............................................................. 46
2.3 Torso Geometry Representations 53
2.4 Solution Methodologies for the Forward problem 53
2.4.1 Surface Methods............................................................................ 54
ix
x Contents
2.4.2 VolumeMethods............................................................................ 58
2.4.3 Combination Methods 61
2.5 Applications of the Forward Problem................................................... 61
2.5.1 Computer Heart Models 62
2.5.2 Effects of Torso Conductivity Inhomogeneities 70
2.5.3 Defibrillation................................................................................ 72
2.6 Future Trends 75
References 75
3 WHOLE HEART MODELINGAND COMPUTERSIMULATION 81
Darning Wei
3.1 Introduction 81
3.2 Methodology in 3DWhole Heart Modeling........................................... 82
3.2.1 Hearttorso Geometry Modeling......................................................... 82
3.2.2 Inclusion of Specialized Conduction System 83
3.2.3 Incorporating Rotating Fiber Directions 85
3.2.4 Action Potentials and Electrophysiologic Properties 89
3.2.5 Propagation Models........................................................................ 94
3.2.6 Cardiac Electric Sources and Surface ECG Potentials 100
3.3 Computer Simulations and Applications 103
3.3.1 Simulation of the Normal Electrocardiogram 103
3.3.2 Simulation of STTWavesin Pathologic Conditions 107
3.3.3 Simulation of Myocardial Infarction 108
3.3.4 Simulation of Pace Mapping 110
3.3.5 Spiral WavesA NewHypothesis of VentricularFibrillation 110
3.3.6 Simulation of Antiarrhythmic Drug Effect 110
3.4 Discussion 111
References 114
4 HEART SURFACEELECTROCARDIOGRAPHIC
INVERSE SOLUTIONS 119
FredGreensite
4.1 Introduction 119
4.1.1 The Rationale for Imaging Cardiac Electrical Function 120
4.1.2 A Historical Perspective 120
4.1.3 Notation and Conventions 123
4.2 The Basic Model and Source Formulations 123
4.3 Heart Surface Inverse Problems Methodology 128
4.3.1 Solution Nonuniqueness and Instability 129
4.3.2 Linear Estimation and Regularization 132
4.3.3 Stochastic Processes and Time Series ofInverse Problems 135
4.4 Epicardial Potential Imaging 138
4.4.1 Statistical Regularization 138
4.4.2 Tikhonov Regularizationand Its Modifications 139
4.4.3 Truncation Schemes 141
Contents xi
4.4.4 Specific Constraints in Regularization 142
4.4.5 Nonlinear Regularization Methodology 143
4.4.6 An Augmented Source Formulation 143
4.4.7 Different Methods for Regularization Parameter Selection 143
4.4.8 The Body Surface Laplacian Approach 144
4.4.9 Spatiotemporal Regularization 145
4.4.10 Recent in Vitro and in Vivo Work 146
4.5 Endocardial Potential Imaging 147
4.6 Imaging Features of the Action Potential 149
4.6.1 Myocardial Activation Imaging 149
4.6.2 Imaging Other Features of the Action Potential 154
4.7 Discussion 155
References 156
5 THREEDIMENSIONAL ELECTROCARDIOGRAPHIC
TOMOGRAPHIC IMAGING 161
Bin He
5.1 Introduction ' " 161
5.2 ThreeDimensional Myocardial Dipole Source Imaging 163
5.2.1 Equivalent Moving Dipole Model 163
5.2.2 Equivalent Dipole Distribution Model 163
5.2.3 Inverse Estimation of 3D Dipole Distribution 164
5.2.4 Numerical Example of 3D Myocardial Dipole Source Imaging 165
5.3 ThreeDimensional Myocardial Activation Imaging 167
5.3.1 Outline of the HeartModel based 3D Activation Time Imaging Approach 167
5.3.2 Computer Heart Excitation Model 168
5.3.3 Preliminary Classification System 169
5.3.4 Nonlinear Optimization System 170
5.3.5 Computer Simulation 171
5.3.6 Discussion 174
5.4 ThreeDimensional Myocardial Transmembrane Potential Imaging 175
5.5 Discussion 178
References 180
6 BODY SURFACE LAPLACIAN MAPPING OF
BIOELECTRIC SOURCES 183
Bin He and lie Lian
6.1 Introduction 183
6.1.1 Highresolution ECG and EEG 183
6.1.2 Biophysical Background of the Surface Laplacian 184
6.2 Surface Laplacian Estimation Techniques 186
6.2.1 Local Laplacian Estimates 186
6.2.2 Global Laplacian Estimates 188
6.2.3 Surface Laplacian Based Inverse Problem 190
6.3 Surface Laplacian Imaging of Heart Electrical Activity 192
xii Contents
6.3.1 Highresolution Laplacian ECG Mapping 192
6.3.2 Performance Evaluation of the Spline Laplacian ECG 193
6.3.3 Surface Laplacian Based Epicardial Inverse Problem 199
6.4 Surface Laplacian Imaging of Brain Electrical Activity 200
6.4.1 Highresolution Laplacian EEG Mapping 200
6.4.2 Performance Evaluation of the Spline Laplacian EEG 200
6.4.3 Surface Laplacian Based Cortical Imaging 206
6.5 Discussion 208
References 209
7 NEUROMAGNETICSOURCE RECONSTRUCTIONAND
INVERSE MODELING 213
Kensuke Sekihara and Srikantan S. Nagarajan
7.1 Introduction 213
7.2 Brief Summary of Neuromagnetometer Hardware 214
7.3 Forward Modeling 215
7.3.1 Definitions 215
7.3.2 Estimation of the Sensor Lead Field 216
7.3.3 Lowrank Signals and Their Properties 219
7.4 Spatial Filter Formulation and Nonadaptive Spatial
Filter Techniques 221
7.4.1 Spatial Filter Formulation 221
7.4.2 Resolution Kernel 222
7.4.3 Nonadaptive Spatial Filter 222
7.4.4 Noise Gain and Weight Normalization 225
7.5 Adaptive Spatial Filter Techniques 226
7.5.1 Scalar Minimumvariancebased Beamformer Techniques 226
7.5.2 Extension to Eigenspaceprojection Beamformer 227
7.5.3 Comparison between Minimumvariance and Eigenspace
Beamformer Techniques 228
7.5.4 Vectortype Adaptive Spatial Filter 230
7.6 Numerical Experiments: Resolution Kernel Comparison between
Adaptive and Nonadaptive Spatial Filters 232
7.6.1 Resolution Kernel for the Minimumnorm Spatial Filter 232
7.6.2 Resolution Kernel for the Minimumvariance Adaptive
Spatial Filter 234
7.7 Numerical Experiments: Evaluation of Adaptive
Beamformer Performance 235
7.7.1 Data Generation and Reconstruction Condition 235
7.7.2 Results from Minimumvariance Vector Beamformer 238
7.7.3 Results from the Vectorextended BorgiottiKaplan Beamformer 238
7.7.4 Results from the Eigenspace Projected VectorextendedBorgiottiKaplan
Beamformer 238
7.8 Application of Adaptive Spatial Filter Technique to MEG Data 243
7.8.1 Application to Auditorysomatosensory Combined Response 243
Contents xiii
7.8.2 Application to Somatosensory Response: Highresolution
Imaging Experiments 245
References 247
8 MULTIMODALIMAGINGFROM NEUROELECTROMAGNETIC
AND FUNCTIONAL MAGNETIC RESONANCE RECORDINGS 251
Fabio Babiloni and FeboCincotti
8.1 Introduction 251
8.2 Generalities on Functional Magnetic Resonance Imaging 252
8.2.1 Blockdesign and EventRelated tMRI 254
8.3 Inverse Techniques 254
8.3.1 Acquisition of Volume Conductor Geometry 255
8.3.2 Dipole Localization Techniques 256
8.3.3 Cortical Imaging 257
8.3.4 Distributed Linear Inverse Estimation 259
8.4 Multimodal Integration of EEG, MEG and tMRI Data 261
8.4.1 Visible and Invisible Sources 261
8.4.2 Experimental Design and Coregistration Issues 262
8.4.3 Integration of EEG and MEG Data 263
8.4.4 Functional Hemodynamic Coupling and Inverse Estimation of
Source Activity 267
8.5 Discussion 275
References 276
9 THE ELECTRICAL CONDUCTIVITY OF LIVING TISSUE: A
PARAMETERIN THE BIOELECTRICAL INVERSE PROBLEM 281
Maria J. Peters, Jeroen G. Stinstra, and IbolyaLeveles
9.1 Introduction 281
9.1.1 Scope of this Chapter 282
9.1.2 Ambiguity of the Effective Conductivity 283
9.1.3 Measuring the Effective Conductivity 284
9.1.4 Temperature Dependence 287
9.1.5 Frequency Dependence 287
9.2 Models of Human Tissue 289
9.2.1 Composites of Human Tissue 289
9.2.2 Conductivities of Composites of Human Tissue 292
9.2.3 Maxwell's Mixture Equation 296
9.2.4 Archie's Law 300
9.3 Layered Structures 307
9.3.1 The Scalp 307
9.3.2 The Skull 308
9.3.3 A Layer of Skeletal Muscle 310
9.4 Compartments 311
9.4.1 Using Implanted Electrodes 311
9.4.2 Combining Measurements of the Potential and the Magnetic Field 312
xiv Contents
9.4.3 Estimation of the Equivalent Conductivity using Impedance Tomography 312
9.5 Upper and Lower Bounds 313
9.5.1 White Matter 314
9.5.2 The Fetus 314
9.6 Discussion 316
References 316
INDEX 321
FROMCELLULAR
ELECTROPHYSIOLOGY TO
ELECTROCARDIOGRAPHY
by Nitish V. Thakor, Vivek Iyer, and Mahesh Shenai
Department of Biomedical Engineering, The Johns Hopkins University, 720 Rutland Ave.,
Baltimore MD 21205
INTRODUCTION
Since many cardiac pathologies manifest themselves at the cellular and molecular levels,
extrapolation to clinical variables, such as the electrocardiogram (ECG), would prove in
valuable to diagnosis and treatment. One ultimate goal of the cardiac modeler is to integrate
cellular level detail with quantitative properties of the ECG (a property of the whole heart).
This magnificent task is not unlike a forest ranger attempting to document each leaf in a
massive forest. Both the modeler and ranger need to place fundamental elements in the
context of a broader landscape. But now, with the recent genome explosion, the modeler
needs to examine the "leaves" at even much greater molecular detail. Fortunately, the rapid
explosion in computational power allows the modeler to span the details of each molecular
"leaf" to the "forest" of the whole heart. Thus, cardiac modeling is beginning to span the
spectrum from DNA to the ECG, from nucleotide to bedside.
Extending cellular detail to wholeheart electrocardiography requires spanning several
levels of analysis (Figure 1.1). The onecell model describes an action potential record
ing from a single cardiac myocyte. By connecting an array of these individual myocytes
(via gap junctions), a linear network (cable), twodimensional (20) network or three
dimensional (3D) network (slab) model of action potential propagation can be constructed.
The bulk electrophysiological signal recorded from these networks is called the local ex
tracellular electrogram. Subsequently, networks representing tissue diversity and realistic
heart geometries can be molded into a whole heart model, and finally, the whole heart
model can be placed in a torso model replicating lung, cartilage, bone and dermis. At each
level, one can reconstruct the salient electric signal (action potential, electrogram, ECG)
from the cardiac sources by solving the forward problem of electrophysiology (Chapter 2).
Simply put, cardiac modeling is equivalent to solving a system of nonlinear differential
(or partial differential) equations, though vigorous reference must be made to numerous
2 N. V. Thakor, V. Iyer, and M. B. Shenai
(lD, 2D)
~
·SG
·100  f   ~   r       ,
~
.a
FIGURE 1.1. Levels of Analysis. Onecell models include the study of compartments and ion channels and
their interactions. The basic electrophysiological recording is the action potential. Network models investigate the
connectivity of onecell units organized in arrays. An electrical measure of bulk network activity is the extracellular
electrogram. Finally, many patches molded into the shape of a whole heart (in addition to torso variables) gives
rise to the ECG. See the attached CD for color figure.
laboratory experiments which aim to determine the nature and coefficients of each equation.
These equations provide a quantitative measure of each channel, each cell, and networks of
cells. As more experiments are done and data obtained, the model can be made more complex
by adding appropriate differential equations to the system. Thus, as more information
about the cellular networks, tissue structure, heart and torso anatomy are obtained, a better
reconstruction of the ECGbecomes possible. Until recently, however, modeling efforts have
primarily focused on accurately reconstructing normal behavior. But with the accumulating
experimental history of cardiac disease (such as myocardial ischemia, longQT syndrome
and heart failure), modelers have also begun to revise and extend the quantitative description
of these models to include important abnormal behaviors.
This chapter will first focus on the theoretical onecell equations, which are only solved
in the time domain. Subsequently, the onecell model will be expanded to represent multiple
dimensions with the incorporation of partial differential equations in space. At each level
of analysis, the appropriate electrical reconstruction is discussed in the context of relevant
pathology to emphasize the usefulness of cardiac modeling.
From Cellular Electrophysiology to Electrocardiography
1.1 THE ONECELL MODEL
3
The origins of the onecell model actually take root from classical neuroscience work
conducted by A.L. Hodgkin and A.F. Huxley in 1952 (Hodgkin and Huxley 1952). In famous
experiments conducted on the giant axon of the squid, they were able to derive a quantitative
description for current flowacross the cell membrane, and the resulting action potential (AP).
This model mathematically formulated the voltagedependent "gating" characteristics of
sodium and potassium ion channels in the nerve membrane. Since similar ion channels exist
in cardiac cells, this HodgkinHuxley formalism was applied to model the Purkinje fiber
action potential by McCallister, Noble and Tsien (McAllister et al.
However, it was determined that the cardiac action potential is considerably more
complex than the neuronal action potential, presumably due to a larger diversity of ion chan
nels present in the cardiac myocyte, the intercellular connections, and its coupling to mus
cular contraction. With the addition of the "slowinward" calcium current in 1976, Beeler
and Reuter (Beeler and Reuter 1976) were able to successfully describe the ventricular action
potential with the characteristic "plateau phase" necessary for proper cardiac contraction.
Since then, numerous ion channels and intracellular calcium compartment dynamics have
been added (DiFrancesco and Noble 1985; Luo and Rudy 1991; Luo and Rudy 1994),
making the current AP model considerably more complex and robust. Nevertheless, many
of these membrane channels still follow the same HodgkinHuxley formalism, reviewed
below for the cardiac myocyte. In addition, the cardiac myocyte contains a prominent
intracellular calcium compartmentthe sarcoplasmic reticulum.
1.1.1 VOLTAGE GATING IONCHANNEL KINETICS (HODGKINHUXLEY
FORMAUSM)
At the most fundamental level of electrophysiology, an ion (K+, Na", Ca2+) must
cross the membrane via the transmembrane ion channel. Typically, the ion channel is a
multidomain transmembrane protein with "gates" that open and close at certain transmem
brane voltages, V
m
V
in
 V
out
). The problem, however, is to characterize the opening and
closing of these gates, a process symbolically represented by the following equation:
(1.1)
where k
1
and
1
are the forward and reverse rates of the process, respectively, and n
open
and nclosed are the percentage of open or closed channels (which is proportional to channel
"concentration"). Thus, by simple rate theory, one would expect the rate of channel opening
(dn/dr) to equal (note that nclosed 1  n
open):
(1.2)
The voltage dependence of these ion channels can be understood if these gates are
treated as an "energybarrier" model, described with Eyring Rate Theory (Eyring et al.
1949; Moore and Pearson 1981). Given the concentration of the charged particle on the
inside and outside ([Cil, [Co]), an energy barrier
o
) located at a relative barrier position
4
V
1l1
N. V. Thakor, V. Iyer, and M. B. Shenai
Extracel lular
K+ Na+
FIGURE 1.2. A BatteryResistorCapacitor model of a generic excitable membrane. Ions flow (current) to and
from the extra and intracellular domains. across a resistor (or conductance). The membrane has an inherent
capacitance, due to its chargeseparating function. The current relates to a transmembrane voltage, V
(8) along the transmembrane route, and a transmembrane voltage (Vm), Eyring Rate Theory
predicts the forward and reverse rates for ion transfer as:
(
_ ~ ) ( ( ). )t FVm )
k = K · e RT • e RT
(
"GO) ( ~ )
k_ = K· e7/T . e RT
(1.3)
where K is a constant, R is the gas constant, T is the absolute temperature, and is the
valence of the ion. While, the solution in Eq. ( 1.3) is an extremely simplified version of
reality, it readily suggests that the forward and reverse rates are voltagedependent (thus
these rates can be represented as and L)
While the "energybarrier" model predicts voltagedependence, it does not account for
the timevarying features in opening and closing channels. model that takes timevariance
into account was developed by Hodgkin and Huxley in 1952 (Hodgkin and Huxley 1952).
The Hodgkin and Huxley model likens the biological membrane to a BatteryResistor
Capacitor (BRC model, Figure 1.2) circuit. The resistor (1/conductance) represents the ion
channel, through which ions pass to create an ionic current (lion). Since the membrane
confines a large amount of negativelycharged protein within the cell, it separates positively
and negatively charged compartments, thus acting as a capacitor
m
). Finally, as ions cross
the membrane and enter (or leave) the intracellular compartment, electrical repellant charge
begins to build that counteracts V
m
. The at which a certain ion is at equilibrium (lion
is termed the Nemst potential
ion
), the "battery" which depends on valence, intracellular
[C] i and extracellular [C] o ion concentrations:
_  RT ([Clo)
E
ion

zF [Cli
(1.4)
Thus , from simple circuit analysi s of Figure 1.2, the ionic current for a certain ion can be
From Cellular Electrophysiology to Electrocardiography
written as:
5
(1.5)
Where g(V, t) is the voltagedependent, timevarying ion channel conductance. To deter
mine the dynamics of an individual ion channel, Hodgkin and Huxley assumed that the
channel was a "gate" as described in Eq. (1.2), which can be rewritten solely in terms of
open probability n
open
or simply, n (the forward and reverse rates, and are
replaced with a(V) and {3(V), respectively) :
dn(t, V)
dt {3(V)[l  n]  a(V)[n] (1.6)
Eq. (1.6) is a firstorder differential equation, which has a particular solution under several
boundary conditions. Following a voltage step LlV(Vm V
rest
LlV) from the resting
membrane potential, n(t) follows an inverted exponential time course with the following
characteristics:
r(v' )
n m  a(Vm) {3(Vm)
(1.7)
The quantity of noo(Vm) represents the steadystate proportion of open channels after a
step voltage has been applied for a nearinfinite amount of time. The variable roo(V
m)
characterizes the time the system takes to reach this noo(Vm). Rewriting Eq. (1.6) in terms
of the quantities derived in Eq. (1.7), gives a differential equation that describes the time
course of the open probability for a channel:
dn noo(V
m)
 n
dt r(Vm)
(1.8)
Using an elegant experimental setup that applied a voltageclamp to a giant squid axon
(Cole 1949; Marmont 1949), Hodgkin and Huxley were able to define regression equations
for noo(V) and rm(V), which represent the gating variables for the potassium channel. To
obtain a suitable fit to experimental data, they arrived at the open channel probability of
n(V, t)4 .Thus, by substituting the open probability into Eq. (1.5), the outward potassium
current can be represented as:
dV 4
l« C n(V, t) (V  E
dt
(1.9)
An analogous equation can be written for the inward sodiumcurrent with the addition of
an inactivation mechanism (Figure 1.3). Following the data fitting, the experimental sodium
channel was represented by Hodgkin and Huxley as three voltageactivated gates similar
to the potassium activation gates described by Eq. (1.8). As with the potassium channel,
increased membrane voltages stochastically increase the probability that these three gates
open. Inactivation follows the same kinetics as Eq. (1.8), except that the inactivation gate
closes with increased voltages (Figure lAc). Thus, the sodium response to an applied
voltage stimulation is biphasic. First, the faster activation gates rapidly open, allowing
K+
Open PrOb3bility
Na+
Open Prob 3bility
Probabi lity aJJ
gates are open
FIGURE 1.3. Idealized ion channels. The potassium channel is generally modeled with four voltageactivation
gates. The sodium channel is represented by three rapidlyactivating voltagesensitive gates, with an additional
slowly acting voltagesenstive inactivation gate. The lumped probability that all potassium gates will be open is
n", while the probability that the activation and inactivation gates of the sodium channel is m'h.
A) I1aJ(V)
09
sc
Volts (mV)
B)
0 9
.c;..
06
....
~
~
0 7·
~
06
~
0$
I::l.,
04
~ ! 03
02
01
.?oo
Volts (mV)
C ) ~ ( V )
09·
~
~
~ 07,
~
~
06'
C
0$·
I::l., 0 4 '
~ 03
C l 02'
0 \ '
so .?oo eo se
Volts (mV)
FIGURE 1.4. Activation curves for (A) potassium channels, (B) activation curve for sodium channel, and
(C) inactivation curve for sodium channel,
From Cellular Electrophysiology to Electrocardiography
inward current to develop. However, with increased voltage, the slower inactivation gates
will close, forcing a decrease in the inward current. There is no conceptual change in the
nature of the current equationthe activation gate n is simply replaced with m and h (though
these gates all differ quantitatively, m and n both increase with more positive V while the
value of h decreases with more positive Vm). The sodium current can be represented as:
(1.10)
The biphasic nature of the inward sodium current is crucial to the rapid elicitation of an
action potential and the characteristic biphasic shape of the action potential.
This simplified approach assumes that the cell membrane contains two distinct types
of voltagegated channels (Na+ and K+) that conducting currents in the opposite direction.
With the addition of other inward and outward channels (see later sections), a generalized
differential equation can be written:
dV
dt
M
INa Iotherchannels I stim)
(1.11)
where I
stim
represents a stimulation current (provided from a stimulating lead or adjacent
cells), and Iotherchannels is provided via many other channels that vary among celltypes
(atrial vs. ventricular cells) and various excitable tissues (heart vs. nervous system). Note
that l«, INa, and other channels are represented by nonlinear terms (i.e. n
4
and m
3
and
are both voltage and timedependent. Thus, Eq. (1.11) coupled with gating equations for
each channel (Eq. (1.8)), represents a system of nonlinear differential equations that must
be solved using techniques of numerical integration.
1.1.2 MODELING THE CARDIAC ACTIONPOTENTIAL
While the model of an action potential was originally described for a neuron, the
methods were quickly adapted to represent the cardiac action potential. Although there
are slight differences in the quantitative description of the sodium and potassium channels
described above, the cardiac myocyte also exhibits a considerable inward calcium current
that is responsible for the distinguishable "plateau" phasewhich coincides with the mus
cular contraction in the ventricular myocyte. Additionally, the cardiac myocyte uniquely
expresses a diverse set of ion channelswhich give unique electrophysiological properties
to different types of heart tissue, in normal and diseased heart function.
Within the heart, there exist a variety of cell types that require different considerations
when developing a model. Pacemaker cells in the sinoatrial node express channels that
allow an autonomous train of action potentials, while Purkinje fibers represents an efficient
conducting system specialized for the fast, uniform excitation of the ventricular myocytes.
Ventricular myocytes express the proper proteome to parlay the electrical excitation into
force generating elements that ultimately produce the cardiac output and blood delivery to
the rest of the body. Even within the ventricle, different models exist for transmural orien
tation (endocardial cells, middlemyocardial cells (Mcells), and epicardial cells). Models
for each type of these cells have been extensively developed and are described in Table 1.1,
and the history of these modeling developments is described below.
8 N. V. Thakor, V. Iyer, and M. B. Shena!
TABLE1.1. Classical and Modern Models of Various Cardiac Cell Types
Classical Models
HodgkinHuxley (1952)
McCallister, Noble, Tsien (1974)
BeelerReuter (1977)
Modem Models
DiFrancescoNoble (1985)
LuoRudy Phase I (1991)
LuoRudy Phase II (1994)
PriebeBeuckelmann (1998)
Zhang et al.
Type
Squid Axon
Purkinje Cell
Ventricular Cell
Purkinje Cell
Ventricular Cell
Ventricular Cell
Human Ventricular Cell
Sinoatrial Nodal cells
Novelty
INa,IK
Ix!,IK2
lSi (slowinward Ica)
INaCa, INaK, ICaL, IcaT
Updated INa, IK
Updated INaCa, INaK, IcaL, IcaT;
Cabuffering
Updated with human data
Updated Ca handling
1.1.2.1 Classical modelsofthe cardiac actionpotential
In 1975, McCallister, Noble and Tsien introduced a prototype numeric model for the
rhythmic "pacemaker activity" of cardiac Purkinje cells by using the voltageclamp method
to study an outward potassium current, I
K2
(McAllister et al. 1975). After repolarization
of the action potential, the deactivation of outward I current allows a net inward current
to produce a diastolic slow wave of depolarization in between action potentials (Figure
1.5). As this slow wave of depolarization brings membrane potential towards threshold, I
K2
is a prominent current in producing the automaticity of pacemaker cells. Additionally, the
McCallister, Noble and Tsien (MNT) model reconstructed the entire action potential, using
a modified HodgkinHuxley sodium conductance for the rapid upstroke phases, while using
voltageclamp methods to describe an lXI, a generalized plateau and repolarization current.
Thus, this landmark model was able to simultaneously describe characteristic pacemaker
activity and rapid conduction velocities associated with Purkinje cells.
However, given the vast diversity of cardiac cell types, the MNT model could not
describe the characteristics of ventricular action potentialsnamely, the prominent plateau
phase that is crucial for forceful contraction. To this end, Beeler and Reuter developed a
numerical model (the BR model) for the ventricular myocyte in 1977 (Beeler and Reuter
1976). This model incorporates an Is component, a slow inward calcium current that is
responsible for the slow depolarization and the prominent plateau phase. This Is current
follows HodgkinHuxley formalism, in that state variables d (activation) and (inactiva
tion) describe timevarying conductances of the slow inward current. However, unlike other
HodgkinHuxley ions, the initial low level of intracellular calcium, [Ca2+]j does not remain
constant with the arrival of the transmembrane Is current. In fact, the range of [Ca2+]i can
range from 1 to 10 M, widely altering the Nemst potential, E
s.
Thus, Beeler and Reuter
modeled the intracellular handling of calcium by assuming it flows into the cell and accu
mulates while being exponentially reduced by an uptake mechanism (in the sarcoplasmic
reticulum). At any given state, the flux of [Ca2+]j can be described by:
_10
7
. Is .07(10
7
 (1.12)
En toto, the model incorporated four major components: the familiar INa current, the
Is calcium current, the timeactivated outward I
Xl
current and I
K1
, a timeindependent
From Cellular Electrophysiology to Electrocardiography 9
McCallister, Noble, Ts ien
(Purkinje Fiber)
Beeler and Reuter
(Ventri cular Fiber)
JOO
(OM<')
'v .......
( 0 __ • __ ........, . __, _ _ ,_ ,
 9)
 x
.. . 8
.

,
t. ou
= ._
. 1 .c .  _: . 
• • •• .. au II:': ' ''
LuoRudy
(membrane)
LuoRudy
(sarcoplasmic reticulum)
.dL JSR
a 1'41 (CSQ. I
I'Ll
" lII)'opIMm
E
..
F
otote I CI "'"
FIGURE 1.5. A comparison of classical (top) and modern models (bottom).
outward potassium current. With this model, Beeler and Reuter began to predict patho
logical phenomena, including determinants of action potential duration, and oscilliatory
behavior in ventricular cells.
1.1.2.2 Modern models of cardiac action potentials
While modem models utilize many of the concepts introduced in the classical mod
els described above, current models now incorporate a larger repertoire of ion channels,
10 N. V. Thakor, V. Iyer, and M. B. Shenai
a richer history of experimentation, and complex intracellular and sarcoplasmic calcium
handling. In addition, improved computational power and numerical techniques can solve
hefty systems of differential equations, allowing a more precise description of cellular elec
trophysiology (onecell) and the interaction of many cells (network models). As a result, the
focus of modeling has shifted from describing normal behavior of myocytes to describing
pathological phenomena.
In 1985, DiFrancesco and Noble described an improved model of the Purkinje action
potential (DN model) (DiFrancesco and Noble 1985), that included the traditional ion
channel formulation, along with improved assumptions on calcium channels (Ltype and
TType) and intracellular calcium handling. Nevertheless, the experimental recording tech
nique at the time was rather limited, and could not account for important arrhythmogenic
phenomena. In 1991, Luo and Rudy published an updated version of the DN model that
included more recent experimental data for the sodium and potassium currents, but omitted
the BR formation of the inward calcium current (lsi), citing a lack of singlechannel and
onecell experimental history (Luo and Rudy 1991). But in 1994, Luo and Rudy published
an updated model which comprehensively updated the DN description of the sarcolemma
Ltype Calcium channel (lea.d, the sarcolemma Na+ICa
H
exchanger, the sarcolemma
Na/K pump, the sarcoplasmic CaATPase, and Ca induced Ca release. Processes not
described in the DN model were also added, such as the buffering of Ca in the my
oplasm, and a nonspecific Calcium current (Luo and Rudy 1994). The model consists of
three compartmentsthe myoplasm, network sarcoplasmic reticulum, and the junctional
sarcoplasmic reticulum. This enhanced model has provided a breakthrough in simulations
of excitationcontraction (EC) coupling and reentrant mechanisms of arrhythmogenesis.
In 1998, the LuoRudy model was updated by substituting animal data in favor of recent
human data (Priebe and Beuckelmann 1998).
While the LuoRudy model describes ventricular action potentials, several other mod
els exist for other cardiac tissues. Recently, Zhang et at. have incorporated recent sinoa
trial data to formulate a modem model of various sinoatrial nodal cells (central nodal and
peripheral nodal cells) (Zhang et at. 2000). Lindblad et at. have used existing biophysi
cal data to simulate a family of action potentials recorded in rabbit atria (Lindblad et at.
1996).
1.1.3 MODELING PATHOLOGICACTIONPOTENTIALS
Currently, there is a comprehensive understanding of basic ionic mechanisms and
their behavior in normal cardiac cells. The various cardiac models listed in Table 1.1 have
widely contributed to this theoretical understanding. However, less is accepted about how
impairments of these ionic mechanisms ultimately predict or provoke gross events, such as
infarction and/or arrhythmogenesis. Among many others, two areas of cardiac pathology,
myocardial ischemia and longQ'I' syndromes (LQTS), are now the focus of intense model
ing research. These studies have contributed not only to a theoretical understanding of the
diseases, but also to electrocardiographic detection and appropriate pharmaceutical inter
vention. Though both myocardial ischemia and longQ'I' syndromes can lead to fatal arrhyth
mias (Wit and Janse 1993; ElSherif et al. 1996), myocardial ischemia does so by shortening
the action potential duration (APD) while LQTS induces arrhythmias by lengthening the
APD.
From Cellular Electrophysiology to Electrocardiography
Acidosis
2. Plateau EADs
Phase3 EADs
2. DADs
Acidosis
2. LQT3
3. Rotors
Phase3 EADs
2. DADs
FIGURE 1.6. Cellular phenomena associated with myocyte ion channel currents. Various ion channels have been
implicated in pathologic phenomena.
To study impaired cells, one must modify existing models of normal behavior. These
modifications may be achieved by: adding novel channels to the existing repertoire of
known membrane channels; (2) altering the quantitative dynamics of known channelsfor
example by altering ionic concentrations or pH; or (3) a combination of new channels and
altered channel dynamics. Figure 1.6 summarizes the various cellular phenomena associated
with myocyte ion channels.
1.1.3.1 Myocardial ischemia
Myocardial ischemia results from a withdrawal of oxygen from myocardial tissue (due
to inefficient or absent perfusion), resulting in disturbances to aerobic respiration and ATP
production. Alterations in intracellular ATP ([ATP]i), can alter the activity of membrane
pumps, and thus the distribution of critical ions (Na+ and K+) that are largely responsible
for the electrophysiological characteristics of myocardiumand proper action potential prop
agation. Thus, ischemia develops at the cellular level, when the amount of oxygen (P
oz
) in
12 N. V. Thakor, V. Iyer, and M. B. Shena!
the vicinity of the mitochondria fails to meet the demand of rephosphorylation in the Kreb's
cycle (Factor and Bache 1998). Myocardial ischemia has at least four cellular sequellae: (1)
hyperkalemia, or an increase in extracellular potassium [K+]o; (2) acidosis , or a decrease in
cellmedium pH (intracellular) or interstitial space pH (extracellular); (3) anoxia, or oxygen
withdrawal that results in a decrease in [ATP]i; and (4) decoupling of cells . The effects of
these individual manifestations on excitability have been widely reported, experimentally
(Kagiyama et al. 1982; Kodama et al. 1984; Kleber et al. 1986; Weiss et al. 1992; Yan et al.
1993) and theoretically (Ferrero et al. 1996; Shaw and Rudy 1997; Shaw and Rudy 1997).
Hyperkalemia
As the intracellular stores of ATP diminish due to reduced aerobic respiration, Na+/K+
pumps responsible for ion distribution also demonstrate reduced activity. Though normally
this pump acts to relocate sodium out of the cell and potassium into the cell, a lethargic
pump performs this process inefficiently. Thus, there is an extracellular accumulation of
potassium, referred to as "hyperkalemia". The electrophysiological consequences of hyper
kalemia are twofold. First, the upstroke velocity (dV/dt
max
) of the action potential can be
diminished. With the increased extracellular potassium, the resting membrane potential
(RMP) becomes more positive, increasing sodium channel inactivation and reducing the
inward sodium current (Weidmann 1955; Morena et al. 1980). This dominating effect is
somewhat mitigated by the increased RMP being closer to the action potential threshold.
Thus, moderate increases (5.4 mmol to 7.5 mmol) in potassium (c 7.5 mmol) may actually
increase upstroke velocity (this is termed "superconduction"), while large increases in ex
tracellular potassium begin to inactivate the sodium current and decrease upstroke velocity.
Even larger increases can prevent the upstroke entirely and produce conduction block (Wit
and lanse 1993; Cascio et al. 1995). Hyperkalemia can also significantly decrease the APD.
This effect is due to exaggerated outward potassium current late in the action potential that
is able to overcome the inward calcium current relatively earlier, reducing the APD (Figure
1.7). Both effects of hyperkalemia, APD shortening and conduction depression, have been
successfully modeled by Shaw and Rudy (Shaw and Rudy 1997; Shaw and Rudy 1997).
41.7
o
65
 100
o 120
FIGURE 1.7. Action potenti al simulations with varying degrees of [K+]o. Increasing extracellular potassium
(hyperkalemia) results in decreasing APD. (From Shaw and Rudy 1997; used by permission)
From Cellular Electrophysiology to Electrocardiography
Acidosis
13
In the absence of aerobic respiration, alternate pathways that attempt to maintain en
ergy production result in the formation of acidic species , thus initially creating intracellular
acidosis. An increase in the intracellular proton concentration leads to proton extrusion
into the extracellular spaceresulting in extacellular acidosis. Changes in acidity can sub
tly change threedimensional protein structures, including ion channels embedded in the
sarcolemma. Most notably, the sodium channel experiences a decrease in maximum con
ductance with extracellular acidosis . Intracellular acidosis reduces the availability of
the Ltype calcium channel (described below) . These changes considerably affect upstroke
velocity (Shaw and Rudy 1997).
Hypoxia
The accumulation of intracellular ADP (at the expense of intracellular ATP) activates
a special KATP channel in the sarcolemma, described by the following equation:
where !ATP is represented by:
! ATP = ( )H
35.8
(1.13)
(1.14)
where is the Hill coefficient that decreases exponentially with [ADPli From Eq. (14),
a decrease in the [ATPli/[ADPJi leads to an increase in the ! ATP coefficient and the out
ward lx_ATP. This outward potassium current supplements the normal potassium current,
enhancing the total outward current and drastically reducing the APD (Ferrero et al.
(Figure 1.8).
Incidentally, the power of computer modeling was used to settle the controversy sur
rounding APD shortening and the role of the KATP channel. Because experiments showed
FIGURE 1.8. AP simulations with varying degrees of fATP. An increase in the fraction of open KATPchannel s
results in profound APD shortening. (From Ferrero et al . 1996; used permission)
14
A.
B.
N. V. Thakor, V. Iyer, and M. B. Shenai
...
FIGURE 1.9. Classifications of afterdepolarizations: (A) Plateau EAD an oscillation during the Phase 2 plateau;
(B) Phase 3 EAD; and (C) DAD an oscillation after complete repolarization.
that anoxia induced a 4060% shortening of the APD while KATP channels demonstrated
only a 1% activation, many investigators felt that the KATP channel was not a major
conducive factor to APD shortening. However, several investigators (Ferrero et al.
Shaw and Rudy 1997) were able to quantitatively model the KATP channel with conduc
tance [srr. being dependent on the amount of intracellular ATP. By adding this individual
channel to the model, they were able to show that even a .4% channel activation can actually
shorten the APD by 50%. Thus, this channel has been implicated as the major factor in APD
shortening and thus may be a crucial factor in arrhythmogenesis.
1.1.3.2 Early afterdepolarizations (EADs) and delayed afterdepolarizations (DADs)
Early afterdepolarizations (EADs) and delayed afterdepolarizations (DADs) (Figure
1.9) are singlecell arrhythmogenic triggering events, typically depending on Ca2+ alter
ations and the interactions between the intracellular and sarcoplasmic compartments within
the myocyte (Marban et al. 1986; Priori and Corr 1990). Because of the dependence on
intracellular calcium, which can accumulate or depreciate from beattobeat, mulitiple beat
models (paced at a basic cycle length) are required to reach a steady state. Simply stated,
From Cellular Electrophysiology to Electrocardiography 15
afterdepolarizations are notches of depolarization that occur after the typical action potential
upstroke. By definition, the EADs occur before the completion of repolarization, whereas
DADs occur after the completion of repolarization.
EADs may occur during the plateauphase (Figure 1.9a) of the action potential (plateau
EADs) (Marban et al. 1986; Priori and Corr 1990) or during the phase3 repolarization
downstroke of the action potential (phase 3 EADs, Figure 1.9b). The plateau EAD is highly
dependent on the Ltype Ca2+ current (also involved in acidosis) (January and Riddle
1989), which is a nonspecific cation channel permeable to Ca Na+, and K+. Briefly, the
formation of this current is the sum of ICa, ICa,K, ICa,Na, each of which are modulated by a
[Ca factor (Luo and Rudy 1994; Luo and Rudy 1994):
fca==
[
2+.]2
rCa ]1
Km,ca
(1.15)
where Km,ca is a halfmaximal constant, equivalent to .6 p.mol/L, As the intracellular
calcium concentration increases.ji, and the Ltype current decrease monotonically. Addi
tionally, the channel is controlled by voltage dependent fgate. During the plateau phase,
when intracellular Ca2+ is elevated, the Ltype Ca2+ channel is relatively inactive due to
a low However, due to rapid intracellular Ca2+ recovery (a phenomenon associated
with longduration action potentialsj.ji, and inward I
Ca
are elevated, resulting in a net
depolarization during an otherwise repolarizing phase.
Unlike the plateau EAD, phase3 EADs and DADs (Figure 1.9c) are dependent on
Na+Ca2+ exchanger and Ins(Ca), the nonspecific calcium current. Like the ICa(L) current,
the Ins(Ca) is permeable to K+ and Na"however, an increase in [Ca2+]i increases the
Ins(Ca). Thus, spontaneous Ca2+release by the SR into the intracellular compartment further
increases the inward current, producing either DADs or EADs (Stem et al. 1988).
Both EADs and DADs produce links between cellular conditions and arrythmogenesis.
For example, simulation studies have reproduced experimental studies demonstrating that
EADs can generate ectopic activity (Saiz et al. 1996). In addition to onecell studies,
afterdepolarizations are studied in the context of linear networks (see Section 1.2) EADs
have also been implicated in the longQT syndrome, as the triggering event to a specific
type of polymorphic reentrant tachycardia, or Torsades de Pointes (TdP) (ElSherif and
Turitto 1999; Viswanathan and Rudy 1999).
1.1.3.3 LongQT syndrome
While myocardial ischemia results in APD shortening, other myocardial pathologies
such as LongQT syndrome may result in APD lengthening. The etiologies of LQTS are
diverse, ranging from various genetic deficiencies at distinct loci, to acquired and iatrogenic
causes. LongQT syndrome is characterized by a prolongation of the QTinterval in the
ECG, presumably due to structurallydeformed potassium and sodium channels. Impaired
outward potassiumflowwould tend to delay the repolarization phase (Phase 3) and increase
the duration of the action potential. Ultimately, this predisposes the patient to fatal cardiac
arrythmias and the unfortunate sequelsudden cardiac death, even at early ages.
Currently, investigations of LQTS are a prototype for blending human genomics with
advanced cardiac modeling. In the early 1990's, a considerable flurry of molecular genetics
16 N. V. Thaker, V. Iyer,and M. B. Shenai
o 50 100 1SO 200
15:1
24:1
=7:1
100 ,
o 50 100 1SO 200
0
S'
§.
·so
·100
so
o
S'
§.
·50
· 100
so
0
S'
§.
·50
A
FIGURE 1.10. The effect of K, (LQTI) and K, (LQT2) mutations on action potential shape and duration of
isolated cells. Each simulation represents a 100% block of the respective current reduction. (From Viswanathan
and Rudy, 2000; used by permission)
studies linked LQTS populations to mutations in three putative genes located on chromo
somes 3, 6 and 11 (LQT1, LQT2, LQT3). The LQT1 and LQT2 genes represent an I
current (potassium delayed rectifier) and I current (potassium slow delayed rectifier), re
spectively (Barhanin et al. 1996; Wang et al. 1996). The LQT3 gene represents an enhanced
(incomplete inactivation) late sodium current. From these groundbreaking bench discov
eries, several modeling studies were able to place molecular genetics in the context of com
prehensive myocyte electrophysiology. For example, Viswanathan and Rudy were able to
show that different myocardial cells (epicardial, midmyocardium (Mcell), and endocardial
cells) respond to LQT gene defects with differing amounts of APD lengthening, producing
a transmural heterogeneity ripe for the formation of EADs (Figure 1.10) (Viswanathan and
Rudy 1999; Viswanathan and Rudy 2000). They modeled LQT1 and LQT2 by reducing
the density ofI and I channels (thereby reducing the maximal channel conductance per
From Cellular Electrophysiology to Electrocardiography 17
cell ). LQT3 was simulated by a right shift in the steadystate inactivation curve, such that
the hand gates demonstrated incomplete inactivation, resulting in a late sodium current.
Truly, longQT syndromes are demonstrating the cuttingedge interaction between
molecular genetics and advanced computer modeling. While the molecular techniques have
been instrumental in identifying the particular channelopathy, computer models have been
successful in placing the channelopathy in the context of other channels and the whole cell ,
producing a quantitative understanding of the disease.
1.2 NETWORK MODELS
1.2.1 CELLCELL COUPLING AND LINEAR CABLE THEORY
While the previous section treats the cardiac myocyte as an isolated element, it actually
exists in a densely interconnected network with other myocytes. This brings up the issue of
howcurrent spreads from one excitable myocyte to a neighboring myocyte . The predominant
model of current spread among excitable elements is termed cable theory (Miller and
Geselowitz 1978; Spach et at. 1981; Plonsey and Barr 1986; Malmivuo and Plonsey 1995).
Cable theory provides the crucial link from onecell to many cells . In its simplest form, it
present s a set of myocytes lined up next to each other and connected by gap junctions
forming a linear "cable" of excitable elements (Figure 1.11).
Cable theory assumes the existence of two compartments, subdivided into differential
compartments, dx. Each differential compartment is connected to its adjacent compartment
by resistances. Each transmembrane resistance, rm, represents a pathway for transmembrane
current, either passive or active (action potential producing elements like ion channels).
the extracellular space, r, represents the resistance between two extracellular differential
elements. The monodomain model assumes that this resistance is negligibly small (=0),
whereas bidomain models assume r
e
to be nonzero. Likewise, r, represents the resistance
between two differential intracellular elements. In bulk, accounts for both the inherent
cytoplasmic resistance and the celltocell resistance (gap junctions).
E: tracellular
FIGURE l.11. A onedimensional cable representing intracellular resistances, transmembrane resistances, and
extracellular resistances.The transmembrane resistance can be replaced with an active component, such as a
voltagesensitive ion channel.
18 N. V. Thakor, V. Iyer, and M. B. Shenai
Given this description, current follows a differential equation that contains both spatial
(in the x direction) and temporal derivatives (Plonsey 1969; Malmivuo and Plonsey 1995):
(1.16)
The right hand of this equation can be derived from (1.11), where the membrane current
is summed from all the ion channels. Numerical methods can then solve this differential
equation to provide a spatial profile of a propagating wavefront.
1.2.2 MULTIDIMENSIONAL NETWORKS
A thin myocardial slab is modeled by extending a linear cable conductor model into
two dimensions (Figure 1.12), assuming a highly conducting external medium. The model
in Figure 1.12 is a monodomain, and does not represent the separate intra and extracellular
resistivities. The propagation across the slab is calculated from a set of the partial differential
equations, derived from continuity and conservation of current at each node (Barr and
Plonsey 1984):
(1.17)
where a is the intracellular conductivity tensor, is the surface area to tissue volume ratio,
C
M
is the membrane capacitance per unit area, and lion (A/m
2
) is the total membrane ionic
current. In two dimensions, the total current flowing into a cell (from adjacent nodes) derives
from Ohm's Law and must equal the total membrane current:
i.. \/:.  1 \/:. 
+ + + = 1
R R R R
(1.18)
where 1 is the membrane current in amperes, and R is the bulk resistance of the cluster
area. The terms on the left represent current that flows from adjacent cells via intercellular
connections, or gap junctions.
Since the myocardium represents a functional syncitium, a single myocyte has proper
ties very similar to a cluster ofmyocytes. Clustering cells (Figure 1.12) allows one to model
larger areas of tissue while minimizing the computational load. In a particular direction
(longitudinal or transverse), the bulk resistances of each cluster, R, and R
y
, are net resis
tances derived from the lumped combination of intracellular resistivities and gapjunctional
resistances in series, and can be calculated by:
~ x
R
ax· h· ~ y
~ y
Ry
v: h· ~ x
(1.19)
where is the slab thickness, ~ x and ~ y represent the cluster dimensions in the longitudinal
and transverse directions, respectively, and a x = .35 Slm, a y = .035 Sim (the electrical
conductivities) at baseline conditions. Equation (1.19) defines resistances not based on
FromCellularElectrophysiology to Electrocardiography
1element
a
'
19
Network of Indi vi <bJ Cells
Cells Oustered
FIGURE 1.12. A myocardial slab with a defined ischemic locality in the center. Each cell in the model can be
interconnected to adjacent cells by gap junctional resistances. To decrease computational time, these cells can be
clustered into a bulk element with bulk conductivities.
20
o
N. V. Thakor, V. Iyer, and M. B. Shenai
*
FIGURE 1.13. A wave of cardiac excitation approaching an observing lead at (*). The inflection is positive as the
wave approaches, becomes rapidly negative at the exact time of incidence between the lead and the wave. Finally,
as the wave travels away from the lead, the recording returns to zero. See the attached CD for color figure.
the individual cell, but directly on spatial dimensions of the myocardial cluster. However,
in clustering an entire patch of myocytes, one assumes that all points within this patch
are isopotential, which defies the assumption of continuous current spread. Thus, while
clustering can increase the slab size at a given computational load, it does compromise the
resolution of the propagating wavefront.
Figure 1.13 depicts normal propagation across a 2D network of LuoRudy cells.
Anisotropy is evident by the preferential spread of current in the horizontal direction, which
reflects the 10:I anisotropy ratio (oja y) in this particular simulation. Figure 1.13 repre
sents only the early portion of activation, namely the upstroke, Phase 2 and early portions
of Phase 3. The procedure to reconstruct a representative electrical signal (the extracellular
electrogram) from this type of activation pattern is discussed in the next section.
1.2.3 RECONSTRUCTIONOF THE LOCAL EXTRACELLULAR ELECTROGRAM
(FORWARD PROBLEM)
While studying a single cell, the obvious candidate signal to study is the action potential.
However, when studying many cells in a network, a large number of action potentials exist,
making it difficult to study each action potential. In addition, action potentials represent
the intracellular potential, are difficult to obtain in gross studies, and are impossible to
obtain clinically. The local extracellular electrogramprovides a representation of all cellular
activity located within the vicinity of a lead. By reconstructing the extracellular electrogram,
an investigator can get a better representation of extracellular experimental recordings.
Ultimately, by solving the entire forward problem, one may be able to reconstruct the entire
ECG (see Chapter 2). For the purposes of examining pathology at the network level, the
reconstruction of the extracellular electrogram is presented.
Given a sheet of cells with time and stimulusdependent transmembrane potentials,
the goal is to reconstruct the extracellular lead potentials at selected points in the
From Cellular Electrophysiology to Electrocardiography 21
FIGURE 1.14. A slab of myocardium with predefined regional ischemia. The "extreme" ischemic zone is
surrounded by a border zone with milder ischemia. Point stimulation initiates a propagating wavefront, allowing
the calculation of the electrogram, and analysis of its features. See the attached CD for color figure.
slab shown in Figure 1.14, with respect to a point at infinity. At any given instant of time,
current injection, 10, into the extracellular space induces at a distance r a potential <1>0, which
is inversely proportional to the distance between the point at which <1>0 is measured and the
point at which 10 is injected:
10
<1>0
471"0"1
(1.20)
For multiple elements of an array, the extracellular potential (at any point in time) is rep
resented by superposition and summation of discrete elements (Plonsey and Collin 1961;
Plonsey and Rudy 1980):
<l>e(XI, vi, zi, t) L...J
all elements 471" a . y, z)
(1.21)
where z) is the transmembrane current at the source element positioned at point
z) and z) is the distance between the element at and the lead position
While Eq. (1.20) uses the transmembrane current to generate the extracellular
potential, many models generate transmembrane voltage, V Thus, the transmembrane
current is derived by (Spach et al.
X _ ~ ( ~ Z») ~ ( ~ Z»)
R,
(1.22)
where R, and R; are the celltocell resistances. This equation assumes that the extracellular
22
_/
•
N. V. Thakor, V.Iyer, and M. B. Shena!
/1
•
~ t ' , , , , ,
't
FIGURE 1.15. (Top) A wavefront propagating in space; (middle) the first spatial derivative of the wavefront
and (bottom) the second spatial derivative of the wavefront used in Eq. 3.18. The latter is directly related to the
extracellular electrogram, when extended multidimensionally.
resistance is approximately 0 (infinitely conductive in the monodomain model) and thus
the extracellular potential is approximately 0 mY. Figure 1.15 conceptually describes how
a passing waveform yields the characteristic shape of the extracellular electrogram. When
applied to the model depicted in Figure 1.13, the use of Eqs. (1.21) and (1.22) result in
a depolarization complex that conforms very well to experimentally recorded unipolar
electrograms under normal and pathological conditions (Gardner et at. 1985; Irnich 1985;
Blanchard et at. 1987) (Figures 1.13, 1.16).
The electrogram can be affected in two primary ways, as demonstrated in Eq. (1.23).
The shape of the action potential (in space or time) can be affected either by the direct
alteration of ionic channel parameters (as in hyperkalemia, acidosis, anoxia or LQTS),
or by the level of coupling resistances R, and R, (decoupling). Taking Eq. (1.22), then
reducing it to one dimension and expanding by the differentiation:
(1.23)
While Eq. (1.22) describes normal current spread, this correlate reveals how pathology may
manifest in the electrogram. In addition to being dependent on the explicit shape of V
m
, 1
m
also depends directly on the spatial derivative of
y
) , which is large near border zones
From Cellular Electrophysiology to Electrocardiography
s
•
~
FIGURE 1.16. A 2D network model of propagation across circular inhomogeneities of various manifestations
of ischemia. Hyperkalemia, anoxia, and decoupling each produce unique features in the QRS morphology of the
reconstructed extracellular electrogram. See the attached CD for color figure.
(the interface of the ischemic and normal tissues). This corresponds with experimental
observations that premature action potentials and ectopic activity may originate in these
border zones (l anse et al. 1980; lanse and Wit 1989).
1.2.4 MODELING PATHOLOGY IN CELLULAR NETWORKS
While the onecell paradigm can study a pathologic ion channel in the context of other
ion channels , networks (10 and 20) are required to understand how pathologic cells interact
with surrounding cells. With the ability to represent both ion channel s and intercellular resis
tivities, networks have been able to lend insight into the dangerous sequelae of myocardial
ischemia and LQTS, particularly arrhythmogenesis. A proarrhythmic mechanism to many
pathologic conditions is reentry, a phenomenon by which an propagating wave circum
vents refractoriness, to reexcite repolarized tissue. Thus, by definition, the study of reentry
requires multicellular networks.
To date, several major classifications of reentry have been defined (Spach 2001) . Clas
sically, peculiarities of membrane channels (densities and inhomogeneous distributions)
may lead to reentry, as in the longQ'T syndrome and myocardial ischemia. Alternately,
N. V. Thakor, V. Iyer, and M. B. Shenai
resisti ve discontinuities (i.e. during ischemia or infarction) and wavefront geometry (spiral
waves) may also be substrates for reentry. The remainder of this section will survey var
ious pathologies, the ionic characteristics relevant to cell network studies, electrographic
reconstructions, and how they may initiate reentry and subsequent arrhythmogenesis.
1.2.4.1 Myocardial ischemia
In the first section, myocardial ischemi a was described at the cellular level. However.
ischemia usually involves the entire spectrum, from a patch of cells (regional ischemia) to
the entire heart (global ischemia). For these more macroscopic manifestations of ischemia,
network models provide tremendous insight into how the cellular level changes alter the
spread of excitation, and appear in the local electrogram, and ultimately, the ECG.
These concepts can be simulated at the 2D network level (Figure 1.16) where the local
electrograms have been reconstructed (Shenai et at. 1999) from a network of LuoRudy
cells . The results from these simulations suggest that different ischemic parameters affect
the propagating wave in distinct ways. For example, hyperkalemia produces a decrease
in both conduction velocity and depolarization (effects derived from Phase 1 alterations ).
However, anoxia does not alter conduction velocity or depolarization, but mainly produces
a more rapid repolarization (a Phase 2,3 phenomenon). Finally, decoupling changes the
conduction velocity only. Likewise, these individual ischemic. mechanisms display several
elements of experimentally recognized signal distortions (baseline elevations, decreased
peaktopeak amplitudes, notching). Decoupled patches , for example, produce local ex
tracellular electrograms that closely corroborate experimental recordings (Figure 1.17)
(Shenai 2000).
In addition to necrosis of nonregenerative myocardium, the electrophysiological al
terations of ischemia may lead to fatal arrhythmogenesis via the mechanism of reentry.
When an excitation wavefront is blocked by an ischemic zone, the propagation is diverted
around the obstacle and may invade the ischemic zone retrogradely. This can create a
"Figure of Eight" 2D pattern (Figure 1.18, top) that may ultimatel y degenerate into ven
tricular fibrillation. In modeling studies, reentry can be simulated with a ringshaped linear
network, by which propagation can simultaneously travel in a diametric and circumferential
path. Ferrero et at. used this ringshaped model of LuoRudy cells, with defined regions
of ischemic impairment (BZ:Border Zone, CZ: Central Zone; Figure 1.18, middle) . With
a premature stimulus at the border zone, they found that the role of acidosis and hypoxia
vary in the establishment of "Figure of Eight" reentry. These models suggest that the I
K

ATP
current, most prominent during hypoxia, may be crucially proarrhythmic in the short term,
with hyperkalemia playing an essential role. However, the arrhythmogenic effect of acidosis
is minimal , and based on preliminary results, it may actually be antiarrhythmic (Ferrero
et at. 2001). Cardioprotective KATP channel modulators (pinacidil) enhance the I
ATP
and desensitize (or precondition) the myocardium to ischemiaderived k ATP enhancement
(Grover and Garlid 2000).
The early stages of ischemia represent a fork in the road to the development of several
different sequellae. While transient ischemia may yield no permanent damage, more acute
cases can lead to infarction, reentry (by Figureof8 or spiral mechanisms) and arrhythrno
genesis. By identifying which cellular ischemic characteristics provoke different sequellae,
and coupling these alterations with the electrographic modeling described above, a unique
paradigm emerges for evaluating sensitive detection algorithms. Cellular level alterations
From Cellular Electrophysiology to Electrocardiography
Experimental
Experimental
Experimental
GaU1er,
25
FIGURE 1.17. Experimental versus modeled waveforms for various ischemic manifestations. See the attached
CD for color figure.
(ischemic and/or proarrhythmic) can manifest in the electrographs and characteristically
alter detection parameters significantly before a critical event. By modeling these ischemic
and proarrhythmic parameters, one can precisely study how features in the electrograms or
ECG reflect these cellular or molecular alterations. With a theoretical understanding of these
changes, clinical monitoring can alarm the patient, when parameters exceed physiological
variation, to seek the proper interventional procedures (angioplasty, bypass) or pharmaco
logical therapies (anticoagulants, antiarrythmics).
1.2.4.2 EADs in and networks
The propagation of an EAD along a linear cellnetwork has been studied extensively
in the context of ectopic beats and arrhythmogenesis. Triggered activity is the initiation
of an abnormal impulse, induced by an EAD or DAD, that may propagate to neighboring
cells. Unlike onecell models, network models allow the use of intracellular coupling, and
observations on how coupling parameters may alter the propagation of the EAD.
tl Z
NZ ...
NZ
BZ
BZ
BZ
pHBZ
FIGURE 1.18. Top: schematic of "figure of eight reentry". BZ: borderzone. ischemic central zone. NZ:
normal zone. Arrows show pattern of propagation. Middle: Ringshaped Idimensional approximation of one of
the reentry circuits. Numbers indicate cell number. Bottom: Various action potential traces between cell #15 and
#315 with defined regions of ischemia =central zone; BZ =border zone, NZ =normal zone) corresponding
to a premature stimulation See the attached CD for color figure. (From Ferrero et al. 200 1; 2001 IEEE)
From Cellular Electrophysiology to Electrocardiography 27
The most suspicious areas for triggered activity occur near the border of normal and
abnormal myocytes. This situation was modeled by Saiz and colleagues (Saiz et al. 1997;
Saiz et al. 1999) with a twocell LR model. An abnormal cell (C1, conditions favorable to
EAD formation) was coupled to a normal cell (C2) by a coupling resistance (R). Their study
suggested that C1's abnormal conditions had a strong influence over EAD formation in C2,
and that APD was heavily influenced by the coupling resistance. Another investigation of
twocell models, by Wagner et al. (Wagner et al. 1995), coupled a sinusoidalgenerating
cell to an LuoRudy cell, and also found that coupling resistance had an instrumental role
in EAD propagation. Similar results implicating the role of coupling resistance in EAD
propagation have been shown in larger networks of cells (Saiz et al. 1996; Nordin 1997).
Larger models are able to incorporate multiple tissue types, while still maintaining fine
ionic detail. For example, the propagation of ectopic activity from the Purkinje network to
the ventricular myocyte network was demonstrated in a 2D model by Monserrat et al.
(Monserrat et al. 2000). In this model, a 2D sheet of ventricular myocytes was coupled
to a Purkinje fiber where EADs were induced. They concluded that the EADs transfer
to the ventricular myocytes only when within a certain range of I blockade and an ICa,f
enhancement.
While many other applications ofEADstudies exist in the context of arrhythmogenesis,
the common themes center around abnormal impulse initiation (onecell ion channel studies
that find cellular conditions favorable to EAD genesis) or abnormal impulse propagation
(cellnetwork models which study how intracellular conditions interact with intercellular
coupling that may elicit ectopic beats).
1.2.4.3 The ionicbasisofspiralwaves andfibrillation
A common cardiac arrhythmia that has been intensely studied with modeling tech
niques is ventricular fibrillationor the degeneration of the normal wave propagation into
irregular and asynchronous events. While a variety of initiating mechanisms may exist,
Davidenko et al. were the first to experimentally observe spiral waves as a mode of reentry
(Davidenko et al. 1992). Conceptually, the "head" of these spiral activation fronts wrap
around to "reenter" and stimulate the "tail" which is already recovering from the action
potential (refractory period). The result is an intriguing, selfsustaining rotor of activity that
may "breakup" and give rise to fibrillation. After the initial observation, many modeling in
vestigations were able to deduce and derive conditions in which a spiral wave may develop.
Several 2D models have been used to successfully demonstrate spiral wave activity, such
as both BeelerReuter and LuoRudy models (Fishler and Thakor 1991; Leon et al. 1994).
Two absolute conditions must be met for the initiation of a spiral wave (and reentry,
in general). First, unidirectional block must allow wave propagation in one direction only.
Second, the time (T
re)
required for the wave to cover the length (L
re)
of the reentrant
path (T
re
Lre/CV) must fall within the relative refractory period of the previous action
potential (the vulnerable window). The effect of cellular coupling and membrane excitability
on unidirectional block is detailed in (Quan and Rudy 1990). These conditions are generally
met by initiating propagation with a rapid series of planar orthogonal S1S2 stimuli. This
rapid pacing produces an APD sufficiently short for reentry. Depending on the modeling
parameters, the spiral wave may remain stationary, or meander around in space (Davidenko
et al. 1992).
28 N. V. Thakor, V. Iyer, and M. B. Shenai
FIGURE 1.19. Initiation of spiral wave activity in an anisotropic cardiac sheet of 2 em x 2 cm using the original
LR model (A) and a modified model exhibiting a short APD (SAPD); B. Numbers at the top of each panel indicate
the time after the S2 stimulus of the crossfield stimulation protocol. The color legend used to map the potential
distribution is shown at the bottom. See attached CD for color figure. (From Beaumont et al. 1998; used by
permission)
While these spiral waves have been observed experimentally, and demonstrated in
various models, the ionic basis for the formation of these spirals remains unclear and a
topic of cuttingedge research. The difficulty in ascertaining this ionic basis lies within
the necessary curvefitting and estimation techniques that define complex models. As a
result, these models cannot provide the accuracy or stability to reproduce realistic spiral
waves. Nevertheless, several groups have begun to demonstrate and deduce ionic roles. For
example, Beaumont et at. were able to show that different spiral wave patterns (stationary,
chaotic, hypocycloidal meandering, epicycloidal meandering) can be defined in different
regions of a parameter space of voltagedependence shift and sodium channel conductance
(Figure 1.19, 1.20) (Beaumont et al. 1998). Qu et at. recently suggested that chaotic spiral
wave meandering and spiral wave breakup are heavily dependent on the Ca2+ and K+
currents (Qu et al. 2000). More recently, Xie et al. were able to demonstrate the effects of
ischemia on the characteristics of spiral wave stability (Xie et al. 2001).
14 1819 13
From Cellular Electrophysiology to Electrocardiography
.§ 0
·10
4.0 5.5 8 0.05 .os .09 .15 .16 .5 1.0 2.0
29
FIGURE 1.20. Representation of various types of spiral activity. and its dependence on G
na
and voltageshift
characteristics. (From Beaumont et al. 1998; used by permission)
By determining the roles of ionic currents in fibrillation generation, putative targets for
defibrillatory drug therapy can be identified ("chemical defibrillation"). Qu et at. studied
the impact of different classes of experimental current blockers (INa, I
K,
and lea) in altering
or terminating the course to fibrillation (Figure 1.20). They found that a combination of
current blockers were most effective in extinguishing a fibrillatory state.
1.2.4.4 Cellnetworks in LongQT Syndrome
Network modeling is beginning to emerge in the study of LongQT syndromes, and
how defined molecular defects ultimately develop into arrhythmias and sudden death.
Viswanathan and Rudy studied a l Dfiber of LuoRudy cells, oriented transmurally with pre
defined regions (endocardial, Mcell, epicardial). Though I
Ks,
I
Kr
and late INa lengthen APD
and EAD formation in isolated cells, this study suggested reduced gapjunction coupling
can alter the location of EAD formation between endocardial and Mcells (Viswanathan
and Rudy 2000).
While the Viswanathan and Rudy model studied preferential EAD formation, another
recent investigation is beginning to lend insight on the role of LQT mutations on arrhyth
mogenesis. Clayton et at. simulated the LQT1, LQT2, and LQT3 mutations in ID and 2D
tissue models of excitation, with the focus of investigating reentrant mechanisms. Using
similar cellular definitions of each LQT mutation (reduction of maximal I
Ks
and IKr currents
and debilitation of sodium inactivation for the late INa current), this study suggested that
LQT mutations did not increase the vulnerability of LQT tissue to reentry (Clayton et at,
2001). Rather, after initiation, LQT tissue is more likely to sustain increased motion of these
reentrant waves (Clayton et at.
1.3 MODELING PATHOLOGY IN THREEDIMENSIONAL AND WHOLE
HEART MODELS
While onecell and network models can lend insight into detailed conduction and
pathologic interaction, these models and electrical reconstructions cannot extrapolate to
30 N. V. Thakor, V. Iyer, and M. B. Shenai
FIGURE 1.21. Cardiac electrical restitution properties and stability ofreentrant spiral waves: a simulation study.
While various classes of current reducers do not terminate activity, a combination of Cachannel reducers and
Class III antiarrhythmics lead to termination (From Qu et al. 1999; used by permission)
variables found in routine clinical settings. The 12lead ECG remains the diagnostic gold
standard for clinical studies , and biochemical assays are primarily used for metabolic and
cytopathic assessment. At a strictly conceptual level, the ECG and assay result s are undoubt
edly linked. Nevertheless, the ECG cannot diagnose cellular pathology, and biochemical
assays cannot convey the global nature of a disease . The central thesis of this chapter is
that modeling can be a valuable tool to quantitatively assess how molecular and cellular
processes are linked to ECG changes.
From Cellular Electrophysiology to Electrocardiography 31
Reconstruction of the ECG requires placement of cells in a realistic whole heart and
torso models . Detailed whole heart and torso models take into account geometry of the
heart and smooth transmural variations in fiber orientation, tissue conduction anisotropy,
distinct tissue types, and volume conduction properties (Ramon et al. 2000; Scollan et al.
2000). While subsequent chapters will offer an extensive treatment of the forward problem
solution (Chapter 2) and whole heart modeling (Chapter 3), the remainder of this chapter
will conceptually focus on molecular or ionic dysfunction that has been extended to the
whole heart, finally bridging cell to ECG.
1.3.1 MYOCARDIAL ISCHEMIA
As discussed in earlier sections, regional myocardial ischemia has been studied at many
levels of analysis: in single myocytes, 2D network models , and 3D tissue slabs. Occlusion
and reentry studies using 3D modeling techniques provide a fundamental understanding of
ischemic localization, and its tendency towards arrhythmogenesis.
A simulation study showcasing whole heart localization of tissue impairment in is
chemia was conducted by Dube, Gulrajani and colleagues (Dube et al. 1996). The authors
successfull y reproduced local ischemia using classical characterizations of ischemic cells
(depolarization of resting potential , reduced action potential upstroke, shortened APD,
and reduced conduction velocity). Realistic threedimensional localization of ischemic tis
sue was simulated by delineating regions of the myocardium subject to various grades of
ischemia, as determined by experimental studies of artery occlusion. The cellular character
istics of the altered action potential studied in onecell model s were thus incorporated into
a whole heart model. Surface potentials were computed to reconstruct the 12lead ECG,
yielding ECG changes similar to clinical findings for each artery occluded (Figure 1.22).
In addition to causing tissue impairment in the myocardium, ischemia and infarction
provide a substrate for arrhythmia. While several two dimensional modeling studies have
shown arrhythmogenesis in ischemia, Leon and Horacek were the first to investigate the
inducibility of reentry in the presence of ischemic regions and infarcted tissue in whole
heart model s (Leon and Horacek 1991). Using a threedimensional model of the heart , the
simulation demonstrated that ischemic conditions may give rise to reentrant activit y.
1.3.2 PREEXCITATION STUDIES
Studies of anomalous excitation in the whole heart often result from deficiencies on a
lower level of analysi s. In WolffParkinsonWhite (WPW) syndrome, abnormal myocardial
tissue formation bridges the fibrous tissue separating the atrium and ventricle, providing
accessory pathways by which reentry is facilitated. Because of its ability to model many
types of tissue, whole heart models have become a tool for studying WPW, which re
quire large scale alterations of multiple tissue types in their cellular manifestation of WPW
syndrome.
On the cellular level, the refractory period, and resultant plateau calciumchannel activa
tion, are prime determinants of susceptibility to ectopic phenomena. Lorange and Gulrajani
were the first to simulate WPW in a 3D whole heart model and to reconstruct body surface
potentials (Lorange and Gulrajani 1986). Following their work, Wei et al. appropriatelyal
tered cellular properties such as intercellular conductivity, tissue anisotropy, and refractory
32 N. V. Thakor, V. Iyer, and M. B. Shenai


a) Nonnal
. ~ . ~ ~ ~
b) LAD
v
v.
r
v
v. va
v .
v.
lJ
~ ~
v
• \.....
d) LCX
'  L A  . . ~ ~
.". .v..
c )
· ~ ~ ~ w ~
FIGURE 1.22. Action potential shape (top) and ECG reconstructions corresponding to simulation of various
coronary artery occlusions (a.b.c.d) in moderate ischemia (bottom). Characteristic features of ischemia, including
leadspecific ST segment depression and alterations in QRS morphology, are reproduced. (From Dube et al.
used by permission)
From Cellular Electrophysiology to Electrocardiography 33
~
R'l P.
Ill
RVF
V /
..
\/IV
~ I
1 ~ ~ _ V _ 2
V
...
__f ~ 6
FIGURE 1.23. Reconstruction of ECG in the simulation of a particular arrhythmia associated with Wolff
ParkinsonWhite syndrome. "Delta" waves. or slow upstrokes leading into QRS (prominent in II. III. and aVF)
are a result of early. nonbundle branch initiation of ventricular excitation. (From Wei et al. 1990; used by
permission)
delays for WPW myocytes (Wei et al. 1990). Regionally, the investigators characterized
cell types in different regions of the heart and included atrioventricular accessory pathways.
Body surface potentials were computed (Figure 1.23) show that the reconstructed 12lead
ECG corresponds closely with clinically recorded WPW ECG traces.
More recently models have been used to study the pharmacological treatment ofWPW.
For example. Fleischmann et al. investigated the effects of verapamil, a calcium channel
blocker at the cellular level, on simulations of WPW preexcitation. Drug administration
significantly affected the formation of reentrant pathways in the study (Fleischmann et al.
1996). WPW preexcitation simulation studies thus offer an example of how 3D models can
provide a useful tool for disease analysis. as well as a theoretical understanding for disease
treatment
:


. \11 ,,\1.
J,,
\ ' ,
",
\',
l
+.
+
J L
N. V. Thakor, V. Iyer, and M. B. Shenai
FIGURE 1.24. Changes in myocardial structure associated with disarray (top) and the reconstructed ECGs
(bottom) in the simulation hypertrophic cardiomypathy. Increases in QRS amplitude are observed in the septal
and leftheart precordial leads, V2 and V3. (From Wei et al. 1999 Figures 2, 56; used by permission)
1.3.3 HYPERTROPHIC CARDIOMYOPATHY
Several modeling studies illustrated that hypertrophic cardiomyopathy (HCM) is a
pathology in which local cellular events, when integrated globally, translate into detectable
ECG changes. HCM is accompanied by many changes on the cellular and regional level. A
fetal mode of genetic upregulation is observed in hypertrophy, through which preexisting
myocytes increase in size while total cell number remains relatively constant. More micro
scopic findings show that the orderly arrangement of myocytes in a parallel, linear layer is
disrupted during HCM, demonstrating the phenomenon of myocardial disarray in the tissue
(Figure 1.24). The faster conduction velocity along the fiber orientation thus degenerates
into isotropic conduction in the myocardium (Varnava et
Wei and colleagues in 1999 incorporated hypertrophied models of myocytes in a
realistic 3D model to arrive at a whole heart model of hypertrophic cardiomyopathy (Wei
et at. 1999). Clinical ECG traces of HCM patients frequently show Q waves with enlarged
magnitude, presumably due to increased contribution to the QRS vector. The salient cellular
features of HCM were implemented in a realistic 3D whole heart model to show that
indicators of this disease are built up from more fundamental cellular causes and reflect
themselves in the ECG. The reconstructed ECGs shown in Figure 1.24 reflect changes
commonly seen in clinical electrocardiography.
While Wei and colleagues investigated the effect of myocardial disarray in hypertrophic
cardiomyopathy, Siregar, et developed a realistic anisotropic cellular automata model
of the heart (Siregar et at. 1998) to study the effects of heart size and wall thickness
on clinically observed parameters. Using a BeelerReuter membrane model, characteristic
From Cellular Electrophysiology to Electrocardiography 35
action potentials were derived for the myocytes, and propagation simulated across the heart.
ECGs were reconstructed that confirmed experimental findings, including increased QRS
amplitude. Thus 3D models were effectively used to explore different aspects of hypertrophy
in the heart.
1.3.4 DRUGINTEGRATION IN THREEDIMENSIONAL WHOLE
HEART MODELS
While pathology studies certainly provide knowledge on the nature of disease, a current
direction of studies is to identify and validate potential drug targets through simulation. The
first regulatory drugtoECG assessment was used to dismiss the potentially proarrhythmic
effect of an antiarrhythmia drug by the American Food and Drug Administration (FDA).
Concerns were raised over mibefradil, a Ttype and Ltype Ca2+ channel blocker, since it
showed Twave ECG perturbations indicative of Torsade de Pointes susceptibility. Models
of the ventricle, placed in a torso, were used to show that the Twave ECG perturbations
actually arise from action potential shortening on the cellular level, and not on potentially
arrhythmogenic action potential lengthening. Thus simulation studies were instrumental
in showing that ECG changes suggesting impairment can have entirely different (indeed,
beneficial) cellular roots.
While the FDA study investigated drug activity from "ECG down", other studies have
built from cellular effects to higher levels of analysis. Recently, promising drug action studies
in 3D ionic models were performed by Garfinkel et al. with bretylium administration, which
has been shown to flatten AP restitution (Garfinkel et ai. 2000). Scroll waves, the three
dimensional correlate to spiral waves, were induced in normal tissue and spontaneous wave
breakup into fibrillatory propagation was observed. When the treated action potentials were
simulated, the scroll waves remained intact, suggesting that treatment is protective against
degeneration into fibrillation (see Figure 1.25). Similar studies showed predictive potential
in investigating pure Ltype channel blockers (Noble et ai. 1999) and Na+ H+ exchanger
blockers used in the treatment of myocardial ischemia (Ch'en et ai.
1.3.5 GENETICINTEGRATION IN THREEDIMENSIONAL WHOLE HEART
MODELS
With the current emergence of a wealth of molecular data, and the concomitant expan
sion in computational resources, genetoECG and genetoheart pathology studies have
recently been investigated by several groups.
One such study was performed by Okazaki et ai., in investigating the link between
Long QT (LQT) syndrome and Torsade de Pointes (TdP) (Okazaki et ai. 1998). Okazaki
et ai. were able to model myocytes with genetic LQT mutations in a threedimensional
model to investigate the whole heart consequences of the syndrome. Simulation of the
diseased action potential led to arrhythmia and the periodic, abnormal ECG characterizing
TdP (Figure 1.26). Future studies promise to similarly carry analysis from gene to ECG.
Winslow's group in 2000 arrived at a model of a failing myocyte based on experimental
findings on genetic regulation in disease (Winslow et ai. 2000). Significant downregulation
of channels carrying the transient outward current Ito! and the fast inward rectifier current I
K1
was observed in endstage heart failure. Reduced expression of SERCA2A (which encodes
the smooth ER calcium pump) and increased expression of NCXl (which encodes the
36
Spontaneous Kroll breakup
Administration 01
Admlnlllratlon 01brelyllum
olelfecll
N. V. Thakor, V. Iyer, and M. B. Shenai
Continued Ilbrillatlon1 ke propagation
Intact Kroll
FIGURE 1.25. Scroll waves show spontaneous degeneration into fibrillatory propagation (top, left). Adminis
tration of class III antiarrhythmic agents steepen the action potential restitution and enhance fibrillationlike
propagation (top, right). Administration of bretylium, which flattens action potential restitution, results in intact
scroll waves (bottom). See the attached CD for color figure. (From Garfinkel et al. 2000; used by permission)
sodium calcium exchanger) was also observed (Winslow et at. 2000). This minimal model
of a failing myocyte was incorporated into a realistic wholeheart model to investigate
whether the resulting action potential prolongation is sufficiently arrhythmogenic on the
whole heart level. Simulations showed waves of uncontrolled propagation in the diseased
heart (see Figure 1.27). Comparison of normal and model ECGs confirms this behavior in
the failing heart.
1.4 DISCUSSION
With the simultaneous explosion of molecular biology and computational power,
paradigms for studying the molecular roles in whole heart pathologies are emerging
From Cellular Electrophysiology to Electrocardiography 37
'A.' . V" ,....·....,.,r JAt' • _ . nO' . • ....,
V WI
'MliWJ4
FIGURE 1.26. Variation in APD according to M cell distribut ion is incorporated into a 3D simulat ion (top). The
reconstructed ECG reflects torsade de pointes (bottom). (From Okazaki et al. 1998; used by permission).
through computer modeling. These physiological models, when coupled with electro
graphic reconstruction techniques can reproduce clinically accessible waveforms. So far,
most studies have spanned only a few levels, from genetocell , from celltonetwork, or
from celltowhole heart. Several reviewers have formally defined these integrative model
ing paradigmsfrom genometophysiome (Rudy 2000) and from genestorotors (Spach
2001). Another domain of modeling is solving the forward problem of electrophysiol
ogy (see Chapters 3), through which activation patterns, and heart/torso geometries are
extrapolated to the ECG. One future of integrative cardiac modeling is to yield a gene
toECG paradigm, by linking genometophysiome models (Sections and 1.2) with
physiometoECG models (Section 1.3). This is an extremely challenging task, requiring a
profound description of the gene/molecular dynamics, intercellular connectivities, diverse
tissue characteristics and heart/torso geometries, all coupled to the forward problem of
electrophysiology. It will require a tremendous amount of experimental and computational
development.
38 N. V. Thakor, V. Iyer, and M. B. Shenai
1.2
1
o.
0.4
J:l
o.z
0
.()
00.%
0 llOO 2llOO ~
0 eoo 2llOO ~
T1_(
FIGURE 1.27. EADs evoke uncontrolled arrhythmic propagation in heart failure (top); reconstructed ECG for
normal tissue (bottom, left) versus failing tissue (bottom, right) confirms erratic excitation See the attached CD
for color figure. (From Winslow et al. 2000; used by permission).
The potential benefits of developing such a paradigm, however, would be enormous.
For example, comprehensive, integrative models that can simulate the cellular level ofdrugs
will emerge to direct efficient pharmaceutical developmentboth testing novel drugs and
testing interactions of infinite drug combinations. Even farther off in the future, clinical
management of heart patients may depend on integrated models customized to a particular
patient's unique set of genetic and acquired deficiencies. Today, however, cardiac models
have found a comfortable niche in interfacing theoretical understanding with experimental
or clinical outcomes, from cell to ECG.
REFERENCES
Barhanin, J., Lesage, E. Guillemare, M. Fink, M. Lazdunski and G. Romey (1996). "K(v)LQTl and IsK (minK)
proteins associate to form the I(Ks) cardiac potassium current." Nature 384: 7880.
From Cellular Electrophysiology to Electrocardiography 39
Barr, R. and R. Plonsey (1984). "Propagation of excitation in idealized anisotropic twodimensional tissue."
Biophys 45: 11911202.
Beaumont, J., N. Davidenko, J. Davidenko and J. Jalife (1998). "Spiral Waves in TwoDimensional Models of
Ventricular Muscle: Formation of a Stationary Core." Biophys J 75: 114.
Beeler, G. and Reuter (1976). "Reconstruction of the action potential of ventricular myocardial fibers." J Physiol
268: 177210.
Blanchard, S., R Damiano, T. Asano, W Smith, R Ideker and 1. Lowe (1987). "The effects of distant cardiac
electrical events on local activation in unipolar epicardial electrograms." IEEE Trans Biomed Eng 34: 539
546.
Cascio, W, T. Johnson and Gettes (1995). "Electrophysiologic changes in ischemic ventricular myocardium:
Influence of ionic, metabolic and energetic changes." Cardiovasc Electrophys 6: 10391062.
Ch' en, E, R VaughanJones, K. Clarke and D. Noble (1998). "Modeling myocardial ischaemia and reperfusion."
PRog Biophys Mol Bioi 69(23): 51538.
Clayton, R, A. Bailey, V. Biktashev and A. Holden (2001). "Reentrant cardiac arrhythmias in computational
models of longQT myocardium." J Theor Bioi 2001 208(2): 215225.
Cole, (1949). "Dynamic electrical characteristics of squid axon membrane." Arch. Sci. Physiol 3: 253258.
Davidenko, 1., A. Pertsov, R Salomonsz, W Baxter and J. Jalife (1992). "Stationary and drifting spiral waves of
excitation in isolated cardiac muscle." Nature 355: 349351.
DiFrancesco, D. and D. Noble (1985). "A model of cardiac electrical activity incorporating ionic pumps and
concentration changes." Philos Trans R Soc Lond B Bioi Sci. 307(1133): 353398.
Dube, B., R Gulrajani, M. Lorange, A. LeBlanc, J. Nasmith and R Nadeau (1996). "A computer heart model
incorporating anisotropic propagation. IV. Simulation of regional myocardial ischemia." J Electrocardiol29:
91103.
EISherif, N., Caref, Yin and M. Restivo (1996). 'The electrophysiological mechanism of ventricular
tachyarrhytmias in the long QT syndrome: tridimensional mapping of activation and recovery patterns." Circ
Res 1996(79).
EISherif, N. and G. Turitto (1999). "The Long QT Syndrome and Torsade De Pointes." PACE 22 (Pt.l): 91110.
Eyring, H., R. Lumry and 1. Woodbury (1949). "Some applications of modern rate theory to physiological systems."
Record Chem. Progr 10: 100114.
Factor, S. and R. Bache (1998). Pathophysiology of Myocardial Ischemia. Hurst's The Heart. R. Alexander, R
Schlant and V. Fuster. New York, McGrawHill: 12411262.
Ferrero, 1., J. Saiz, J. Ferrero and N. Thakor (1996). "Simulation of action potentials from metabolically impaired
cardiac myocytes: role of ATPsensitive K current." Circ Res 79: 208221.
Ferrero, J., V. Torres, E Montilla and E. Colomar (2001). "Simulation of Reentry During Acute Myocardial
Ischemia: Role of ATPsensitive Potassium Current and Acidosis." Computers in Cardiology.
FishIer, M. and N. Thakor (1991). "Amassively parallel computer model of propagation through a twodimensional
cardiac syncytium." Pacing Clin Electrophysiol14(11 pt 2): 16949.
Fleischmann, P., G. Stark and P. Wach (1996). "The antiarrhythmic effect of verapamil on atrioventricular reentry
in the WolffParkinsonWhite syndrome: a computer modle study." Int Biomed Comput 41: 125136.
Gardner, P, P. Ursell, J. Fenoglio and A. Wit (1985). "Electrophysiologic and anatomic basis for fractionated
electrograms recorded from healed myocardial infarcts." Circulation 72: 596611.
Garfinkel, A., Kim, O. Vorshilovsky, Z. Qu, J. Kil, M. Lee, Karageuzian, J. Weiss and P. Chen (2000).
"Preventing ventricular fibrillation by flattening cardiac restitution." Proc Natl Acad Sci 97(11): 60616.
Grover, G. and K. Garlid (2000). "ATPSensitive potassium channels: a review of their cardiprotective pharma
cology." Mol Cell Cardiol32: 67795.
Hodgkin, A. and A. Huxley (1952). "A Quantitative description of membrane current and its application to
conduction and excitation in nerve." Physiol117: 500544.
Irnich, W (1985). "Intracardiac Electrograms and Sensing Test Signals: Electrophysiological, Physical and Tech
nical Considerations." PACE 8: 870888.
Janse, M., E v. Capelle, H. Morsink, A. Kleber, E WilmsSchopman, R. Cardinal, C. d' Alnoncourt and D.
Durrer (1980). "Flow of "injury" current patterns of excitation during early ventricular arrythmias in acute
regional myocardial ischemia in isolated porcine and canine hearts. Evidence for two different arrhythmogenic
mechanisms." Circ Res 47(2): 151165.
Janse, M. and Wit (1989). "Electrophysiological mechanisms of ventricular arrythmias resulting from myocar
dial ischemia and infarction." Phys Rev 69: 10491152.
40 N. V.Thakor, V.Iyer, and M. B. Shenai
January, C. and J. Riddle (1989). "Early afterdepolarizations: mechanism of induction and block, a role for Ltype
Ca2+ current." Circ Res 64: 977990.
Kagiyama, Y.,1. Hill and Gettes (1982). "Interaction of acidosis and increased extracellular potassium on action
potential and conduction in guinea pig ventricular muscle." Circ Res 51: 614623.
Kleber, A., M. Janse, F. WilmsSchoprnann, A. Wilde and R. Coronel (1986). "Changes in conduction velocity
during acute ischemia in ventricular myocardium of isolated porcine heart." Circulation 73: 189198.
Kodama, I., Wilde and M. Janse (1984). "Combined effects of hypoxia, hyperkalemia, and acidosis on membrane
action potential and excitability of guineaypig ventricular muscle." Mol Cell Cardioll6: 247259.
Leon, and Horacek 99\). "Computer model of excitation and recovery in the anisotropic myocardium. III.
Arrhythmogenic conditions in the simplified left ventricle." J Electrocardio124( I): 3341.
Leon, F. Roberge and Vinet (1994). "Simulation of twodimensional anisotropic cardiac reentry: Effects of
the wavelength on the reentry characteristics." Annals Biomed Eng 22: 592609.
Lindblad, D., C. Murphey, J. Clark and W.Giles (1996). "A model of the action potential and underlying membrane
currents in a rabbit atrial cell." Am Physio1241(4 Pt 2): HI66696.
Lorange, M. and R. Gulrajani (1986). "Computer simulation of WolffParkinson White preexcitation syndrome
with a modified MillerGeselowitz heart modle," IEEE Trans Biomed Eng 33(9): 862873.
Luo, C. and Rudy (1991). "A model of the ventricular cardiac action potential: depolarization, repolarization,
and their Interaction." Circ Res 68: 15011526.
Luo, C. and Rudy (1994). "A dynamic model of the cardiac ventricular action potential: I. Simulations of ionic
currents and concentrations." Circ Res 74: 10711086.
Luo, C. and Y.Rudy (1994). "Adynamic model of the cardiac ventricular action potential: II. Afterdepolarizations,
triggered Activity, and potentiation." Circ Res 74: 10971113.
Malmivuo, 1. and R. Plonsey (1995). Bioelectromagnetism. New York, Oxford, Oxford University Press.
Marban, E., S. Robinson and W. Wier (1986). "Mechanisms of arrhytmogenic delayed and early afterdepolariza
tions in ferret ventricular muscle." J Clin Invest 78: 1185: 1192.
Marmont, G. (1949). "Studies on the axon membrane. I. A new method." J Cell Comp Physio150: 140111.
McAllister, R., D. Noble and R. Tsien (1975). "Reconstruction of the electrical activity of cardiac Purkinje fibres."
J Physio1251: I59.
Miller, W. and D. Geselowitz (1978). "Simulation studies of the electrocardiogram. I. The normal heart." Circ Res
43: 301315.
Monserrat, M., J. Saiz, J. Ferrero, 1. Ferrero and N. Thakor (2000). "Ectopic activity in ventricular cells induced
by early afterdepolarizations developed in Purkinje cells." Ann Biomed Eng 28: 134351.
Moore, J. and R. Pearson (1981). Kinetics and Mechanisms. New York, Wiley.
Morena, H., M. Janse, J. Fiolet, W. Krieger, H. Crijns and D. Durrer (1980). "Comparison of the effects of regional
ischemia, hypoxia, hyperkalemia and acidosis on intracellular and extracellular potentials and metabolism
in the isolated porcine heart." Circ Res 46: 634646.
Noble, D., J. Levin and W. Scott (1999). "Biological simulations in drug discovery." Drug Discov Today 4(1):
1016.
Nordin, C. (1997). "Computer model of electrophysiological instability in very small hereogeneous ventricular
syncytia." Am J Physioll72: HI8381856.
Okazaki, D. Wei and Harumi (1998). "A simulation of Torsade de Pointes with M cells." Electrocardiol
31(Suppl): 14551.
Plonsey, R. (1969). Bioelectric Phenomena. New York, McGrawHill.
Plonsey, R. and R. Barr (1986). "A critique of impedance measurements in cardiac tissue." Ann Biomed Eng 14:
30722.
Plonsey, R. and R. Collin (1961). Principles and applications of electromagnetic fields. NY, McGrawHill.
Plonsey, R. and Y. Rudy (1980). "Electrocardiogram sources in a 2dimensional anisotropic activation model."
Med Biol Eng Comp 18: 8794.
Priebe, and D. Beuckelmann (1998). "Simulation study of cellular electric properties in heart failure." Circ Res
82(11): 12061223.
Priori, S. and P. Corr (1990). "Mechanisms underlying early and delayed afterdepolarizations induced by cate
cholamines," Am J Physio1258: HI796HI805.
Qu, Z., F.Xie, A. Garfinkel and J. Weiss (2000). "Origins of spiral wave meander and breakup in a twodimensional
tissue model." Ann Biomed Eng 28: 75571.
FromCellularElectrophysiology to Electrocardiography 41
Quan, and Rudy (1990). "Unidirectional block and reentry of cardiac excitation: a model study." Cire Res
66: 367382.
Ramon, C; Wang, J. Huaeisen, P. Schimpf, S. Jaruvatanadilok and A. Ishimaru (2000). "Effect of myocardial
anisotropy on the torso current flow patterns, potentials and magnetic fields." Phys Med BioI. 45(5): 1141
1150.
Rudy, (2000). "From genome to physiome: integrative models of cardiac excitation." Ann Biomed Eng 28(8):
945950.
Saiz, J., J. Ferrero, M. Monserrat, J. Ferrero and N. Thakor (1997). From the cell to the body surface. Electrocar
diology '96.1. Liebman. NJ, World Scientific Publishing: 209212.
Saiz, 1., J. Jr, M. Monserrat, J. Ferrero and N. Thakor (1999). "Influence of electrical coupling on early
afterdepolarization in ventricular myocytes." IEEE Trans Biomed Eng 46(2): 138147.
Saiz, 1., M. Monserrat, J. JM Ferrero, J. Ferrero and N. Thakor (1996). "Ectopic activity generated by early
afterdepolarizations in ventricular tissue. A computer simulation study." Computers in Cardiology 1996.
Scollan, D., A. Holmes, J. Zhang and R. Winslow (2000). "Reconstruction of cardiac ventricular geometry and
fiber orientation using magnetic resonance imaging." Ann Biomed Eng 28(8): 93444.
Shaw, R. and Y. Rudy (1997). "Electrophysiologic effects of acute myocardial ischemia: A mechanistic investi
gation of action potential conduction and conduction failure." Circ Res 80: 124138.
Shaw, R. and Rudy (1997). "Electrophysiologic effects of acute myocardial ischemia: a theoretical study of
altered cell excitability and action potential duration." Cardiovascular Res 35: 256272.
Shenai, M. (2000). Myocardial Ischemia Detection: A timefrequency investgation of intraQRS changes in the
endocardial electrogram. Dept. of Biomedical Engineering. Baltimore, The Johns Hopkins University.
Shenai, M., B. Gramatikov and N. Thakor (1999). "Computer models of depolarization alterations induced by
myocardial ischemia: the effect of superimposed ischemic inhomogeneities on propagation in space and
timefrequency domains." Journal of Biological Systems 7(4): 553574.
Siregar, P., 1. Sinteff, N. Julen and P. LeBeux (1998). "An interactive 3D anisotropic cellular automata model of
the heart." Comput Biomed Res 31: 32347.
Spach, M. (2001). "Mechanisms of the Dynamics of Reentry in a Fibrillating Myocardium. Developing a Genes
toRotors Paradigm." Circ Res 88: 753755.
Spach, M., Miller and D. Geselowitz (1981). "The discontinuous nature of propagation in normal canine cardiac
muscle: evidence for recurrent discontinuities of intracellular resistance that affect the membrane currents."
Circ Res 48: 39.
Spach, M., Miller and E. MillerJones (1979). "Extracellular potentials related to intracellular action potentials
during impulse conduction in anisotropic canine cardiac muscle." Circ Res 45: 188204.
Stern, M., M. Capogrossi and E. Lakatta (1988). "Spontaneous calcium release from the sarcoplasmic reticulum
in myocardial cells mechanisms and consequences." Cell Calcium 9: 247256.
Varnava, A., P. Elliot, S. Sharma, McKenna and M. Davies (2000). "Hypertrophic cardiomyopathy: the inter
relation of dissarray, fibrosis, and small vessel disease." Heart 84: 476482.
Viswanathan, P. and Rudy (1999). "Pause induced early afterdepolarizations in the long QT syndrome: a
simulation study." Cardiovascular Research 42: 530542.
Viswanathan, P. and Y. Rudy (2000). "Cellular Arrhythmogenic Effects of Congenital and Acquired LongQT
Syndrome in the Heterogeneous Myocardium." Circulation 101: 1192.
Wagner, M., Gibb and M. Lesh (1995). "A model study of propagation of early afterdepolarizations." IEEE
Trans Biomed Eng 42(10): 991997.
Wang, Q., M. Curren, 1.Splawski, T. Burn, J. Millholland, T. VanRaay, J. Shen, Timothy, G. Vincent, T. d. Jager,
P. Schwartz, J. Towbin, A. Moss, D. Atkinson, G. Landes, T. Connors and M. Keating (1996). "Positional
cloning of a novel potassium channel gene: KVLQTI mutations cause cardiac arrhythmias." Nature Genet.
12: 1723.
Wei, D., N. Miyamoto and S. Mashima (1999). "A computer model of myocardial disarray in simulating ECG
features of hypertrophic cardiomyopathy." Jpn Heart J 40(6): 819826.
Wei, D., G. Yamada, T. Musha, H. Tsunakwa and K. Harmumi (1990). "Computer simulation of supraventricular
tachycardia with the WolffParkinsonWhite Syndrome using threedimensional heart models." J Electrocar
dioI23(3): 261273.
Weidmann, S. (1955). "The effect of the cardiac membrane potential on the rapid availability of the sodiumcarrying
system." J Physiol127: 213224.
42 N. V. Thakor, V.Iyer, and M. B. Shenai
Weiss, 1., N. Venkatest and S. Lamp (1992). "ATPsensitive K+ channels and cellular K+ loss in hypoxic and
iscaemic mammalian ventricle." J Physiol (Lond.) 447: 649673.
Winslow, R., D. Scollan, J. Greenstein, C. Yung, W. Baumgartner, G. Bhanot, D. Gresh and B. Rogowitz (2000).
"Mapping, modeling, and visual exploration of structurefunction relationships in the heart." IBM Systems
Journal 40(2): 118.
Winslow, R., D. Scollan, A. Holmes, C. Yung, 1. Zhang and M. Jafri (2000). "Electrophysiological Modeling of
Cardiac Ventricular Function: From Cell to Organ." Ann Rev Biomed Eng 2: 119155.
Wit, A. and M. Janse (1993). The Ventricular Arrhythmias of Ischemia and Infarction: Electrophysiological
Mechanisms. Mount Kisko, Futura Pub. Co.
Xie, E, Z. Qu, A. Garfinkel and J. Weiss (2001). "Effects of ischemia on spiral wave stability." Am Physiol
Heart Cire PhysioI280(4): HI66773.
Yan, G., Yamada, A. Kleber, J. McHowat and P. Corr (1993). "Dissociation between cellular K+ loss, reduction
in repolarization time, and tissue ATPlevels during myocardial hypoxia and ischemia." Cire Res 72: 560570.
Zhang, H., Holden, 1. Kodama, H. Honjo, M. Lei, Varghese and M. Boyett (2000). "Mathematical models of
action potentials in the periphery and center of the rabbit sinoatrial node." Am J Physiol Heart Circ Physiol
279(1): H397421.
THE FORWARD PROBLEMOF
ELECTROCARDIOGRAPHY:
THEORETICAL UNDERPINNINGS
ANDAPPLICATIONS
Ramesh M. Gulrajani
Institute of Biomedical Engineering, Universite de Montreal
2.1 INTRODUCTION
The forward problem of electrocardiography refers to the calculation of the potentials on the
body surface due to the heart sources, using the theoretical equations of electromagnetism.
As a prerequisite for this calculation, suitable representations of the heart sources and of
the torso geometry are needed. The former is usually assumed to be a current dipole, which
may be taken to be a current source and sink of equal magnitude / separated by a very small
distance 8. The dipole is then represented as p /8. The bold font indicates that p is a
vector, whose magnitude is /8 and whose direction is that of the vector 8, namely along
the line joining sink to source. The rationale behind representing the heart sources with
a current dipole is taken up in Section 2.2 below. In a second approach, the question of an
adequate representation of the heart sources is circumvented by calculating the torso surface
potentials using the actual potentials on the heart's epicardial surface (or more correctly
on the surrounding pericardial sheath) as the starting representation. This second approach
will also be described.
Torso geometry is nowadays modeled as a threedimensional computer representation
of the external torso surface and its internal inhomogeneities of differing conductivities.
Prior to the advent of the computer, the torso was often modeled as a sphere or a cylinder, and
analytic expressions for the potential due to a current dipole within such a sphere or cylinder
used to compute the surface potential. We do not consider such analytic solutions of the
forward problem here, nor do we consider other early analog solutions, in which the torso
Address for Correspondence: Ramesh M. Gulrajani, Institute of Biomedical Engineering, Universite de Montreal,
P.O. Box 6128, Station Centreville, Montreal, (Quebec) H3C 317, CANADA. Telephone: (514) 3435705,
Fax: (514) 3436112, Email: gulrajan@igb.umontreal.ca
43
44 R. M. Gulrajani
was represented by a realisticallyshaped physical analog and potentials due to an inserted
artificial dipole actually measured on the surface of this analog. Reviews of these analytic
and analog solutions are to be found in Gulrajani et al. (1989). Thus this chapter focuses
exclusively on numericallycomputed solutions of the forward problem and the theoretical
equations underlying these solutions. Generally, it is the farfield forward problem that
is considered, where the desired potentials on the body surface are assumed far enough
from the heart sources to permit the use of simple dipole models for the latter. A more
challenging problem occurs when nearfield potentials on or within the heart are desired.
There is much current interest in at least being able to obtain a first approximation to the
epicardial potential distribution on the heart surface, and this estimation is also considered.
The chapter concludes with a section describing the more common applications of the
forward problem. Other recent reviews of the forward problem and its applications are to
be found in Gulrajani (1998a; 1998b).
2.2 DIPOLE SOURCE REPRESENTATIONS
Prior to discussing the dipole source representations used in the forward problem,
we present the fundamental equations that form the biophysical underpinning of potential
calculations.
2.2.1 FUNDAMENTAL EQUATIONS
The biological current sources in the heart are the ionic currents that flow across the
surface membrane of the individual heart cells into the extracellular space. If we denote by
the net source current density (in A/m at a point characterized by the spatial vector
r, then we can write the equation
(2.1)
Equation (2.1) states that the total current density J is expressed as the sum of the source
current density J
s
, if present, and the conduction current density a E, where E is the electric
field and the conductivity. assumes that quasistatic conditions apply whereby capaci
tive, inductive and propagation effects are all neglected (Plonsey, 1969), and field quantities
at a given instant are determined by just considering the source currents J, existing at the
instant in question. Under these quasistatic conditions, the divergence of the total current
V . 0, so that taking the divergence of Eq. (2.1) yields
J, (2.2)
Using the relation E V<1>, where denotes the potential, Eq. (2.2) can be transformed
to the fundamental equation that governs the relationship between electrocardiographic
potentials and heart sources, namely
Js (2.3)
The Forward Problem of Electrocardiography
p
FIGURE 2.1. Aheartof volume andsurroundingepicardialsurface existsinaninfinitemediumofuniforrn
conductivity a. The potentialis sought at an arbitrarypoint P characterized by the positionvector r. See text for
additional details.
If now the conductivity is assumed constant everywhere, i.e. the medium is infinite and
homogeneous, Eq. (2.3) reduces to Poisson's equation with the solution
<I>(r) Js(r')
41Ta
VH
(2.4)
Equation (2.4) gives the potential at a fixed observation point, P,characterized by the position
vector r. It entails performing a spatial integration over the heart volume Y (Fig. 2.1), and a
dummy variable of integration that traverses the source coordinates has been introduced.
The primes on V' and dY'are used to reinforce the point that it is that is the variable and
that all spatial derivatives need to be evaluated with respect to
The membrane source currents can be expressed in an alternative form by the relation
Isv(r') == Js(r') (2.5)
where I
sv
denotes the source volume current density in Nm This relation follows from
Gauss' law for the current flux that leaves a source I
sv.
We have
I,w(r')dY' odS'
V
H
SE
(2.6)
where 0 is the unit normal to the epicardial surface S that surrounds the heart volume Y
Applying the divergence theorem to the righthand side of (2.6), and substituting for
V' . (o E) from Eq. (2.2), we immediately obtain the equivalence relation of Eq. (2.5).
Accordingly, Eq. (2.3) can also be written as
sv
(2.7)
46 R. M. Gulrajani
By using the expression for the divergence of the product of a scalar and a vector,
namely V . (¢A) V¢ . A ¢ (V . A), we can rewrite Eq. (2.4) as
<l>(r) If J s(r' ) . dV'  f Js(r' ) ) dV'] (2.8)
4na
H ~
The divergence theorem can be applied to the second integral on the right converting it to
a surface integral over (Fig. 2.1), where the heart sources J, vanish. Thus the second
integral on the right is zero, and we have
<l>(r) = _1_ f Js(r') . V' (_1) dV'
4na
V
H
(2.9)
Now the potential due to a current dipole p situated at r' in an infinite homogeneous medium
of conductivity a is given by
1 , ( 1 )
<l> (r)
4na
(2. 10)
Comparing Eqs . (2.9) and (2. 10), we see that the heart current sources generate a dipolar field
and that Jscan also be interpreted as a current dipole density. Note that this interpretation
hinges on the validity ofEq. (2.4), which in turn is only true if a is homogeneous everywhere
and Eq. (2.3) reduce s to Poisson ' s equation.
2.2.2 THE BIDOMAIN MYOCARDIUM
A major difficulty occurs in estimating the ionic current density J, which is clearl y
impossible to do in a multicellular preparation like the heart . This leads to the notion of
determining an equivalent source J
eq
for the heart, one that can replace the true source
Js at least as far as the calculation of the farfield torso surface potentials is concerned. A
necessary prerequisite to obtaining this equivalent source is a more macroscopic view of the
heart , one that smooths out the detail of the individual heart cells and their ionic currents.
The individual heart cell is an elongated structure with approximate dimensions of 15 x
15 x 100 J.1m (Fig. 2.2). is made up of "sarcomeres" (from Z line to Z line), within each
of which are the interdigitating contractile myofibrils. The cells are electrically connected
to one another in the longitudinal direction by gap junctions present in the intercalated
disks separating the cells (Fig. 2.2), in effect forming long fibers. The cells also branch at
irregular intervals, and this result s in electrical continuity ofthe intracellular space in the two
transverse directions also. In effect, the entire intracellular space forms a continuum or, in
biological terms , a syncytium. Surrounding the invididual cells is the extracellular or, what is
termed in cardiac electrophysiology, the interstitial space. This interstitial space too forms
a second syncytium. The macroscopic view of the heart mentioned earlier then consists of
these two syncytia or domain s, that are both assumed to occupy the same volume (Schmitt,
1969; Tung, 1978; Miller and Geselowitz, 1978). The two domains do not exist in isolation,
The Forward Problem of Electrocardiography
Mitochondria
Capillary
47
FIGURE 2.2. Diagram of cardiac muscle fibers illustrating the characteristic branching, the intercalated disks,
and the internal myofibrils. Individual cells are made up of sarcomeres. A sarcomere occurs from Z line to Z line,
with the M lines lying at the midpoint of each sarcomere. Cells are separated longitudinally by the intercalated
disks. Other indicated structures are the blood capillaries and the mitochondrial cells that provide the energy
required by the contracting fibers. Reproduced, with permission, from Pilkington and Plonsey (1982). ©IEEE.
Modified, with permission, from Berne and Levy (1977).
however, but are coupled at every point in space by the continuity of the transmembrane
current of the individual cells which flows out of the intracellular domain and into the
interstitial one.
2.2.2.1 Equations for an Isotropic Bidomainthe Uniform Dipole Layer
Due to the infrequent branching of the cardiac cells in the transverse directions, the
electrical conductivity in the transverse directions is much less than that in the longitudinal
direction, and the myocardium is profoundly anisotropic. Nevertheless, for simplicity, in
this subsection we assume the myocardium to be electrically isotropic. This restriction will
be removed in the next subsection.
The assumption that both intracellular and interstitial domains occupy the same total
heart volume leads to the notion of effective conductivities. In the intracellular space, the
effective conductivity gi is given by gi I.«. where a, is the real intracellular conductivity
and is the fraction of the total crosssectional area occupied by the intracellular space. This
reduced effective conductivity ensures that the intracellular conductance remains the same
despite the increase in assumed intracellular crosssectional area. Similarly, the effective
interstitial conductivity ge )(Je, where (Je is the real conductivity of the interstitial
space.
With these effective conductivities, we may write the following equations for the
intracellular and interstitial domains, respectively:
v .(gi V<Pi)
mv
V· (geV<Pe) I
mv
(2.11)
(2.12)
Equations (2.11) and (2.12) are of the form of Eq. (2.7) and govern the intracellular and
48 R. M. Gulrajani
interstitial potentials, and
e
, respectively. A volume source current formulation has
been used so that I
mv
is the membrane current (in Nm that flows out of the intracellular
domain and into the interstitial one. It therefore acts as a sink in the intracellular domain
and is considered negative in Eq. (2.11), but is a source of equal magnitudethat appears in
the interstitial domain and is taken as positive in Eq. (2.12). This macroscopic membrane
current,
mv
, which is impossible to measure, may be eliminated by combining Eqs. (2.11)
and (2.12) to yield
(2.13)
Upon adding V .
e
) to both sides of Eq. (2.13), and utilizing the usual definition of
the transmembrane potential
m e
, we get a second form ofEq. (2.13), namely,
(2.14)
where g s. ge' A third form is obtained by multiplying Eq. (2.14) by gelg so that
(2.15)
(2.16)
Eq. (2.15) becomes
(2.17)
The similarity between Eqs. (2.17) and (2.3) suggests that Jeq can serve as an equivalent
source for computing the interstitial potential
e
. Note that J
eq
only exists whenever the
spatial gradient of Ym is nonzero, i.e., only in regions of the myocardiumthat are undergoing
excitation or repolarization. Furthermore,
eq
, unlike the true sources
s
' can be computed
if the effective conductivities of intracellular and interstitial domains can be estimated and
if the spatial distribution of the transmembrane potential in the myocardium is known.
Assuming for the moment an infinite homogeneous myocardium, the solution to Eq. (2.17)
is similar to that of Eq. (2.3) and is given by (compare with Eq. 2.4)
<P(r) Jeq(r')dY'
e
V
H
Exactly as Eq. (2.4) was rewritten as Eq. (2.9), we may rewrite Eq. (2.18) as
<P(r) _1_ f Jeq(r') . V' (_1)
e
V
H
(2.18)
(2.19)
The Forward Problem of Electrocardiography 49
thereby identifying Jeq as an equivalent currentdipole density. Once again this interpretation
only holds if the interstitial conductivity ge is homogeneous. Alternative formulations for
Jeq based on Eqs. (2.13) and (2.14) are also possible. For the former,
(2.20)
and acts in an interstitial domain of homogeneous effective conductivity for the latter,
(2.21)
and acts in an interstitial domain of homogeneous effective conductivity
The above field equations were written assuming an infinite homogeneous extent for
the bidomain myocardium. With the finite heart in the torso, this is evidently not the case.
In particular, we have a bidomainmonodomain interface between heart and torso, and
deciding which boundary conditions apply at this interface is not immediately evident.
Three boundary conditions are needed. The first two are the continuity of the interstitial
potential <I>e at the epicardial surface to the torso potential just outside the heart, and of
the normal component of the total current that crosses over from the heart to the torso.
mathematical terms, these two conditions at the hearttorso interface may be expressed as
<l>e =
o<l>e
gi  ge =0'0
on on on
(2.22)
(2.23)
where the normal derivative o<l>jon denotes the component of the gradient V<I>· n along
the outward normal n, and 0'0 denotes the torso conductivity just outside the interface. The
third boundary condition generally used is that the intracellular current stops at the heart
surface (Tung, 1978; Krassowska and Neu, 1994), so that we have,
(2.24a)
With this condition, only the normal component of the interstitial current in Eq. (2.23)
crosses over into the torso. An alternative third boundary condition (ColliFranzone et al.,
1990) that has sometimes been employed is
ae,
g =0'0
on on
(2.24b)
This formulation is particularly convenient if
eq
is given by Eq. (2.21), since the interstitial
medium in which this
eq
acts has conductivity and Eq. (2.24b) then simply expresses the
continuity of the normal component of the current from this equivalent interstitial medium
into the torso.
Equations (2.16), (2.20) and (2.21) all identify Jeq as an equivalent currentdipole den
sity per unit volume that exists wherever a spatial gradient oftransmembrane or intracellular
potential is present, for example, in the vicinity of a propagating excitation wavefront. An
alternative equivalent surface currentdipole density that is placed on this excitation front
50 R. M. Gulrajani
FIGURE 2.3. Diagram illustrating the spatial distribution of the transmembrane potential as an excitation wave
front sweeps the myocardium along the direction indicated by the unit vector n. Reproduced, with permission,
from Gulrajani (l998a).
can be derived using Eq. (2.16). Figure 2.3 shows such a propagating action potential wave
front of spatial extent d. Ahead of this front, the myocardial cells are at rest with V
m
V"
where V denotes the resting transmembrane potential. Behind the front, the cells are de
polarized with V
m
V
d,
where V
d
denotes the depolarized transmembrane potential. we
apply Eq. (2.16) to the transition region between the dashed lines, the total dipole moment
p, associated with the wavefront of area is
p Jeq x Volume geqVV
m
x Ad
~ geq(V Vr)An geq VpAn (2.25)
In Eq. (2.25), the gradient VV
m
has been approximated by (Vp/d)n, where V
p
== Vd  V
r
is
the amplitude of the propagating action potential and the unit vector n denotes its direction
of propagation. Equation (2.25) holds in the limit that the propagating wavefront is assumed
infinitely thin and allows us to identify a surface dipole layer associated with the wavefront
whose density is geq Vpn. Dipole orientations within the layer are everywhere normal to the
wavefront, and, if the action potential amplitude is uniformeverywhere, then so is the dipole
density, whence the term uniform dipole layer used to describe this equivalent source.
2.2.2.2 Equations for an Anisotropic Bidomainthe Oblique Dipole Layer
Equations (2.11) and (2.12) are rewritten for a homogeneous anisotropic bidomain by
replacing the scalar effective conductivities gi and ge by their tensor equivalents G and G
e,
respectively. We get,
v .
mv
(GeV<I>e) I
mv
(2.26)
(2.27)
The Forward Problem of Electrocardiography 51
G; and G
e
are now the diagonal intracellular and inter stitial effect ive conductivity tensors,
respectively, and can be written in 3 x 3 matrix form,
(2.28)
where it is assumed that the cardiac fibers are oriented along the z axis. Although not strictly
true (see Hooks et al. , 2002), symmetry about this axis is generally assumed so that g;x g;y
and gex gey' Note that G; V<1>; and G
e
V<I>e in Eqs . (2.26) and (2.27), respectively, are
vectors, being the product ofa 3 x 3 tensor matrix with a 3 x 1column matrix representation
of the gradient vector.
A major complication arises on account of the fiber rotation present in the real heart,
which leads to an inhomogeneous myocardium. Under these circumstances, the conductivity
tensors are diagonal only in a local coordinate system, characterized by the unit vectors
e" e2, e3, where e3 is always oriented along the fiber direction. These local unit vectors
may be expressed in terms of fixed global unit vectors e. , e.g.,
(2.29a)
and reciprocally, the global unit vectors expressed in terms of the local unit vectors, e.g.,
(2.29b)
Accordingly, a rotation matrix A, characterizing the transformation from local to global
coordinates, may be written as
(2.30)
with the transposed matrix characterizing the reverse transformation from global to
local coordinates. From the mathematical definition of a Cartesian tensor (Fung, 1977),
it can be shown that the diagonal conductivity tensors G; and G
e
in the local coordinate
system transform to the symmetric conductivity tensors G; and G ~ , respectively, in the
global coordinate system that are given by the matrix products
(2.31)
Equations (2.26) and (2.27) continue to hold in global coordinates, but with the primed
conductivity tensors G; and G ~ replacing the unprimed tensors G; and G
e
. We have,
V · (G;V<I>; ) =
mv
V . ( G ~ V<I>e)
mv
(2.32)
(2.33)
52 R. M. Gulrajani
Note that the primed conductivity tensors vary from point to point in the myocardium with
the change in fiber orientation.
As for the isotropic case, the volume current density
mv
may be eliminated from
Eqs. (2.32) and (2.33), to yield
(2.34)
Upon adding V .
e
) to both sides ofEq. (2.34), we get
(2.35)
where G' G; G ~ is the bulk myocardium conductivity tensor. Equations (2.34) and
(2.35) are the equivalents ofEqs. (2.13) and (2.14) for the isotropic myocardium. Since G;
and G ~ are tensors, there is no equivalent of Eq. (2.15). Deducing that Jeq G;V<Pi is
an equivalent currentdipole density that may be used to compute the interstitial potential
<Pe is only valid, however, under the assumption that G ~ is constant, since it is only then
that Eq. (2.34) simplifies to Poisson's equation. Similarly, from Eq. (2.35), we obtain
eq
G;VV as an equivalent currentdipole density that can be used to compute <P
under the assumption that the bulk conductivity G' is constant. Despite their obvious
invalidity, the assumptions that G ~ and/or G' are constant are used for want of anything
better; indeed often they are not only assumed constant, but isotropic as well. Only
with these assumptions, can we consider J
eq
G;V<Pi and/or J
eq
G;VV as the
respective equivalent currentdipole densities.
When G; is an invariant scalar, then as shown in the previous subsection, Jeq
V reduces to a uniform dipole layer on the action potential wavefront that is normal
to the wavefront (along the direction 
m
) . In general, however, with G; a tensor, we
obtain a dipole layer that is oblique to the wavefront. Moreover, the strength of this dipole
layer is not uniform, but varies in both magnitude and orientation from point to point along
the wavefront due to the variation in G;. The boundary conditions applicable at the heart
torso interface remain the same as before, and the equivalent forms of Eqs. (2.22), (2.23),
(2.24a) and (2.24b) are, respectively,
<Pe <Po
n G;V<Pi n G ~
e
n GoV<Po
0
n
e
n GoV<Po
(2.36)
(2.37)
(2.38a)
(2.38b)
In Eqs. (2.37) and (2.38), each of the terms is a triple matrix product, involving a row
matrix a 3 x 3 conductivity matrix, and a gradient column matrix. That these Equations
do express the continuity of the normal components of currents is evident if we note, for
example, that the triple matrix product n GoV<Po can also be written as the scalar product
n . (GoV<po) of two vectors n and GoV<Po.
More insight regarding the oblique dipole results if we look at its components in
local coordinates. The oblique dipole now becomes
eq
VV
m
where is diagonal.
Assuming again transverse symmetry of the zoriented cardiac fibers so that gix giy gil
The Forward Problem of Electrocardiography 53
and where and are the transverse and longitudinal effective conductivities,
then in a manner similar to that in which Eq. (2.25) was derived, we can show that the dipole
moment p associated with a wavefront of area A is given by
where n nz and n3are the components of the normal n to the wavefront in local coordinates.
The above equation can be rewritten as
(2.39)
revealing that the oblique dipole consists of a component normal to the wavefront plus a
second axial component along the fiber direction (Corbin and Scher, 1977; ColliFranzone
et al., 1982; 1983).
2.3 TORSO GEOMETRYREPRESENTATIONS
Torso geometry may be represented either by a numerical discretization of its different
interfaces, or by a full threedimensional discretization of its entire volume. The choice is
dictated by the solution methodology to be used (see Section 2.4). Thus in Figure 2.4, which
shows a stylized torso, with interface discretization only the outer torso surface So, the lung
surfaces and the outer heart surface S3, and the bloodfilled heart cavities and
are triangulated. The intervening regions are assumed of constant isotropic conductivity.
the anisotropicconductivity skeletal muscle layer underlying the skin (not shown in
Fig. 2.4) is to be included, it is first converted to an approximatelyequivalent isotropic
layer (McFee and Rush, 1968) prior to the triangulation of its interfaces. With volume
discretization, the entire threedimensional torso volume is modeled either in pointwise
fashion as a collection of equispaced points, or in piecewise fashion by a combination of
tetrahedral and hexahedral (brickshaped) elements. The major advantage of these volume
representations is the ability to accurately model not only anisotropic conductivity regions,
but also regions of continuously varying conductivity.
2.4 SOLUTIONMETHODOLOGIES FORTHEFORWARD PROBLEM
Calculation of the torso potentials from heart source dipoles is done via one of two gen
eral approaches, namely surface methods or volume methods. As may be surmised, surface
methods employ torso models with only the interfaces discretized, and obtain the potentials
only on these interfaces. Volume methods, on the other hand, use full threedimensional
discretizations of the torso volume and obtain the potential everywhere. The number of
potentials to be evaluated is accordingly much greater with volume methods, leading to
large coefficient matrices. However, the potential at each point is expressed only in terms of
its nearest neighbors, so that the matrices are sparse and may be inverted via sparse matrix
solvers. With surface methods on the other hand, while the coefficient matrices are much
smaller, the potential at any interface point is coupled to the potential at every other interface
54 R. M. Gulrajani
FIGURE 2.4. Torso with multiple regions of differing isotropic conductivity. See text for more details.
point, with the result that the coefficient matrix is fully populated and sparse matrix routines
are inapplicable. Surface methods are also often termed "boundaryelement methods" in the
literature.
2.4.1 SURFACE METHODS
2.4.1.1 Solutions from Equivalent Dipoles
Surface methods usually employ integral equations for the potentials to be calculated.
These integral equations may be derived by applying Green's second identity to the torso
geometry shown in Fig. 2.4, to obtain for the potential at the observation point on
interface (Barr et al., 1966),
The Forward Problem of Electrocardiography 55
where Jeq(r') denotes the equivalent dipole sources present in the heart myocardium
the torso interfaces S/ extend from 0 to (with internal and external conductivities (f/ and
(f/+, respectively), and dQrr' denotes the solid angle subtended at the observation point r
by an element of the surface integral d S' at (dQ
rr
, I ~ ~ ~ ; . tsd The term involving
the socalled "auto solid angle" (dQrr), representing the solid angle subtended at r by the
surface element containing r, is excluded from the summation. The contribution of this
auto solid angle term is 2rr if the surface around the observation point r is smooth, and has
already been incorporated in the derivation of Eq. (2.40). The point to note is that the first
term on the righthand side in Eq. (2.40) is proportional to the infinitemedium potential
due to the equivalent sources Jeq(r') (see Eq. 2.9). The second term explicitly represents the
effect of the different torso interfaces.
The simplest assumption with triangulated torso interfaces is to consider the potential
as constant over each triangle face. With this constant potential assumption, the observation
point r may be placed at the centroid of each triangle, and an equation such as Eq. (2.40)
can be written as r is moved from triangle to triangle. The ensemble of NT equations, where
NT is the total number of triangles, may be written in compact matrix form as
(2.41)
where is now an NT x column matrix containing the triangle potentials, G is an NT x
matrix containing the first terms on the right ofEq. (2.40), and A is an NT x NT weighted
solid angle matrix (with diagonal terms zero on account of the elimination of the auto
solid angle term from the summation on the righthand side in Eq. 2.40) that depends only
on torso geometry and conductivities. Equation (2.41) can be rewritten as G,
where is an NT x NT identity matrix. A straightforward solution would entail inverting the
coefficient matrix to get = Unfortunately, this is not possible as the
matrix (I  A) is singular and does not possess an inverse. The singularity can be demon
strated mathematically (Lynn and Timlake, 1968), but it can be understood by realizing that
potential measurements are never absolute but always with respect to a reference. Thus the
potential is always indeterminate up to the constant value chosen for the reference potential,
and mathematically the singularity of the coefficient matrix ensures that a unique solution
for is not possible unless a choice for the reference potential is first made. In practice, the
problem is circumvented by replacing A by a "deflated" matrix A (the deflation procedure
entails removing an eigenvalue x 1 of A), so that we end up solving A G.
The coefficient matrix (I  A is no longer singular, and accordingly this equation can be
inverted. Note, however, that is not the original potential on account of the changed
coefficient matrix. can be shown that for triangles on the outer torso surface
but for internal interfaces and differ by a constant. can also be shown that the
deflation procedure sets the sum of the triangle potentials on to zero, in effect establishing
a zero reference for the potential equal to the mean of the triangle potentials on
A more accurate procedure with triangulated torso interfaces is to assume a linear
variation of the potential over each triangle interface, so as to better represent the varying
interface potential. This leads to placing the observation point s r at triangle vertices rather
than at triangle centroids. One immediate advantage is that for closed triangulated surfaces,
the number of vertices is approximately half the number of triangles, so that the number
of unknown potentials to be determined is also approximately half, leading to smaller
56 M. Gulrajani
matrix sizes . Alternatively, if the computer power is already present, we can use a finer
triangulation doubling the number of triangle vertices and still keeping the number of
unknown potentials the same. This vertex approach entails a slight modification of the
governing integral equation (Eq. 2.40) as the auto solid angle is no longer 2][. This is because
with an observation point r at a vertex, any small selected neighborhood around r is no
longer smooth, but subtends a difficultto calculate auto solid angle Q , s,at r . Equation (2.40)
may be rewritten, but this time explicitly in terms of Q, s, . We get,
[a
k
Q,sJ  ak+Q, s,l<f>k(r) Jeq(r' ) . V' (Ir ~
V
H
(2.42)
As before, the summation on the righthand side excludes the auto solid angle term, as this
has already been moved over to the lefthand side. An equation such as Eq. (2.42) can be
written as the observation point moves from vertex to vertex, and the ensemble of equations
combined in matrix form. The terms on the lefthand side of this ensemble will eventually
form the diagonal terms of the coefficient matrix that multiplies cPo Since the auto solid
angle Q, s, is difficult to compute, and it only occurs in these diagonal terms, a common
approximation is to set each diagonal term to be equal to the negative sum of the other
terms in its row. Thi s introduces a linear dependency between matrix columns, rendering
the coefficient matrix singular, as mandated by the nonunique nature of the potential. Other
approximations for Q,s, are discussed in Meijs et (1989), Heller (1990) and Wischmann
et (1996) . As with the earlier centroid option, deflation is also needed with this vertex
opt ion for potential sites.
An interesting variant ofthe traditional approach of expressing J
eq
as the gradient ofthe
transmembrane potential was derived by Geselowitz (1989) for the special situation of equal
anisotropy, i.e., when the intracellular and interstitial conductivity tensors are proportional.
In particular, for a heart with isotropic intracellular and interstitial conductivities (which is a
degenerate case of equal anisotropy), Geselowitz showed that the equivalent dipole surface
density could be represented as
(2.43)
where is a unit dipole perpendicular to all heart surfaces, epicardial as well as the
endocardial surfaces adjacent to the blood masses (surfaces 5
3
,5
4
and 55 in Fig. 2.4).
Using this equivalent source representation simply converts the volume integral on the right
hand side ofEq. (2.40) to a surface integral over these three heart surfaces. we denote the
union of these three surfaces by 5 we get
(2.44)
The Forward Problem of Electrocardiography 57
A similar conversion of the volume integral on the righthand side of Eq. (2.42) also results
when Eq. (2.43) is used as an equivalent source.
2.4.1.2 Solutionsfrom Epicardial Potentials
As mentioned in the Introduction, the forward problem can also be solved using the
heart's epicardial potentials as the starting point (Barr et al., 1977). The governing equations
may be determined by applying Green 's theoremto a torso model containing only the heart's
epicardial surface S3and the outer torso surface So (Fig. 2.4). By allowing the observation
point r to first approach So, and then S3, the following two equations are obtained:
f 1 , ,
<1>B (r ) <1>E . DdS
4n r
BE
<1>EV  , . DdS   <1>BV , . DdS
4n r
BE
4n r
BB
~ ~
1 1 , ,
<1> E(r)  ,V <1> E . DdS
4n r
EE
<1>EV  , . DdS   <1>BV  , . DdS
4n r
EE
4n rEB
~ ~
(2.45)
(2.46)
In Eqs. (2.45) and (2.46) , we have explicitly denoted epicardial and body surface po
tentials by the SUbscripts and respectively, and the corresponding surfaces by
SE(= S3) and SB(= So), respectively. The unit normals to these surfaces are always
outward. The scalar distance [r  r' 1is now denoted by r ~ E ' etc. , with the first subscript
denoting the location of r and the second that of and the prime on r ~ E is used to simply
reinforce the fact that the variable of integration is r'. Two sets of matrix equations result as
the observation point r is moved from triangle to triangle on the body and heart surfaces,
and these may be written in compact fashion as
A B B ~ B + A B E ~ E +BBEr
E
0
A E B ~ B A E E ~ E BEErE 0
(2.47)
(2.48)
In Eqs. (2.47) and (2.48), ~ B and ~ E are column matrices of body surface and epicar
dial potentials, T is a column matrix of epicardial potential gradients, and the A and B
coefficient matrices depend solely on integrations involving epicardial and body surface
geometries. The first subscript on A (or B) indicates the surface on which the observation
points are selected, and the second subscript whether the integration is over the epicardial
or body surface. Equation (2.47) may be used to obtain an expression for which when
substituted in Eq. (2.48) yields:
(2.49)
The elements of T
BE
=[A
BB
 BBE(BEE)IAEBr l[BBE(BEE)IAEE  ABE] are the
transfer coefficients relating the potential at a particular epicardial point to that at a particular
58 R. M. Gulrajani
torso surface point. As was the case for forward solutions using current dipoles , these points
may be selected at triangle centroids, implying a potential that is constant over each triangle,
or at triangle vertices with a potential that varies linearly over each triangle.
2.4.2 VOLUME METHODS
2.4.2.1 FiniteDifference Method
The finitedifference method represents the torso volume by a threedimensional array
of regularlyspaced points or nodes that are connected to each other by intervening resis
tors, whose values are selected to best reflect the intervening resistance between the points.
Kirchhoff's current law is written for each node, resulting in a large set of equations relating
the potentials between adjacent nodes . In effect, the method represents a discrete approx
imation to the governing differential equation (Eq. namely
sv
with
sv
= J
eq,
where Jeq denotes the equivalent heart sources . The solution of the resulting
set of equations evidently depends on the fineness of the node spacing, and the accuracy
with which the resistors represent the torso resistances. The equations are usually solved
by GaussSeidel iteration with successive overrelaxation. The main drawback is the slow
convergence. As with other volume methods , the finitedifference method can handle vary
ing anisotropic conductivities. A good illustration of its application to electrocardiography
is to be found in Walker and Kilpatrick
2.4.2.2 FiniteElement Method
Here, the torso volume is represented by contiguous threedimensional tetrahedra
and/or hexahedra (bricklike elements). The finiteelement method also solves Eq.
with
s v
V .
eq
. The starting assumption is that the potential within each element can
be approximated by
where <l>i denotes the potential at an element node, is an appropriate interpolation poly
nomial (usually linear in x, y, and z, for tetrahedral elements), and r is the number of nodes
for the element. Each is equal to unity at node and is zero at all other element nodes.
we substitute Eq. into Eq. then on account of the approximation, we get
I
s v
R
where a is the element conductivity and R denotes a residual. The technique of weighted
residuals (Brebbia and Dominguez, is now invoked in an attempt to reduce R to zero,
but in a "weak form" by reducing the set of weighted integrals below to zero:
Isv]WidV = RWidV =
v v
= r
The Forward Problem of Electrocardiography 59
In Eq. (2.52), each Wi is a weighting polynomial, and the integration is over the element in
question. Often Wi is set equal to the interpolating polynomial f3i, and we get the socalled
Galerkin weightedresidual formulation, namely
(aV'¢)]f3i f3;Isv 0
The first integral is now integrated by parts to obtain
i r (2.53)
f3i(aV'¢). (aV'¢)· V'f3i f3;Isv
S v v
i r (2.54)
where the surface integral is over the bounding surfaces of the element and 0 is the unit
outward normal. Consider, initially, internal volume elements that do not abut the outer
torso surface So. The contribution of the surface integral in Eq. (2.54) will eventually be
cancelled by similar terms from contiguous elements on account of the continuity of the
normal component of the current, and because the f3i'S are selected such that the potential
at a common interface is only determined by nodes on that interface. This last is done to
ensure the continuity of the potential across the interface between elements. Both volume
integrals in Eq. (2.54), however, need to be considered, with that on the righthand side only
contributing for elements where I
sv
is nonzero. Assuming tetrahedral elements and linear
polynomials for the f3i, the set of r equations in Eq. (2.54) may be written in linear matrix
fonn
(2.55)
where A is an r x r coefficient matrix, and are r x 1 column matrices, and the
superscript is used to denote that Eq. (2.55) holds for a particular element. The matrix
ep(e)containsthe element potentials to be computed, and is the matrix representation of
the source terms and surface integrals of Eq. (2.54).
The coefficient matrices from the different elements, internal as well as those that abut
the outer torso surface, may be combined to result in a global matrix equation
(2.56)
where now A is m x m if there are m unknown potentials in the torso to be determined, and
where epand F are each m x 1. Surface integrals from contiguous elements will now cancel.
For elements that abut the outer torso surface So, the surface integrals over the sides that
form a part of will remain uncancelled. These uncancelled surface integrals yield the set
of surface integrals over
f3i(a S, i m (2.57)
60
the boundaryconditions applicableon are of the mixedform
= a on
DdS = indS on
R. M. Gulrajani
(2.58a)
(2.58b)
where a denotes the knownpotential on a portion of and in is the injectednormal
currentdensityovertheremainingportion thensincethefJi reducetozeroover there
beingnounknownpotentialsthere, Eq. (2.57) needonlybe integratedover Substituting
Eq. (2.58b) in Eq. (2.57), the uncancelledsurfaceintegralsbecome
1,2, . . . . m (2.59)
Only the integralsfor correspondingto surfacenodes over contribute. In buildingEq.
(2.56), theyformanother sourcetermdue to the injectedcurrent, andaddto the appropriate
term in the matrix F. Thus, the socalled Neumannboundarycondition of Eq. (2.58b)
enters naturally into the finiteelement formulation. The Dirichlet boundary condition
of Eq. (2.58a), however, has to be introduced explicitly into the global matrix equation
(Eq. 2.56). Thus if the potential at node is a, then all matrixelements in row are set
equal to zero, except akk, which is set equal to unity; in addition fk is set equal to This
also renders A nonsingular. Since A is large, solutions to Eq. (2.56) may be obtained by
iterative techniques, though many finiteelement packages have direct solvers that exploit
the sparse nature of A.
Finally, the finiteelement methodcan alsobe usedtocomputebody surfacepotentials
from epicardial potentials, in which case since the heart region is excluded, the volume
integral on the righthand side of Eq. (2.54) drops out. The uncancelledsurface integrals
of Eq. (2.57) are also zero since no current leaves the torso. Only the volume integral on
the lefthand side of Eq. (2.54) remains, and leads to the global matrix equationA<I? O.
The matrix <I?, however, also contains the knownepicardial potentials and these Dirichlet
boundary conditions at the epicardial surface need to be introducedexplicitly, exactly as
explainedabove. Ineffect, weend up solvingfor thebodysurfacepotentialsusingthegiven
epicardial potentials as source functions.
2.4.2.3 FiniteVolume Method
The finitevolume methodwas firstappliedto thebioelectricproblembyAbboudet at.
(1994).It is a closerelativeof the finiteelement method.The governingequationis nowthe
integral form of Eq. (2.7), which integral formafter applicationof the divergence theorem
to the lefthandside may be writtenas
DdS IsvdV
s v
(2.60)
Whereasthefiniteelement methodsoughtto satisfyEq. (2.7),albeitinweakform,inaglobal
manner over the entire torso volume, the finitevolume method aims to satisfy Eq. (2.60)
The Forward Problem of Electrocardiography 61
locally, over each torso element. These local elements are usually selected to be small cubic
volumes or cells, over each of which Eq. (2.60) must be satisfied. In addition, continuity
of the normal component of the current between cells must be ensured. Rosenfeld et al.
(1996) describe the mechanics of solving Eq. (2.60), and of ensuring that the continuity
condition is satisfied, by approximating the gradient V<I> required in Eq. (2.60) by its integral
definition
(2.61)
The use of Eq. (2.61) leads to Eq. (2.60) being approximated over each cell by a linear
equation involving the unknown potentials, at the center of the cell in question, and at
the centers of its nearest neighbors. Eventually, when the equations from all the cells
are combined, a linear matrix equation of the type Aep F results, where the coefficient
matrix A is again sparse. The nonunique nature of the potential is handled by applying the
additional condition that Gauss' flux theorem should be satisfied over the entire volume. A
better approximation than Eq. (2.61) for V<I> can be obtained if, similar to the finiteelement
method, we start with an approximating equation such as Eq. (2.50) for the potential (Harrild
and Henriquez, 1997).
2.4.3 COMBINATION METHODS
Surface and volume methods may be combined such that the former is used where
the volume conductor is isotropic and the latter where it is anisotropic. This aims to take
advantage of the reduced computational load of surface methods, and at the same time be
able to accurately represent anisotropic conductivities. Such a combination was illustrated
by Stanley and Pilkington (1989), where torso potentials were computed from epicardial
potentials in the presence of an anisotropic skeletalmuscle layer. A transfercoefficient
matrix between epicardial potentials and the inner surface of the skeletal muscle, using
Eq. (2.49), was first determined from a boundaryelement discretization ofthe torso up to this
inner layer. The finiteelement method was then used to represent the anisotropic skeletal
muscle and proceed from the potentials on this inner layer to the torso surface potentials. The
methodology of a combination method employing higherorder interpolation, capable of
matching the potential as well as potential gradients across the elements, has been described
by Pullan (1996).
2.5 APPUCATIONS OF THE FORWARD PROBLEM
Three categories of applications ofthe forward problemare described below. The first is
its obvious use with computer heart models to calculate torso (and in some cases, epicardial)
potentials. The second is to gauge the effects of torso conductivity inhomogeneities on
electrocardiographic potentials. The final application is the reciprocal problemof obtaining
the currents traversing the heart due to currents injected at the body surface. One further
application of the forward problem that is only mentioned here in passing is its use in the
62 R. M. Gulrajani
inverse problem of electrocardiography. Almost all inverse solutions entail a prior forward
problem calculation.
2.5.1 COMPUTER HEART MODELS
Presentday computer heart models start by storing a realistic threedimensional repre
sentation of the heart anatomy that, in tum, is properly positioned within a second realistic
threedimensional representation of the human torso. A full solution would then entail the
solution of the two inhomogeneousanisotropy bidomain equations (Eqs. 2.32 and 2.33)
applicable in the heart region, Laplace's equation applicable in the torso region, plus the
boundary conditions at the hearttorso interface and at the torsoair interface. Equations
(2.32) and (2.33) also imply a description of
mv
, the membrane currents of the cardiac
cell. These should be obtained from one of the more recent models for cardiac membrane
currents (e.g., Luo and Rudy, 1991). In essence, the solution to this mix should be the
intracellular and interstitial potentials, <1>; and computed everywhere in the heart, as
well as the torso potential <1>0, computed everywhere in the torso. The potentials <1>; and
then permit determination of the transmembrane potential in the heart, and thus a
description of the heart's excitation. In addition, the values of at the heart surface will
automatically yield the epicardial potential distribution.
To date, on account of its complexity, only one group (Lines et al., 2003a) has achieved
this complete solution for realistic heart and torso geometries. They showa single simulation
of a transmembrane potential distribution in the heart along with a simultaneouslycomputed
body surface distribution (Lines et al., 2003b). A second group (Buist and Pullan, 2002)
has described two solution methodologies for the complete solution, but only illustrate
their results with a twodimensional model of a realisticgeometry heart slice placed inside
a torso slice. Both these groups have pioneered the numerical techniques needed for the
complete solution, but undoubtedly computer memory and speed limitations have prevented
exploiting the techniques to the full. Earlier work by many other groups used a simplified
methodology that entailed splitting the problem into two subproblems. The first entails
solving Eqs. (2.32) and (2.33) in the heart region, assuming that the heart is insulated from
the torso, i.e., neither intracellular nor interstitial currents cross over into the torso. This
would solve the subproblem of determining the excitation of the heart. The second sub
problem would then attempt to calculate torso surface potentials and/or epicardial potentials
from the nowknown excitation pattern of the heart, most commonly by using equivalent
dipole representations for this excitation. Both subproblems, as well as a review of the
many simulations realized, are described briefly below.
2.5.1.1 Determining the Excitation Pattern ofthe Heart
Historically, two categories of heart models may be defined, those that assume a
fixed activation pattern for the heart, based on the activation isochrones first measured
in the human heart by Durrer et al. (1970), and those that incorporate their own activation
algorithm. The best example of the first type of heart model, with fixed activation isochrones
corresponding to normal excitation, was the one developed by Miller and Geselowitz (1978).
No solutions for the heart's excitation pattern are needed here; instead a predetermined
The Forward Problem of Electrocardiography 63
form of the cardiac action potential is triggered at each model point at its corresponding
isochrone time. This then serves to determine the spatial and temporal distribution of
m
.
Clearly, heart models possessing an intrinsic activation algorithm are much more ver
satile and capable of simulating both normal and abnormal excitation. Early examples of
such heart models were all of the socalled "cellular automaton" type. Here, as opposed to
solving Eqs. (2.32) and (2.33), activation is determined by a set of rules that define the propa
gation velocities between model cells, the excitation and refractory states, and the form of the
action potential. Such cellular automaton heart models originated with Okajima et al. (1968),
and some of the better known ones are those of Solomon and Selvester (1971; 1973), Horacek
and van Eck (1972), Lorange and Gulrajani (1993), and Werner et at. (2000). One interesting
way to solve the excitation problem in more realistic fashion is to derive an approximating
"eikonal" equation (ColliFranzone and Guerri, 1993; Keener and Panfilov, 1995) for the
activation wavefront from Eqs. (2.32) and (2.33). Solution of this simplified eikonal equa
tion then yields just the activation time at a given myocardium point or, equivalently, the
wavefront position at a given time. A predetermined action potential is again triggered at
each point corresponding to its activation time. This eikonal approach offers an approximate
solution for the wavefronts, and was developed for want of the requisite computing power
available at the time to solve the anisotropic bidomain equations. Another approach, also
dictated by the lack of computing power, was that of Leon and Horacek (1991). These investi
gators, by assuming equal anisotropy, combined Eqs. (2.32) and (2.33) into a single equation
for the transmembrane potential
m
. We rewrite these Equations below in slightly modified
form
v . f3lm  f3Istim
( G ~ 131
m
(2.62)
(2.63)
where 13 is the surfacetovolume ratio of the cardiac cells, and an intracellularlyinjected
stimulating current I
stim
(in Nm has been introduced to permit excitation sites for the
model. The surfacetovolume ratio permits the conversion of the current per unit volume
Umv, expressed in Nm
3
) to the current per unit membrane surface area Um, expressed in
Nm via the relation I
mv
= 131
m.
With equal anisotropy, i.e., G; = ~ G ~ , and expressing
the membrane current as the sum of the capacitive and ionic currents
(2.64)
Eqs. (2.62), (2.63) and (2.64) may be combined to result in a single governing reaction
diffusion equation:
a
m
[ ~ 13]
at  ~ 13 ~ sum
(2.65)
Leon and Horacek solved Eq. (2.65) for the subthreshold case, i.e. by assuming that the ionic
current was passive and given by lion GmVm, where Gmis the constant resting membrane
conductance per unit area. Once
m
reached a fixed threshold value, a predetermined
action potential waveform was triggered. Thus, the Leon Horacek model was a hybrid, with
64 R. M. Gulrajani
correct subthresholdexcitation, but with cellular automaton characteristics above threshold.
A comprehensive review of most of these heart models, and of the electrocardiographic
simulations realized with them, has been provided by Wei (1997).
Later work focused on the solution of Eq. (2.65), but with both sub as well as
suprathreshold representations for lion. Simulations with a simple FitzHughNagumo rep
resentation for lion have been described by Berenfeld and Abboud (1996) and by Panfilov
(1997). Huiskamp (1998) described a particularly impressive study that solved Eq. (2.65)
for an 800,000 point model of the dog ventricles, using modified BeelerReuter equations
(Drouhard and Roberge, 1986) for lion' This dog model was developed by Hunter et al.
(1992) and incorporated measured fiber directions at every point. Recent work by our
group (Trudel et al., 2001) has employed a multiprocessor computer to solve Eq. (2.65)
for a 12rnillion point highresolution version of the earlier Lorange and Gulrajani (1993)
humanheart model, with analyticallyintroduced fiber rotation, and with a Luo and Rudy
(1991) membrane model for lion. The ventricular activation isochrones for normal excitation
obtained in these simulations by Trudel et al. are shown in Figure 2.5.
Current work by our group focuses on doing away with the equal anisotropy assump
tion and solving Eqs. (2.62) and (2.63) together, without their combination into the single
Eq. (2.65). This is now feasible with our 12rnillion point heart model, again due to a
newer generation multiprocessor computer now available to us. A second advantage of the
solution of Eqs. (2.62) and (2.63) without their combination would be the automatic deter
mination of the interstitial distribution
e
, and hence of the epicardial potential distribution
of the isolated heart. Other groups are attacking the same problem but with different and
novel approaches (Penland et al., 2002; Buist et al., 2003).
2.5.1.2 Calculating Torso and/orEpicardial Potentials
Once the spatial and temporal distribution of m has been determined by one of the
abovedescribed excitation methodologies, subsequent calculation of the body surface po
tentials has usually assumed the myocardium to be isotropic. Accordingly, most investi
gators have used an equation such as Eq. (2.16), namely Jeq geq
m
, to compute an
equivalent dipole density. This leads to individual dipoles at each model point, which are
then combined vectorially for individual heart regions to realize a "multipledipole" heart
model, e.g., the MillerGeselowitz model used 23 such regional dipoles to represent ac
tivation of its ventricles. A surface methodology can then be used to compute the body
surface potentials due to each regional dipole, which potentials are then added to get the
final body surface potential. Thus, Figure 2.6 shows the normal 12lead electrocardiogram
corresponding to the activation isochrones of Fig. 2.5, computed from 58 regional dipoles
of the Trudel et al. heart model, using the integral equation formulation of Eq. (2.40).
Huiskamp (1998), on the other hand, used the variant equation J
eq
gi VmDH
(Eq. 2.43) described by Geselowitz (1989), to determine dipoles on the epicardial and
endocardial surfaces of his dog heart model, prior to using these dipoles as sources to com
pute the potential on the torso surface. This approach may also permit the approximation of
epicardial potentials by computing potentials on a surface just outside the real epicardium,
as demonstrated by Simms and Geselowitz (1995). The disadvantage with this approach
is that for an injured heart containing dead tissue, this dead tissue forms an additional
interface on which equivalent dipoles need to be placed. The more traditional route of
40
30
msec
msec
FIGURE 2.5. Isochrones corresponding to normal activation of the ventricular heart model employed by Trudel
et al. (200 I). A transverse section (top) and a longitudinal section (bottom) are depicted. The colors indicate the
time of activation as per the color bars on the right. Isochrones start at 5 ms after ventricular activation and are
spaced at 5 ms intervals. See the attached CD for color figure. IEEE.
66
'I
M. Gulrajani
V1 2
FIGURE 2.6. Thenormal I2lead correspondingtotheactivation isochronesof Figure2.5.Reproduced, with
permission, fromTrudel et al. (2001). ©IEEE.
lumping dipoles computed by Eq. (2.16) into regional dipoles and then using Eq. (2.40) does
not yield sufficiently accurate potential distributions just outside the epicardium, largely be
cause of a loss of spatial resolution, due to the lumping, that shows up at close distances.
However, as shown by Hren et al. (1998), reasonablyaccurate approximations to the epi
cardial distribution can be obtained if the dipoles at the individual points are not combined
but used individually in the computations. Hren et al. also used the oblique dipole model to
compute their epicardial potentials. They also assumed that the oblique dipoles existed in
a homogeneous, isotropic myocardium. Nevertheless, the computed epicardial potentials
(Fig. 2.7) for paced stimulation at different intramural depths in the right ventricular wall of
Hren et al .'s heart model revealed both the characteristic oneminimumtwomaxima pattern,
33. 50/5.52
F
·15.70/4.75
c
E
·32.84/5.08
31.68/4.85
29.48/4.68
26.95/4.71
8.24/4.96
6.12/4.93
4.36/5.00
1.93/5 .29
FIGURE 2.7. Simulated potential maps on a patch of epicardium 10 ms after the onset of activation, for pacing
at different intramural depths in the right ventricular free wall. Pacing sites were 0.5 mm apart, progressing from
the epicardium (panel A) to the endocardium (panel The epicardial projection of each pacing site is indicated
by the black dot. Isopotentiallines are plotted for equal intervals, with no zero line; solid contours represent the
positive and broken contours the negative values of the potential; the magnitudes of the minimum and maximum
are given (in mV) at the bottom of each map. Note that the axis joining the two maxima rotates counterclockwise
with increasing pacing depth following the transmural counterclockwise rotation of fibers from epicardium to
endocardium. Figure reproduced, with permission, from Hren et al., 1998.
68 R. M. Gulrajani
as well as the rotation with depth of stimulation, observed experimentally by Taccardi et al .
(1994).
The use of the oblique dipole model, along with the approximation of a homogeneous
isotropic myocardium, in order tocompute torso surface potentials may be more contentious.
In a simulation study involving a layered myocardium block placed inside a larger volume
conductor block, Thivierge et al. (1997) showed that the effect of the axial dipole component
on surface potentials is reduced both on account of its orientation along the highconductivity
fiber direction as well as due to a wellknown effect described by Brody (1956), whereby
dipoles oriented tangential to the highconductivity blood masses are dimini shed in so far as
torso surface potentials are concerned, whereas dipoles oriented radial to the blood masses
are enhanced. In the heart, the fibers are in the main oriented tangential to the blood masses.
During normal activation of the heart from endocardium to epicardium, the component
of the oblique dipole normal to the wavefront ends up being radial to the blood masses
and is therefore enhanced, whereas the axial component along the fibers is diminished
both due to the Brody effect as well as due to its orientation along the highconductivity
direction of the myocardium. Furthermore, the axial dipole is small to begin with during
normal activation. This is because the endocardialtoepicardial spread of normal activation
ensures that the largest component of the transmembrane potential gradient VV
m
lies in
the endocardiumtoepicardium direction perpendicular to the fibers, and that the axial
component of VV
m
parallel to the fibers is very small. Mathematically, this translates to
a very small value for n 3 in Eq. (2.39), and hence a small axial component for p. All
of these reasons may well explain why torso potential simulations that have ignored the
axial component have often successfully reproduced clinical electrocardiograms (ECGs).
If using the oblique dipole model for computing torso potentials is being contemplated, it
may be necessary to include both the highconductivity intraventricular blood masses as
well as the myocardial anisotropy into the torso model so as to correctly reduce the effect
of the axial component. these inclusions are not possible, may actually be preferable
to use the uniform dipole model and an isotropic myocardium. On the other hand, for
epicardial potentials, as the simulations by Hren et al. (1998) show, it is essential to include
the axial component. It may also be essential to include this component for body surface
computations during repolarization, which being less organized than depolarization, may
have a larger axial component. Also abnormal stimulation of the ventricles at isolated
sites will certainly lead to initial propagation along the fiber direction, and larger axial
components of VV
m
• Intramural stimulation, such as that depicted in Fig. 2.7, therefore
may warrant use of the oblique dipole model not only for epicardial potentials but also for
torso surface potentials.
All of the above computations of torso and/or epicardial potentials described above
have been via surface methodologies. Fischer et al. (2000) recently described a mixed
boundary elementfinite element route to obtaining both epicardial and torso potentials . The
starting point was the transmembrane potential distribution
m
, calculated by Huiskamp
(1998). The heart region of Huiskamp's model was then represented via finite elements ,
with the known
m
distribut ion acting as a source. Outside the heart, a boundaryelement
surface representation for the torso was employed. The governing equations were Eq. (2.35)
in the heart , namely V · (G' V<l>e)  V · (G; VV
m),
and Laplace' s equation in the torso.
Since V is known, only two boundary conditions are needed at the hearttorso interface.
The Forward Problem of Electrocardiography 69
The ones selected were Eqs. (2.36) and (2.37), the latter rewritten in terms of the known
potential and the unknown potentials and <1>0 being sought, namely,
(2.66)
Solution of this mixed finiteelement and boundaryelement problem yielded and <1>0;
the former then gives the epicardial potential distribution and the latter the torso surface
potential distribution. Two important points to note are that due to
m
being used directly
as the source and due to the finiteelement methodology used for the heart, an accurate
solution results for e and <1>0 without the need to draw on the oblique dipole interpretations
used with surface methodologies. Yet the computed solution is the equivalent of a surface
methodology that employs an oblique dipole in an inhomogeneous anisotropic heart, i.e.,
with both G; and G' varying tensor quantities. Fischer et al. went on to compare their results
with those computed assuming G' to be an isotropic constant, i.e., equivalent to an oblique
dipole but acting in an isotropic heart, and with those assuming both G; and G' isotropic
constants, i.e., equivalent to a uniform dipole layer acting in an isotropic heart. Relative
errors in the body surface potentials were approximately 34% for G' constant and 43% for
both G; and G' constant, reflecting the importance of including myocardial anisotropy in the
body surface computations. Qualitatively, however, the surface potential distributions were
similar (Fig. 2.8). Interestingly, the solution with G' constant, i.e, equivalent to an oblique
dipole in an isotropic heart, overestimated the potential variations, while the solution with
both G; and G' constant, i.e., equivalent to a uniform dipole layer in an isotropic heart,
underestimated the variations. This reinforces the finding by Thivierge et al. (1997) that the
axial component of the oblique dipole does tend to get reduced by its orientation along the
highconductivity fiber direction. The highconductivity blood masses were not included
in the study by Fischer et al. so the Brody effect did not come into play in the potential
distributions of Fig. 2.8. would be interesting to see if the presence of blood masses
<l> [mV)
. t 5
FIGURE 2.8. Body surface potential maps on the anterior torso model 48 ms after the onset of activation: (a)
oblique dipole layer model in an isotropic heart, (b) full anisotropic myocardium model, (c) uniform dipole layer
model in an isotropic heart. Contours are plotted in steps of 0.5 mV. Field patterns are in qualitative agreement.
but quantitative differences are large, even though the isotropic conductivities were chosen to realize the smallest
difference. Figure reproduced, with permission, from Fischer et al., 2000.
70 M. Gulrajani
within the heart would bring the three torso surface distributions of Fig. 2.8 closer together.
On the other hand, the computed epicardial distributions (not shown here) under the three
conditions clearly revealed that the approximation of a uniform dipole layer in an isotropic
heart was unable to reproduce, even in qualitative fashion, the correct epicardial potential
distribution.
2.5.2 EFFECTS OFTORSO CONDUCTIVITYINHOMOGENEITIES
The effects of torso conductivity inhomogeneities have long been of interest to re
searchers. Early interest was kindled simply by the desire to know the intrinsic effect of
the major torso inhomogeneities such as the intracardiac blood masses, lungs, ribs, spine,
skeletal muscle, and subcutaneous fat on the electrocardiographic potentials. A second
motivation today is to gauge which of the above inhomogeneities has the largest perturb
ing effect on inverse solutions, and therefore needs to be carefully included in the torso
volume conductor used in computing these inverse solutions. Many early forward prob
lem studies used physical analogs of the torso, into which materials that mimicked the
torso inhomogeneities could be inserted to see their effect on the surface potentials gen
erated by a current dipole source placed within the analog. The most informative early
studies were, however, analytical, employing spherical or cylindrical models of the heart
and torso. We have already alluded to the Brodyeffect which was deduced using the
model of a spherical ventricular cavity of infinite conductivity that represented the blood
masses, situated in an otherwise infinite homogeneous medium that represented the my
ocardium (Brody, 1956). The Brody model was extended by Rudy and coworkers (Rudy
et aI., 1979; Rudy and Plonsey, 1980) who used an "eccentricspheres" model in which the
torso was represented by two systems of spheres. The inner system which mimicked the
heart and bloodfilled cavity was eccentric with respect to the outer system which repre
sented lungs, skeletal muscle, and subcutaneous fat. Besides confirming the Brodyeffect
for radial dipoles, Rudy and coworkers described the effects of varying the conductivities
of the lungs, skeletal muscle and fat regions. A review of their work has been written by
Rudy (1987). The Brody effect has been revisited recently by van Oosterom and Plonsey
(1991). They reiterate that, while generally correct, the magnitude of the Brody effect
is dependent on the field point, and, if the medium is bounded, also on the location of
the reference electrode. The Brody effect can also be negative under certain conditions,
where the insertion of the blood masses causes the potential to shift from positive to
negative. This corresponds to shifts of the zero isopotential line in the surface potential
distribution.
With the development of numerical techniques for computing the torso surface po
tentials, attention shifted to utilizing realisticgeometry torso models to study the effects
of torso inhomogeneities. One such study was done by Gulrajani and Mailloux (1983) us
ing a boundaryelement torso model that comprised intraventricular blood masses, lungs
and a skeletal muscle layer. This last was approximated as an isotropic layer of increased
thickness as first suggested by McFee and Rush (1968). No subcutaneous fat layer was
used, the approximated skeletalmuscle layer extending all the way to the surface. Equation
(2.40) was used to compute the surface potentials due to the 23 individual current dipoles
of the MillerGeselowitz heart model as the inhomogeneities were introduced onebyone
into an otherwise homogeneous torso. Most of Gulrajani and Mailloux's findings were in
The Forward Problem of Electrocardiography 71
accordance with earlier work employing realistic torso models (Barnard et al., 1967;
Selvester et aI., 1968; Horacek, 1971). Apart from qualitatively confirming the Brody
effect , by activating the 23 MillerGeselowitz dipoles in concert to generate normal activa
tion Gulrajani and Mailloux could gauge the effects of the inhomogeneities on the normal
ECG and the whole body surface potential map (BSPM). The major qualitative effects were
restricted to a smoothing of notches in the ECG and of isopotentials in the BSPM due to,
in descending order of importance, the blood masses, muscle layer and lungs. However,
although qualitative pattern changes in the ECG and BSPM were limited to these smoothing
effects, there were large quantitative changes in both, notably magnitude increases due to
the blood masses and magnitude decreases due to the muscle layer. The latter is due to the
increased distance of the torso surface from the dipoles on account of the increased effective
thickness of the muscle layer, and the former is due to the Brodyeffect enhancement on the
predominantly radial orientation of the current dipoles associated with normal activation.
The advent of powerful computers has seen more elaborate finiteelement and finite
difference torso models being used for the study of inhomogeneity effects. One example is
the finiteelement model developed at the University of Utah (Johnson et al., 1992), which
was constructed on the basis of magnetic resonance images of the torso, and incorporated
lungs, an anisotropic skeletalmuscle layer, subcutaneous fat, as well as secondary inho
mogeneities such as epicardial fatpads, bloodfilled major arteries and veins, sternum, ribs,
spine, and clavicles. This model has been used in a study by Klepfer et aI. (1997) on the
effects of the inhomogeneities and anisotropies on a known, fixed, epicardial potential dis
tribution. Klepfer etaI. estimated 11 to 15%changes in the BSPM due to addition or removal
of either the lungs, anisotropic skeletalmuscle layer, or subcutaneous fat. No major BSPM
pattern changes were noted. Klepfer et al. , however, may have lessened the impact of the
inhomogeneities since the starting epicardial distribution was kept fixed and not "loaded"
by the changing inhomogeneities. Changes in the amplitudes of epicardial potentials with
changes in the torso conductivity have been reported by MacLeod et al. (1994) in an ex
perimental study that used measured epicardial and surrogate torso potentials from a dog
heart placed inside a humanshaped torso tank. Data from the Utah model was also used
in a study by Bradley et (2000) that used boundary elements for the epicardial surface,
lung surface and torso cavity, but finite elements for the anisotropicconductivity skeletal
muscle and subcutaneous fat layers. Bradley et at. used a single timevarying current dipole
derived from a Frank vectorcardiographic signal to represent the excitation of the heart.
This dipole was placed within the heart region and served as the source. Consequently,
the calculated epicardial potentials are subject to loading effects as the inhomogeneities
are removed or added. Again pattern changes in the BSPM were not noted, but up to 30%
magnitude changes in the BSPM could be seen. Bradley et al. found that the effect of the
subcutaneous fat to be more important than that of the skeletal muscle. This was in contrast
to the earlier studies of Gulrajani and Mailloux (1983) and of Stanley and Pilkington (1986)
who found that the effect of the skeletal muscle layer was important. However, in both these
earlier studies the skeletal muscle layer extended all the way to the outer torso surface,
there being no subcutaneous fat layer. It could well be that it is the layer that abuts the outer
torso surface that needs to be correctly represented for more accurate torso potential magni
tudes. On the other hand, Hyttinen et aI. (2000) in a study involving a finitedifference
model of the torso, constructed on the basis of the US National Library of Medicine's
Visible Human Man data, gauged the effect of a 10% increase in the conductivities of the
M. Gulrajani
individual inhomogeneities on torso surface potentials due to a heart dipole. They found
that the effect of the heart muscle was largest, followed by the intracardiac blood, skeletal
muscle, lungs and subcutaneous fat, in that order. This protocol of a 10% conductivity
change, by standardizing the extent of the change, may be a better indication of the effect
of a particular inhomogeneity, than adding or removing a homogeneity altogether.
While the inhomogeneities mentioned above only affected potential magnitudes with
little effect on BSPM patterns, Bradley et al. also found that the position and orientation
of the heart in the torso made the most difference for both torso potential magnitudes and
distributions. This is in accordance with recent work by Ramanathan and Rudy (2001a) that
used measured epicardial and surrogate torso potentials from a dog heart placed inside a
humanshaped torso tank. They found that torso inhomogeneities have a minimal effect on
torso patterns computed from the measured epicardial potentials, but that for a good match
with measured torso potentials it was essential that heart and torso geometry be accurately
represented. In an accompanying paper (2001b), they show that even a homogeneous torso
can be used in the inverse computation of epicardial potential distributions, provided heart
and torso geometry are correctly represented. This was also shown in the study by Hyttinen
et al. (2000) cited earlier who, having access to heart geometries corresponding to both sys
tole and diastole, found that the error in inverselycomputed dipoles increased significantly
if the wrong heart model was used in the inverse computations.
Clearly the best way to judge the effect of torso inhomogeneities on surface potentials
is to use a fullycoupled complete hearttorso solution. Buist and Pullan (2003) have done
such a study but with their twodimensional hearttorso slice mentioned earlier. They report
that in none of their tested situations did the twostep equivalent dipole approach completely
reproduce the fullycoupled results further supporting the above assertion.
2.5.3 DEFIBRILLATION
Defibrillation consists of applying a highenergy shock to the fibrillating heart , the idea
being to simultaneously depolarize all the ventricular cells thereby halting the fibrillatory
activity. Upon recovery from the shock, the sinus node often regains control of the heart and
a normal heartbeat ensues. This is the "total extinction" hypothesis for defibrillation, first
put forward by Wiggers (1940). Later, Zipes et al. (1975) proposed the "critical mass" hy
pothesis, whereby halting the fibrillatory activity in a certain critical mass of myocardium,
thought to be greater than 75% of the total mass was sufficient for successful defibrillation.
The remaining mass would then be incapable of sustaining the residual fibrillatory activ
ity. More recently, Chen et al. (1986) suggested that even if a shock was strong enough
to halt the fibrillatory activity everywhere or in a critical mass of myocardium, it could
still reinitiate fibrillation upon removal, and therefore had to be somewhat larger. This
hypothesis was based on two observations, first that fibrillation was only induced in dogs
for shocks between a lower and an upper limit, and second that shocks above this upper
limit of vulnerability never failed to defibrillate an already fibrillating heart, presumably
because it never reinitiated fibrillation. Chen et al.'s hypothesis has come to be known as the
"upper limit of vulnerability" hypothesis. Whichever of the above three hypotheses holds,
it is clear that a certain minimum level of excitation is required at the heart for success
ful defibrillation which translates to a minimum value of the applied current density (or
The Forward Problem of Electrocardiography 73
voltage gradient) everywhere in the heart. The number used in modeling studies is between
12.5  35 mNcm or 56 V/cm (Karlon et aI.,1994; Panescu et al., 1995; Min and Mehra,
1998). At the same time, it is necessary that the current density anywhere in the heart does
not exceed approximately 500 mNcm since at these densities tissue damage is likely to
occur. Thus, it is important that the defibrillation electrodes ensure a reasonably uniform
current distribution in the heart. For transthoracic defibrillation, the shock is applied via ex
ternal paddle electrodes, but in internal defibrillation intraventricular catheters or epicardial
patches are used for shock application.
The finiteelement method is the most direct way to calculate the current density in the
heart due to the defibrillatory shock. The governing equation is obtained by replacing the
surface integral in Eq. (2.54), by the uncancelled component of this integral over S02 that
remains once Eq. (2.54) is applied to all volume elements. This uncancelled component is
given by Eq. (2.59), so that we get,
(O'V<l». Vf3i
dV
f3i
lsvdV
f3;Jn
dS
v V ~ 2
i 1,2, . . . . m (2.67)
Note that the number of equations spans the number of unknown potentials in the entire
volume conductor. Since during these current density calculations, the heart is treated as
passive, the first volume integral on the righthand side is zero. Moreover, during defibrilla
tion it is incorrect to assume that the electrodes inject a uniform current density In' Thus, it
is more appropriate to use a Dirichlet boundary condition, setting the potential at the nodes
corresponding to the electrodes equal to the applied electrode voltage. This eliminates the
surface integral in Eq. (2.67) also. Only the lefthand side remains in Eq. (2.67), which
may then be reduced to global matrix form. Due to the complex torso models used, the
matrices are large, and Ng et al. (1995) discuss the use of parallel computers in obtaining
the solution. Once all unknown node potentials are calculated, the current density within the
elements representing the heart is obtained from the gradient of the potential distribution
<1>. By using a finiteelement representation for the intraventricular catheter electrode, the
finiteelement approach can also be used for studying the heart current densities due to
implantable defibrillators (Jorgenson et al., 1995).
External or transthoracic defibrillation can also be studied with the boundaryelement
formulation (Claydon et al., 1988; Oostendorp and van Oosterom, 1991; Gale et al., 1994;
Gale, 1995). While matrix sizes are smaller, the major disadvantage is that anisotropic
conductivity variations, especially those in the myocardium, cannot be taken into account.
The governing equation for the potential <I>(r) anywhere in the heart is again obtained from
an application of Green's second identity to the torso geometryofFig. 2.4 and is (Oostendorp
and van Oosterom, 1991):
~ f In(r') ,
41l'O'(r)<I>(r) = (0'/  )<I>(r )dQ
rr,
dS
/=0
SI So
(2.68)
In Eq. (2.68), O'(r) is the conductivity at the observation point, and In(r') is the normal
74 R. M. Gulrajani
component of the injected current density at the defibrillation electrode. Oostendorp and
van Oosterom discuss the numerical solution of Eq. (2.68) for the unknown potentials and
current densities in terms of the known potentials at the defibrillation electrodes.
Simulation studies of defibrillation have concentrated on determining the optimal
positioning and size of defibrillation electrodes in order to ensure an adequate and
approximatelyuniform current density everywhere in the heart (Gale et aI., 1994; Gale,
1995; Camacho et al., 1995; Panescu et al., 1995; Schmidt and Johnson, 1995). More re
cent work has focused on determining the defibrillation threshold, namely the electrode
voltage or energy needed for a given percentage of the myocardial mass (usually 95%)
to attain a voltage gradient of at least 5 V/cm (Aguel et al., 1999; De Jongh et al., 1999;
Eason et 1998; Kinst et 1997; Min and Mehra, 1998). This is in line with the crit
ical mass hypothesis for defibrillation. One early study focused on the sensitivity of the
currentdensity distribution in the heart to variations in skeletal muscle anisotropy (Karlon
et al., 1994). It was found that in transthoracic defibrillation, the anisotropy made little
difference to current flow patterns in the heart, but simply affected current magnitudes. On
the other hand, the same study showed that other inhomogeneities such as the lungs, ribs
and sternum affected both magnitudes and current patterns. Another study (Eason et aI.,
1998) found that the voltage defibrillation threshold, for internal defibrillation between a
right ventricular catheter electrode and the defibrillator can in the pectoral region of the
left chest, differed by only 4.5% if the realistic fiber architecture in the heart model was
replaced with an isotropic conductivity myocardium.
It is being widely recognized today that defibrillation is an immensely complex phe
nomenon, and that the above simulation studies yield, at best, an estimate of the extracellular
potential gradient and current density inside a monodomain passive heart. What really counts
is the transmembrane potential distribution in the heart, and the response of the active cells
in the heart to this transmembrane distribution. Theoretical analysis has revealed that even
with extracellular stimulation of the passive bidomain heart, the unequal anisotropies of
the intracellular and interstitial space result in contiguous regions of large transmembrane
potential depolarization and hyperpolarization in the heart (Sepulveda et al., 1989). Local
currents flowing between these regions have the ability to reinitiate fibrillation and negate
the findings of studies that assume a monodomain heart. Experimental work and theoretical
simulations both tend to support this reinitiation mechanism(Efimov et al., 2000; Skouibine
et al., 2000). A second perturbing factor that affects real transmembrane voltage gradients
in the heart is the effect of local conductivity discontinuities (e.g., gap junctions, fiber cur
vature, random clefts). Sobie et at. (1997) have shown how, using a "generalized activation
function," the effects of both extracellular stimulation and of conductivity discontinuities
can be taken into account in determining the transmembrane potential distribution. This
generalized activation function suggests that more than the gradient of the extracellular
potential, it is the second spatial derivative of the extracellular potential that determines the
transmembrane potential change. Although Sobie et at. illustrated their approach in passive
myocardium, it can just as easily be applied to active myocardiumif sufficient computational
resources are available. This represents the final, and most important step, in translating an
applied extracellular potential to the actual response of the cardiac cells. Already, Skouibine
et at. (2000) illustrate such active responses in simulations in a twodimensional bidomain
sheet of active myocardium subjected to a defibrillation shock.
The Forward Problem of Electrocardiography
2.6 FUTURE TRENDS
75
This review of the forward problem of electrocardiography has presented its theoreti
cal underpinnings, the solution methodologies employed, and finally its major applications.
The greater accessibility of multiprocessor computers means that we shall likely see more
accurate heart models capable of simulating more complex heart pathologies, with simul
taneous computation of heart activation and torso potentials via a complete solution. Torso
representations are also likely to be more accurate with better accounting of anisotropic con
ductivity, and with the effect of torso inhomogeneities gauged with a complete fullycoupled
threedimensional hearttorso solution. Finally, successful defibrillation is more than just
a matter of knowing the extracellular potential or current density everywhere in the heart.
Rather it is the interaction of this potential or current density with the cardiac cell, and the
subsequent effect on the cellular action potentials that determines whether the defibrilla
tion shock succeeds or fails. While the defibrillation simulations described above may be
useful in electrode design, a better understanding of defibrillation necessitates simulations
in which the heart is not treated simply as a passive conductor, but as an active bidomain.
Already, Efimov et al. (2000) have described the transmembrane potentials generated in a
threedimensional realisticgeometry rabbitheart model with varying fiber directions, fol
lowing uniform electricfield application. Although they represented the myocardium with
passive bidomain equations, it is only a matter of time before such wholeheart defibrillation
simulations will be realized with an active membrane representation for the heart's ionic
currents.
ACKNOWLEDGMENT
Work supported by the Natural Sciences and Engineering Research Council of Canada.
REFERENCES
Abboud, S., Eshel, Y, Levy, S., and Rosenfeld, M., 1994, Numerical calculation of the potential distribution due
to dipole sources in a spherical model of the head, Comput. Biomed. Res. 27: 441455.
Aguel, E, Eason, J. Trayanova, N. A., Seikas, G., and Fishier, M. G., 1999, Impact of trans venous lead position
on activecan ICD defibrillation: A computer simulation study, PACE 22 [Pt. II]: 158164.
Barnard, A. C. L., Duck, I. M., and Lynn, M. S., 1967, The application of electromagnetic theory to electrocardi
ology. II. Numerical solution of the integral equations, Biophys. 7: 463491.
Barr, R. Pilkington, Boineau, J. P., and Spach, M. S., 1966, Determining surface potentials from current
dipoles with application to electrocardiography, IEEE Trans. Biomed. Eng. 13: 8892.
Barr, R. Ramsey, M., III, and Spach, M. S., 1977, Relating epicardial to body surface potential distributions
by means of transfer coefficients based on geometry measurements, IEEE Trans. Biomed. Eng. 24: 111.
Berenfe1d, and Abboud, S., 1996, Simulation of cardiac activity and the ECG using a heart model with a
reactiondiffusion action potential, Med. Eng. Phys.18: 615625.
Berne, R. M., and Levy, M. N., 1977, Cardiovascular Physiology, Mosby, St. Louis, chapter 4.
Bradley, C. P., Pullan, A. J., and Hunter, J., 2000, Effects of material properties and geometry on electrocardio
graphic forward simulations, Ann. Biomed. Eng. 28: 721741.
Brebbia, C. A., and Dominguez, J., 1992, Boundary Elements. An Introductory Course, 2nd ed., WIT Press,
Southampton, U.K., chapter 2.
76 R. M. Gulrajani
Brody, D. A., 1956, A theoretical analysis of intracavitary blood mass influence on the heartlead relationship,
Circ. Res. 4: 731738.
Buist, M., and Pullan, A., 2002, Torso coupling techniques for the forward problem of electrocardiography, Ann.
Biomed. Eng. 30: 12991312.
Buist, M., and Pullan, A., 2003, The effect of torso impedance on epicardial and body surface potentials: a modeling
study, IEEE Trans. Biomed. Eng. 50: 816824.
Buist, M., Sands, G., Hunter, P., and Pullan, A., 2003, A deformable finite element derived finite difference method
for cardiac activation problems, Ann. Biomed. Eng. 31: 577588.
Camacho, M. A., Lehr, 1. L., and Eisenberg, S. R., 1995, A threedimensional finite element model of human
transthoracic defibrillation: Paddle placement and size, IEEE Trans. Biomed. Eng. 42: 572578.
Chen, P.S., Shibata, N., Dixon, E. G., Martin, R. and Ideker, R. E., 1986, Comparison of the defibrillation
threshold and the upper limit of ventricular vulnerability, Circulation 73: 10221028.
Claydon, F. J., III, Pilkington, T. C, Tang, A. S. L., Morrow, M. N., and Ideker, R. E., 1988, A volume conductor
model of the thorax for the study of defibrillation fields, IEEE Trans. Biomed. Eng. 35: 981992.
ColliFranzone, P., and Guerri, L., 1993, Models of the spreading of excitation in myocardial tissue, in: High
Performance Computing in Biomedical Research (T. C. Pilkington, B. Loftis, J. F. Thompson, S. L.Y Woo,
T. C. Palmer, and T. F. Budinger, eds.), CRC Press, Boca Raton, FL, pp. 359401.
ColliFranzone, P., Guerri, L., and Rovida, S., 1990, Wavefront propagation in an activation model of the anisotropic
cardiac tissue: Asymptotic analysis and numerical simulations, Math. Biol. 28: 121176.
ColliFranzone, P., Guerri, L., and Viganotti, 1983, Oblique dipole layer potentials applied to electrocardiology,
Math. Biol. 17: 93124.
ColliFranzone, P.•Guerri, L., Viganotti, C., Macchi, E., Baruffi, S., Spaggiari, S., and Taccardi, 8., 1982, Potential
fields generated by oblique dipole layers modeling excitation wavefronts in the anisotropic myocardium.
Comparison with potential fields elicited by paced dog hearts in a volume conductor, Circ. Res. 51: 330346.
Corbin, L. V., II, and Scher, A. M., 1977, The canine heart as an electrocardiographic generator, Circ. Res. 41:
5867.
De Jongh, L., Entcheva, E. G., Replogle, 1. A., Booker, R. S., III., KenKnight, B. H., and Claydon, F. J., 1999,
Defibrillation efficacy of different electrode placements in a human thorax model, PACE 22 [Pt. II]: 152157.
Drouhard, J. P., and Roberge, F. A., 1986, Revised formulation of the HodgkinHuxley representation of the
sodium current in cardiac cells, Comput. Biomed. Res. 20: 333350.
Durrer, D., van Dam, R. T., Freud, G. E., Janse, M. J., Meijler, F. L., and Arzbaecher, R. C, 1970, Total excitation
ofthe isolated human heart, Circulation 41: 899912.
Eason, J., Schmidt, J., Dabasinskas, A., Seikas, G., Aguel, F., and Trayanova, N., 1998, Influence of anisotropy
on local and global measures of potential gradient in computer models of defibrillation, Ann. Biomed. Eng.,
26: 840849.
Efimov, I. R., Aguel, F., Cheng, Y, Wollenzier, B., and Trayanova, N., 2000, Virtual electrode polarization in the
far field: implications for external defibrillation, Am. 1. Physiol. Heart Circ. Physiol. 279: HI055Hl070.
Fischer, G., Tilg, B., Modre, R., Huiskamp, G. J. M., Fetzer, 1., Rucker, w., and Wach, P., 2000, A bidomain model
based BEMFEM coupling formulation for anisotropic cardiac tissue, Ann. Biomed. Eng. 28: 12291243.
Fung, Y., 1977, A First Course in Continuum Mechanics, 2nd ed., PrenticeHall, Englewood Cliffs, NJ, chapter 2.
Gale, T. J., 1995, Modelling the electric field from implantable defibrillators, Ph.D. thesis, University of Tasmania,
Hobart, Tasmania, Australia.
Gale, T. J., Johnston, P. R., Kilpatrick, D., and Nickolls, P. M., 1994, Implantable defibrillator electrode comparison
using a boundary element model, in: Proc. 17th Ann. Intl. Con! IEEE Eng. Med. Biol. Soc., IEEE Press, New
York, pp. 3132.
Geselowitz, D. B., 1989, On the theory of the electrocardiogram, Proc. IEEE 77: 857876.
Gulrajani, R. M., 1998a, Bioelectricity and Biomagnetism, Wiley, New York. Chapter 7.
Gulrajani, R. M., 1998b, The forward and inverse problems of electrocardiography, IEEE Eng. Med. Biol. Magazine
17 (5): 84101.
Gulrajani, R. M., and Mailloux, G. E., 1983, A simulation study of the effects of torso inhomogeneities on
electrocardiographic potentials, using realistic heart and torso models, Cire. Res. 52: 4556.
Gulrajani, R. M., Roberge, F. A., and Mailloux, G. E., 1989, The forward problem of electrocardiography, in:
Comprehensive Electrocardiology, Volume I (P. W. Macfarlane and T. D. V. Lawrie, eds.), Pergamon Press,
New York, pp. 197236.
Harrild, D. M., and Henriquez, C. S., 1997, A finite volume model of cardiac propagation, Ann. Biomed. Eng. 25:
315334.
The Forward Problem of Electrocardiography 77
Heller, L., 1990, Computationof the return current in encephalography: The auto solid angle, in: Digital Image
Synthesis and Inverse Optics. Proc. SPIE 1351 (A. E Gmitro, P. S. Idell, and 1.1. LaHaie, eds.), pp. 376390.
Hooks, D. A., Tomlinson, A., Marsden, S. G., LeGrice,1.1., Smaill, B. H., Pullan, A. 1., and Hunter, P.1., 2002,
Cardiac microstructure. Implications for electrical propagationan defibrillationin the heart, Cire. Res.
331338.
Horacek, B. M., 1971, The effect on electrocardiographic lead vectors of conductivity inhomogeneities in the
humantorso, Ph.D. thesis, Dalhousie University, Halifax, Canada.
Horacek, M., and RitsemavanEck, H. 1., 1972,The forwardproblemof electrocardiography, in: The Electrical
Field of the Heart , (P. Rijlant, ed.), Presse Acad. Eur., Bruxelles, p. 228.
Hren, R., Nenonen, J., and Horacek, B. M., 1998, Simulated epicardial potential maps during paced activation
reflect myocardial fibrous structure, Ann. Biomed. Eng. 26: 10221035.
Huiskamp, G., 1998, Simulations of depolarization in a membraneequationsbased model of the anisotropic
ventricle, IEEE Trans. Biomed. Eng. 45: 847855.
Hunter,P.1., Nielsen, P.M., Srnaill, B. H., LeGrice, 1., and Hunter,1. W., 1992,An anatomical heart model with
applicationsto myocardial activation and ventricularmechanics, CRC Crit. Rev. Biomed. Eng. 20: 403426.
Hyttinen, 1., Puurtinen, H.G., Kauppinen, P., Nousiainen, I., Laame, P., and Malmivuo, 1., 2000, On
the effects of model errors on forward and inverse ECG problems, Int. Bioelectromagnetism
http://www.ee.tut.fi/rgilijbem
Johnson, C. R., MacLeod, R. S., and Ershler, P. R., 1992, A computer model for the study of electrical current
flowin the human thorax, Comput. BioI. Med. 22: 305323.
Jorgenson, D. B., Hayner, D. R., Bardy, G. H., and Kim Y., 1995, Computational studies of transthoracic and
transvenous defibrillation in a detailed 3Dhuman thorax model, IEEE Trans. Biomed. Eng. 42: 172184.
Karlon, W. I. , Lehr, 1. L., and Eisenberg, S. R., 1994, Finite element models of thoracic conductive anatomy:
Sensitivity to changes in inhomogeneity and anisotropy,IEEE Trans. Biomed. Eng. 41: 101G1017.
Keener, 1. P., and Panfilov, A. V., 1995, Threedimensionalpropagationin the heart: The effects of geometryand
fiber orientationon propagation in myocardium, in: Cardiac Electrophysiology. From Cell to Bedside. 2nd
ed. (D. P. Zipes and1.lalife, eds.) B. Saunders, Philadelphia, chapter 32.
Kinst, T. E, Sweeney, M. Lehr, J. L., and Eisenberg, S. R., 1997, Simulated internal defibrillation in hu
mans using an anatomicallyrealistic threedimensional finite element model of the thorax, Cardiovasc.
Electrophy siol. 8: 537547.
Klepfer. R. N., Johnson, C. R., and MacLeod, R. S., 1997, The effects of inhomogeneities and anisotropies on
electrocardiographic fields: A 3D finiteelement study, IEEE Trans. Biomed. Eng. 44: 706719.
Krassowska, w., and Neu, 1. C., 1994, Effectiveboundaryconditions for syncytial tissues, IEEE Trans. Biomed.
Eng. 41: 143150.
Leon, L.l., and Horacek, B. M., 1991, Computer model of excitationandrecoveryin the anisotropicmyocardium.
Rectangularand cubic arrays of excitableelements, Electrocardiol. 24: 115.
Lines, G. T.. Gratturn, P., and Tveito, A., 2oo3a, Modelingthe electrical activity of the heart: A bidomainmodel
of the ventriclesembedded in a torso, Comput. Visual. Sci. 5: 195213.
Lines, G. T., Buist, M. L., Grettum, P., Pullan, A.l., Sundnes, J., and Tveito, A., 2oo3b, Mathematicalmodels and
numerical methodsfor the forwardproblemin cardiac electrophysiology, Comput. Visual. Sci. 5: 215239.
Lorange, M., and Gulrajani, R. M., 1993, Acomputer heart model incorporatingmyocardial anisotropy.1. Model
constructionand simulationof normal activation,1. Electrocardiol. 26: 245261.
Luo, C., and Rudy, Y., 1991, A model of the ventricular cardiac action potential. Depolarization, repolarization,
and their interaction, Circ. Res. 68: 15011526.
Lynn,M. andTirnlake, P., 1968, The useof multipledeflationsin the numerical solutionof singularsystems
of equations with applicationto potential theory, SIAM Numer. Anal. 5: 303322.
MacLeod, R. S., Taccardi, B., andLux, R. L., 1994, The influenceof torsoinhomogeneitiesonepicardialpotentials,
Computers in Cardiol. , IEEEPress, New York, pp. 793796.
McFee, R., and RushS., 1968, Qualitativeeffects of thoracicresistivity variations on the interpretationof electro
cardiograms: The lowresistance surfacelayer, Am. Heart 76: 4861.
Meijs, 1. W. H., Weier, O. w., Peters, M. J., and van Oosterorn, A., 1989, On the numerical accuracy of the
boundaryelement method, IEEE Trans. Biomed. Eng. 36: 10381049.
Miller, T., III, and Geselowitz,D. B., 1978, Simulation studies of the electrocardiogram. 1.The normal heart,
Cire. Res. 43: 301315.
Min, X., andMehra, R., 1998, Finiteelement analysis of defibrillationfieldsin a humantorsomodel for ventricular
defibrillation, Progr. Biophys. Molec. BioI. 69: 353386.
78 R. M. Gulrajani
Ng, K. Hutchinson, S. A., and Gao, S., 1995, Numerical analysis of electrical defibrillation. The parallel
approach, Electrocardiol. 28 (supp!.): 1520.
Okajima, M., Fujino, Kobayashi, and Yamada, K., 1968, Computer simulation of the propagation process
in excitation of the ventricles, Circ. Res. 23: 203211.
Oostendorp, and van Oosterom, A., 1991, The potential distribution generated by surface electrodes in inho
mogeneous volume conductors of arbitrary shape, IEEE Trans. Biomed. Eng. 38: 409417.
Panescu, D. Webster, J. G., Tompkins, W. J., and Stratbucker, R. A., 1995, Optimization of cardiac defibril
lation by threedimensional finite element modeling of the human thorax, IEEE Trans. Biomed. Eng. 42:
185192.
Panfilov, A. v., 1997, Modelling of reentrant patterns in an anatomical model of the heart, in: Computational
Biology of the Heart (A. Panfilov and A. Holden, eds.), Wiley, New York, pp. 259276.
Penland, R. C; Harrild, D. M., and Henriquez, C. S., 2002, Modeling impulse propagation and extracellular
potential distribution in anisotropic cardiac tissue using a finite volume element discretization, Comput.
Visual. Sci. 4: 215226.
Pilkington, and Plonsey, R., 1982, Engineering Contributions to Biophysical Electrocardiography, IEEE
Press, New York, chapter 3.
Pullan, A., 1996, A highorder coupled finite elementlboundary element torso model, IEEE Trans. Biomed. Eng.
43: 292298.
Ramanathan, and Rudy, Y, 2001a, Electrocardiographic imaging: Effect of torso inhomogeneities on body
surface electrocardiographic potentials, Cardiovasc. Electrophysiol. 12: 229240.
Ramanathan, C., and Rudy, Y, 2001b, Electrocardiographic imaging: II. Effect of torso inhomogeneities on non
invasive reconstruction of epicardial potentials, electrograms, and isochrones, Cardiovasc. Electrophysiol.
12: 241252.
Rosenfeld, M., Tanami, R., and Abboud, S., 1996, Numerical solution of the potential due to dipole sources in
volume conductors with arbitrary geometry and conductivity, IEEE Trans. Biomed. Eng. 43: 679689.
Rudy, Y, 1987, The effects of the thoracic volume conductor (inhomogeneities) on the electrocardiogram, in:
Pediatric and Fundamental Electrocardiography (1. Liebman, R. Plonsey and Y Rudy, eds.), Martinus
Nijhoff, Boston, pp. 4972.
Rudy, Y, and Plonsey, R., 1980, A comparison of volume conductor and source geometry effects on body surface
and epicardial potentials, Circ. Res. 46: 283291.
Rudy, Y, Plonsey, R., and Liebman, J., 1979, The effects of variations in conductivity and geometrical parameters
on the electrocardiogram, using an eccentric spheres model, Circ. Res. 44: 104111.
Schmidt, 1. A., and Johnson, C. A., 1995, DefibSim: An interactive defibrillation device design tool, in: Proc. 17th
Ann. IntI. Con! IEEEEng. Med. Bioi. Soc., IEEE Press, New York, pp. 305306.
Schmitt, O. H., 1969, Biological information processing using the concept of interpenetrating domains, in: Infor
mation Processingin the Nervous System(K. N. Leibovic, ed.), SpringerVerlag, New York, pp. 325331.
Selvester, R. H., Solomon J. C., and Gillespie, 1968, Digital computer model of a total body electrocardio
graphic surface map. An adult maletorso simulation with lungs, Circulation 38: 684690.
Sepulveda, N. G., Roth, B. J., and Wikswo, J. P., Jr., 1989, Current injection into a twodimensional anisotropic
bidomain, 1. 55: 987999.
Simms, H. D., Jr., and Geselowitz, D. B., 1995, Computation of heart surface potentials using the surface source
model, Cardiovasc. Electrophysiol. 6: 522531.
Skouibine, K., Trayanova, N., and Moore, P., 2000, Success and failure of the defibrillation shock: insights from
a simulation study, Cardiovasc. Electrophysiol. 11: 785796.
Sobie, E. A., Susil, R. and Tung, 1997, A generalized activation function for predicting virtual electrodes
in cardiac tissue, Biophys. 73: 14101423.
Solomon, J. and Selvester, R. H., 1971, Myocardial activation sequence simulation, in: Vectorcardiography 2
Hoffman, ed.) NorthHolland Publishing Company, Amsterdam, pp. 175182.
Solomon, 1. C., and Selvester, R. H., 1973, Simulation of measured activation sequence in the human heart, Am.
Heart 1. 85: 518523.
Stanley, and Pilkington, T. 1989, The combination method: A numerical technique for electrocardio
graphic calculations, IEEETrans. Biomed. Eng. 36: 456461.
Stanley, P. Pilkington, T C., and Morrow, M. N., 1986, The effects of thoracic inhomogeneities on the
relationship between epicardial and torso potentials, IEEETrans. Biomed. Eng. 33: 273284.
Taccardi, B., Macchi, E., Lux, R. Ershler, P. E., Spaggiari, S., Baruffi, S., and Vyhmeister, Y, 1994, Effect of
myocardial fiber direction on epicardial potentials, Circulation 90: 30763090.
The Forward Problem of Electroeardiography 79
Thivierge, M., Gulrajani , R. M., and Savard, P., 1997, Effect s of rotational myocardi al anisotropy in forward
problem computations with equivalent heart dipoles, Ann. Biomed. Eng. 25: 477498.
Trudel, M.C. , Gul rajani , R. M., and Leon, J., 200I , Simulati on of propagat ion in a reali sticgeometry computer
heart model with parallel processing, in: Proc. 23rd Ann. Inti. Con! IEEE Eng. Med. BioI. Soc., IEEE Press,
New York, available on CDROM.
Tung, L., 1978, A bidomain model for describing ischemic myocardial doc potentials, Ph.D. thesis, Massachusetts
Institute of Technol ogy, Cambridge, MA.
van Oosterom, A., and Plonsey, R., 1991, The Brody effect revisited, Electrocardiol. 24: 339 348.
Walker, S., and Kilpatrick, D., 1987, Forward and inverse electrocardiographic calculations using resistor network
models of the human torso, eire. Res. 61: 504513.
Wei, D.. 1997, Wholeheart model ing: Progress, principles and applications, Progr. Biophys. molee. Bioi. 67:
17fJ6.
Werner, C. D., Sachse, 8., and Dossel, 2000, Electrical excitation propagation in the human heart, Int.
Bioeleetromagnetism 2(2): hitpu/wwwee.tut.fi/rgi/ijbem
Wiggers, C. J., 1940, The physiologic basis for cardiac resuscitation from ventricular fibrillation Method for
serial defibrillation, Am. Heart 1. 20: 413422.
Wischmann, H.A., Drenckhahn, R. , Wagner, M., and Fuchs, M., 1996, Systematic distribution of the auto solid
angle and related integrals onto the adjacent triangles for the node based boundary element method,
BioL. Eng. Comput. 34 (Suppl. 1, Part 2): 245 246.
Zipes, D. P., Fischer, J., King, R. M., Nicoll, A. D., and Joll y, W. w. , 1975, Termination of ventricular fibrillation
in dogs by depolarizing a critical amount of myocardium, Am. Cardiol. 36: 3744.
3
WHOLE HEART MODELING AND
COMPUTER SIMULATION
DamingWei
Graduate Department of Information System,The Universityof Aizu, Japan
3.1 INTRODUCTION
Bioelectrical models of the heart are studied in three levels: the singlecell model, the
cellnetwork (tissue) model and the whole heart model (Wei, 1997). The singlecell model
describes ionic current flow across myocardial cell membranes. The cellnetwork model
describes ionic current flow between aggregates of myocardial cells in temporal and spatial
domains. Details of these model s have been described in the previous chapters. The devel
opment and spread of ionic currents throughout the heart and body volume conductor result
in electrical potentials that can be measured on the body surface, called electrocardiogram
(ECG). A whole heart model describes three major mechanisms: the propagation of activa
tion in the heart , the cardiac electrical sources, and the extracellular potentials within and
on the body surface. Thi s kind of model is able to relate the body surface ECG waveforms
to the action potential, conduction velocity of cardiac tissue and other electrophysiological
properties of the heart and, thus, yield clinically comparable ECG waveforms. For this rea
son, the whole heart model offers a unique means to bridge the clinical applications with
the singlecell or cellnetwork models .
Conventionally, a whole heart model refers to a 3dimentional (3D) hearttorso model
that contains realistic geometry of the heart and torso. Some heart models contain both atria
and ventricles to represent the entire heart , but some only contain a portion of the heart such
as the ventricles, or just the left ventricle. A computer heart model is usually represented
by a 3D voxel array. The voxel in each grid is called an element. As the array increases in
size, the model exhibits higher resolution. In most of the existing whole heart models, the
element is of an order of 1 mm in diameter. It is usually several thousand times larger than
a true cardiac cell . In simulation studies, the element is usually treated as a cardiac cell,
called a model cell in this chapter, but it is important to keep in mind that the element or
model cell in a whol e heart model is actually a lumped model of tissue.
Tsuruga, Ikkimachi, AizuWakamatsu City, Fukushima 9658580, Japan, Tel. +81242372602, Fax. +81242
372728, dmwei@uaizu.ac.jp
81
82 D.Wei
Whole heart modeling and computer simulation are typical topics associated with
the electrocardiographic forward problem. Since the main theory and methodology for the
forward problem have been described in detail in Chapter 2, this chapter will focus on howto
construct a 3D hearttorso model, how a whole heart model can be applied to understanding
mechanisms of bioelectric phenomena of the heart, and to relating the electrical activity of
the heart to the surface ECG.
3.2 METHODOLOGY IN3D WHOLE HEART MODELING
The main issues in whole heart modeling shall include: hearttorso geometry modeling,
specialized conduction systems, rotating fiber direction and myocardial anisotropy, action
potential and propagation, cardiac electric sources, and body surface potential calculation.
These topics are described in detail in the following sections.
3.2.1 HEARTTORSO GEOMETRY MODELING
Descriptions of heart anatomy can be found in a vast number of textbooks. For heart
modeling purposes, one should acquire some basic knowledge of the heart's anatomic
features. First of all, we should have a quantitative concept of the heart's size. One point
is that the left ventricle including the left free wall and the interventricular septum (wall
thickness 911 mm) constitute the major mass of the heart. The right ventricular wall is
about 1/3 the thickness of the left and, therefore, about 1/27 of the total myocardium in
volume. The wall of the atria is even thinner. This means that the left ventricular myocardium
provides the major contribution to the generation of ECG.
To construct a heart model with a realistic shape, the first step is to map the heart
geometry onto a matrix of a regular 3D grid. In the early stage of model development
(Okajima et al., 1968), the heart geometry was constructed using traditional digitization
techniques. With such a method, a heart is first embedded in a gelatin solution and then
frozen. After that, the frozen heart is sliced and photographed, and then pictures of each
slice are digitized and input into a computer.
For construction of highresolution heart models, it is necessary to use CT imaging
technology. In the study of Rorange et al. (1993a), a human heart obtained from autopsy
was first inflated with air at enddiastolic pressures and submerged in liquid nitrogen. The
frozen heart was then CTscanned in 1 mm intervals resulting in 132 slices of 512*512
image data. These data were processed to extract the heart edges and eventually to obtain
a 3D matrix of about 250,000 points spaced 1 mm apart.
Recently, several databases provide accurate and anatomically detailed images of
the human body available for model studies. One of the most commonly used databases
for this purpose is the Visible Human Project of the U.S. National Library of Medicine
(http://www.nlm.nih.gov/research/visible/visibleJmman.html). It provides 1,871 section
images of a human cadaver at 1 mm intervals. The original images are 1048 by
1216 pixels with 24bit color. These data have been used in recently published models
(Balasubramaniam et al., 1997; Kauppinen et al., 1999; Ramanathan and Ruddy 2001a).
The main techniques to model the hearttorso shapes include image segmentation and
element generation. To the author's knowledge, there is no full automation technology
Whole Heart Modeling and Computer Simulation 83
currently available for image segmentation for the purpose. In most studies, the segmen
tation is performed in a semiautomatic manner, which includes steps of digital filtering,
image enhancement, region growing processing, and segment decision (Heinonen et al.,
1996; Hsiao and Kao, 2000). After these procedures, the boundary contours of distinct
anatomical regions are obtained. Based on these data, the anatomical regions are filled with
discrete elements representing distinct cardiac cell types. The cell types usually include
cells of the atria, ventricle and specialized conduction system. For an inhomogeneous torso
model, the volume is usually divided into piecewise homogeneous regions that may include
the heart with blood mass, the lung, the fat, and the bones, with distinct regions being as
signed different values of electrical conductivity. Cell classification can only be performed
manually with the current techniques.
An important issue in whole heart modeling is spatial and temporal resolutions of
the model. Qualitatively, the model's resolution mainly affects the propagation details,
especially if rotating fiber directions are taken into account. For surface ECG calculations,
the resolution of most existing heart models is sufficient, considering the fact that many
studies reduce the total number of elemental dipoles to a few multiple dipoles before
calculating the surface ECG potentials. A heart model of 1 mm spatial resolution has
enabled the simulation of reentrant propagation such as spiral waves (Panfilov 1993, 1995).
In a recent model of Hren and Horacek (1997, 1998), a high resolution 0.5 mm is adapted
with 1.7 million cells. The heart model has enclosed more anatomic details including the
myocardial wall, trabecular tissue, and papillary muscles.
3.2.2 INCLUSION OFSPECIALIZED CONDUCTIONSYSTEM
For studying bioelectrical phenomena, the heart can be simply thought of as an electric
generator comprising a specialized conduction system (SCS) and excitable myocardial
tissues, as diagrammed in Fig. 3.1. Understanding the SCS is important for whole heart
modeling because it is one of the key issues that affect the propagation sequence of activation
in normal and abnormal hearts. In the following paragraphs, preparative knowledge for the
activation of the heart and the role of the SCS are briefly introduced while referring to Fig.
3.1. Detailed references can be found in other references (Guyton, 1986; Van Dam, 1989).
For a normal heart, the electric pulse is spontaneously generated by the sinus node,
the pacemaker (located in the upper right atrium as shown in Fig. 3.1). The pulse stimulates
the neighboring atrial cells to activate them. The activating cells in tum stimulate their
neighboring cells so that the excitation propagates throughout the atria. Although some
studies suggest that there may exit a specialized conduction system in the atria, there is no
definite evidence to support this assumption.
As clearly illustrated in Fig. 3.1, there is no direct "electric connection" between the
atria and the ventricles, except in some abnormal hearts such as those suffering from the
WPW syndrome (described later in Section 3.4). For normal hearts, the only pathway for
conduction of the excitation between the atria and the ventricles is the atrioventricular node
(AVnode) and a specializedconduction system consisting ofthe His bundle, the left and right
bound branches, and the Puikinje network, as illustrated in Fig. 3.1. The most important fea
ture of the specialized conduction system is the high conduction velocity in comparison
to that of normal myocardial fiber. The typical conduction velocity in the His bundle and
the main branches is 2 mS while that in the ventricular myocardium is 0.5 mS The
84
sinus
node
FIGURE 3.1. A schematic diagram of the heart and the specialized conduction system.
D.Wei
distribution of the Purkinje network influences the heart excitation sequence the most, and
therefore is very important to the model study. The Purkinje fibers penetrate the septal my
ocardium near the apex and are distributed in the subendocardium. In general, the Purkinje
network is distributed on the apical half of the subendocardium and absent on the basal half.
This configuration avoids outflow obstruction.
In whole heart modeling, distributing the Purkinje network in the ventricular model
greatly influences the excitation process of the heart model and the resulting body surface
ECG. Because it is not possible to identify the Purkinje fiber from image data, modeling
the Purkinje network is usually based on the excitation sequence of the ventricles observed
by experiments. The most cited literature regarding human heart excitation is that of Durrer
et al. (1970). The measured excitation isochrones in isolated human hearts are shown in
Fig. 3.2. According to Durrer et aI., the early excited areas in the left ventricle were observed
in three endocardial sides: high on the anterior paraseptal wall just below the attachment
of the mitral valve, central on the left surface of the interventricular septum, and posterior
paraseptal about one third of the distance from the apex to the base. Early excitation of the
right ventricle was found near the insertion of the anterior papillary muscle. Septal activation
was found to start at the middle third of the junction of the septum and posterior wall. These
data are useful for arranging the Purkinje fiber in a heart model and for evaluating the
simulation results of the model.
Inclusion of the SCS to a heart model is usually a timeconsuming task in the con
struction of a model. The positions and distributions are usually adjusted repeatedly until
Whole Heart Modeling and Computer Simulation
o
8S
FIGURE 3.2. Excitation sequence of the human heart. (Reproduced with permission from Durrer et aI., 1970).
simulated excitation isochrones fit the experimental data. In the adjustment, the distribution
of the Purkinje network, the position of joint points connecting bundle branches to the
Purkinje network, and the activation time arriving at these points have the most important
effects on simulation results.
In most whole heart models, the Purkinje network is represented by a onelayer sheet
and the bundle branches are represented by cables (Aoki et al., 1987; Rorange et al., 1993).
The sheet model of the Purkinje fiber network is supported by some experimental studies.
In the model of Aoki et aI., the left bundle branch terminates at three points: the central region
of the septum, the anterobasal region, and the posteroapical region of the left endocardium.
The right bundle branch terminates at one point on the anteroapical region of the right
endocardium. A more detailed model of the specialized system can be found in AlNashash
and Lvov (1997), where the HisPurkinje electrogram is the target of simulation. In this
study, a HisPurkinje model with a 3D curvature resembling the ventricular endocardial
surface is built and the HisPurkinje system electrogram is simulated using the volume
conductor theory.
3.2.3 INCORPORATING ROTATING FIBER DIRECTIONS
Myocardial anisotropy is an important issue in studying cardiac phenomena. Nowa
days, a whole heart model without inclusion of myocardial anisotropy would hardly
be accepted. Inclusion of myocardial anisotropy in a heart model involves three as
pects: anisotropic geometry of the myocardial muscle, anisotropic propagation, and
anisotropic cardiac sources. Anisotropy geometry means that a heart model should have
86 D.Wei
locationvarying fiber directions for all discrete elements, based on the fact that the ventric
ular myocardium has a spiral structure with fiber orientations rotating from the epicardial
surface to the endocardial surface in a total angle of 90
0
to 120
0
(Streeter et aI., 1969). This
is known as rotational anisotropy. Anisotropic propagation requires a directiondependent
conduction velocity in controlling the excitation process of the model. The anisotropic
cardiac source arises from the fact that both the intracellular and extracellular domains
are anisotropic, and the anisotropic ratios are essentially different from point to point in
myocardial muscle.
There are several ways to incorporate rotating fiber directions in a heart model. The
simplest way is called stylized representation, as used in Lorange et al. (1993). With this
method, a family of nested ellipsoids of revolution extending from the endocardium to the
epicardium is used, where the fiber angle varies 120
0
from the endocardial to the epicardial
ellipsoids. The fiber direction at a model point is obtained by determining which ellipsoid
passes through the point. A similar method is used in other models (Adam et al., 1987;
Leon et aI., 1991a, b). The advantage of this method is the simplicity in calculating the
pointbypoint fiber direction. The disadvantage is that the fiber directions are too simple
in comparison with actual heart anatomy.
The most precise way to incorporate fiber direction is through microscopic determi
nation of fiber orientation as reported in Panfilov et al. (1993). In this study, detailed data
of the ventricular geometry and fiber orientation were microscopically measured on an
intact canine heart to produce a finite element model (Nielson, 1991). These data were
then mapped onto a regular 93*93*93 grid with 1 mm of distance between the grid points.
Compared to the stylized representation, this method provides more geometric details of
the heart.
Microscopic determination of fiber orientation is not always possible for model studies.
Wei et al. (1989, 1995) proposed a discrete method tocalculate pointbypoint fiber direction.
This method can generate intermediate precision between that of microscopic determination
and stylized representation. This method uses the following assumption based on Streeter
et al. : (1) the myocardial fibers of the ventricles have a layered structure, (2) all fiber
orientations are parallel to each other in one layer but different from layer to layer, (3) the
fiber orientations rotate counterclockwise over 90° to 120
0
with increasing depth from the
epicardium to the endocardium.
Fig. 3.3 (a) shows a longitudinal crosssection of the heart model. The spatial configu
ration of the elements is illustrated in Fig. 3.3(b). To assign each element a fiber direction,
the model is layered as shown in Fig. 3.3(c). The myocardial fibers in one layer are math
ematically described as intersection curves by cutting the layer with a group of parallel
planes, called fiber planes, as illustrated by Fig. 3.3(d) and (e). Obviously, the fiber planes
in one layer have a common normal direction, called fiber plane direction (FPO). Follow
ing experimental data (Spaggiari , 1987), an FPO perpendicular to the geometric heart axis
along the apextobase direction is assigned to the epicardial layer. Then, the FPO is rotated
counterclockwise in the septal plane to determine FPOs for all layers in the epicardiumto
endocardium sequence. The rotating angle for a layer with respect to that of the FPO of
epicardium is given by
A
= (3.1)
WholeHeart Modeling andComputer Simulation 87
FIGURE 3.3. Heart model of Wei et al. (a) Longitudinal crosssect ion of the heart model. P denotes the Purkinje
fiber. (b) 3D configuration of model elements . (c) Model elements layered for solving pointbypoint fiber direc
tions. (d) Illustration of rotational fiber orientations in the layers. (e) Mathematical relationships of fiber plane,
fiber plane direction, and fiber plane rotation. FPo: fiber plane of the outermost layer; FPj: fiber plane of the ith
layer; FPDj: fiber plane direction of the ith layer; ali): rotating angle of layer PS: septal plane (Reproduced with
permi ssion from Wei et al., 1995).
88 D.Wei
.........
... .... .....
• If • ••
• # •• • • • • ••• • •••
•••• If • • I t • • • • ••
• ••••• • • • ••• 10 •••
.
.. .
..... ............................ ..
..
.. .. ..
:: : : ::::::::: :::::::::: :::::::: :::.:::: ::::
layer
lnyei ]
FIGURE 3.3.
where N is the total number of layers of the model and A is the total rotating angle between
the epicardial and endocardial layers, having a value between 90° and 120°.
In implementation, the heart model is layered from the epicardium to the endocardium
by applying simultaneous stimuli to all model elements on the outermost layer (the epi
cardium) of the model and solving the propagation sequence of "virtual excitation" for the
entire model. The septum of the heart model is specially treated in the propagation process
so as to make the septal fibers natural extensions of the left ventricular fibers. As a result,
Whole Heart Modeling and Computer Simulation 89
FIGURE 3.3.
the sequence number of such a propagation process corresponds to the layer number of
the model element. Then, the outer product of two unit vectors determines the local fiber
direction at each element. One is the fiber plane direction of the layer to which the unit
belongs, and the other is the normal direction of the layer at that element. If an element
i . belongs to layer and the layer has a fiber plane direction of and a normal
direction at i . the local fiber direction is given by
x i . (3.2)
To confirm the rotating fiber directions of the heart model , the model is stimulated at the
left ventricular wall at three depths : at the epicardial layer, the intramural layer, and the
endocardial layer. The isochrones are shown in Fig. 3.4. The long axes ofthe isochrones show
approximately 90° of total rotation from the epicardial (Fig. 3.4b), intramural (Fig. 3.4c),
and endocardial (Fig. 3.4d) layers. For comparison, propagation in the isotropic model is
shown in Fig. 3.4a.
Determination of fiber direction can also be performed in terms of analytical methods.
In the study of Hren and Horacek (1997), the epicardial and endocardial surfaces are ana
lytically obtained with sampling data from CT images using surface harmonic expansion
(Hren and Stroink, 1995). Then, the principal fiber direction at a point lying in a tangential
plane at an element is determined so that the fiber direction rotates counterclockwise from
the epicardial to the endocardial surface.
3.2.4 ACTIONPOTENTIALS AND ELECTROPHYSIOLOGIC PROPERTIES
In simulation studies , the model cell mimics the actual cardiac cell to reconstruct the
cardiac process . Like an actual cell , the model cell should own the action potentials and
other electrophysiologic properties. These properties are the main input data for simulation
studies.
..
.
• ••• • : • • • • • • • • . • . • :::•.• • : : . • : ••••• •• • : • .• , 0 ••••••• •• '0' '"0 0°: •
.. ... ..... . . . .......... .. .. ..... .... ..... . .... .. .. .. . . ,0 0 ••
........ . . . .. ... .. .. .. . . 0 . 0 0
.. .. . . . .. . . . . . . . 0.0 ..
.... .. .. .. .. .. .. .. .. .
•••••• • 0 • • •• ..
' 0' ••• ••••• 0 0 •
FIGURE3.4. Simulated isochrones created by applying stimuli at different depths in the left ventricle of the heart
model. The stimulus is denoted by Time sequence is identi fied by the thickness of the circle and its center
point: the thicker the circle and the center point, the more advanced the time sequence. Each increase in thickness
of the circle represent s 6 ms, and that of the center point represent s 36 ms (circles with no center points represent
the first 36 ms).
(a) Epicardi al isochrones of the left ventricle with the isotropic propagation
(b) Epicardial isochrones of the left ventricle with anisotropic propagation
(c) Intramurallayers beneath the left epicardium with anisotropic propagation
(d) Intramural isochrones 9 layers beneath the left epicardium with anisotropic propagation
The isochrones from (b) through (d) show rotating fiber directions in the heart model. (Reproduced with permission
from Wei et al., 1995).
Whole Heart Modeling and Computer Simulation
~ t : " l
. .. .. ......... . . .... . .......... .... . .. . .. .. .... .. .... . . . .
... ... .. . . . . .... . .... .. .. .. . .. . .. .. .. . . . .. .. .. ....... ..
.. .... ... ... .. .. .... .... .. . .. ..
.. .. . .. . .. .... .... .. . . . . . . . . . . . ...... . .. .. .. ...... ...
.. .
. ... . .. .. .. .. .. . ... . ..... .. . ...... . .. .. . .. .. . .. ..
.. . . .. . . . .. . .. . .
.. .
91
. ....... .. .... . ..
.. .. . . .. .. .. .. .. .. ..
i: ·
(d) ~
FIGURE 3.4.
The ideal way to assign action potentials to the model cell is to use the singlecell
model to generate action potential as a function of time and distances. However, since the
calculation of action potentials with singlecell models is computationally demanding and
the scale of a whole heart model is quite large, most existing whole heart models do not
directly calculate the action potentials, but assign predefined action potentials to the model
cells.
For an electrophysiological simulation study, one should take into account at least
the following basic nature of action potential in modeling. First, different kinds of cardiac
cells have different action potentials. Second, at least for the ventricular cells, the action
potential duration (APD) is locationdependent in the heart, longest on the endocardium and
the base, and shortest on the epicedium and the apex (Harumi et al., 1964). Third, the APD
92 D.Wei
is timevarying during premature excitation, depending on the restitution property with the
coupling interval (Harumi et al., 1989a).
In many studies, the cardiac cell is represented by finite state automaton. This is called
a cellular automata (CA) model (Siregar et aI., 1996). In fact , the CA model was first used
in an early study by Moe et al. (1964) to simulate atrial fibrillation . In this study, a discrete
time step of 5 ms and a unit diameter of 4 mm were used. One time step was the propagation
time in a fully recovered unit , corresponding to a normal conduction velocity of 80 cmls
for the atrial tissue. The excitability of the units was represented by 5 states. The first state
corresponded to the absolute refractory period, having a duration of R KJC, where C
was the preceding cycle length and K was a property of the unit, different from unit to
unit. States 2, 3 and 4 together corresponded to the relative period. Dividing the relative
period into different states incorporated the timevarying conduction velocity during the
propagation process. From state 2 to state 4, the conduction velocity increased from 1/4
to 1/2 of the normal conduction velocity. In state 5, the conduction velocity returned to
normal. State 5 lasted until the next excitation and corresponded to the recovery period.
Because the model of Moe et al. was aimed at studying the propagation of activation,
the waveform of the action potential was not needed. For models that need to calculate ECG
potentials, the waveform of the action potential should be defined. A simple representation is
the piecewise linear approximation as used in Siregar et al. (1996). In this model , the muscle
cell is resented by four states : resting, depolarization, absolute refractory, and absolute
refractory. The function of the action potential, Y, is expressed by
(3.3)
where e and i denote model cell and state, respectively, a and define a line for the cell
and state. As functions oftime, a and are contextdependent so that the action potential is
timevarying according to the prematurity of the incoming impulse. This model is capable
of simulating typical types of arrhythmias.
In assigning action potentials to the model cells, there is a technical difficulty due to
the fact that a 3D heart model usually has a huge number of cells. each model cell has
to be accompanied with a dataset containing an action potential and other parameters, the
memory space for modeling would be extremely large. One method for solving the memory
problem used by Wei et al. (1995) is to assign the action potentials and electrophysiologic
parameters to cell types, instead of individual cells. In this way, model cells are organized
in limited cell types. Action potentials and other electrophysiologic properties are linked to
cell types so that limited memory is required. There are sixteen cell types available in the
model of Wei et al. Each of them is associated with a parameter table containing the action
potential waveform and other electrophysiological parameters as listed in Table 3.1. The
available cell types include normal cardiac cells like the atria, ventricles, and a specialized
conduction system. Besides, any special cell types, such as ischemia, infarction, or ectopic
beats, can be defined by parameter settings suitable to the study.
The following paragraphs show how the action potential and other electrophysiological
details are assigned to the model cell s in Wei et al. ( 1990, 1995). See Table 3.1 for meanings
and abbreviations for the parameters, and see Fig. 3.5 for the action potential definition.
The basic parameter relative to action potential is action potent ial duration (APD) in
ms, taken as the sum of phases through 4. The parameters relative to time are given as a
TABLE3.1. Electrophysiologic Parameters
Parameter
Action Potential
APD
TO
Tl
T2
T3
VO
VI
V2
GRD
DVT
DC
Conduction:
CVL
DC
Automaticity:
ICL
PRT
DLY
ACC
Pacing:
BCL
BN
ICN
Definition
action potential duration (ms)
duration of action potential phase 0 of APD)
duration of action potential phase 0 of APD)
duration of action potential phase 0 of APD)
duration of action potential phase 0 (% of APD)
duration of action potential phase 0 (% of APD)
potential of phase 0 (resting potential) (mv)
maximum action potential (mv)
potential of phase (mv)
potential after full recovery (mv)
gradient of the APD distribution (mslIayer)
deviation of APD for random distribution (mv)
APD change to coupling interval
conduction velocity along the fiber axis (m1sec)
anisotropic ratio (m1sec)
conduction velocity after ARP (m1sec)
intrinsic cycle length (ms)
protected or nonprotected (yes/no)
maximum delay of the phase response of ICL)
maximum acceleration of the phase response (% of ICL)
break point position of the phase response (%)
basic cycle length of pacing (ms)
beat number of pacing for the simulation
increment of cycle length per cycle of BCL)
YO
T2
4
FIGURE 3.5. Definition of action potential used in the simulation (Reproduced with permission from Wei et aI.,
1995).
94 D.Wei
percentage of the APD. Lines specified by time intervals and voltages are used to represent
phases 0, I, 2 and 4 of the action potential waveform. Phase 3 is defined by a curve through
interpolation (second order Lagrange interpolation) of sample data. The action potential
waveform is linked to each cell type and used as a lookup table during simulation.
To distribute the predefined action potential waveforms over the ventricles so that
the APD lengthens from the epicardial to the endocardial and from the base to the apex,
a parameter GRD (see Table 3.1) is defined to specify the gradient interval along the
epicardialtoendocardial and basetoapex sequence. Thus, the value of APD for a cell at
location is given by
APD(i, i. k) APo, GRD· SQ(i, k) (3.4)
where APDd is the defined value of APD for the cell type, and SQ(i, i. k) is a sequential
number of the cell sorting along the epicardialtoendocardial and basetoapex directions.
In this model, a positive GRD value of about 5 ms per sequence yields a normal T wave in
the simulated ECG. Adjusting the parameter GRD in a simulation study is useful to simulate
T wave abnormalities.
To make the APD adaptive to the coupling interval so that it is dynamically modified
during the simulation, a parameter called dynamic coefficient (DC) is defined as the ratio
of the change in APD to the change in coupling interval. During simulation, the value of
APD at any time, t, is dynamically modified by
APD(t) = APD(t  DC . l(t) (3.5)
where is the change in coupling interval at time t.
In addition to the action potentials, many other electrophyosiologic properties are
defined with parameters. These parameters are concerned with conduction velocity, pacing,
and automaticity. The conduction velocity is made adaptive by assuming a piecewise linear
function of the coupling interval. During the absolute refractory period, the conduction
velocity is zero. After full refractory, the conduction velocity is assumed constant specified
by a parameter, CVL (see Table 3.1). During the relative refractory period, the conduction
velocity is a linear incremental function of the coupling interval. To define a cycle length
and pacing times, anyone or more cells can be paced in the simulation. The pacing rate
can be increased or decreased with the changing rate specified by a parameter INC, which
is defined as a percentage increment to the basic cycle length. The parameters relative to
automaticity define the electrotonic interaction between normal and ectopic pacemakers
and is used to simulate cardiac arrhythmias (Wei, 1992).
3.2.5 PROPAGATIONMODELS
There are three types of propagation models used in whole heart modeling. These are
propagations based on Huygens' principle, HodgkinHuxley (HH) formulism (Hodgkin
and Huxley, 1952; Plonsey, 1987; Leon et al., 1991a, b; Beeler and Reuter, 1977; Luo and
Rudy, 1991, 1994a, b), and the FitzHughNagumo (FN) model (FitzHugh, 1961; Panfilov
and Keener 1995; Berenfeld, 1996). The model using Huygens' principle provides the
simplest way to simulate the propagation of the action potentials. It requires calculations to
Whole Heart Modeling and Computer Simulation 95
generate wavelets around activating cells at each discrete instant. Cells within the wavelets
are to be activated at the next instant. Constructing the wavelets for each instance yields
the excitation sequence of the heart model. The effectiveness and the details of this type of
model depend on the way the action potential and other electrophysiologic properties are
defined. the definitions contain sufficient electrophysiologic details, the heart model is
efficient enough to describe the dynamic process of the heart such as the cardiac arrhyth
mias. The propagation model based on HH formulism is theoretically most correct. It has
the capability to respond to ionic current, membrane potentials, and other factors to provide
detailed information that cannot be provided by the former. However, the computation com
plexity of such a model limits the application in largescale models. While most published
studies using HH formulism in simulating propagations are twodimensional models, many
researchers currently show increasing interest in applying the HH type propagation to 3D
whole heart modeling. A recent work is discussed in Chapter 2. Propagation using the
FitzHughNagumo (FN) model is a compromise between the implementation complexity
and the electrophysiologic details. The FN type cellular dynamics is an electrodynamic
model that reconstructs the action potential by reaction state equations. Compared to HH
type equations, numerical solutions for FN type propagation need much less computation
time.
3.2.5.1 Propagation model ofHuygens' type
As an example, the propagation of the Huygens' type is introduced with the model of
Wei et al. In this model, the excitability of model cells is one of two basic types: conductive
or nonconductive. Nonconductive activation is concerned with pacemaker cells. For these
cells, the activation is obligatory whenever the firing time comes. Conductive activation is
treated in three steps. The first step is to calculate the extent of propagation in one time step
around each excited model unit by constructing an ellipsoidal wavelet based on local fiber
direction as shown in Fig. 3.6. The long serniaxis of the wavelet is along the fiber direction,
and the other two short serniaxes are along the two transversal directions. The lengths of
the long and short serniaxes are
and
= Vt(t)· (3.6)
(3.7)
respectively, where T is the time step, Vt(t) is the longitudinal conduction velocity of the
cell at time t, and is the conductivity ratio. The extent of propagation in one time step is
described by
(3.8)
where n, are the distances along and across the fiber direction in the local coordinate
system established at the excited element.
96
Fiber direction
D.Wei
(b)
FIGURE 3.6. An illustration of the propagation wavelet.
Cells within the wavelet can be activated if they are excitable at that time. The ex
citability is checked by the principle of refractoriness as
(3.9)
where pre is the starting time of the previous excitation of the model unit and
A
RP(t) is
the absolute refractory period at time t.
the model cell is recognized as excitable, the third step is to assign a conduction
velocity to it for propagation at the next time step. Note that the conduction velocity in this
model is timevarying, depending on the coupling interval.
In this model , the automat icity of a pacemaker cell can be either protected or un
protected, depending on the parameter setting. For unprotected automaticity, stimuli from
neighboring cells unconditionally activate the unit and reset automatically after activa
tion. For the protected automaticity, the model unit is not directly activated by surrounding
stimuli, but its intrinsic cycle length is modulated based on a socalled phase response
curve (PRe). Fig. 3.7(a) shows a PRC assigned to the cells of an ectopic pacemaker to
simulate the ectopic firing of the ventricular parasystole. The biphasic line of the PRC was
Whole Heart Modeling and Computer Simulation
Delay (%)
SO
40
30
DLY
10
0
10
20
30
40
SO
(b)
97
FIGURE 3.7. (a) The phaseresponse curveassignedtocells of an ectopic pacemakerin simulatingectopicfiring
of ventricularparasystole(Reproduced with permissionfrom Wei et aI., 1995). (b) SN: sinus node, the normal
pacemaker; PVC: the ectopic pacemakercausing premature ventricular contraction. (c) From top to bottomare
simulated ECGsof bigeminy, trigeminy, quadrigeminy and an operativeratio of the 5:I type.
98
~ . o m v
O.2sec
D.Wei
FIGURE 3.7.
used to approximate the experimental data (Jalife, 1976; Moe, 1977). By assigning proper
values of intrinsic cycle length to sinus and ectopic pacemaker cells (Fig. 3.7(b», ECG
waveforms of bigeminy, trigeminy, quadrigeminy, and an operative ratio 5: 1 type were sim
ulated as shown in Fig. 3.7(c) . The simulated ECG waveforms are comparable with those
in clinical findings. They are also in agreement with a theoretical model where the cardiac
dynamics are represented by difference equations (Ikeda, 1983). The example demon
strates that a Huygens' type model is able to dynamically reproduce the electrophysiologic
process of the heart when adaptive action potential and other properties are used in the
simulation.
3.2.5.2 Propagation ofHodgkinHuxley type
The HH equation is expressed as
(3.10)
where L; is the transmembrane current density; V
m
<1>;  e is the transmembrane poten
tial; is the membrane capacitance; gk, gN
a,
g/ and Ei ; E
Na
, E/ represent the conductivity
and Nemst potential of potassium, sodium and leakage, respectively. Originated from the
HH model, the membrane ion kinetics has been modified based on new experimental find
ings in recent years, leading to a wide range of singlecell models such as the sinus cell
model (Noble, 1989), the atrial cell model (Earm and Noble, 1990), the Purkinje fiber
model (DiFrancesco and Noble, 1985), and the ventricular model (Beeler and Reuter, 1977;
Luo and Rudy, 1991, 1994a, b). For simulating 2D or 3D propagations of activation in the
ventricular myocardium, the Beeler and Reuter model (referred to as the BR model) and
the Luo and Rudy (referred to as the LR model) are widely used. The BR and LR models
are general mammalian ventricular cell models derived from experimental data and mathe
matically represented by HH type formulism. These models are expressed in physiological
parameters so that they are well suited for simulation studies. Use of BR model takes a
Whole Heart Modeling and Computer Simulation 99
smallercomputationloadthanthe LRmodeldoes, but the LRmodel has incorporatedmore
recent informationand provides more options such as the abilityto chang the extracellular
ion concentrations.
To model HH type propagation, it is necessary to link the transmembrane current
density, to an excitable tissue network arising from the electrical structure. For a one
dimensionaltissue model, the propagation of the actionpotential is described(Plonseyand
Barr, 1987)by
av,
lm=
Lnair, r
e
) az
(3.11)
wherea is the radius of the fiber, r, and r; are the intracellularand extracellularresistances
per unit length along the axial coordinate of z. The equationexpandedto 20 and 3D with
anisotropy canbefoundinPlonseyandBarr(1987).Usually, thepartialdifferentialequation
is solvednumerically with discrete steps in time and space domains.
One difficulty in the implementation of HH type propagation is that the numerical
solutionfor Vm(x, y, z, t) in a 3Dmodel sometimestakes an impracticalcomputationload.
This is the reason why early studies of HH type simulationused 20 tissue models (Virag
et al., 1998). On the other hand, however, HH type propagation can relate surface ECG to
cellularinformation whenusedin3Dwholeheart modeling. Recently, someresearchgroups
are tryingto realize HHtype propagation in 3Dwhole heart modelingin different ways. A
Japanese group is establishingsuch a model using a supercomputer(Suzuki et aI., 2001).
Gulrajani and coworkershave recently simulatedthe surface ECGsuccessfullyduring the
normal cardiac cycles using a 3D whole heart model employing the LR model using the
parallel computingtechnique(see Chapter 2 for details).
An alternative way to apply HH formulism to a 3D heart model is the combination
of HH and Huygens' type models, as reported by Leon et al. (1991). In this study, when
transmembrane potential is less than a potential threshold, the propagation in progress
withinthe model cell is governedby
av
= DVv  i;on(v) i
app
at
Vth (3.12)
wherev is the transmembrane potential;i;on,the ioniccurrent; i
app,
the appliedcurrent; Cm,
the membrane capacitance; and the conductivity tensor.
When the transmembrane potential is greater than the potential threshold, the model
reduces to a conventional Huygens' type and the propagationprocess is controlledby
V=!Ct,T,C) (3.13)
where! is a preassignedactionpotentialwaveformfor a cell type r andits valuevarieswith
time t, and the coupling interval C. In this manner, both subthresholdand suprathreshold
phenomenawere simulatedat a whole heart level.
100 D.Wei
3.2.5.3 Propagation using FitzllughNagumo model
Taking the model of Berenfeld (1996) as an example, the action potential using the
FitzHughNagumo model is represented by state equations of V and U:
dv
 =c(U +bV  V
3)+
z
dt
 = +a
dt c
(3.14)
(3.15)
where and are constants, and parameter c is an adaptive function representing the
reaction to the cellular mechanism of the depolarization and repolarization processes. As in
this equation, the variable of state V is able to represent the property of membrane potential.
It can be linked to the membrane current density by giving z the physiologic meaning of
stimulus intensity, and thus obtain
dv 3
c(  V' .
dt
(3.16)
where D is the diffusion tensor dependent on fiber orientation. The term z V' .
behaves the same as with the transmembrane current term in the cable equation. Equations
(3.16) and (3.14) constitute a socalled reactiondiffusion system. See details in Berenfeld
(1996) , where a complete cycle of ECG is simulated with the reactiondiffusion system
incorporating rotational anisotropy.
3.2.6 CARDIAC ELECTRIC SOURCES AND SURFACE ECG POTENTIALS
With any of the propagationstrategies described above, we can estimate transmembrane
potential Vm(x, y, z, t) at any voxel of the heart model. Therefore, we are able to determine
the cardiac sources based on the spatial distribution of transmembrane potential at any instant
of simulation (see Chapter 2 for details). is usually convenient to calculate elemental
sources at individual cells of the model, and then integrate them throughout the heart .
From the cardiac sources, the ECG potentials on the surface of the torso model can be
computed. Simulation of surface ECG potentials is one of the main purposes for whole heart
modeling.
The basis to determine the cardiac electric source is the bidomain theory. Solving the
ECG potentials on the torso model is a typical volume conductor problem. Both of these
topics are fully covered in Chapter 2. Therefore, the general principles and methodology will
not be repeated in this chapter. Instead, we introduce the algorithms and implementations
in our own model study as an example.
We start from the Miller and Geselowitz (1978) formula that is based on bidomain
theory:
(3.17)
where (J is the conductivity, ¢ is the membrane potential , and subscript refers to the
intracellular domain. says that the current dipole density is proportional to the spatial
Whole Heart Modeling and Computer Simulation 101
gradient ofintracellular potential distribution. In fact, the cardiac sources can be equivalently
expressed in terms of either intracellular or extracellular membrane potentials depending
on the form of effective interstitial conductivity used in simulation. Equation (3.17) can
also be modified (Geselowitz, 1989) as
(3.18)
where represents transmembrane potential. In this case, the effective interstitial conduc
tivity becomes a = a.; called bulk conductivity. Because the spatial gradient can be
approximated by the potential difference between a cell and its neighbors, this expression
is convenient for the implementation in whole heart models with discrete elements.
The cardiac source expression of Miller and Geselowitz applies to isotropic media.
the anisotropy in both intracellular and interstitial domains is taken into account, the
solution will become complicated. In Wei et al. (1995), the anisotropy was simplified by
assuming an anisotropic intracellular domain and an isotropic interstitial domain. With this
assumption, the expression of the cardiac dipole source has a similar form to (3.17):
(3.19)
where is the intracellular conductivity tensor.
The simplification of assuming an anisotropic intracellular domain and an isotropic
interstitial domain is reasonable if we analyze the measured values of the conductibility
anisotropy. The conductivity ratios measured by Clerc (1976) are approximately 9.0 for the
intracellular space and 2.4 for the extracellular space. Those measured by Roberts et al.
(1979) are approximately 5.8 and 1.5, respectively. Clearly, as compared to the intracellular
conductivity ratio, the extracellular conductivity ratio is quite small. In other words, the
effect of conductivity anisotropy is predominantly caused by intracellular anisotropy.
To calculate the conductivity tensor, it is convenient and reasonable to assume an
axially symmetric anisotropy. Thus, if we establish a local coordinate system with one axis
being along the fiber direction, and the other two perpendicular axes, and tz, being
within the transversal plan, the conductivity tensor can be expressed as
(3.20)
where a/ and at represent intracellular conductivity along and perpendicular to the fiber
direction. Then D, is expressed in the global system as
(3.21)
where and ti are unit vectors (column vectors) along each axis in the local coordinate
system. If we further assume an equal conductivity ratio and let the ratio along and across
the fiber axis be r, the conductivity tensor can be simplified as
(3.22)
102
where
T
D.Wei
(3.23)
In computing the surface ECG potentials , we (Aoki et al., 1987; Wei et al., 1995)
developed an algorithm that transforms Poisson equation to Laplacian equation to simplify
the volume conductor problem, and used the boundary element method (BEM) for solutions.
The problem originated from the Poisson equation with respect to the surface potential
¢ is expressed as:
(3.24)
Since the surface potentials are results of the primary source owing to the current dipoles
in the heart , and the secondary source owing to the boundary effect, we introduce an
intermediate variable ¢  ¢o with the boundary condition of
a1/J a¢o
an an
(3.25)
where ¢o is the potentials in an infinite medium, and n represents normal direction to the
body surface. Applying the BEM with respect to leads to
where G is the Green function of
G= _
4nlr 
(3.26)
(3.27)
where r is the distance from a dipole source and the integration is performed with respect
to ; ' . Discretization of the integral equation (3.26) leads to linear equations
N N
L
hij1/J
j = LgijqJ
fori = 1,2, . . . (3.28)
where hij and gij are coefficients depending on the torso geometry, and N is the number of
elemental triangles which approximates the torso surface. To ensure a unique solution, an
(N 1)th equation
'L
a
j1/Jj = 'Laj¢J
(3.29)
is added to (3.28), where aj is proportional to the area of an elemental triangle which has
an indexj. This condition is used to define a potential reference so that the surface potential
Whole Heart Modeling and Computer Simulation 103
integral is zero. Rewrite the (N simultaneous equations of (3.28) and (3.29) in a matrix
notation, we have
Hlp _GQo  A
The final solution for surface potentials is obtained as
where
and
3.3 COMPUTER SIMULATIONS ANDAPPliCATIONS
(3.30)
(3.31)
(3.32)
(3.33)
we look at the details of publications in the past, it is very clear that whole heart
models and simulations have been extensively employed in research investigating cardiac
mechanisms. Part of the significance of whole heart modeling is the capability of linking the
undergoing pathology to the ECG features used in clinical applications. Typical examples
are models of abnormal STT waves (Harurni, 1989b; Dube, 1996; Hyttinen et aI., 1997;
Abildskov and Lux, 2000), myocardial infarction (StarttiSelvester et al., 1989, Zenda et
aI., 2000), WPW syndrome (Lorange et aI., 1986; Wei, 1987, 1990), hypertrophy and car
diomyopathy (Harumi, 1989b; Wei et aI., 1999), reentrant and scroll waves in ventricular
arrhythmias and fibrillations (Panfilov, 1993; Gray and Jalife, 1996; Okazaki et aI., 1998;
Clayton et al., 2001a, b). Some new topics in the past five years are long QT syndrome
(Clayton et al., 2001, Okazaki and Lux, 1999), Twave altemans (Abildskov and Lux, 2000),
and late potentials (Yamaki et aI., 1999). Furthermore, recent studies have extended model
applications to the development ofnewinstrumentation techniques. Examples are pace map
ping (Xu et al., 1996, Hren and Horacek, 1997), epicardial mapping (Ramanathan and Rudy,
2001), Laplacian ECG mapping (Wuet aI., 1998; He and Wu, 1999; Wei and Mashima, 1999;
Wei, 2001), artificial heart (Zhang et aI., 1999), and impedance CT (Kauppinen et al., 1999).
In the following sections, some typical applications of the whole heart model are intro
duced. In each section, the undergoing pathology is briefly introduced before the description
of the model study. For details of pathology, please refer to medical textbooks. Three ma
jor reference books used in the description are Goldman (1986), MacFarlane and Lawrie
(1989), and de Luna (1993).
3.3.1 SIMULATION OF THE NORMAL ELECTROCARDIOGRAM
A correct simulation of the normal ECG is usually the first step for whole heart
modeling, before applying the model to special pathologic conditions. The normality of
104
TABLE 3.2. Definition of l 2lead Electrocardiogram
D.Wei
Lead
Bipolar limb leads
Augmented unipolar limb leads
aVR
aVL
aVF
Precordial leads
Vj(I = 1, 2, .. . , 6)
Definit ion
ER
II EF  ER
EF 
aVR (3j2)(ER  Ewe'>
aVL = (3j2)(EL  Ewe, )
aVF (3j2)(EF  Ewe, )
Ewo! is called Wilson Central Terminal.
Ewct (EI.+ ER+ EF)/3 .
E denotes potential, subscripts R, and F denote left arm, right arm
and left foot, respectively. See Fig. 3.8for electrode positions of V;.
TABLE 3.3. Electrode Positions for Recording 12lead Electrocardiogram
Elect rode
Left arm (R)
Right arm (L)
Left foot (F)
Right foot
VI
V2
V3
V4
V5
V6
Position
Left wrist
Right wrist
Left ankle
Right ankle
Right sternal margin, fourth intercostal space
Left sternal margin, fourth intercostal space
Midway betwee n V2 and V4
Left midclavi cul ar line, fifth intercostal space
Left anteri or axi llary line, V4 level
Left midaxi llary line, V4 and V5 level
the simulation results can be evaluated with the excitation sequence of the heart model, the
simulated vectorcardiogram (VCG) , l2lead ECG, and body surface isopotential maps. The
12lead ECG is the most popular lead system used in clinical practice. To evaluate models
and simulations with different pathologies, the results are usually compared with clinical
recorded l2lead ECGs. Recording the l2lead ECG requires 10 electrodes. Four electrodes
are placed on limbs to record six limb leads and the other six are placed on the precordial
chest wall to record the precordial leads. The definition of l2lead ECG is summarized in
Table 3.2, and the electrode positions for recording are shown in Fig. 3.8(a) and described
in Table 3.3 (Horacek, 1989). In model studies, torso models usually do not include the
limbs . The limb leads are moved to the closed positions on the torso. Because the potential
difference on the limb is sufficiently small , this does not significantly change the simulation
results. Most torso models are represented by polygon meshes, and the nodal point s may
not exactly overlap the electrode position. In this case, interpolation is usually needed to
calculate either the electrode position s from positions of the surrounding nodal points, or
the ECG potentials from the potentials on surrounding nodal points.
A typical ECG waveform is illustrated in Fig. 3.8(b). The waves of the ECG are
designated by Einthoven as P, Q, R, S, T, as shown in this figure. It is well known that the P
Whole Heart Modeling and Computer Simulation 105
~
,
V2·. · · .
1.0
u
S T s e ~ e n t
0.4 0.6 0,8 0.2
1+1+++4
8,
~
~
FIGURE 3.8. (a) Electrode positions for the precordial leads. (b) A typical electrocardiogram.
wave corresponds to atrial depolarization, the QRS complex to ventricular depolarization,
and the T wave to ventricular repolarization. The U wave is occasionally recorded after the
T wave, and its mechanism remains unclear (di Bernardo and Murray, 2002). The criteria
of normality for the ECG include time and amplitude standards. Clinically, the following
parameters are evaluated for the normality of ECG:
• Cardiac rhythm and heart rate
• PR interval and segment
• QRS interval
• QT interval
• Pattern and amplitude of P wave
• Pattern and amplitude of QRS complex
106
• Pattern and amplitude of T wave
• ST segment and T wave
• Mean electrical axis (frontal plane)
D.Wei
The limit of normal ECG can be found in MacFarlane and Lawrie (1989) and the
textbook of de Luna (1993). The following are typical values: P wave interval 100
ms, QRS interval 90 ms, PR interval 150 ms, and QT interval 380 ms. Because
most heart models do not yield absolute values for ECG potentials, the amplitudes can be
evaluated by patterns and relative amplitudes. For example, the aVR wave should have an
inversed waveform. The R wave should gradually increase from VI to V4 or V5 and fall
from V5 or V6. In simulation studies, the use of VCG is very convenient in evaluating
the simulation results. By summing the cardiac dipole sources in all model cells at each
instance into a single dipole and plotting it in the frontal, horizontal and saggital planes,
the VCG is obtained. It is theoretically close to the clinical VCG. From these figures, the
mean electrical axises can easily be estimated. The normal P, QRS, and T axises should be
between 0° to +90 30° to + 110°, and 0° to +90°. The body surface isopotential maps are
also easily used for evaluation. The typical patterns for a normal heart can be found in many
articles and textbooks (e.g., MacFarlane and Lawrie, 1989). Fig. 3.9 shows the simulated
results of a normal heart mode by Wei at al. In Fig. 3.9(c), potential distributions on the
torso during the QRS period are shown by isopotential countour maps. In most instances,
the distributions show dipole fields with one positive maximum on the anterior chest and
one negative potential minimum on the back at the early QRS and gradually reversed in the
late QRS. In a short period of the middle QRS (time 183189), we can find a multipole field
represented by one potential maximum and two negative minimums. This is a typical pattern
in the isopotenials maps representing the right ventricular breakthrough, as experimentally
observed by Taccardi (1963).
Simulation of a normal heart is generally a repetitive procedure to adjust the model
until simulated excitation isochrones fit the experimental data, and the simulated vector
cardiogram and surface ECGs fall within the normal range. Generally, obtaining a normal
P wave is not difficult. The excitation from the sinus node spreads leftward and downward
does reproduce a normal P wave in the simulated ECG. Getting a normal QRS lead in the
12lead ECG is more difficult. To keep normal waveforms in most leads on the surface
of the torso model, including the normal time periods and normal amplitude relationships
among the leads, it is important to correctly mount and adjust the specialized conduction
system in the heart model.
The simulationofthe T wave requires a deep understanding of the Twave mechanism
the reason why positive T waves are measured in most of the 12lead ECG. Suppose the
action potential is uniform at all ventricular myocardium, the depolarization and repolar
ization would be along the same direction, and thus the T waves would have the opposite
polarity as that of the QRS waves. But actually, positive T waves are observed in most leads.
The mechanism can be interpreted with a theoretical model that assumes longer action po
tential duration in the sides of the endocardium and apex than toward the epicardium and
the base (Harumi et al., 1964). In this way, the repolarization spreads along the opposite
direction to the depolarization. As a result, the T wave polarity is the same as that of the QRS
wave. The T wave model is consistent to the experimental measurement on the epicardium
(Spack et al., 1977). A detailed description of a T wave mechanism can be found in Barr
Whole Heart Modeling and Computer Simulation
avtv
vt vt vl v:k
107
Horizontal frontal sagittal
(b)
114 117 ms
e
8
195 198 201 204 201
(@) [ @) 01
FIGURE 3.9. Simulation results of a normal heart model: (a) ECG; (b) VCG; and (c) body surface isopotent ial
maps. In (c), and "" show positions of potential maximum and minimum, respectively. Lines of light black
represent contours of zero potential. The left area of six tenths correspond s to the anterior torso chest, and the
right of four tenths corresponds to the back. Time is counted from the onset of the P wave. (Figures of a and bare
reproduced with permiss ion from Wei et al., 1995).
(1989). The algorithmused in simulationto distributethe action potential in the 3Dwhole
heart model is introducedin Section 3.2.4.
3.3.2 SIMULATION OF STT WAVES IN PATHOLOGIC CONDITIONS
Studyingthepathologicchangesin theSTsegmentandthe Twavesare typical andim
portantapplications of wholeheart models. This is becausethe STTchangesareassociated
withseriousheart diseases such as myocardial ischemia, hypertrophy, andcardiomyopathy.
108
J. 1St zoo :&SO
(ms)
D.Wei
FIGURE 3.10. Predefined act ion potentials assigned to cells in normal and middle, moderate , and severel y
ischemic regions in the simulation of Dube et al. (Reproduced with permission. from Dube et al., 1996).
Myocardial ischemias arise from insufficient blood flow due to an occlusion of coro
nary arteries. In recent years , the percutaneous trasluminal coronary angioplasty (PTCA)
provides an opportunity to precisely confirm the relationship between the ECG features and
the sides of the blood block by controlling the balloon inflation during the PTCA operation.
Dube et al. (1996) simulated clinical body surface potential maps and ECGs using the PTCA
protocol. Because the action potential change of the ischemic tissue is the direct cause of the
ST segment changes in the ECG, the way of setting action potential in the heart model is a
key to the simulation. Three transmural zones, middle, moderate and severe ischemia, were
set to the heart model, located in the vicinity of the left anterior descending, left circumflex,
and right coronary arteries . The action potentials, as shown in Fig. 3.10, were set to these
regions, representing action potentials under middle, moderate and severe ischemias. The
simulation produced ECG maps quantitatively similar to clinical maps.
Fig. 3.11 shows another example that simulates "giant negative T waves" known as
the main feature of the apical hypertrophic cardiomyopathy (Harumi, 1989b). The simu
lated ECG and VCG give surprisingly similar results to clinical findings. The results were
obtained by modifying the APD gradient and the conductivity value for the pathologic
zone. Unlike the ischemia, the heart with hypertrophic cardiomyopathy is impossible to
make with experimental animals . In this sense, computer simulation is the only way of in
vivoexperimentation to study the unknown mechanism.
3.3.3 SIMULATION OF MYOCARDIAL INFARCTION
Myocardial infarction is a typical concern in heart modeling. The example introduced
in this section demon strates that model study is not only useful for understanding the
mechanism, but also it is useful for helping to develop a diagnosis tool for clinical practice.
Whole Heart Modeling and Computer Simulation
Frontal Sagittal
109
FIGURE 3.11. Simulated ECGand VCG in the heart model of apical myocardial cardiomyopathy (Reproduced
with permission from Harumi et aI., 1989).
Myocardial infarction is caused by the occlusion of the coronary artery. The devel
opment of myocardial infarction is usually classified in three phases by ECG patterns (de
Luna, 1993). The early phase of ischemia is characterized by T wave changes. The later
phase of injury is characterized by ST segment changes. The final phase of necrosis is char
acterized by wave changes. The ST and T wave changes can be simulated with a whole
heart model by changing the action potentials for the model cell s. The action potentials
during the reduced blood flow can be measured experimentally.
The location and size of myocardial infarction is an important aspect in clinical diag
nosis. In clinical practice, the locat ion and size of infarction are qualitatively interpreted
with the theory of vectorcardiogram. With a whole heart model , StarttlSelvester et al. (1989)
systematically simulated the infarcts due to three major coronary artery distributions, and
expected Q waves were obtained in each case. They found that in any case the degree of
QRS change was proportional to the degree of local infarction. Dividing the ventricles into
four walls by 12 segments, they developed a quantitative method to estimate the location
and size of the infarction. The result led to a practical tool known as the ECGNCG scoring
system, where each point scored was set up to represent 3% of the left ventricular my
ocardium. The scoring system predicted a distribution of damage in the 12 left ventricular
110 D.Wei
segments in good correlation with the average of planometric pathology found in the same
subdivisions. They further developed a monogram based on the same simulation study,
which relates the VCG changes to the infarct size. the duration and magnitude of QRS
deformity are measured before and after infarction, the infarct size can be simply found on
the monogram. Details can be found in Startt/Selvester et al. (1989).
3.3.4 SIMULATION OF PACEMAPPING
Pace mapping is a new technique that is used to speed up the procedure for localizing ec
topic focus in catheterization (SippensGroenewegen et aI., 1993). In SippensGroenewegen
et aI., body surface potential maps in patients with cardiac arrhythmias but with no evidence
of structural heart disease were recorded during ectopic beats by catheter stimulation at dif
ferent endocardial sites in the ventricles. Based on these data, they were able to classify the
QRS integral map patterns with respect to the location of ectopic beats.
The same procedure was reproduced by Xu et aI. (1996) with the heart model of
Lorange et aI. In the simulation, 38 selected endocardial sites (25 on the left ventricle
and 13 on the right ventricle) corresponding to SippensGroenewegen et aI. were paced
to initialize the excitation process of the heart model. With a more detailed heart model
(0.5 mm spatial resolution), Hren and Horacek (1997) generated a database of 155 QRS
integral maps by pacing the epicardial surfaces in the left and right ventricles. This database
would be useful in catheter pace mapping during treatment of ventricular arrhythmia. The
simulation of pace mapping is a good example to show the clinical usefulness of whole
heart models.
3.3.5 SPIRAL WAVESA NEW HYPOTHESIS OF VENTRICULAR
FIBRILLATION
Reentrant excitation is recognized as a mechanism of lifethreatening arrhythmias.
Among several hypotheses to explain the reentrant excitation, spiral wave is the latest one
and is getting more and more support from experimental studies (Gray, 1995, 1996). In
addition to experimental studies, computer simulation of the spiral wave using the whole
heart model with realistic geometry and anatomy is a useful tool because it is capable of
producing comparable results with experiments. Fig. 3.12 shows an example of spiral waves
simulated with a whole heart model (Gray and Jalife, 1996). The simulated isochrones and
ECGs (right) are comparable with the measured results.
3.3.6 SIMULATION OF ANTIARRHYTHMICDRUG EFFECT
The example introduced in the following demonstrates how the drug effect is confirmed
by a simulation study with a whole heart model (Wei et aI., 1992). The study was based
on an experiment that investigated the relationship (Harumi, 1989a) between the restitution
of premature action potential and the stimulation coupling interval for Purkinje fiber and
ventricular muscle in dogs before and during the infusion of antiarrhythmic drugs. In the
simulation, lines through linear regression as shown in Fig. 3.13(a) were obtained to approx
imate the experimental data. The slopes of these lines corresponding to a parameter called
dynamic coefficient (DC) were input to the heart model. Note that, the figure shows differ
ent slopes for the ventricular muscle and the Purkinje fiber before and during the infusion
Whole Heart Modeling and Computer Simulation
FIGURE 3.12. Simulation of spiral wave. (Reproducedwith permission, fromGray and Ja1ife, 1996).
111
of antiarrhythmic drugs. In the simulation, ten successive extrastimuli of 170 ms interval
started at 300 ms after the first sinus pacing were applied to the epicardium of the ventricle.
The simulated ECG shown on the top of Fig. 3.13(b) corresponds to the case where the APD
changes follow the lines before the drug infusion. In this case, the stimulation caused two
tachycardialike waveforms followed by sustained VF. When APD changes follow the lines
during the drug infusion, normal waveforms are restored after the stimulation, as shown by
the waveform in the bottom of Fig. 3.13(b). Figure 3.13(c) shows a picture of animation
developed for visualizing the propagation of excitation during the fibrillation. A number of
propagation wavefronts developed due to a number of reentries can be seen in this picture.
This simulation supported the assumption that different ratios of restitution in premature
APD between the ventricular muscle and the Purkinje fiber may playa role in the induction
of VF. also demonstrated the antiarrhythmic drug effect in suppressing VF.
3.4 DISCUSSION
Principles, methodology and applications in 3D whole heart modeling are described in
this chapter. The significance of 3D whole heart models is that, as compared to experimental
studies, computer simulations with whole heart models are always in vivo so as to provide
information relating intracardiac events to the body surface electrocardiogram in different
<::)
<::)
<::)
Q
~ < : : )
<::)
<::)
( b)
Lo
0 2
FIGURE 3.13. (a) Approximation of experimental results describing restitutions of premature action potential
to the stimulation coupling interval for Purkinje fiber and the ventricular muscle in dogs before (left) and during
(right) the infusion of antiarrhythmic drug. (b) Simulated ECGs. The waveform on the top shows ventricular
fibrillation induced by applying successive stimuli to the left ventricular wall. The waveform on the bottom shows
that the fibrillation stops after the infusion of the antiarrhythmic drug. (c) An image of 3D animation showing
excitation propagation during ventricular fibrillation.
Whole Heart Modeling and Computer Simulation
pathologic processes. Such information is sometimes difficult to obtain by experimental
studies. Other advantages arise from the facts that the computer simulation is lowcost , fast
and repeatable. As an experimental tool, a whole heart model can be used in laboratories
for research purposes and used in classroom for computeraided instruction in medical
education.
The main limitation of current whole heart models is the insufficiency of the elec
trophysiological details, due to the large scale of the whole heart and the computational
capability related to the problem. In the present stage, the heart model s in singlecell, cell
network and whole heart levels are still separately studied. More time is needed to realize
and spread the 3D whole heart model s of nextgeneration that can generate action potentials
based on cellular mechanisms, that can simulate propagation based on HH type formulisrn,
and reconstruct clinically comparable body surface electrocardiograms. This kind of model
would be able to directly relate basic experimental findings to clinical applications.
In a search of related abstracts on Medline in the U.S. National Library for the two
decades till the end of 1980s, Malik and Camm (1991) found that both the absolute and
relative numbers of publications in computer modeling and simulation have increased much
faster than the total increase in cardiology as a whole. We have searched the same database
for publications in the 1990s using the keywords "heart model simulation computer."
The result is shown in Fig. 3.14. Publication activity continues to increase.
Nowadays, computer modeling and simulation have become more attractive to re
searchers than ever before. Because the biological system is so complex, no single experi
ment can synthesize the system as a whole. As the computation ability of computers becomes
more and more powerful , computer modeling and simulation technologies are making it
possible to study the biological system of humans as a whole. This is recently known
as "Physiome" ("physi " means life, "orne" means whole, see http://www.physiome.orgl).
Whole heart modeling is one of the common concerns in Physiome.
Key wordr'Hea:t+model+S1mulahon+computer'
100
'0 40
,Q
§ 20
i
o
1990 1991 1992 1993 1994 1 9 9 ~ 1996 1997 1998 1999 2000
YE:U'
FIGURE 3.14. Publicati ons searched on MedPub. (Keyword "heart model simulation comput er").
REFERENCES
D.Wei
Adam, D., and Barta, E., 1987, The effect s of anisotropy on myocardial activation, in: Simulation and Cont rol of
the Cardiac System, Volume III (S. Sideman, and R. Beyar, eds. ), CRC, Boca Raton, Florida, pp. 1327.
Abi ldskov, J. A., and Lux, R. 2000, Mechanisms in Twave alternans caused by intraventricular block,
Electrocardi ol . 33(4):311319.
AlNashash, H., and Lvov, 8. . 1997, Thr eedimensional model for the simulation of the HPS electrogram, Biomed.
Mat er. Eng. 7(6):40 1410.
Aoki, M., Okamoto, Y., Musha, T., and Harumi, K., 1987, Threedimensional simulation of the ventricular depo
larization and repolari zation processes and body surface potentials: Normal heart and Bundle Branch Block,
IEEE Trans. Biomed. Eng. 34:454462.
Balasubramaniam, C; Gopakumaran , B., and Jagadeesh, J. M., 1997, Simulation of cardiac conduction system in
distributed computer environment, Biomed. Sci. Instrum. 33: 138 .
Barr, R. C, 1989, Gene sis of the electrocardiogram, in: Comprehensive Electrocardiology (P. W. MacFarlane and
T. D. V. Lawrie, eds. ), Pergamon Press, New York, pp. 129151.
Beeler, G. W., and Reuter, H., 1977, Reconstruction of the action potent ial of ventricular myocardial fibres,
Physiol.268(1):I7721O.
Bercnfeld, and Abboud, S., 1996, Simulation of cardiac act ivity and the ECG using a heart model with a
reactiondiffusion action potential, Med. Eng. Phys. 18:61525.
Clayto n, R. H., Bailey, A., Biktashev, V. N., and Holden, A. V., 200 1a, Reentrant cardiac arrhythmias in compu
tational mode ls of long QT myocardium, 1. Theor. Bioi. 208:21525.
Clayt on, R. H., 2001b, Computational models of normal and abnorma l action potentia l propagation in cardiac
tissue: linking experi mental and clinical cardiology, Physi ol. Mea s. 22: R15 R34.
Cler c, 1976, Directional differences of impulse spread in trabec ular muscle from mammalian heart . Physiol.
255:335 345.
De Luna, A. B., 1993, Clinical Electrocardiograph y: A Textbook, Futura Publ ishing Company Inc., New York.
di Bernardo, and Murray, 2002, Origin on the elect rocardiogram of Llwaves and abnormal Uwave
inversion, Cardiova sc. Res. 531 :202 208.
DiFrancesco, and Noble, 1985, A model of cardi ac electrical activity incorporating ionic pumps and
concen tration changes, Phil. Trans. R. Soc. Lond. 8307:353 98.
Dubc, B., Gulrajani, R. M., Lorange, M., LeBl anc, A. R., Nasmi th, J., and Nadeau , R, A. , 1996, A computer heart
mode l incorporating anisotropic propaga tion. IV. Simulation of regional myocardial ischemia, Electrocar
di ol. 29:91103.
Durrer, D., van Dam, R. T., Freud, G. E., Janse, M. J., Meijler, F. and Arzbaec her, R. C; 1970, The total
excita tion of the isolated human heart , Circ. 41:899912.
Earrn, E., and Noble, 1990, A model of the single atrial cell: betwee n calci um current and calcium release,
Proc. Roy Soc. 240:83 96.
FitzHugh, R., 1961, Impu lses and physiological states in theoret ical models of nerve membrane. Biophys.
1:445466.
Gczelowi tz, D., and Miller W. T., 1983, A bidomain model for ani sotropic cardiac muscle, Ann. Biomed. Eng.
11:191206.
Gese lowitz, D. B., 1989, Theory and simulat ion of the Electrocardiogram, in: Comprehensive Electrocardiology
(P. W. MacFarlane and T. D. V. Lawrie, eds .), Pergamon Press, New York, pp. 181 195.
Goldman, M. J., 1986, Principles of Clini cal Electroca rdiograph y, Lange Medical Publicat ions, Los Altos,
Califo rnia.
Gray, R. A., Jalife, J. , Panfilov, A. v.. Baxter, W. T., Cabo, Davidenko, J. M., and Pertsov, A. M., 1995,
Mechanism of cardiac fibrillation, Science 270: 12221225.
Gray, R. A., and Jalife, J., 1996, Spi ral waves and the heart , lnternational of Bifurca tion and Chaos 6:
415435.
Gulrajani, R. M., and Mailloux, G. E., 1983, A simulation study of the effects of torso inhomogeneities on
electroca rdiographic potentials, using realistic heart and torso models, Circ. Res. 52:45 56.
Guyton, A. 1986, Textbook ofMed ical Physiology, W. 8. Saunders, London.
Harumi. K., Burgress , M. J., and Abildskov, 1. A., 1964, A theoret ical model of the T wave, Circ. 34:657668.
Harumi, K., Tsutsumi , T., Sato, T.. and Seki ya, S., 1989a, Classification of antiarrhythmic drugs based on ventricular
fibrillat ion threshold, Amer. Cardiol. 64: IOJ1 4J.
Whole Heart Modeling and Computer Simulation 115
Harurni, Tsunagawa, H., Nishiyama, G., Wei, D., Yamada, G., Okamoto, Y, and Musha, T., 1989b, Clinical
application of electrocardiographic computer model, 1. ElectrocardioI22(Suppl):5463.
He, B., and Wu, D., 1999, Lapacian Electrocardiography, Critical Reviews in Biomed. Eng. 27:285338.
Heinonen, T., Eskola, H., Kauppinen P., and Malmivuo J., 1996, Tissue segmentation of visible human man data
using lARD method, Med. Biol. Eng. Comput. 34(Suppl 1, Part 1):239240.
Hodgkin, 1., and Huxley, A. E, 1952, A quantitative description of membrane current and its application to
conduction and excitation in nerve, J. Physiol. 117:500544.
Horacek, B. M., 1989, Lead theory, in: Comprehensive Electrocardiology (P. W. MacFarlane and T. D. V. Lawrie,
eds.), Pergamon Press, New York, pp. 291352.
Hren, R., and Stroink, G., 1995, Application of the surface harmonic expansions for modeling the human torso,
IEEE Trans. Biomed. Eng. 42:52124.
Hren, R., and Horacek, B. M., 1997, Value of simulated body surface potential maps as templates in localizing
sites of ectopic activation for radiofrequency ablation, Physiol. Meas. 18:373400.
Hren, R., 1998, Value of epicardial potential maps in localizing preexcitation sites for radiofrequency ablation.
A simulation study, Phys. Med. Bioi. 43:144968.
Hsiao, C. H., and Kao, T, 2000, Constructing a 3D mesh model for electrical cardiac activity simulation, Comput.
Biomed. Res. 33:2342.
Hyttinen, J., Viik, 1., Lehtinen, R., Plonsey, R., and Malmivuo, 1., 1997, Computer model analysis of the relationship
of STsegment and STsegmentfheart rate slope response to the constituents of the ischemic injury source,
Electrocardiol.30:16174.
Ikeda, N., Yoshizawa, S., and Sato, T, 1983, Difference equation model of ventricular parasystole as an interaction
between cardiac pacemakers based on the phase response curve, 1. Theo. Biol. 103: 43965.
Jalife, J. and Moe, G. 1976, Effect of electrotonic potentials on pacemaker activity of canine Purkinje fibers
in relation to parasystole. Cire. Res. 39:801808.
Kauppinen, P., Koobi, T, Kaukinen, S., Hyttinen, J., and Malmivuo, J., 1999, Application of computer modelling
and lead field theory in developing multiple aimed impedance cardiography measurements, 1. Med. Eng.
Technol. 23(5): 16977.
Klepfer, R. N., Johnson, C. R., and Macleod, S., 1997, The effects of inhomogeneities and
anisotropies on electrocardiographic fields: a 3D finiteelement study, IEEE Trans. Biomed. Eng. 44(8):
70619.
Leon, 1. 1., and Horacek, B. M., 199Ia, Computer model of excitation and recovery in the anisotropic myocardium.
I. Rectangular and cubic arrays of excitable elements, 1. Electrocardiol. 24:115.
Leon, 1. J., and Horacek, B. M., 1991b, Computer model of excitation and recovery in the anisotropic myocardium.
II. Excitation in the simplified left ventricle, Electrocardiol. 24: 173
Lorange, M., and Gulrajani, R. M., 1986, Computer simulation of the WolffParkinsonWhite preexcitation syn
drome with a modified MillerGeselowitz heart model, IEEE Trans. Biomed. Eng. 33:86373.
Lorange, M., and Gulrajani, R. M., 1993a, A computer heart model incorporating anisotropic propagation. Part I.
Model construction and simulation of normal activation, Electrocardiol. 26:245261.
Lorange, M., and Gulrajani, M., 1993b, A computer heart model incorporating anisotropic propagation. Part
II. Simulation of conduction block, Electrocardiol. 26:263278.
Luo, C. H., and Rudy, Y, 1991, A model of the ventricular cardiac action potential, Circ Res., 68:150126.
Luo, C. H., and Rudy, Y, 1994a, A dynamic model of the cardiac ventricular action potential. I. Simulations of
ionic currents and concentration changes, Cire. Res. 74:107196.
Luo, C. H., and Rudy, Y, 1994b, A dynamic model of the cardiac ventricular action potential. II. Afterdepolar
izations, triggered activity, and potentiation, Circ. Res. 74:1097113.
MacFarlane, P. W. and Lawrie, T. D. V. (eds.), 1989, Comprehensive Electrocardiology, Pergamon Press, New
York.
MacFarlane, P. W., and Lawrie, T. D. V., 1989, The normal electrocardiogram and vectorcardiogram, 1989, in:
Comprehensive Electrocardiology (P. W. MacFarlane and T D. V. Lawrie, eds.), Pergamon Press, New York,
pp.407457.
Malik, M., and Camm, A. J., 1991, Cardiac electrophysiological experiments in Numero, Part I: Concepts and
strategies of mathematical and computer models, Pace 14: 1492I502.
Miller, W. T. III and Geselowitz, D. B., 1978a, Simulation studies of the electrocardiogram. I. The normal heart,
Circ. Res. 43:301315.
Miller, W. T. III and Geselowitz, D. B., 1978b, Simulation studies of the electrocardiogram II. Ischemia and
Infarction, Circ. Res. 43:315323.
D.Wei
Moe, G. K., Rheinbolt, W. and Abildskov, J. A., 1964, A computer model of atrial fibrillation, Am. Heart 1.
67:200220
Moe, G. K., Jalife, J., Mueller, W.J., and Moe, 8., 1977, A mathematical model of parasystole and its application
to clinical arrhythmias, Circ. 56:96879.
Nielson, P. M. E, LeGrice, 1.J., Smail, B. H., and Hunter, P. J., 1991, A mathematical model of the geometry and
the fibrous structure of the heart, Am. 1. Physiol. 260:HI36578.
Noble, D., DiFrancesco, D., and Denyer, J. C, 1989, Ironic mechanism in normal and abnormal cardiac pacemeker
activity, in: Cellular and Neuronal Oscillators (Jacklet, J. w., ed.), Dekker, New York, pp. 5985.
Okajima, M., Jujino, T., Kobayashi, T. and Yamada, K., 1968, Computer simulation of the propagation process in
the excitation of the ventricles, Circ. Res. 23:203211.
Okazaki, Yamauchi, Y, Kashida, M., Izumo, K., Akatsuka, N., Ohnishi, S., Shoda, M., Nirei, T., Kasanuki,
H., Ebato, M., Mashima, S., Harumi, K., and Wei, D., 1998, Possible mechanism of ECG features in patients
with idiopathic ventricular fibrillation studied by heart model and computer simulation, J. Electrocardiol.
30(Suppl):98104.
Paine, R., 1988, Generationand Interpretationof the Electrocardiogram, Lea Febiger, Philadelphia.
Panfilov, A. V., and Hogeweg, P., 1993, Spiral wave breakup in a modified FitzHughNagumo model, Phys. Lett.
AI76:295299.
Panfilov, A. v., and Keener, J. E, 1995, Reentry in an anatomical model of the heart, Chaos, Solitons and Fractals
5(3):681689.
Plonsey, R., 1969, Bioelectric Phenomena, McGrawHill, New York.
Plonsey, R., and Barr, R. C; 1987, Mathematical modeling of electrical activity of the heart, 1. Electrocardiol.
20:219226.
Plonsey, R., 1989, Introductory physics and mathematics, in: ComprehensiveElectrocardiology (P. W. MacFarlane
and T. D. Lawrie, eds.), Pergamon Press, New York, pp. 4176.
Ramanathan, C, and Rudy, Y., 2001a, Electrocardiographic imaging: 1. Effect of torso inhomogeneities on body
surface electrocardiographic potentials, J. Cardiovasc. Electrophysiol. 12(2):22940.
Ramanathan, C., and Rudy, Y, 200lb, Electrocardiographic imaging: II. Effect of torso inhomogeneities on non
invasive reconstruction of epicardial potentials, electrograms, and isochrones, J. Cardiovasc. Eleetrophysiol.
12(2):24152.
Roberts, D. E, Hersh, T., and Scher, A. M., 1979, Influence of cardiac fiber orientation on wavefront voltage,
conduction velocity, and tissue resistivity, Cire. Res. 44: 701712.
Rudy, Y, and Plonsey, R., 1979, The eccentric spheres model as the basis for a study of the role of geometry and
inhomogeneities in electrocardiography, IEEE Trans. Biomed. Eng. 26:392399.
Rudy, Y, and Plonsey, R., 1980, A comparison of volume conductor and source geometry effects on body surface
and epicardial potentials, Cire. Res. 46:283291.
Shahidi, S. v., and Savad, P., 1994, Forward problem of electrocardiography: Construction of human torso models
and field calculations using finite element models, Med. BioI. Eng. Comput. 32:S2533.
SippensGroenewegen, A., Spekhorst, H., van Hemed, N. M., Kingma, J. H., Hauer, R. N. w., Grmbergen, C. A.,
de Bakker, J. M. T., Janse, M. J., and Dunning, A. J., 1993, Localization of the site of origin of postinfarction
ventricular tachycardia by endocardial pace mapping, Cire. 88: 2290306.
Siregar, P., Sinteff, J. P., Chahine, M., and Lebeux, P., 1996, Acellular automata model of the heart and its coupling
with a qualitative model, Comput. Biomed. Res. 29:22246.
Spaggiari, S., Baruffi, S., Arisi, G., Macchi, E., and Taccardi, B., 1987, Effect of intramural fiber orientation on
epicardial isochrone and potential maps (abstract), Cire76(suppl IV): 241.
StarttiSelvester, H., Wagner, G. S., and Ideker, E., 1989, Myocardial infarction, in: ComprehensiveEleetro
cardiology (P. W. MacFarlane and T. D. Lawrie, eds.), Pergamon Press, New York, pp. 565629.
Spack, M. S., Barr, R. Lanning, C. E, and Tucek, E C., 1977, Origin of body surface QRS and Twave potentials
from epicardial potential distributions on the intact chimpanzee, Cire. 55: 28678.
Streeter, D. D., Spotnitz, H. M., Patel, D. P., Ross, J. Jr., and Sonnenblick, E. H., 1969, Fiber orientation in the
canine left ventricle during diastole and systole, Cire. Res. 24:339347.
Suzuki, T., Ashihara, T., Inagaki, M., Toda, S., Namba, T., Ikeda, T., Suzuki, T., Inada, H., Sugimachi, M., and
Nakazawa, K., 2001, Reconstruction of electrocardiograms during sinus rhythm in the LuoRudy ventricular
shaped model (in Japanese), Japan Soc. ME BE 39: 176.
Taccardi, B., 1963, Distribution of heart potentials on the thoracic surface of normal human subjects, Cire. Res.
12:341352.
Whole Heart Modeling and Computer Simulation 117
Tung, 1978, A bidomain model for describing ischemic myocardial DC potentials, Ph.D. dissertation, MIT,
Cambridge, MA.
Van Dam, R. Th., and Janse, M. J., 1989, Activation of the heart, in: Comprehensive Electrocardiology (P. W.
MacFarlane and D. V. Lawrie, eds.), Pergamon Press, New York, pp. 101127.
Virag, N., Vesin, J. M., and Kappenberger, 1998, A computer model of cardiac electrical activity for the
simulation of arrhythmias, Pacing Clin. Electrophysiol. 21:236671.
Wei, D., Aoki, M., Okamoto, Y, Musha, and Harumi, K., 1987, Computer Simulation of the WolffParkinson
White syndrome utilizing a human heart model, Japanese Heart J. 28:707718.
Wei, D., Yamada, G., Hasegawa, Nakamura, and Musha, 1989, Evaluation of the rotating anisotropy of
the ventricular myocardium: A simulation study, IEEE Proceed. of the Annual International Conference of
the IEEE EMBS 11:186187.
Wei, D., Yamada, G., Musha, T, Tsunagawa, T, Tsutsumi, T, and Harumi, K., 1990, Computer simulation
of supraventricular tachycardia with the WolffParkinsonWhite syndrome using threedimensional heart
models, J. Electrocardiol. 23:261273.
Wei, D., Harasawa, E., Okazaki, Harumi, and Musha, 1992, Computer simulation of cardiac arrhythmias
by incorporating phase response curve to a threedimensional heart model. in Computers in Cardiology, Los
Alanitos, IEEE Computer Society, pp. 327330.
Wei, D., Okazaki, Harumi, Harasawa, E., and Musha, 1992, Inducing and suppressing ventricular
fibrillation: A simulation study with threedimensional heart model and experimental data, IEEE Proceed.
of the Annual International Conference of the IEEE EMBS 14:400401.
Wei, D., Okazaki, 0., Harumi, Harasawa, E., and Hosaka, H., 1995, Comparative simulation of Excitation and
body surface electrocardiogram with isotropic and anisotropic computer heart models, IEEE Trans. Biomed.
Eng. 42:343357.
Wei, D., 1997, Whole heart modeling: Progress, principles and applications, Prog. Biophys. Mol. Biol. 67: 1766.
Wei, D., and Mashima, S., I999a, Prediction of accessory pathway locations in WolffParkinsonWhite syndrome
with body surface potentials Laplacian mapsA simulation study, Japan Heart J. 40:451459.
Wei, D., Miyamoto, N., and Mashima, S., 1999b, A computer model of myocardial disarray in simulating ECG
features of hypertrophic cardiomyopathy, Japan Heart 1., 40:8 I9826.
Wei, D., 200 I, Laplacian electrocardiograms simulated using realistically shaped heart torso model during normal
and abnormal ventricular depolarization, Electromagnetics, 21:593605.
Wu, D., Schablowski, M., Ono, Hosaka, H., and He, B., 1998 "A Simulation Study of Laplacian ECG in a Re
alistically Shaped Torso Volume Conductor: Myocardial Infarction," Bioelectrochemistry and Bioenergetics,
Vol. 47, 231235.
Xu, Z., Gulrajani, R. M., Molin, E, Lorange, M., Dube, B., Savard, P., and Nadeau, R. A., 1996, A computer
heart model incorporating anisotropic propagation. Part III. Simulation of ectopic beats, J. Electrocardiol.
29:7390.
Yamaki, M., Kubota, and Tomoike, H., 1999, Simulation of late potentials and arrhythmias by use of a three
dimensional heart model: Casuality of periinfarctional slow conduction in ventricular fibrillation, Elec
trocardiol. 32: 11521.
Zanda, N., Tsutsumi, Sato, M., Takeyama, Y, Harumi, and Wei, D., 2000, Computer simulation of notches
on initial part of QRS complex in patients with anterior myocardial infarction, in Electrocardiology 2000,
Proceedings ofthe XXVII international congress on Electrocardiology D. Ambroggi, ed.), University of
Milan, Italy, pp. 117120.
Zhang, B., Masuzawa, T, Tatsumi, E., Taenaka, Y, Uyama, Takano, H., and Takamiya, M., 1999, Three
dimensional thoracic modeling for an anatomical compatibility study of the implantable total artificial heart,
Artif. Organs 23:22934.
HEART SURFACE
ELECTROCARDIOGRAPHIC
INVERSE SOLUTIONS
Fred Greensite
1
Department of Radiological Services
University of California, Irvine
4.1 INTRODUCTION
In this chapter, we will review the problemof noninvasive and minimially invasive imaging
of cardiac electrical function. We use the term "imaging" in the sense of methodology which
seeks to spatially resolve distributed properties of cardiac muscle electrophysiology such
as extracellular potential, or features of the action potential. Thus, we do not consider the
problems of computing properties ofan "equivalent" cardiac multipole, moving dipole(s), or
any other source model that does not satisfy such criteria. We will further restrict ourselves
to resolving such electrophysiological features on the epicardial or endocardial surfaces
a reasonable restriction, since measurements currently accessed by invasive procedures
are obtained on these surfaces, and also because the spatial dimension of the "source"
domain then nominally matches the spatial dimension of the data domain. Thus, we will
not consider the earliest distributed source model, representing intramural current density
imaging (Barber and Fischman, 1961; Bellman et al., 1964), on which work continues (e.g.,
see (He and Wu, 2001), or the recent heartexcitationmodel based 3D inverse imaging
approach (Li and He, 2001) in Chapter 5 in this book).
Following an historical perspective, we will discuss in some detail the inherent dif
ficulties of this imaging problem (principally mathematical), and strategies developed to
circumvent them. We will not attempt to comprehensively cite the voluminous work done
on these formulations of the inverse electrocardiography problem. Excellent reviews for the
period prior to 1990 exist (Gulrajani et al., 1989; Rudy and MessingerRapport, 1988), and
more recent shorter reviews can be found in (MacLeod and Brooks, 1998; Gulrajani, 1998).
Mailing Address: Fred. Greensite, Department of Radiological Sciences, University of CaliforniaIrvine Medical
Center, Trailer 11, Route 140, 101 The City Drive South, Orange CA 92868 USA. Email: fredg@uci.edu.
Telephone: (714) 4567404, FAX: (714) 4566380.
119
120 F. Greensite
Our principal objectives are to provide a meaningful presentation of the issues involved in
this inverse problem, and to survey some work of the last several years.
4.1.1 THE RATIONALE FOR IMAGING CARDIAC ELECTRICAL FUNCTION
Coordination of the heart's mechanical functioning is accomplished electrically. In fact,
there is a tight coupling of mechanical and electrical function both at the microscopic cellu
lar and macroscopic organ level. Global contraction of the heart proceeds from a sequence of
local changes in cell membrane conductances to various ionic species (the opening and clos
ing of voltagegated speciesspecific membrane channels). These membrane conductance
changes, resulting in transmembrane currents, generate local source currents which lead to
Ohmic currents throughout the remainder of the body volume conductor. In this manner,
the sequential cardiac muscle membrane changes, which characterize the heart's electrical
functioning, are reflected in electrical potentials that can be measured at the body surface.
The local changes in intracellular ionic concentrations, resulting from the transmem
brane electrical current, trigger the local subcelluar mechanical events (the "sliding" of
filaments over each other) which on a global scale summate to coordinated contraction
of the muscle mass. Since the local mechanical events must be coordinated satisfactorily
with each other to produce an effective organic contraction, a corollary is that a mechan
ical catastrophe can only be avoided if the inciting local electrical events are themselves
coordinated properly on a global scale. Even when mechanical function appears adequate,
the spatiotemporally distributed electrical functioning of the heart is of the greatest interest
medically, since well characterized disturbances of cardiac electrical functioning are known
to predispose to such mechanical catastrophes, and are known to be amenable to various in
terventions (pharmcological, catheterbased, or surgical). Evidently, it is useful to consider
how such cardiac electrical distubances might be effectively diagnosed noninvasively.
Although indications of electrophysiological dysfunction may be apparent from the un
adorned clinical electrocardiogram (tracings of electrical potential at certain body surface
locations), successful treatment often requires "imaging" of such disturbancesusually
accomplished by laborious transvascular electrode catheter manipulations within a cardiac
chamber. The prospect that such imaging might be effected in a less invasive and effi
cient manner, is the principal motivation for present efforts to develop a practical clinical
technology for imaging cardiac electrical functioning.
4.1.2 A HISTORICAL PERSPECTIVE
In our context, imaging could be looked at as a global depiction of local features. Thus,
it can be understood in terms of microscopic and macroscopic cardiac electrophysiology.
Microscopic: Action Potential
Sensoryneuromuscular tissue is distinguished by its ability to generate and conduct
action potentials. The latter represent cellular transmembrane potential evolution during the
characteristic functioning of the tissue (figure 4.1). The concept was inherent in the work of
Bernstein in the late 1800s, wherein functional neuronal cellular electrical activity was at
tributed to sequential membrane permeability changes. Muscle is particularly distinguished
Heart Surface Electrocardiographic Inverse Solutions 121
90
FIGURE 4.1. The action potential in heart Purkinje cells resembles a pulse (i.e., square wave). The upstroke
(phase 0, "activation") corresponds to rapid intracellular influx of sodium. A limited fast repolarization (phase 1),
and subsequent plateau (phase 2), correspond to an interval for which the location is refractory to stimulation. The
downstroke (phase 3) is over an interval for which stimulation will lead to weakened activation. The horizontal
axis can be either time (action potential duration is on the order of hundreds of milliseconds) or space (although
simultaneous recording of transmembrane potential along a path in the heart muscle at a single time instant would
usually only reveal the full action potential shape in the setting of reentry type arrhythmias). The 90 millivolt
baseline defines phase 4.
in that arrival of an action potential at a given location triggers a cascade of events leading
to contraction. Wilson et al. (1933) expressed great insights into the significance of this
transmembrane functioning in cardiac muscle, in a seminal work from the 1930's.
All aspects of the action potential are potentially of interest. For example, phase 0
(signifying local activation) initiates the events leading to the local contraction (sliding
of filiments). At the phase 0 time, the location "depolarizes" (transmembrane potential
changes from roughly 90 millivolts to roughly zero millivolts). The amplitude of phase
ohas important implications as regards local muscle integrity. During Phase 1 (transient,
limited, fast repolarization) and Phase 2 (plateau), the location cannot be further stimulated
(absolute refractory period). Phase 3 (progression through to "repolarization") is a relative
refractory period, during which attenuated responses to further stimulation are possible.
The durations of the absolute and relative reftractory periods govern when the location will
be susceptible to being triggered again (for the next heartbeat), and how strong that next
contraction might be locally. During phase 4, the location is fully repolarized, i.e., maintains
the baseline 90 millivolt transmembrane potential difference.
Given the significance of these features, an obvious goal would be to noninvasively
image (spatially resolve) the action potential at every location in the heart muscle (i.e.,
reconstruct the time series of the transmembrane potential at each location). However,
articulation of such a goal has been long in coming.
Macroscopic: Electrocardiogram
Recognition of the electrical functioning of muscle predates the nineteenth century
(and possibly the eighteenth century) (Malmivuo and Plonsey, 1995). By the midnineteenth
bn_ _ 
F. Greensite
FIGURE 4.2. An illustration adapted from (Waller, 1889). The cardiacgenerated potential field depicted is
essentially dipolar in nature. [Adapted from figure 5, p. 186: Waller, A., 1889, On the electromotive changes
connected with the beat of the mammalian heart, and of the human heart in particular, Phil. Trans. R. Soc. Lond.
180: 169194. Used by permission.]
century, the epicardial potentials had been invasively accessed. Later in the century, Waller
had the brilliant insight that the limbs could be viewed as electrode leads emanating from the
heartso that coupling them to an appropriate electrical apparatus should allow noninvasive
assessment of the previously established cardiac electrical functioning. From his published
illustrations, it appears that he conceived of modeling the heart as a current dipole source,
begging the question of inversely "imaging" this source from his body surface potential
measurements (Waller, 1889) (figure 4.2).
Interestingly, Waller (1911) was not optimistic that such information would be of much
use medically. However, nearly coincident with his cautionary remarks, Einthoven (1912)
was engaged in the work of demonstrating the clinical efficacy of improved instrumentation
to accomplishjust such a goal. Indeed, early in the twentieth century Einthoven popularized
(via the "Einthoven triangle") the concept of the cardiac dipole as (qualitatively) estimated
from measurements at standardized electrode locationsthe electrocardiogram. In ensuing
decades, a number of investigators attempted to make such calculations more quantitative
(e.g., (Wilson et al., 1947; Frank, 1954». In this context, Wilson's earlier work (Wilson
et al., 1933) provided the conceptual link between the microscopic and macroscopic. In
particular, was observed that macroscopic cardiac activation wavefronts, arising from
cellular membrane events, effect measurable remote potentials whose magnitude is roughly
Heart Surface Electrocardiographic Inverse Solutions 123
proportional to the solid angle subtended by the activation wavefront and the electrode
location.
However, the implication that cardiac electrical functioning might be remotely imaged
had to await the diffusion of computer technology to the biomedical community in the early
1960s (Gerlernter and Swihart, 1964; Bellman et al., 1964; Barnard et al., 1967).
4.1.3 NOTATION AND CONVENTIONS
The imaging formulas for cardiac electrophysiology derive from linear partial differ
ential equations. When they are "discretized" for numerical solution (see Chapter 2 in this
book), one encounters large scale algebraic manipulations. We will find the following vector
and matrix conventions useful: We denote a matrix composed of zeros as 0, and an identity
matrix as Superscript t applied to a matrix denotes the transpose, so when it is applied to
a column vector it therefore denotes the corresponding row vector (and viceversa). Thus,
for column vectors a, b, the inner product will be written as either a . b or at b, while the
outer product is written as either a ® b or ab' (the outer product of a and b is the matrix
whose entry is the product of the component of with the component of
A vector in three dimensions will be denoted by a bold face font (e.g., n will denote a unit
normal to a surface, andj will denote a current density). Matrices will be denoted by capital
letters, e.g., The row of will be denoted as and the column will be
denoted M:
j
• Random variables (and random vectors and random matrices) will sometimes
be denoted with a superscripted bar, e.g., The probability (or probability density) of some
event will be denoted p(.), so that p(alb) is the (conditional) probability of event a given
that event has been observed.
As is customary in physics, elemental sources (e.g., electrical monopoles) are expressed
via the Dirac delta function )a convenient shorthand for the limiting notion of a
function whose nonzero portion is localized to a tiny region, but whose integral over all
space is unity (Jackson, 1975). We note that such entities are rigorously treated in the context
of the theory of distributions (e.g., see (Keener, 1988) for details).
4.2 THE BASICMODEL AND SOURCE FORMULATIONS
In this section, we will derive imaging equations relating body surface potentials to
heart surface potentials. We do not present the analogous equations for magnetic field data,
which can be derived in a similar manner from the equations of magnetostatics.
As noted, description of muscle electrophysiology can be organized around the con
cept of an action potential, which reflects the opening and closing of ionic channels in
the cell membranewhich in tum lead to local intracellular ionic environmental changes
that trigger the local mechanical function, and the return to the resting state. By defi
nition, an action potential is the transmembrane potential that occurs in sensoryneuro
muscular tissue during its characteristic functioning. encompasses the phenomena of
selfpropagation to adjacent locations, meaning that the form of the transmembrane poten
tial at a fixed location as a function of time is recapitulated in a tracing of transmembrane
potential at a fixed time as a function of location (along the path of activation). Propa
gation of the action potential coordinates local contraction over the full threedimensional
124 F. Greensite
extent of the muscle, and implies its link to the global mechanical functioning of the
tissue.
Of course, the action potential is measured historically via microelectrodes inserted
into cells. Our objective is to resolve features of the transmembrane potential remotely from
the tissue. Thus , we must formulate the relationship between cardiac muscle action potential,
and remote measurements (such as on the body surface) . This is conveniently accomplished
via the socalled "bidornain" model , a concept first suggested by Schmitt (1969), and later
given mathematical form by others (reviewed by Henriquez (1993)) . Although the model is
discussed in Chapter 2, we represent it here as we wish to derive our source formulations
from "first principles", as well as provide a consistency of notation within the present
chapter.
At a level of resolution appropriate to this problem, it can be assumed that every body
location consists of a small amount of intracellular space and a small amount of extracellular
space. We can write
(4.1)
where ¢e, and ¢m, are intracellular potential, extracellular potential, and transmembrane
potential, respectively. In the context of our remote sensing problem, it has been shown that
capacitive, inductive, and electromagnetic propagative effects are "negligible" (the "quasi
static assumption") (Plonsey, 1969). Therefore, the intracellular or extracellular current
densities at the bidomain point are each linearly dependent on the gradient ofthe intracellular
or extracellular potential (Ohm' s Law). Thus, these respective current densities are given
by
je G
e
V¢e,
(4.2)
(4.3)
where G, and represent conductivity in the intracellular and extracellular components
of the bidomain point, and these conductivities are (in practical terms) independent of the
potential (the medium in linear) (Plonsey, 1969). is important to note that cardiac muscle
tissue is "anisotropic", in that it is composed of (interconnected) muscle fibers whose
intracellular conductivity along the fiber direction is much greater than the conductivity
normal to the fiber direction. Due to the geometrical constraints imposed by the intervening
fibers, the extracellular conductivity is also anisotropic. Therefore, G and Ge are tensors (in
this case, symmetric 3 x 3 matrices) which thereby linearly map a gradient of potential (at
a point) to a current density vector. G and Ge were proportional to each other (implying
that the principal axes, and the conductivity ratios for different pairs of principal axes, are
the same for intracellular space as for extracelluar space), we could speak of there being
"equal anisotropy".
According to Eq. (4.2) and Eq. (4.3) , total current density je at any of the
bidomain "points" must satisfy
(4.4)
The quasistatic assumption and charge conservation imply that any excess current (nonzero
Heart Surface Electrocardiographic Inverse Solutions 125
divergence) appearing in the extracellular component of a bidomain point must come from
the intracellular component of the bidomain (via the cell membrane), and vice versa. That
is, V' . Thus, Eq. implies
(4.5)
The divergences on both sides of the above equation express the location's role as a source
of extracellular current (there is net transmembrane current if the divergence is nonzero).
Transmembrane currents capable of influencing body surface potentials only occur in ex
citable tissue (i.e., sensoryneuromuscular tissue)the predominant one being the heart.
Thus, outside the heart there is no propagation of action potentials, and no source current,
since G is zero there. Writing the extracellular potential e, as simply 4>, then using Eq. (4.1)
we can rearrange Eq. (4.5) as
 V' . [(Ge G V' 4>] V' . [G V' (4.6)
We have the further condition that no current leaves the body, so the component of current
density normal to the body surface is zero. Thus, for y on the body surface and a unit
normal to the body surface at we have
y
.
e
o. uniformly zero boundary
condition (such as this) is referred to as "homogeneous".
Equation (4.6) is a partial differential equation, specifically, Poisson's equation. Let
us imagine that we are given V' . [G V' (the "source"), and we wish to compute the
resulting potential 4>the socalled "Forward Problem" (see Chapter of this book). One
very important feature of the Poisson equation is its linearity. That is, if a solution to
 V' . [(Ge Gi) V' 4>] satisfying the homogeneous boundary condition is known as 4>
and a solution to  V' . [(Ge Gi) V' 4>] g satisfying the homogeneous boundary condition
is known as then it is easy to verify that the solution to a solution to V'.
e
GJV'4>] c.] is given by (for Cl, C2 constants), and this solution also
satisfies the homogeneous boundary condition. This means that if we know the solutions
to Eq. (4.6) for a source localized to any single location in the heart, then to determine the
solution for any more geometrically complex source within the heart volume we only
need to add up (integrate) the solutions that would be obtained for each single location
comprising the complex source.
Thus, consider Eq. (4.6) where its righthandside (the source) has unit strength when
integrated over all space, but is zero everywhere except at x (i.e., the source is an electric
monopole). Denoting the solution as (where as a function of satisfies the
homogeneous boundary condition), we have
V' .
e
(4.7)
where the divergence and gradient operators are with respect to the field point y in three
space. Thus, is the potential at point in the body that would be induced by a unit
strength source that was zero everywhere in the heart except at the location x (a unit strength
source localized to a point is mathematically represented by a delta function). is
known as a Green's function. Since Poisson's equation is linear, we can now write the
126 F. Greensite
solution to Eq. (4.6) for the geometrically complicated source (on its righthandside) as
(4.8)
where V is the heart volume.
Following Yamashita and Geselowitz (1985), integration by parts applied to Eq. (4.8)
(i.e., application of the Divergence Theorem and the identity V . (fV fV . V V
leads to
L [G;
where S is the surface surrounding the heart muscle volume Equation (4.10) follows
firstly because the surface integral in Eq. (4.9) vanishes, i.e., n, is zero for on
the heart surface (G; V is the source current, and therefore confined to the heart , so that it
will have no component normal to the heart surface). The integrand on the righthandside
of Eq. (4.10) results because of the symmetry of G; (i.e., for vectors a, b and symmetric
matrix C, a' Cb b'Ca). A second integration by parts, now applied to Eq. (4.10), gives
Given that y is not in (e.g., we typically consider y to be on the body surface), the
volume integral on the righthand side ofEq. (4.11) is zero if is proportional to
e
(equal
anisotropy). This is because in that case we would have for some scalar a
(4. 12)
where the second equality ofEq. (4.12) follows from Eq. (4.7) since y on the body surface
is external to the source domain In this case, the imaging equation (4.11) becomes
L[G;V1fr (X,
(4.13)
Since satisfies Eq. (4.7), as a function of it can be thought of as the field generated
by a monopole at Thus, n, =V1fr [G;llx] can be thought of as
the field generated by a current dipole at x pointing in the G;llx direction. This can be
verified by introducing the second source where is a point close to with the
line between x and x' oriented as G;llx' This monopole of opposite polarity is associated
with a second Green 's function so that the composite of monopole sources of
opposite sign at x and x' approach a dipole . The appropriate limiting procedure leads to a
field determined by (G; n) . V1fr as in the integrand above.
Heart Surface Electrocardiographic Inverse Solutions 127
Thus, we have dervied a linear relationship between transmembrane potential at the
cardiac surface S (endocardium plus epicardium), and measurable body surface potential.
In the practical setting where the geometry is discretized, the function ¢ over body surface
points is expressed as a column vector whose components are measured potentials at various
electrode sites, while ¢mis a vector whose components are transmembrane potential at some
set of locations on the heart surface, and
x
(4.14)
is a matrix.
The forward problem associated with the above transmembrane potential formulation
requires knowledge of the anisotropic conductivity of the heart in construction of the op
erator Eq. (4.14). For example, solution of Eq. (4.7) for any given source point x requires
knowledge of G; throughout the body volume, including in the heart (where we
have to know G; and G
e
as tensors). There is an additional complication in that the equal
anisotropy assumption is not accurate, so one must also consider the second integral on the
righthandside of Eq. (4.11).
However, in the portion of the body external to the heart muscle we have that G; V¢;
0, since G; is zero outside the heart. In that volume, we have from Eq. (4.5) that
(4.15)
i.e., Laplace's equation. The relevant volume is bounded by the epicardium and the body
surface. Thus, the boundary conditions are divided into two parts: 1) the zero normal
component of current density at the body surface, 2) the (unknown) epicardial potentials.
Suppose we know the solution to this equation for the situation where the epicardial potential
is identically zero except for having unit strength concentrated at location xand call this
solution From the linearity of Eq. (4.15), we can then find the solution for any
geometrically complex epicardial potential distribution by simply adding together such
elemental solutions, just as was done for Poisson's equation. This again defines a linear
relationship as
(4.16)
where S is now the epicardial surface and ¢ep; is the epicardial potential. Thus, we have the
linear relationship between epicardial potentials and (measured) body surface potentials.
The Green's function is provided by solution of the forward problem (see Chapter
2 in this book). This formulation avoids having to consider the anisotropic myocardium in
the construction of since the volume under consideration for the partial differential
equation does not include the heart volume.
The last source formulation we will consider is that of the endocardial potentials.
we design a transvenous catheter such that its tip is embedded with many electrodes
(the "probe"), pass it into a cardiac chamber, and register its location with respect to the
endocardium (e.g., via ultrasound or electronic means), then we can consider the volume
between the catheter probe and the endocardium. Laplace's equation (4.15) still holds for
this volume (it is sourcefree). The boundary conditions are the zero component of current
128 F. Greensite
density normal to the probe surface, and the (unknown) endocardial potentials. Analogously,
we can again use Green's functions to derive a linear relationship between the endocardial
potentials and the probe electrode potentials, and we again have an equation of the form as
above, i.e.,
i
x
• (4.17)
That is, S is the endocardial surface, is potential at point on the endo
cardium, and
2
( X' , is the solution to Eq. (4.15) subject to the endocardial potential as
described by Data is the potential measured at point the electrode probe
surface.
Figure 4.3 presents a diagram which underlines the similarities and differences, and
relative advantages and disadvantages, of the endocardial and epicardial potential recon
struction methods.
Equation (4.13) (for transmembrane potential), Eq. (4.16) (for epicardial potential)
and Eq. (4.17) (for endocardial potential), constitute the basic imaging equations for the
techniques we will consider. As noted, there are very important distinctions between the
above source formulations, in that computation of the Green's function for the transmem
brane potential source formulation on the righthandside of Eq. (4.6) requires knowledge
of the anisotropic conductivity of the heart as well as knowledge of the torso tissue conduc
tivities external to the heart. In this regard, we note that strategies exist for patientspecific
tissue conductivity tensor imaging via MRI (Reese et al., 1995; Ueno and Iriguchi, 1998),
which might possibly be applied to the forward problem associated with Eq. (4.6). In the
epicardial potential formulation Eq. (4.15), only knowledge of tissue conductivity external
to the heart is requiredalthough this still requires knowledge of the heterogeneous torso
conductivity. However, recent work by Ramanathan and Rudy (2001) suggests that it is
unnecessary to precisely include the torso tissue conductivities for the epicardial potential
imaging problem (the opposite conclusion was reached by Huiskamp and van Oosterom
(1989) in the setting of activation imaging). In the endocardial potential formulation, one
only needs to have a value for the blood conductivitya striking simplification (that must,
however, be paid for by the invasiveness of the approach).
After discretization of the relevant surfaces and volumes (see Chapter 2 in this book),
Eq. (4.13), Eq. (4.16), and Eq. (4.17) become equations of the form
h Fg, (4.18)
where h and g are data and source vectors, respectively, and F is a (transfer) matrix.
4.3 HEARTSURFACE INVERSE PROBLEMS METHODOLOGY
Some significant mathematical issues are involved in providing optimal solutions for
the imaging equations of the last section. Firstly, one must deal with the concept of an "ill
posed problem" (intuitively, a problem whose nominal solution is unstable to small changes
in datae.g., unstable to small noise variations). Accordingly, it is necessary to introduce
"regularization" formalisms. Since the problem is inextricably bound up with noise, we
Heart Surface Electrocardiographic Inverse Solutions
myocardium
= 0
cavit y G, a const ant
= 0
_ ~ O
an p,.obe
endocardium:
epicardium:
body G, complicat ed
129
' body su rface: ¢b.
FIG URE 4.3. A diagram containing the elements used to define the endocardial versus epicardial potential
imaging problems . Epicardial potential imaging problem: The outer box contai ns the body. The outermost
ellipse represents the epicardial surface, on which potential
e
exists. Electrical potential satisfies Laplace' s
equation in the body volume external to the epicardium. The mixed boundary condit ions of the (unknown)
epicardial potential
e
and absent current component normal to the body surface, fully determine potential in
the body volume external to the heart. Green's second identity can be used to derive a linear dependence between
potential at the body surface and potential at the epicardium. However, computation of this linear dependence (the
forward problem) is dependent on knowledge of the extracellular conductivity Ge  which is very heterogeneous,
e.g., due to the lungs, fat, and muscle (however, there is recent evidence that the impact of these inhomogeneities on
an inverse solution may be small (Ramanathan and Rudy, 2001». Endocardial potential imaging problem: The
innermost ellipse represent s the surface of a catheter electrode probe, and the next ellipse represents the endocardial
surface. The region between these two surfaces (the bloodfilled lumen of a cardiac chamber) contains no current
sources, so Laplaces equation holds in this volume. The boundary conditions are the endocardial potential, and the
absent component of current normal to the probe surface. Again, a linear relationship can be derivedthis time
between measured probe potentials, and the (unknown) endocardi al potential. The latter linear operator is much
easier to compute than the corresponding operator for the epicardial potential problem, since the conductivity in
the relevant volume is uniform (simply being the conductivity of blood), and the geometry is easily measured
(i.e., it does not require CT or MRl ). However, the attendant advantages are tempered by the fact the the method
is invasive.
emphasize the statistical approach, from which other methods (e.g., those of Tikhonov) can
be interpreted as special cases . Secondly, one must be prepared to deal optimally with a
time seriesof such problemsii.e., stochastic processes.
4.3.1 SOLUTION NONUNIQUENESS AND INSTABILITY
As noted in the last section, we are required to deal with equations of the form h F
where the components of vector h are our noninvasive or minimally invasive measurements
at various spatially acces sible locations, F is a transfer matrix (supplied by the forward
130 F. Greensite
problem solution), and g is the image of transmembrane, epicardial, or endocardial po
tential. Assuming that has an inverse, one would presume that we could supply the
image as
To understand the naivity of such an approach, we can consider that our situation is
similar to being given a highly blurred image of a scene (our electrodes are located remote
from the sources, and all the sources potentially contribute to what is measured at each
electrode). We have some knowledge of the "point spread function", and can accordingly
attempt some "image enhancement" with the objective of "deconvolving" the source image
from the point spread function. But a blurring operator attenuates the information responsi
ble for higher resolution. Thus, the signal power of the high resolution information relative
to the power of the low resolution information is much smaller in the blurred image than in
the unblurred source image. At the same time, noise (not described by the blurring operator)
is also present in the resulting blurred image. That is, the noise (being added in addition
to the blurring) will not be blurred away (much of it can be thought of as being added
after the blurring operation, e.g., due to electrode noise or imprecisions in computation of
the transfer matrix Thus, it will be typical that the noise power exceeds the signal power
in the high resolution subspaces of the data. If one naively applies the inverse of the blurring
operator, the noise will continue to dominate the high resolution subspacesthus assuring
continued absence of identifiable high resolution features. Furthermore, since the blurring
operator severely attenuates high resolution, its inverse must involve a marked amplification
relevant to the high resolution subspaceswhich then also markedly amplifies the noise
in these subspaces, so that noise will dominate the entire solution (i.e., contribute most of
the power to the resulting image). This means that a nonsense solution estimate will be
obtained, also characterized by its instability to small changes in the data. Because of the
dominance of noise in the high resolution subspaces, one must simply forgo restoration
of the high resolution subspaces and be content with restoring those subspaces where the
blurred signal outweighs the noiseunless physiologically meaningful and valid extrinsic
constraints can be imposed.
In other medical imaging modalities, one does not have this difficulty. In MR!, for
example, the spins (the hydrogen protons in the body water or fat) are induced to produce a
signal (picked up by an antenna) whose frequency reflects their spatial location. Thus, the
amplitude ofthe signal at a particular frequency reflects the number of spins at corresponding
locations. In fact, following selective excitation of the spins in a single slice of the body,
the spins at a particular location in the slice are cleverly given two frequencies in different
bands (one from their spatiallyvarying NMR gyroscopic precessional frequency resulting
from a magnetic field gradient applied along one spatial direction, and another one via
manipulations of their relative phase via application of magnetic field gradient pulses in the
direction orthogonal to the first). The relationship of frequency to magnetic field strength
(which varies in space due to the applied gradients) means that an image of the tissue can be
obtained by applying a twodimensional Fourier Transform to the antenna data, so that the
magnitude of the resulting function is an image of spin density in the tissue slice. Unlike a
blurring matrix, a (discrete) Fourier Transform does not attenuate information in any source
subspace more or less than in any other subspace. Thus, inverting the effects of the operator
does not involve any differential noise amplification.
Heart Surface Electrocardiographic Inverse Solutions 131
To intelligently approach our dilemma, one ultimately needs to make the above notions
more quantitative. For this, it is useful to introduce the singular value decomposition (SVD)
of a matrix. A matrix represents a linear transformation, mapping a domain vector to a
range vector. For any linear transformation, it can be shown that there exists a particular
orthogonal coordinate system in the domain space, and a particular orthogonal coordinate
system in the range space, such that a vector pointing along a coordinate axis of the domain
space, is mapped to a vector pointing along a coordinate axis of the range space, and whose
magnitude is amplified by a nonnegative scalar depending only upon which domain axis it
was pointing along. Since any vector in either space can be written as a linear combination
of unit vectors in the above coordinate axis systems, the preceding wordy statement is
equivalent to the assertion that any matrix can be written as the SVD,
F USV
where the columns of matrices V and U are the requisite orthogonal bases of the domain
and range coordinate systems alluded to above, and
S diagts}, ... ,
is a diagonal matrix whose diagonal entries (the singular values) are the amplification
constants referred to above (they are arranged in order from largest to smallest). and
are each orthogonal matrices, and S is referred to as the singular value matrix. Note that
each singular value s, is associated with corresponding onedimensional domain and range
subspaces (the ith columns of V and U). By convention, we will take U to be an x
matrix, to be an x matrix, so that S is a x matrix.
We are now in a position to understand the severe mathematically determined difficulty
of our problem. In any noninvasive imaging technique for cardiac electrophysiology, is
always severely illconditionedbecause the field ¢ diminishes with distance from the
source, and the field at a point has contributions from all sources (there is "blurring"). As
a result, the ratio of its largest and smallest positive singular values is "large" (the value of
the smallest positive singular value is "small" compared to the value of the largest singular
value). That is, F is "illconditioned". In particular, the noise in the data in many of the
singular subspaces (columns of dominates the signal in those subspaces. However,
(assuming the inverse exists). A solution of the form thereby entails
application of l/si to the data component of h in the Uu subspace. If this is one of the
many subspaces for which 1 is very large and in which the noise dominates the signal,
we can appreciate that the noise is this subspace will be markedly amplified in the solution
estimate (this will also imply that the solution estimate will be very unstable to small noise
perturbations). Intuitively, we would thereby expect that it will be necessary to somehow
attenuate the solution components associated with many (or most) of the subspaces
meaning that there will be a severe limit on the number ofdegrees of freedom in a meaningful
estimate for g in Eq. (4.18) (essentially given by the number of singular values large enough
not to attenuate signal components below the noise amplitude in the subspace defined by the
corresponding column of Thus, much of the structure of an estimate for g must come in
the form of priori constraintsseither by default (imposed as artifacts of the regularization
procedure), or by design (constraints that truly reflect the class ofphysiologically meaningful
132 F. Greensite
solutions). Without such constraints, the solution estimate would be nonuniquesince the
addition of any vector in the supressed high resolution subspace to any solution estimate
gives a new estimate that is also consistent with the accessed data.
The field of Inverse Problems typically deals with situations where one is given data
reflecting the effect of some operator on a "source" we would like to estimate, but where
the inversion procedure (undoing the effect of the operator) is inherently unstable (e.g.,
highly noise amplifying), and (in practical terms) solution estimates are not unique. Such
problems are loosely referred to as "illposed" (the latter term has a quite precise meaning
in general Hilbert space settings, that we will not go into further).
4.3.2 LINEAR ESTIMATIONAND REGULARIZATION
Taking noise vector v into account, Eq. (4.18) becomes
(4.20)
where it is required to estimate g given hand
It might first occur to us that a useful estimate for g would be the one which maxi
mizes the choice that maximizes the conditional probability (density) that the
measured h would occur given a particular candidate for This is known as the "maximum
likelihood estimate". Assuming has an inverse
1
, and the noise has zero mean, this
is given by gml
1
hbecause the likeliest value for the noise (its mean) is the zero
vector (under conditions of zero mean Gaussian noise, the maximum likelihood solution
coincides with the leastsquares solution, i.e., if does not exist, is replaced by the
pseudoinverse). F is illconditioned, we have seen that this is in general a very poor
estimate for and is very unstable to noise variations.
But one could alternatively take the estimate for to be the choice which maximizes
the choice that maximizes the chance that a particular will be present given
the observed noisy data This is referred to as the "maximum a estimate,
gmap. The relationship between the two conditional probability densities above is given by
a version of Bayes Theorem:
gmap has the great advantage of being stable to noise variationsbut the disadvantage
that its calculation requires one to first supply nontrivial information concerning statistical
properties of (in fact, the entire "posterior" probability density such is available
and reliable, the methodology is referred to as "Bayesian". one supplies the statistical
properties as "drawn out of the air", or perhaps estimated from the given data h itself
("noninformative"), the methodology is referred to as "empirically" Bayesian. The general
approach is also referred to as "Satatistical Regularization" (evidently, in discrete settings,
the maximum likelihood estimate is equivalent to the Bayesian approach in a minimum
information setting where every realization of g is considered equally likely to occur).
All of this suggests that, once noise is introduced as in Eq. (4.20), it is useful to
carefully consider statistical notions. We will endeavor now to make these a little more
precise. Specifically, each component of noise vector v is a "realization" (outcome) of a
"random variable", the composite of which defines random vector For our purposes,
a random variable is an entity that associates a probability density value with every real
number. From this, we can compute the probability that some realization of the random
Heart Surface Electrocardiographic Inverse Solutions
variable (outcome of a measurement) will yield a value falling in some given interval.
Accordingly, the expectation tTl of some expression involving a random variable is the
integral of the expression over all possible values of the random variable weighted by the
probability density associated with each value (a zero mean random variable ais such that
0). Furthermore, a "Gaussian" random variable has a Gaussian probability density
(the familiar bellshaped curve), and is fully characterized by its particular expectation £[
and variance Similarly, a zero mean Gaussian random vector Wis a column
vector of zero mean jointly Gaussian random variables Wi, i.e., W (WI, ... , A zero
mean random vector is further characterized by its autocovariance matrix
t
], which
describes the dependence between all different pairs of components of W (note that the
product of jointly Gaussian random variables is Gaussian). Similarly, the crosscovariance
matrix of zero mean random vectors v, w is given by
t
], and describes the mutual
dependence of ii and W.
Just as v is a realization of a random vector v, so too can g be considered to be an
(unknown) realization of random vector For notational simplicity, in this subsection we
will suppress the superscript"  ", and denote a random variable and its realization by the
same symbol. However, we will resume the notational distinction in the next subsection.
In approaching Eq. (4.20), a good objective is to find gopt such that gopt 11
is minimum (this being the "minimummeansquareerror" estimate). If g and v are re
alizations of zero mean Gaussian random vectors, gopt is obtained via the Wiener filter.
Under these conditions, the maximum a posteriori estimate gmap is equivalent to gopt. This
linear estimation procedure develops as follows.
A linear estimate of is given by application of an "estimation matrix" M
est
to data
h, i.e., gest Mesth. Ideally, we desire the solution estimate
(4.21)
such that Mopth is minimum. Thus, it is sufficient to calculate M
opt'
The way to
proceed follows from the "Orthogonality Principle", which asserts that gopt minimizes the
meansquareerror when
(4.22)
i.e., when the crosscovariance matrix of the "error of the estimate" (g  gest)and the "data
vector" h is the zero matrix (so that the error and the data have no dependence). Intuitively,
the Orthogonality Principle assures that every bit of useful information is extracted from
the data h in making the solution estimate gopt. Substitution of Eq. (4.20) and Eq. (4.21)
into Eq. (4.22) immediately gives
(4.23)
Thus, assuming that g and v are independent (i.e., £[gv
t]
= 0), Eq. (4.23) can be written
as
where C == £[ggf] and C; == £[vv
t]
are the autocovariance matrices of signal g and
noise v. Hence,
F. Greensite
(4.24)
Thus, the optimal solution estimate for Eq. (4.20) is provided by Eq. (4.24) and Eq. (4.21)
assuming we know the autocovariance matrices of signal and noise.
is interesting to express this estimation matrix as a modification of
assuming the latter exists. We have from Eq. (4.24) that
where Ci;
v
) is the autocovariance matrix of h (as is seen via Eq. (4.20)).
Thus, M involves an initial preprocessing of the data h via (the classical
Wiener filter (Papoulis, 1984)) followed by application of (note that providing
nominally requires knowledge of C although one can attempt toestimate Ch using the given
measurement h itselfessentially the problem of spectral estimation (Papoulis, 1984)).
C
v
a; I (i.e., if the noise is white), Eq. (4.24) becomes
2 1)1
rrs (4.26)
since If using the SVD,
USV we can write Eq. (4.26) as
(4.27)
An alternative applicable formulation for treating Eq. (4.20) is provided by the
maximum likelihood method in concert with a deterministic constraint (such as that the
signal power is equal to some a priori vaule Maximum likelihood then corresponds to
minimizing
2
subject to E (assuming Gaussian noise). This also leads to
a linear estimation matrix of similar form to the above. Ultimately, the distinction between
the Bayesian and constrained maximum likelihood methods are that the latter treat only the
noisy measurements in a statistical fashion, while the former also treats the underlying sig
nal statistically. Thus, the Bayesian methods potentially allow (or require) introduction of a
larger class of a priori information. A third alternative, Tikhonov regularization (Tikhonov,
1977), could be viewed as either a hybrid or an empirical Bayesian method. The linear
estimation matrix supplied by it can typically be interpreted as resulting from the assump
tion of white noise and selection of a term proportional to /a in Eq. (4.26), where the
proportionality scalar (the regularization parameter) is selected using either a deterministic
or datadependent constraint.
Comparing the righthandside of Eq. (4.27) to
1
= SI (assuming the in
verse exists), we can appreciate the violence that estimation theory does to the notion of a
high resolution reconstruction of Reversal of the effects of would require application
of 1/s to the data component in the U; subspace. Instead, for the regularized estimate,
the factor sd(s; is applied. As s, becomes small, this factor bears no resemblance
to 1/SiSO there is no attempt at faithful reconstruction of components of the associ
ated source subspaces. Another way of looking at the situation, is that a squareintegrable
Heart Surface Electrocardiographic Inverse Solutions 135
(i.e., wellbehaved) function (or image) must be such that its higher order Fourier coeffi
cients tend to zero (we imagine the Fourier coefficients to be with respect to the SVD domain
coordinate system of given by the columns of in the discretized approximation). In
the presence of white noise, whose Fourier coefficients therefore do not tend to zero, it is
clear that higher order Fourier coefficients of data h are hopelessly noisecorrupted. The
Wiener filter, and Tikhonov regularization, achieve stable results by removing any attempt
at meaningful reconstruction of the high resolution components.
There is only one way out of the dilemma of resolution loss. If physiological constraints
exist which effectively reduce the dimension of the solution space to be commensurate with
the number of useful data Fourier coefficients, one can anticipate that it will be possible to
preserve spatial resolution. For example, for the inverse electroencephalography problem
(where one wishes to image the brain sources of the scalp electrical potentials), it might be
true that only a single focus is responsible for inciting an epileptic seizure, and that this focus
can be modeled as a single current source dipole located at some unknown location in the
brain. In that case, one is searching for an entity with six degrees of freedom (reflecting its
location, orientation, and magnitude). High spatial resolution could conceivably be possible
assuming there are six or more data Fourier coefficients (with respect to the transfer matrix
SVDderived coordinate system) that are not dominated by noise. At first blush, such an
obvious constraint does not appear to be physiological in the heart, since the heart is not
faithfully modeled as a single dipole. However, a deeper look at the geometry reveals
that such constraints do in fact apply (in principle) for the "critical points" of ventricular
activationfrom which an activation map can be fashioned (see Section 4.6).
4.3.3 STOCHASTIC PROCESSES AND TIME SERIES OF INVERSE PROBLEMS
The data available in our problem are distributed in time as well as space. Thus,
Eq. (4.20) would be more appropriately written as
hi Vi, (4.28)
i 1,2, ... , n, where i indexes the time instants at which measurements are made (note
that we leave open the possibility that the transfer matrix is timevarying, thus we write it
as Underlying Eq. (4.28) are the time series of random vectors, hi, gi, and vii.e.,
stochastic processes. The important additional feature of a stochastic process is that the i th
random vector may have correlations with the jth random vector for =1= i. However, a
"state variable model" which embodies known correlations between and for =1= is
not explicitly available in our problems. Given the lack of such explicit accurate constraints,
it is usual to adopt a "minimum information" perspective. Though such an approach seems
reasonable, and suggests that the equations ofEq. (4.28) might best be treated independently
of each other (by simply applying the methods of the prior section to each one), the reality
is more subtle.
For convenience, let us define matrices G, such that Hi hi,
i
gi, Ni
ViSO that Eq. (4.28) becomes
Hi Fi
G:
i Ni'
(4.29)
136 F. Greensite
i 1, ... , n. We then have the underlying random matrices as G, N such that their
ith columns are respectively. The usual assumption is that the entries of are
independent and identically distributed random variables (and similarly for the entries of
This is equivalent to the statement that all row autocovariance matrices are proportional
to the identity matrix (with the same proportionality constant), and all row crosscovariance
matrices are the zero matrix (this is also equivalent to the statement that all column auto
covariance matrices are proportional to the identity matrix, with the same proportionality
constant, and all column crosscovariance matrices are the zero matrix). This minimum
information assumption would imply that each member of equation sequence Eq. (4.28)
can be treated independently of every other member of the sequence.
However, if we leave open the possiblity that there are correlations between the dif
ferent the members of the equation sequence can no longer be considered necessarily
independentand an optimal processing of the data is subject to specification (or identifi
cation) of appropriate choices of the crosscovariance matrices of the columns of Thus,
suppose we continue to assume that the row crosscovariance matrices of are the zero
matrix and the row autocovariance matrices of are identical, but that the latter are not
necessarily proportional to the identity matrix (thus, we will be rejecting the minimum in
formation approach). This means that the column autocovariance matrices are proportional
to the identity matrixbut we still have not specified the column crosscovariance matrices.
Estimates for these will be derived from the data, i.e., empirically. This is actually not a
radical thing to do, since even in the minimum information approach one typically derives
the signal power (or signaltonoise ratio) from the given data (thus, the minimum informa
tion approach is by no means "pure" in this respect). In fact, under the present conditions,
there is a favored nontrivial choice of each crosscovariance matrix 0
For the purposes of linear estimation, Eq. (4.29) can be equivalently written in block
matrix form as
~ : l ) = [ ~ ' .. ~ ....~ .. :] ( ~ l ) ( ~ : 1
H: n 0 0 . . . G'; Nn
(4.30)
The Wiener filter (detailed in the last subsection) supplies an optimal estimate of the entries
of G as given by
(4.31)
where diag(F is the block diagonal matrix on the righthandside of Eq. (4.30), and
c., =
i
0 s.»;
(4.32)
(4.33)
Heart Surface Electrocardiographic Inverse Solutions
i.e., C is the block matrix whose (i, block entry is the crosscovariance matrix of ifi
with c.; etc. Thus, C is the (large) autocovariance matrix ofthe random vector consisting
of the entries of Equation (4.31) is simply the composite of Eq. (4.21) and Eq. (4.24)
applied to Eq. (4.30).
In the "Standard Method" (the minimum information approach), one takes 0
G:
j
] to be the zero matrix when i # In the "New Method", for i # one takes
(4.34)
assuming the trace of is not zero. can be shown that the meansquareerror in the
resulting estimate of signal autocovariance matrix C is smaller than the estimate used in
the Standard Method (Greensite, 2002).
For the case of white noise, and assuming the are identical (i.e., for all
it can also be shown that the New Method reduces to the following procedure: Instead of
individually treating the equations
Hi FG:i Ni,
1, ... , n, we instead individually treat the equations
HX: FGX: NX:
(4.35)
(4.36)
i 1, ... , n, where the columns of n x n matrix X are the eigenvectors of Denoting
the solution to the ith equation ofEq. (4.36) as (the ith column of a matrix
we take the solution estimate for G to be
(4.37)
The method generalizes to the case where there are nontrivial spatial correlations,
and also to the case where a priori constraints are available regarding time correlations.
However, if nonwhite characteristics of the noise are known, the route of Eq. (4.37) is
unavailable, and one is left with the computationally complex method resulting directly
from Eqs. (4.30)(4.34) (Greensite, 2002).
Underlying the New Method is the recognition of a fundamental asymmetry regard
ing G That is, the signal G undergoes a spatial transformation, but does not
undergo a time transformation. For an equation of the form h F v, where we con
sider g, v to be spatial vectors, it is quite reasonable in a filtering context to impose a
signal autocovariance matrix proportional to the identity matrix. This would imply an
autocovariance matrix for the noiseless portion of as given by implying nontrivial
filtering of noisy h (see the second equality in (4.25)). However, for an equation of the form
h h v, where the h and are considered time series, it is quite unreasonable to set the
signal
o
) autocovariance matrix proportional to the identity, since the resulting (Wiener)
filter is no filter at all (assuming white noise). Thus, since is a spatial transformation, while
G is spatiotemporal, one cannot simply impose the minimum information condition that
the entries of are independent and identically distributedassuming that one wishes to
138 F. Greensite
effectively filter in the time domain. In the setting of "minimum constraints" as opposed to
"minimum information", the New Method is a means of performing spatiotemporal filtering
in a manner dictated by the brokensymmetry of the problem, and the desire to minimize
meansquareerror in the utilized signal autocovariance matrix.
4.4 EPICARDIAL POTENTIAL IMAGING
The source formulations for the inverse problem of electrocardiography have included
those of a single moving dipole (Gabor and Nelson, 1954), two moving dipoles (Gulraj ani
et al., 1984), dipole arrays (Lynn et al., 1967; Barber and Fischman, 1961; Bellman et al.,
1964; He and Wu, 2001), multipole expansion coefficients (Geselowitz, 1967), and a heart
excitation model (Li and He, 2001). But in a very influential letter to the editor, Zablow
(1966) asserted the need to reconstruct an actual anatomicallybased entity that was already
being accessed invasivelyso that artifacts of the source model might be minimized, and
the result could be thought of as representing some sort of verifiable physiological truth.
In essence, he noted that a linear relationship existed between the epicardial potentials
and measurable potentials at the body surface, and suggested the former as the source
formulation to be reconstructed. This was particularly attractive, since physiologists were
already engaged in measuring the epicardial potentials invasively, and these were deemed
useful. Over the next several decades many investigators pursued the objective of epicardial
potential imaging. The many proposed refinements of technique can be divided into those
pertaining to
• Statistical regularization,
• Tikhonov regularization,
• Truncated SVD regularization,
• Constrained least squares regularization,
• Nonlinear regularization methods ,
• Augmented source formulation ,
• Different methods for selecting regularization parameters,
• Preprocessing the data,
• Introduction of spatiotemporal constraints.
Before embarking on a discussion of these refinements, we observe that there is no
consensus regarding which methods are the most worthy of employmentand we do not
attempt such value judgements here. A comprehensive approach to this question is itself a
sizable objective that has not yet been achieved. Ultimately, the difficulty is in the experi
mental setup requiredi.e., the need for (ideally) simultaneous collection of body surface
and epicardial data, with coincident anatomical imaging and electrode registration, in a
series of animals and human subj ects with a variety of pathological conditions (Nash et al.,
2000).
4.4.1 STATISTICAL REGULARIZATION
In what was apparently the first serious treatment of the epicardial potential source for
mulation , Martin and Pilkington (1972) reported on the dismal prospects of "unconstrained"
Heart Surface Electrocardiographic Inverse Solutions 139
inverse epicardial potenti al imaging , identifying implications of the problem illposedness
discus sed in Section 4.3. They subsequently applied the Weiner filter, and reported more
encouraging results in followup simul ations (Martin et aI., 1975). This approach requires
estimates of both the signal and noise autocovariance matrices. While the noise might be
considered white (ignoring inaccuracies in the forward problem construct a choice for
the signal autocovariance matrix is less obvious. They proposed two ways of choosing one.
The first was based on estimating the spatial autocovariance from time ensembles of epicar
dial potential maps supplied from a representative set of activation sequences. The second
was a Monte Carlo method, whereby each epicardial location was given some a priori
probability of being activated at any given time, and epicardial maps were then generated
by random numbers assigned to each locationthus leading to a computation for the signal
autocovariance matrix.
Following the innovat ions of Barr et al. (1977) on the forward problem, Barr and Spach
(1978) reported on inverse calculation of epicardial potential in twelve dogs with chronically
implanted epicardial electrodes. In applying the Wiener filter, they simply opted to take the
signal autocovariance matrix as proportional to the identity (i.e., as random variables , the
epicardial potential s at all locations on the epicardium were presumed to be independent
and identically distributed). They concluded that some features of the epicardial potential
distribution through time can be imaged, particularly in dogs for which detailed geometry
measurements.were available (postmortem).
Recently, van Oosterom (1999) has reexamined the statistical regularization approach,
concluding that impressive improvements in accuracy (compared with other regularization
methods ) are possible if a nontrivial accurate signal covariance matrix is available. He
suggested that the signal autocovariance matrix could be based on prior estimation of the
activation sequence via other techniques (e.g., as in Section 4.6).
4.4.2 TIKHONOV REGULARIZATIONAND ITS MODIFICATIONS
Despite the presence of noise v in Eq. (4.20), we are still tempted to view our problem
as one of applying a kind of "inverse" of Indeed, given that the magnitude of v is "small",
we are even tempted to pretend that we are dealing with h F Our prior discussion has
surely revealed that one cannot expect to apply the actual inverse of to because of
instability problems, but it is also useful to specifically address the (typical) setting where
doesn 't even have an inverse (which will always be the case if is not square). Firstly,
it could be that is not even in the range of In that case, we might (naively) choose the
solution estimate g est that minimizes
(4.38)
over all point s g in mspace (the expres sion Eq. (4.38) is known as the "residual" or
"discrepancy"). But a second problem could be that the residual may not have a unique
minimizer (as occurs when the dimension of h is smaller than the dimension of One
could then ask for the estimate gest that is of minimum norm among all the minimizers of
the residual. From convexity arguments, it can be shown that the minimizer of the residual
Eq. (4.38) of smallest norm is unique . In fact, it can be shown that the minimumnorm
140
leastsquares solution for gin h Fg is given by
F. Greensite
where is the "pseudoinverse" of For the SVD of as in Eq. (4.19), the pseudoinverse
is given by
where is the diagonal matrix whose ith diagonal entry 1/s, if s, 0, and zero otherwise.
Intuitively, it is easy to see why this works: The null space of is the subspace orthogonal
to the range of F. Thus, h does not burden the estimate with any component that doesn't
contribute to fitting the data. Otherwise, simply undoes the attenuation s, that components
of g experience when F is applied.
However, the above is simply a fix for the situation where F does not have an inverse.
From Section 4.3, we know that, even if F has an inverse, we are faced with solution
estimate instability if F is illconditionedbecause of the subspaces corresponding to small
positive singular values. This consideration obviously will still hold for the pseudoinverse
based solution. We have already encountered the Wiener filter regularization approach,
which requires priori estimates of at least the form of the signal and noise autocovariance
matrices.
ColliFranzone et at. (1985) introduced the Tikhonov regularization method to the
epicardial potential imaging problem, ostensibly avoiding the problem of providing esti
mates for the signal autocovariance and noise autocovariance matrices. This approach skirts
the usual notions of stochastic processes, and instead begins with the desire to minimize
Eq. (4.38). But instead of simply searching for a solution estimate gest which minimizes the
residual (which would lead to an estimate that is exceedingly noise sensitive and unstable),
in the Tikhonov approach one searches for the solution estimate that minimizes
(4.39)
where R is some matrix. Thus, one seeks an estimate for which the residual II Fgest  h is
"small", while at the same time some other property of the estimate, measured by II Rg
est
[[2,
is also small. For example, R might be the identity, in which case one is looking for an
estimate with small residual as well as a small norm (unstable nonphysiological solutions
will tend to have large norms). Alternatively, R could be such that II Rg
est
11 reflects the first
or second spatial derivative of the solution estimateso that the regularized estimate would
have to be relatively "smooth". These approaches require selection of a "regularization
parameter" y, which regulates how strong an influence the cominimized second property
has in determining the solution estimate.
In the Tikhonov approach, the regularized estimate is given by
obtained by setting to zero the derivative of Eq. (4.39) with respect to g (a gradient),
assembling the simultaneous equations into a matrix equation, and solving for (note
Heart Surface Electrocardiographic Inverse Solutions 141
that Eq. (4.39) is a function of the variable ga point in mspace; thus, its gradient is
a vector in mspace). In the sense of Section 4.3, Eq. (4.40) evidently describes a linear
estimation methodthe estimation matrix being the expression in brackets on the right
handside of Eq. (4.40). Many alternative regularization operators can be used. With
"zeroorder Tikhonov", the estimation matrix results from the choice R Iso the tech
nique corresponds to statistical regularization under the assumption that the signal and
noise covariance matrices are both proportional to the identity matrix, with the regulariza
tion parameter presumptively being the inverse of the square of the signaltonoise ratio
(i.e., compare Eq. (4.40) to the first equality in Eq. (4.27)). Firstorder and secondorder
Tikhonov regularization correspond to a choice of R derived from the gradient and Laplacian
operators. Thinking of these in the context of statistical regularization (compare Eq. (4.40)
to Eq. (4.26)), these symmetric higherorder Tikhonov reguarization operators correspond
to a signal autocovariance matrix that is a smoothed version of the sharp "ridge" repre
sented by the identity matrix (with the assumption of white noise). That is, there is now
a nonzero covariance between spatially proximate locationsinstead of these being taken
to be independent (as with a signal autocovariance matrix proportional to I). Although
ColliFranzone et at. (1985) suggested that firstorder Tikhonov regularization was more
accurate in in vitroexperiments, MessengerRapport and Rudy (1988) found no significant
differences in the results obtained with zeroorder, firstorder, or secondorder Tikhonov
regularization.
When expressed in terms of an SVD, the zeroorder Tikhonov solution
estimate is given by
(4.41)
which employs a linear estimation matrix comparable to that in Eq. (4.27). A socalled
"regional regularization" scheme was suggested by Oster and Rudy (1997), whereby the
solution is given as
where is a diagonal matrix whose diagonal elements take on a few different valuesin
effect, the diagonal values of represent multiple regularization parameters. Again, this
can be interpretted as an attempt to supply the signal autocovariance matrix.
The "spatial regularization" method (Velipasaoglu et al., 2000) selects R in a manner
inspired by the fact that the noisy data fails to satisfy the Discrete Picard Condition (Hansen,
1992; Throne and Olsen, 2000). The latter condition asserts that a stable solution requires
that the squares of the Fourier coefficients of the data with respect to the eigenvectors of
F should on average decay faster than the eigenvalues of FFt. A means of modifying
the data to conform to this Condition is the basis of this approach.
4.4.3 TRUNCATION SCHEMES
As noted earlier, the minimumnorm leastsquares solution for gin h Fg is given
by gest Ft h Y h, where is the diagonal matrix whose ith diagonal entry I/si
if s, 0, and zero otherwise. With this in mind, one could consider a regularized solution
142
estimate given by
F. Greensite
(4.42)
where is the diagonal matrix whose ith diagonal entry is 1/ s, if s, and 0 otherwise.
This regularization method is known as Truncated SVD, or TSVD regularization (Hansen,
1992). The value functions as a regularization parameter. TSVD performance is usually
very similar to that of zeroorder Tikhonov regularization.
The truncation idea is also incorporated in the "Generalized Eigensystems" approach
of Throne and Olson (Throne and Olsen, 1994), which is relevant to a finite element dis
cretization of the body (rather than a boundary element discretization). Instead of truncating
a solution expanded in the singular vectors of the transfer matrix (as in Eq. (4.42)), they
consider a set of generalized eigenvectors defined over the entire finite element mesh, hav
ing the properties that each generalized eigenvector satisfies the boundary conditions on the
forward problem, as well as Laplace's equation within the volume, and the "subvectors"
consi sting of the components on the epicardial surface are orthogonal. One then constructs
a linear combination of the generalized eigenvectors such that the body surface potential
data h is fitted by the components that correspond to locations on the body surface. The
components corresponding to the epicardial potential locations then take on values deter
mined by this linear combinationwhich would be the presumed inverse solution desired .
However, this will nominally lead to an unstable noisedominated solution. Therefore, one
truncates the linear combination, using only the generalized eigenvectors associated with
the largest generalized eigenvaluesthus achieving a stable solution estimate . Instead of
expanding (and truncating) the solution series in terms of the eigenvectors of FF' as with a
TSVD (a sequence which most efficiently represents the effects of for a given (truncated)
number of terms), one is truncating a series derived from a set of field vectors that most
efficiently pack the power of the field over the entire body volume a given (truncated) set
of components.
Truncation is also employed in the "local regularization" scheme of Johnson and
MacLeod (Johnson, 2001). In this approach, it is recognized that is expressed in terms
of the "inverses" of three different submatrices, when a finite element discretization of the
forward problem is employed. Since these matrices have much different condition number,
the implication is that they should each be receive different degrees of regularization (e.g.,
individualized SVD truncation).
4.4.4 SPECIFIC CONSTRAINTS IN REGULARIZATION
The Tikhonov type formulation (4.39) also suggest s a way in which other types
of constraints could be formulated . For example, suppose one has knowledge that the
solution shares some features with a preliminary estimate grough' One could then suggest
the minimization of
(4.43)
The explicit expression for gesl is then obtained by setting the gradient of the above ex
pression to 0, and solving for g. This is known as the Twomey method (Oster and Rudy,
1992).
Heart Surface Electrocardiographic Inverse Solutions 143
Iakovidis and Gulrajani (1992) introduced a method whereby a deliberately over
regularized estimate (to be used to estimate the location of the epicardial zeropotential
line, but otherwise having too few interesting features ) was used to constrain the solution
for what would otherwise be an underregularized estimate (i.e., the second regularization
parameter would have been too small if the constraint had not been present, in the sense
that the estimate obtained would have been unstable and "noi sy").
4.4.5 NONLINEAR REGULARIZATIONMETHODOLOGY
Overwhelmingly, linear estimation methods have been applied to obtain regularized
(i.e.•noisestable) estimates of linear formulations of the inverse electrocardiography prob
lems. However, nonlinear (e.g., information theoretic) methods have also been applied (a
particularly recent example is provided in (He et al., 2000)). We will not describe such
approaches here.
4.4.6 ANAUGMENTED SOURCE FORMULATION
one treats Eq. (4.5) via a Green's function approach along the lines of what was
done with Eq. (4.6) (i.e., performing two integrationsbyparts analogous to Eq. (4.9) and
Eq. (4.11)), one obtains an expression relating body surface potentials to the composite
of epicardial potentials and the normal component of epicardial current density (as in
(Greensite, 2001, p. 151152),
Noting that such a formulation occurs as an intermediate step in the development of the
forward problem method expounded by Barr et al. (1977) , Horacek and Clements (1997)
investigated solutions obtained where both the epicardial potential and epicardial normal
current density are inversely computed. They suggested that this problem might be slightly
better posed than the traditional epicardial potential formulation, and they also investigated
refinements in the regularization technique.
4.4.7 DIFFERENTMETHODS FOR REGULARIZATION PARAMETER
SELECTION
The Tikhonov estimate Eq. (4.40) requires selection of a value for the regulariza
tion parameter. Some guidance in this regard is provided by the "Discrepancy Principle"
(Hansen, 1992). Here the reasonable assumption is made that the discrepancy (or resid
ual) IIF h11 is not zerobecause of the noise present. Rather, the discrepancy (for the
true solution is most likely to be the noise power. Thus, it makes sense to look for a
solution estimate which produces this discrepancy. So, consider the constrained minimiza
tion problem: "minimize
2
such that
2
Again,
2
and
2
are
each functions of (which varies over points in mspace). We know from Calculus that
minimization of the first function subject to the constraint on the second function implies
(under rather general conditions) that the gradients of the two functions at the constrained
minimum gm in lie on the same rayi.e., the gradients are proportional at gmi n' Another
144 F. Greensite
way of saying this is that there exists a scalar y such that the gradient of Eq. (4.39) is zero
at gmin (in this case R 1). As we know, the requisite gmin is given by the righthandside
of Eq. (4.40), where the regularization parameter y was to be determined. But now, the
regularization parameter is simply given as that which produces a Tikhonov solution esti
mate gmin satisfying the discrepancy expression gmin  h
2
E, where E is the noise
power.
However, the error in the data (noise power E) is not known (being a composite of
electrode noise and modeling errors in There are actually several other methods for
regularization parameter selection. In the socalled "Lcurve method" (Hansen, 1992), one
computes a loglog plot of the first term in Eq. (4.39) versus the second term in Eq. (4.39)
(the residual versus the solution estimate seminorm). The solution estimate is chosen as the
one corresponding to the "comer" of the above Lshaped graph (a balance between small
seminorm and small discrepancy). should be noted, however, that a comer on the Lcurve
does not always exist.
The Composite Residual and Smoothing Operator (CRESO) method (Colli et al., 1985)
chooses the smallest positive value of the regularization parameter for which the second
derivative of the first term in Eq. (4.39) with respect to the regularization parameter equals
the second derivative of the second term in Eq. (4.39) with respect to the regularization
parameter. The crossvalidation method (Whaba, 1977) is another important regularization
parameter selection method, though it has not been prominently applied in the context of
the inverse electrocardiography problem.
4.4.8 THE BODY SURFACE LAPLACIANAPPROACH
As we have previously noted (see equation (4.25)), statistical regularization can be
viewed as the composite of a particular "preprocessing step" followed by application of the
standard inverse (if such exists). This suggests the possible application of other preprocess
ing steps. He and Wu (1997) proposed preprocessing the data to be in the form of body
surface Laplacian measurements, and solving the inverse problem using this processed data
as input. The resulting transfer matrix is of a different nature (evidently, it results from
left multiplying the usual transfer matrix by the same matrix that is used to preprocess the
body surface data), in that the surface Laplacian of potential at a body surface location is
significantly influenced by many fewer source locations than potential itself (i.e., there is
less "blurring"). As a tradeoff, the effect of the source on the "data" falls of more rapidly
(with the fourth power of the distance from the source, rather than the second power of the
distance, as with body surface potential), and there is a theoretical noise amplification in
numerical differentiation of the body surface potential.
Ignoring the curvature of the body surface, the body surface Laplacian of body surface
potential h with respect to an body surface coordinate system is given by
Ultimately, we can write this as where is the Laplaciana linear operator. After the
problem is discretized for numerical treatment, is simply a matrix. Thus, v
Heart Surface Electrocardiographic Inverse Solutions
becomes
Lh (L F) g L v,
145
Our data is now Lh, and we now need to "invert" (L F) to estimate epicardial potential
The regularization tools remain the same as before. is a differential operatorand
thus can be thought of as akin to a high pass filter. With respect to a Fourier expansion of
h, application of L amplifies the high frequency terms. In particular, the high frequency
components of the noise are greatly amplified (in the nondiscretized setting, the noise is
"unboundedly" amplified). However, the latter is in principle taken care of by the fact that LF
is more "singular" than Fmeaning that it's "inverse" will be smoother (i.e., application
of the inverse is more stable, and tends to smooth noise). If one has an expectation of
somehow reducing the Lv contribution to data Lh prior to application of the "inverse",
one might expect greater stability and fidelity in the estimate of than that obtainable with
the direct treatment of h Fg v. Such an expectation could be reasonable with use of
Laplacian electrodes (He and Cohen, 1992), though it has not been established that these can
be accurately designed in practice. In the absence of such, investigations have proceeded
with direct application of L to measured data h. The approach seems to have potential in
identifying and spatially distinguishing cardiac sources close to the body surface electrodes
(Johnston, 1997; He and Wu, 1997).
4.4.9 SPATIOTEMPORAL REGULARIZATION
Ultimately, one is faced with a time series of problems
(4.44)
i = 1, 2, ..., where the subscript now refers to the source and data at the ith time point
in the cardiac cycle. Oster and Rudy (1992) suggested using preliminary (e.g., zeroorder
Tikhonov) estimates at time points i  I and/or i 1 to constrain the regularization at
the ith time step. This was done using a Twomey regularization formalism via equat ion
Eq. (4.43).
On the other hand, Eq. (4.44) evidently describes a stochastic process, which begs the
question of applying Kalman filter theory. This requires that a stochastic model be applied,
defining the presumed interdependence of the epicardial potentials between different times
(Joly et al., 1993)is itself a not entirely trivial problem.
Temporal and spatial constraints can also be joined by the the method of Brooks
et al. which employs two or more regularization parameters in a traditional constrained
minimization format (Brooks et al., 1999). Thus , Eq. (4.44) is written as
(4.45)
One now writes a functional to be minimized, consisting of the residual (for this augmented
146 F. Greensite
problem), a spatial regularizing operator (e.g., expressing the sum of the norms of the
solution estimates at each time point), and a temporal regularizing operator (e.g., the mag
nitude of the discretized "time derivative" of the solution estimates over all the time points)
where the latter two operators are given their own regularization parameters. The solution
estimate is ultimately expressed as
where is the block matrix on the righthandside ofEq. (4.45), is the discretized
version of a temporal differential operator, and Yl, Yz are the two regularization parameters.
The "admissible solution" approach of Ahmed et al. (1998), posits that any solution
satisfying a sufficiently robust composite of constraints is deemed satisfactory, and such
constraints can include those related to time. The solution algorithm requires the constraints
to be convex, e.g., the "ball" of vectors satisfying is an example of a convex set
(any line joining any two members of the set consists only of members lying in the set). The
need for regularization parameters is replaced by the need for bounds defining the required
convexity.
Finally, the approach of Greensite (1998; 2002) (described in Section 4.3.3) effec
tively replaces the original sequence of (nonindependent) Eq. (4.35) (or Eq. (4.44)) with
a smaller number of mutually independent equations Eq. (4.36)without the imposition
of any extrinsic temporal constraints (or temporal regularization parameter). The method
derives from the recognition that there is something intrinsically wrong with the assumption
that the entries of G are realizations of independent and identically distributed random vari
ables. Indeed, this symmetry condition is broken once one poses the problem described by
Eq. (4.44). Given the assumption that the rows of G are independent and have identical au
tocovariance matrices, the solution mechanism uses a more accurate signal autocovariance
matrix estimate than the other methods (in a meansquareerror sense).
4.4.10 RECENTIN VITRO AND IN VIVO WORK
There has been a significant amount of in vitro work done with the Utah torso tank. The
experimental setup employs a heart suspended in a torso shaped electrolytic tank, perfused
by an anesthetized dog external to the tank (Oster et al., 1997). Electrodes are present on the
outer margin of the tank, and also in proximity to the epicardium. Such work has addressed
inverse reconstruction of epicardial potentials and activation in the setting of sinus rhythm,
pacing, and arrhythmias (Burnes et al., 2001), and has also been used to assess the impact
of torso inhomogeneities (Ramanathan and Rudy, 2001).
Perhaps the most impressive demonstration of the potential promise of the epicardial
potential imaging formulation is in the quasiin vivo work where densely sampled epicardial
potential data was accessed from infarcting dogs, and used in simulations with numerical
human torsos to test the fidelity of the zeroorder Tikhonov inversion under realistic condi
tions of modeling and electronic noise (figure 4.4) (Burnes et al., 2000). This approach has
also been applied to demonstrate the feasibility of reconstructing repolarization properties
of interest (e.g., increased dispersion of repolarization (Ghanem et al., 2001)).
On the other hand, earlier results of a different group using invasive data from patients
undergoing arrhythmia surgery, suggested that the usual epicardial potential regularization
Heart Surface Electrocardiographic Inverse Solutions 147
FIGURE 4.4. Maps of epicardial potent ial at two different times during ventricular activation. in the study of
(Burnes et al., 2000). The invasively measured epicardial potential data (from dogs) is used to forward generate
body surface potentials on a numeri cal human torso surface. Geometri cal and "electronic" noise is added to these
body surface potentials , and inversely reconstructed epicardial potential maps are computed. [From: Burnes, J. E.,
Taccardi , B.• Macleod, R. S., and Rudy, Y, 2000, Noninvasive ECG imaging of electrophysiologcially abnormal
substrates in infarcted hearts, a model study, Circulation. 101: 533 540. Used by permission.]
methodology was able to usefully image epicardial potential during the QRS inteval only
in its initial portions (Shahidi et ai., 1994).
Finally, a report by Penney et ai. (2000) (extending work by MacLeod et al. (1995))
identified local changes in inversely computed epicardial electrograms in patients whose
data was accessed during coronary catheterization, preceeding and following angioplasty
balloon catheter inflation. In the eighteen study patients, the predicted region of ischemia
following balloon inflation correlated with the expected region of perfusion deficit based
on the vessel occluded.
4.5 ENDOCARDIAL POTENTIAL IMAGING
Interventional cardiologists employ transvenous catheter procedures to treat arrhyth
mogenic foci and aberrant conduction pathways. Such treatment first requires mapping the
endocardial potential. This initial invasive imaging is cumbersome, tedious, and lengthy.
Typically, a roving probing trasvenous electrode catheter is brought into contact with many
endocardial locations over the course of many heartbeats, and a depiction of an endocardial
activation map is thereby inferred (an improved technology along these lines is given by
Gepstein et ai. (1996)). The number of sites accessed is limited, and there is no accounting
for beattobeat activation variability when reconstructing the maps from the many beats.
Recently introduced expandible basket electrode arrays (Schmitt et ai. , 1999) have their
own problems related to limited numbers of electrodes, the need to contact (and perhaps
148 F. Greensite
irritate) the endocardium, and the possibility of difficulties in collapsing the basket at the
end of the acquisition.
These problems can be potentially addressed by the use ofa transvenous catheter whose
tip is studded with multiple electrodes, and which is placed somewhere in the midst of a
cardiac chamber (without contacting the endocardium). Once the catheter location relative
to the endocardium is registered, it becomes theoretically possible to inversely compute
the endocardial potentials from a single heartbeatindeed, to follow dynamic isopotential
maps within a single beat, as well as beattobeat changes in activation maps. For this
inverse problem, the volume is bounded by the endocardial surface and the multielectrode
probe surface. Laplace's equation holds in this volume, and the boundary conditions are the
(unknown) endocardial potentials, and the zero normal current density at the multielectrode
probe surface. As in Section 4.2, a linear relationship is derived between the endocardial
potentials and the catheter electrode potentials.
Notwithstanding the inconvenience of the required cardiac catheterization, there are
two very significant advantages of this formulation over the technique of imaging the
epicardial potentials from the body surface. First, the electrodes are relatively close to all
portions of the surface to be imaged (e.g., as opposed to the distance between body surface
electrodes and the posterior wall of the heart). Second, the relevant volume is composed only
of blood in the lumen of the cardiac chamber. Therefore, the geometric modeling required
for estimation of the transfer matrix is vastly less, and the uncertainties in the values of key
components of the model (i.e., tissue conductivities) are markedly diminished (the blood
has uniform isotropic conductivity).
The initial proposal and work on a multielectrode noncontact array, placed in a cardiac
chamber for purposes of accessing endocardial potentials, was due to Taccardi et al. (1987).
In the past few years there has been much significant work reported on successors to this
idea. For example, in experiments on dogs, Khoury et al. (1998) used a 128 electrode
catheter, inserted via a purse string suture in the left ventricular apex, and showed that
faithful renditions of endocardial activation, both with paced and spontaneous beats, was
possible by solving the inverse problem. Ischemic zones were also well defined. A spiral
catheter design has also been investigated (Jia et al., 2000).
An impressive series of experiments has been performed with a competing system,
developed by Endocardial Solutions, Inc. In addition to a 64electrode 7.5 ml inflatable bal
loon catheter, a second transvacular catheter is passed and dragged along the endocardium.
As it is dragged, a several kHz signal is passed between it and the electrode catheter, lo
calizing its position with respect to the electrode catheter. In this way, a rendition of the
endocardium with respect to the electrode catheter is produced. Following construction of
a "virtual endocardium" via a convex hull algorithm applied to the above anatomical data,
the inverse problem is then solved, generating several thousand "virtual electrograms" on
the virtual endocardium (figure 4.5 and figure 4.6).
The literature on this subject is growing rapidly, and we site only a few of examples.
Overall, very impressive utility and fidelity is being established. For example, a report by
Schilling et al. (2000) describes the classification of atrial fibrillation in humans in terms
of numbers of independent reentrant wavefronts identified. A report by Strickberger et al.
(2000) describes the successful ablation of fifteen instances of ventricular tachycardial
guided by this catheter system. A recent report by Paul et al. (2001) describes the utility of
the system in directing catheter ablative therapy in subjects with atrial arrhythmias refractory
to pharmacologic therapy.
Heart Surface Electrocardiographic Inverse Solutions 149
r
..
R
v,·r....._v.... ......./___.f'
J\ J '\ r
R
J
R "J""/\V''\jl \,/"''"'\
FIGURE 4.5. Surface ECG from lead I (ECGI), endocardial electrogram via a contact electrode (C), and inversely
reconstructed electrogram using input from a noncontact multielectrode probe in the atrium (R), with C and R
from the same location, in three patient s with atrial fibrillation, in the study of (Schilling et al., 2000, [From:
Schillin g, R. J., Kadish, H., Peters, N. S., Goldberger, J., Wyn Davies, D., 2000, Endocardial mappin g of atrial
fibrillation in the human right atrium using a noncontac t catheter, European Heart Journal . 21: 550564. Used
by permission of the publisher, WB Saunders .]
4.6 IMAGING FEATURES OF THE ACTIONPOTENTIAL
4.6.1 MYOCARDIAL ACTIVATIONIMAGING
Epicardial and endocardial potential imaging addresses the need to reconstruct some
thing that is currently accessed invasively, and is thus of evident interest. However, such
potentials are not themselves a clinical endpoint. Ultimately, clinicians are interested in the
action potentialor at least , feature s of the action potential. The most important features
of the action potential are activation time (time of arrival of phase zero at every location,
the aggragate of which globally describe conduction disturbances), phase zero amplitude
(reflecting ischemia), and action potential duration (reflecting refractory periods, potentially
associated with propensity for reentrant arrhythmias).
The marker for activation in an electrogram (a tracing of epicardial or endocardial po
tential at a given cardiac site) is the "intrinsic deflection"defined as the steepest downward
deflection of the electrogram. Recall that the source is the gradient of transmembrane poten
tial (e.g., Eq. (4.10» , and that during cardiac activation this is usually appreciably nonzero
only at the locus of points undergoing action potential phase O. Thi s locus is approximately
a surface (the interface between depolarized and nondepolarized muscle). Electrically, this
behaves approximately as a propagating surface of dipole moment density (a double layer).
There is a discontinuity of potential as the double layer is crossed. Ideally, as an extracellular
location is passed over by the activation wavefront, there will then be a sharp downward
deflection in the extracellular (electrogram) potentialthe intrinsic deflection. However,
150 F. Greensi te
FIGURE 4.6. Time sequential views of a portion of the "virtual endocardium" depiction of the atria in the
study of (Paul et al., 2001), showing isopotential maps at six successive times. Spreading endocardial activation
wavefronts can be appreciated (e.g., two of these collide in E and F) during an atrial reentrant tachycardia. See
the attached CD for color figure. [From: Paul, T., WindhagenMahnen , B., Kriebel, T., Bertram, H., Kaulitz, R..
Korte, T., Niehaus, M., and Tebbenj ohanns, 1., 200I, Atrial Reentrant Tachycardia After Surgery for Congenital
Heart Disease Endocardial Mapping and Radiofrequency Catheter Ablation Using a Novel, Noncontact Mapping
System, Circulation. 103: 22662271. Used by permission.]
the reality is that it is not infrequent that there is more than one reasonable candidate for the
intrinsic deflection within a given location's electrogram. Furthermore, the intrinsic deflec
tion is often rather lengthy, so the selection of a single activation time within the intrinsic
deflection is to some extent arbitrary (Ideker et al., 1989; Paul et al., 1990). The activation
time is presumably the inflection point of the deflection (which itself is poorly defined in
the noisy setting). To a large extent, these problems are inherent in the source formulation:
The epicardial (or endocardial) potential at a location actually reflects contributions from
elect rical activity at all surrounding locations , when in fact we desi re to resolve results of
the membrane function at a single locationi.e., the local action potential. In this section
we examine work done on imaging the myocardial activation feature of the action potential ,
rather than the epicardial potential.
Enthusiasm for immedi ately attacki ng the problem of reconstruct ing the transmem
brane potential (x), or its gradient, is tempered by recogniti on of a dimensionality prob
lem: our measurements are confined to a surface (of the body), while the source V . (G V
permiates a volume (the heart). Inherentl y, we are faced with a "projection" of the three
Heart Surface Electrocardiographic Inverse Solutions 151
dimensional source on the two dimensional body volume (a further exacerbation of the
already described illposedness of the problem). However, building on the work of Wilson
et at. (1933), Frank (1954) noted that the source during the QRS inteval was roughly a
double layer (i.e., a surface), which tends to mitigate the above dimensionality problem.
While Frank was interested in quantifying the inaccuracy of the single moving dipole model
of the heart via forward computations (rather than imaging the double layer), two decades
later Dotti (1974) made an interesting observation: Neglecting anisotropic conductivity of
the heart, assuming uniform action potential amplitude, and recognizing the fact that the
gradient of transmembrane potential propagates as a dipolar wavefront (double layer), he
noted that the source surface at any time is electrically equivalent (as regards points external
to the heart) to a double layer consisting of the portions of the endocardium and epicardium
already depolarized (figure 4.7).
This is a consequence of the wellknown fact from electrostatics that a closed uni
form double layer in an isotropic medium generates no external potential. This means that
one can derive a relationship between the body surface potential at a given time, and the
locus of points on the cardiac surface that have been activated. Thus, the dimensionality
problem resolves, and the surface of interest is actually fixed. Dotti presented a very small
scale two dimensional simulation illustrating this concept. Similar observations were made
independently by Salu (1978) a few years later.
However, the concept can be said to have been formally introduced in a more complete
engineering context by Cuppen and van Oosteroom in the early 1980s. They presented the
imaging equation as
¢(y, t) = iA(x, y)H(t  r(x»dS, (4.46)
where S is the composite of the endocardial and epicardial surfaces, and the action potential
(during the QRS interval) is modeled using the Heaviside function (zero for t 0,
unity for t 0). Thus, the action potential (figure 4.1) is taken to be the step function
a(x) b(x)H(t  r(x». (4.47)
The action potential amplitude b(x) is assumed to be constant over the ventricles, and is
subsumed into the transfer function A(x, y). The offset a(x) is also assumed constant, and
thus has no effect since Is A(x, y)dS 0 (i.e., a uniform closed double layer generates
no external potential). Note that in the absence of reentrant arrhythmias there is no repo
larization during the QRS interval, so the action potential can then be modeled as a step
function in that interval. Thus, Eq. (4.46) is fully consistent with Eq. (4.13) (dervied from
the bidomain). Using Eq. (4.46), one wishes to determine rex), the time that point x on the
surface surrounding the heart undergoes action potential phase zero. Note that the equation
is nonlinear.
Equation Eq. (4.46) achieves a superficially satisfying form upon integration over the
activation QRS interval,
f
¢(y, t)dt A(x, y)f H(t  r(x»dtdS
x
A(x, y)r(x)dS
x
. (4.48)
QRS 1s QRS 1s
152 F. Greensite
FIGURE 4.7. An electri cal double layer in an infinite homogeneous volume conduct or of infinite extent generates
potential at a point proportional to the solid angie subtended by the point and the double layer. The above diagram
depicts how ventricular intramural depolarization wavefronts generate potential equivalent to that generated by
"virtual" double layers on the epicardium and/or endocardium. [From: van Oosterom, A., 1987, Computing the
depolar ization sequence at the ventricular surface from body surface potentials, in: Pediatric and Fundamental
Electrocardiography , (1. Liebman, R. Plonsey, and Y. Rudy, eds.), Martinus Nijhoff, Zoetermeer, The Netherlands,
pp. 7589. Used by permissi on.]
Apparently, the imaging of myocardial activation is also a linear problem. But attempts
to solve Eq. (4.48) soon run up against the problem that the computed activation times
are entirely unrealisticbecause regularization schemes typically favor solution estimates
with lower norm (even with higher order Tikhonov regularization). Thus, the computed
Heart Surface Electrocardiographic Inverse Solutions 153
QRSinterval becomeshighlycontracted. Furthermore, there is the impressionthat one has
been wasteful of the temporallyresolved (dynamical) informationinherent in Eq. (4.46),
by integratingit all away in Eq. (4.48). This is unacceptablein an already very illposed
problem.HuiskampandvanOosterom(1988)addressedthis objectionablefeaturebyusing
the regularizedsolution to Eq. (4.48) as a seed for a quasiNewton routine for solving a
regularizedversionof the full nonlinear expressionEq. (4.46). As with the basic Newton
procedurefrom Calculus, which extracts the root of a nonlinear function nearest the seed,
the quasiNewton procedureapplied here is a means of finding a root (i.e., the appropriate
for aregularizedversion Is O.However,aswith
the basic Newtonmethod, one is dealing with an intrisicallylocal procedure that does not
performa global optimization. The solution estimate obtainedis highly influencedby the
initial seedfromEq. (4.48). On the other hand, thereis noreasonwhya global optimization
routine such as simulatedannealing, could not be used (in fact, this is proposed in a very
recent paperon activation timeandactionpotentialamplitudeimaging(Ohyuet al., 2002».
However, a furtherproblemis that Eq. (4.46) is validonlyunder the assumptionthat cardiac
musclehas isotropicconductivity, or satisfies equal anisotropy.
Adifferentapproachwas takenby Greensite(1994; 1995).The general idea is that the
myocardial surface activation function like (nominallydifferentiable) function,
is greatly characterized by its relative extremae.g., its relative maxima and minima.
Predominantly, these are the epicardial breakthrough points and activation sinks of the
transmuraldepolarization wavefront. Indeed, since (x) is definedover a compact domain
(the heart surface), and has a finite range (the QRS interval), knowledgeof these "critical
points" reduces the space of admissible solutions to that of a compact set of functions.
The problemof reconstructingthe rest of from Eq. (4.48) is nominally a wellposed
problem.In simpleterms,if therelativemaximaandminimaof r are known,theproblem
of determining the rest of becomes simply a matter of optimized interpolationfor
which the constraints embodiedby Eq. (4.46) should be sufficient. An efficientmeans for
computing the critical points was given in the Critical Point Theorem (Greensite, 1995).
Consider the "data operator"
1
QRS
In the practical setting, ¢ is a spacetimematrix, each of whose rows is the body surface
potential time series (ECG) at a particular electrode location. The Critical Point Theorem
states that is a critical point of if and only if is in the space spanned by
the eigenfunctions of ¢¢t. In fact, the Theoremholds even in the case of an anisotropic
myocardium. Complications ensue once noise is added to the formulation, but an effi
cient algorithmemployingthese ideas in a noisy context was proposed in (Huiskampand
Greensite, 1997).
Oostendorpet at. at the University of Nijmegen/University of Helsinki haveproduced
work evaluating the latter approach both in vitro (Oostendorp et aI., 1997) and in vivo
(Oostendorp and Pesola, 1998) (validation in hearts removedat the time of cardiac trans
plantation, figure 4.8).
Work on invasive validation of these latter ideas has also recently been undertaken
by a group at the Technical University of Graz, (Tilg et aI., 1999; Modre et aI., 2001a,
154 F. Greensite
FIGURE 4.8. One of a series of four hearts, removed at transplantation, in the study of (Oostendorp and Pesola,
1998). The two upper images of the anterior and posterior ventricular epicardium show the activation maps obtained
at the time of surgery (prior to cardiac transplantation) via application of an epicardial electrode sock (epicardial
electrode locations indicated by circles). The lower two images show the corresponding preoperative activation
map, inversely computed from body surface potential electrode data. [From: Oostendorp, T., and Pesola, K., 1998,
Noninvasive determination of the activation time sequence of the heart: validation by comparison with invasive
human data, Computers in Cardiology. 25:313316. Copyright IEEE. Used by permission].
Wach et ai., 2001; Tilg et ai., 2001; Modre et ai., 2001b), and a group at the University of
Auckland/University of Oxford (Pullan et ai.,
4.6.2 IMAGING OTHER FEATURES OF THE ACTIONPOTENTIAL
Let tl be a time during the TP interval of the ECGi.e., a time during which all
ventricular locations are in action potential phase 4 (fully repolarized). Let t: be a time
during the ST interval of the ECGi.e., a time during which all ventricular locations are
in phase 2 (fully depolarized). From Eq. (4.13) and the action potential model Eq. (4.47),
t2)  t1) = t2)  tl»dS
nxb(x)dS. (4.49)
Note that if b(x) is a constant, both sides of the above equation will be zero (e.g., a closed
uniform double layer generates no external potential). Indeed, the body surface potential
Heart Surface Electrocardiographic Inverse Solutions 155
during the TP and ST segments have the same value in healthy subjects. However, in the
case of cardiac ischemia, the action potential amplitude is spatially varying. In that setting,
one can imagine solving the above integral equation to obtain the spatiallyvarying action
potential amplitudeup to a spatial constant (the null space of the operator is the space
of constant functions). Since the phase 0 amplitude in healthy myocytes is already known
to be approximately 90 millivolts, one can then (in principle) image the action potential
amplitude fully.
Reflecting on the approach of Cuppen and van Oosterom (1984), Geselowitz (1985)
noted that it would be possible to image the area under the action potential (i.e., the integral
with respect to the baseline of the action potential) by simply extending the time interval of
integration in Eq. (4.48) to be the interval (encompassing the time period of activation and
repolarization). Thus,
JQRST JQRST Js
Js JQRST
n
where is the area under the action potential at
Now one can use knowledge of and to create an image of action potential
duration as Thus, one might anticipate imaging action potential attributes such
as action potential amplitude and action potential duration, in addition to phase 0 time
(activation imaging). Apparently, this joining of the prior two paragraphs has not been
investigated, and the practicality of such manipulations is speculative.
4.7 DISCUSSION
Among the many engineering challenges posed by the imaging problem treated in
this chapter, the necessity of a proper mathematical understanding of the computational
difficulties (and their optimal treatment) has some preeminence. In this regard, there
are lively controversies regarding which is the favored source formulation to be imaged
(epicardial/endocardial potentials versus action potential features), the possible role of pre
processing the raw signals (e.g., Laplacian electrocardiography), the reductionist role in
activation imaging (e.g., the Critical Point Theorem), and the desirability of integrating
the temporal data from a stochastic processes standpoint. Recent history has shown that
there is surely room for improvement in algorithmic technique. Methodological refinements
continue to be proposed by many different groups.
At the same time, the biophysical understanding, technical apparatus, and mathematical
methodology, are clearly already in place to create images of extracellular potential and
action potential features on the epicardial and endocardial surfaces. The principal question
is whether the resulting images are either too blurred to be of much use, or are otherwise
unreliable and misleading (e.g., due to the inherent illposedness of the problem, lack
F. Greensite
of sufficiently powerful mitigating constraints, or insufficiently accurate forward problem
solutions due to uncertainties in knowledge of body tissue conductivities and anisotropies).
Thus, image validation is presently a major question of interest in this field. Such validation is
fairly well advanced in the case ofthe minimally invasive techniques of imaging endocardial
potential via transvascular catheter probe electrode arrays. However, for the epicardial
imaging approaches, it is particularly difficult to address the validation question adequately,
because validation ideally requires the simultaneous acquisition of epicardial signals, body
surface signals, and anatomical body imaging (via CT or MRI). That is, the body imaging
should be conducted closed chest (otherwise, the body surface signals would be subject to
an unrealistic transfer matrix), despite the simultaneous need for "gold standard" invasively
obtained epicardial potentials for the validation. Nevertheless, the latter validation goal is
being aggressively pursued by a number of groups worldwide.
It is likely that the field is maturing to the extent that the next several years will see
clarification of the true potential and promise of the methodologies discussed in this chapter.
Accordingly, the era of Noninvasive Imaging of Cardiac Electrophysiology (NICE) could
soon be at hand.
REFERENCES
Ahmed, G. E, Brooks, D. H., and MacLeod, R. S., 1998, An admissible solution approach to inverse electrocar
diography, Ann. Biomed. Eng. 26:278292.
Barber M. and Fischman, E. J., 1961, Heart dipole regions and the measurement of dipole moment, Nature.
192:141142.
Barnard, A. C. L., Duck, J. M., Lynn, M. S., and Timlake, W. P., 1967, The application of electromagnetic theory
to electrocardiography, II, Biophys. 7:463491.
Barr, R. Ramsey, M., and Spach, M. S., 1977, Relating epicardial to body surface potential distributions by
means of transfer coefficients based on geometry measurements, IEEE Trans. Biomed. Eng. 24: III.
Barr, R., and Spach, M., 1978, Inverse calculation of T epicardial potentials from body surface potential
distributions for normal and ectopic beats in the intact dog, IEEE Trans. Biomed. Eng.
Basser, P., MattieIlo, J., and LeBihan, D., 1994, MR diffusion tensor spectroscopy and imaging, Biophys.
66:259267,1994.
Bellman, R., Collier, C, Kagiwada, H., Kalaba, and Se!vester, R., 1964, Estimation of heart parameters using
skin potential measurements, Comm. ACM. 7:666668.
Brooks, D. H., Ahmad, G., MacLeod, R. S., and Maratos, G. M., 1999, Inverse electrocardiography by simultaneous
imposition of multiple constraints, IEEE Trans. Biomed. Eng. BME46:318.
Burnes, J. E., Taccardi, B., MacLeod, S., and Rudy, Y, 2000, Noninvasive ECG imaging of electrophysiolog
cially abnormal substrates in infarcted hearts, a model study, Circulation. 101:533540.
Burnes, J. E., Taccardi, B., Ershler, R., and Rudy, Y, 2001, Noninvasive ECG imaging of substrate and intramural
ventricular tachycardia in infarcted hearts, 1. Am. Co!' Cardio!' in press.
ColliFranzone, P., Guerri, L., Tentoni, S., Viganotti, Baruffi, S., Spaggiari, S., and Taccardi, B., 1985, A
mathematical procedure for solving the inverse potential problem of electrocardiography. Analysis of the
timespace accuracy from in vitro experimental data, Math. Biosci. 77:353396.
Cuppen, and van Oosterom, A., 1984, Model studies with inversely calculated isochrones of ventricular depo
larization, IEEE Trans. Biomed. Eng. BME31:652659.
Dotti, D., 1974, A spacetime solution of the inverse problem, Adv. Cardiol., 10:231238.
Einthoven, W., 1912, The different forms of the human electrocardiogram and their signification. Lancet.
1912 :853861.
Foster, M., 1961, An application of the WienerKolmogorov smoothing theory to matrix inversion, SIAM.
9:387392.
Frank, E., 1954, The images surface of a homogeneous torso, Amer. Heart 1. 47:757768.
Heart Surface Electrocardiographic Inverse Solutions 157
Gabor, D., and Nelson, C 1954, Determi nation of the resultant dipole of the heart from measurements on the
body surface, Applied Physics. 25:41 3416.
Gepstein, Hayam, G., and BenHaim, S. A., A novel method for nonfluoroscopic catheterbased elec
troanatomical mapping of the heart: in Vitro and in vivo accuracy results, Circulation. 95:1611
1622.
Gelemter, H. and Swihart, J. C , 1964, A mathematicalphysical model of the genesis of the electrocardiogram,
Biophys. 4:285301.
Geselowitz, D. B., 1967, On bioelectric potenti als in an inhomogeneous volume conductor, Biophys. 7:11 1.
Geselowitz, D. B., 1985, Use of time integrals of the ECG to solve the inverse problem, IEEE Trans. Biomed. Eng.
BME·32:7375.
Ghanem, R. N., Burnes, 1. E., Waldo, A. and Rudy, R., 200 I, Imaging dispersion of myocardial repolarization.
II, Circulation, 104: 13061 312.
Golub. G., and van Loan, C , 1996, Matrix Computations, 3rd ed., Johns Hopkins University Press, Baltimore.
Greensite, E, 1994. Wellposed formul ation of the inverse problem of electrocardiography, Ann. Biomed. Eng.
22:172183.
Greensite , E, 1995. Remote reconstruction of confined wavefront propagation , Inverse Problems . 11:361 370.
Greensite, E, and Huiskamp, G., 1998, An improved method for estimating epicardial potenti als from the body
surface, IEEE Trans. Biomed. Eng.. BME45:17.
Greensite, E, 2001. Myocardial Activation Imaging, in: Computational Inverse Problems in Electrocardiography,
(P. Johnston, ed.), WIT press, Brisol, pp. 143190.
Greensite, E, 2002, A new treatment of the inverse problem of multivariate analysis, Inverse Problems. 18:
363379.
Gulrajani, R., Roberge, E, and Savard, P., 1984, Moving dipole inverse ECG and EEG solutions, IEEE Trans.
Biomed. Eng. BME31:903 910.
Gulrajani, R. M., Roberge, E A., and Savard, P., 1989, The inverse problem of electrocardiography, in: Compre
hensive Electrocardiology, Volume I ( P. Macfarlane, and T. Veitch Lawrie, eds. ), Pergamon, Oxford,
pp. 237288.
Gulrajani, R. M., 1998, The forward and inverse problems of electrocardiograprhy,lEEE Eng. Med. Bio!' 17:84
\0 1.
Hansen, P. C , 1992, Numerical tools for analysis and solution of Fredholm integral equations of the first kind,
Inverse Problems . 8:849872.
He, B., and Cohen, J., 1992, Body surface Laplacian ECG mapping, IEEE Trans. Biomed. Eng. 39:1179119 1.
He, B., and Wu, D., 1997, A bioelectric inverse imaging technique based on surface Laplacians, IEEE Trans.
Biomed. Eng. BME16: 133 138.
He, R., Rao. Liu, S., Yan, w., Narayana, P. A., and Brauer, H., 2000, The method of maximum mutual information
for biomedical electromagneti c inverse problems, IEEE Transaction on Magnetics. 36: 17411744.
He, B., and Wu, D., 2001, Imaging and visualization of 3D cardiac electric activity, IEEE Trans. Inf. Techno!'
Biomed.5:181186.
Henriquez, C , 1993, Simulating the electrical behavior of cardiac tissue using the bidomain model, Crit. Rev.
Biomed. Eng. 21:1 77.
Horacek, B. M., 1997, The inverse problem of electrocardiography: a solution in terms of single and doublelayer
sources on the epicardial surface, Math. Biosci. 144:119 154.
Huiskamp, G., and van Oosterom, A., 1988, The depolariz ation sequence of the human heart surface computed
from measured body surface potentials, IEEE Trans. Biomed. Eng. BME·35: 1047 1058.
Huiskamp, G., and van Oosterom, A., 1989, Tailored versus real istic geometry in the inverse problem of electro
cardiography, IEEE Trans. Biomed. Eng. BME36:827835.
Huiskamp, G., and Greensite, E, 1997, A new method for myocardial activation imaging, IEEE Trans. Biomed.
Eng. BME44: 433446.
Iakovidis,l., and Gulrajani, R. M., 1992, Improving Tikhonov regularization with linearlyconstrained opt imization:
appl ication to the inverse epicardial potential solution, Math. Biosci. , 112:55 80.
Ideker, R. E., Smith, M., Blanchard, S. M., Reiser, S. Simpson, E. V., Wolf, D., and Danieley,
N. D., 1989, The assumptions of isochronal cardiac mapping, PACE. 12:456478.
Jackson, J. D., 1975. Classical Electrodynamics, Wiley, New York.
Jia, P., Punske, B., Taccardi, B.. and Rudy, Y, 2000, Electrophys iologic endocardial mapping from a noncontact
nonexpandable catheter, Cardiovasc. Electrophysio!., 12381251 .
158 Greensite
Johnston, P. 1997, The Laplacian inverse problem of electrocardiography: an eccentric spheres study, IEEE
Trans. Biomed. Eng. 44:539548.
Johnson, 2001, Adaptive finite element and local regularization methods for the inverse problem of electro
cardiography, in: Computational Inverse Problems in Electrocardiography, (P. Johnston, ed.), WIT press,
Brisol, pp. 5188.
Joly, Goussard, Y, and Savard, P, 1993, Timerecursive solution the inverse problem of electrocardiography:
a modelbased appraoch, in: Proc. 15th Ann Int conf IEEE Eng. Med. Bio!. Soc., IEEE Press, New York, pp.
767768.
Kadish, A., Hauck, 1., Pederson, B., Beatty, G., and Gornick, C, 1999, Mapping of Atrial Activation With a
Noncontact, Multielectrode Catheter in Dogs, Circulation. 99:19061913.
Keener, 1., 1988, Principles ofApplied Mathematics, Addison Wesley, Redwood City, CA, pp. 135146.
Dirar S. Khoury, PhD; Keith Berrier, BS; Shamim M. Badruddin, MD; William A. Zoghbi,
Khoury, D. S., Berrier, K. Badruddin, S. M., and Zoghbi, W.A., 1998, ThreeDimensional Electrophysiological
Imaging of the Intact Canine Left Ventricle Using a Noncontact Multielectrode Cavitary Probe: Study of
Sinus, Paced, and Spontaneous Premature Beats, Circulation. 97:399409.
Leder, U., Pohl, H., Michaelson, S., Fritschi,T, Huck, M., Eichhorn, 1., Muller, S., and Nowak, H., 1998, Nonin
vasive biomagnetic imaging in coronary artery disease based on individual current density maps of the heart,
Int. Cardio!. 64:8392.
Li, G., and He, B., 2001, Localization of the site of origin of cardiac activation by means of a heartmodelbased
electrocardiographic imaging approach, IEEE Trans. Biomed. Eng. 48:660669.
Lynn, M. S., Barnard, A. C L., Holt, J. H., and Sheffield, L. T., 1967, A proposed method for the inverswe problem
in electrocardiography, Biophys. 7:925945.
MacLeod, S., Gardner, M., Miller, M., and Horacek, B. M., 1995, Application of an electrocardio
graphic inverse solution to localize ischemia during coronary angioplasty, Cardiovasc. Electrophys. 6:
218.
MacLeod, R. S., and Brooks, D. H., 1998, Recent progress in inverse problems of electrocardiograprhy, IEEE
Eng. Med. Bio!. 17:7383.
Malmivuo, J., and Plonsey, 1995, Bioelectromagnetism: principles and applications of bioelectric and bio
magneticfields, Oxford University Press, New York.
Martin, R. and Pilkington, T 1972, Unconstrained inverse electrocardiography: epicardial potentials, IEEE
Trans. Biomed. Eng. BME19:276285.
Martin, R. Pilkington, T C., and Morrow, M. N., 1975, Statistically constrained inverse electrocardiography,
IEEE Trans. Biomed. Eng. BME22:487492.
MessingerRapport, and Rudy, Y., 1988, Regularization of the inverse problem of electrocardiography: a
model study, Math. Biosci., 89:79.
Modre, R., Tilg, B., Fischer, G., and Wach, P., 2001, An iterative algorithm for myocardial activation time imaging,
Computer Methods and Programs in Biomedicine 64:17.
Modre, Tilg, B., Fischer, G., Hanser, E, Messnarz, B., Wach, P., Pachinger, Hintringer, E, Berger, T,
AbouHarb, M., Schoke, M., Kremser, C, and Roithinger, E, 2001, Stability of activation time imaging from
single beat data under clinical conditions, Biomedizinishe Technik 46:213215.
Nash, M. P, Bradley, C P., Cheng, Pullan, A. J., and Paterson, D. J., in press, An invivo experimental
computational framework for validating ECG inverse methods, Int!' Bioelectromagnetism.
Ohyu, S., Okamoto, and Kuriki, S., 2001, Use of the ventricular propagated excitation model in the magne
tocardiographic inverse problem for reconstruction of electrphysiological properties, IEEE Trans. Biomed.
Eng. in press.
Oostendorp, T, MacLeod, R., and van Oosterom, A., 1997, Noninvasive determination of the activation sequence
of the heart: validation with invasive data, Proc. 19th Annual Int. Con! IEEE EMBS, CDROM, 1997.
Oostendorp, T, and Pesola, 1998, Noninvasive determination of the activation time sequence of the heart:
validation by comparison with invasive human data, Computers in Cardiology. 25:313316.
Oster, H., and Rudy, 1992, The use of temporal information in the regularization of the inverse problem of
electrocardiography, IEEE Trans. Biomed. Eng. BME39:6575.
Oster, H. S., and Rudy, 1997a, Regional regularization of the electrocardiographic inverse problem: a model
study using spherical geometry, IEEE Trans. Biomed. Eng. 44:188199.
Oster, H., Taccardi, B., Lux, R., Ershler, P, and Rudy, 1997, Noninvasive electrocardiographic imaging,
Circulation. 96:10121024.
HeartSurfaceElectrocardiographic Inverse Solutions
Papoulis, A., 1984, Probability, Random Variables, and Stochastic Processes, McGrawHill, New York.
Paul, Moak, J. P., Morris, C., and Garson, A., 1990, Epicardial mapping: how to measure local activation,
PACE. 12:285292.
Paul, WindhagenMahnert, B., Kriebel, T., Bertram, H., Kaulitz, Korte, Niehaus, M., and Tebbenjohanns,
J., 2001, Atrial Reentrant Tachycardia After Surgery for Congenital Heart Disease Endocardial Mapping and
Radiofrequency Catheter Ablation Using a Novel, Noncontact Mapping System, Circulation. 103:22662271.
Penney, C. 1., Clements, J. C., and Horacek, B. M., 2000, Noninvasive imaging of epicardial electrograms during
controlled myocardial ischemia, Computers in Cardiology 2000.27:103106.
Plonsey, R., 1969, Bioelectric Phenomena, McGrawHill, New York.
Pullan, J., Cheng, L.K., Nash, M.P., Bradley, c.P., Paterson, DJ., 2001, Noninvasive electrical imaging of the
heart: theory and model development. Ann. Biomed. Eng. 29:817836.
Ramanathan, C., and Rudy, Y, 2001, Electrocardiographic Imaging: II. Effect of torso inhomogeneities on non
invasive reconstruction of epicardial potentials, electrograms, and isochrones. Cardiovasc. Electrophysiol.
12:242252.
Reese, Weisskoff, R., Smith, Rosen, B., Dinsmore, and Wedeen, 1995, Imaging myocardial fiber
architecture in vivo with magnetic resonance, Magnetic Resonance in Medicine. 34:786791.
Rudy, Y, and MessingerRapport, B. J., 1988, The inverse problem in electrocardiography: solutions in terms of
epicardial potentials, Crit. Rev. Biomed. Eng. 16:215268.
Salu, Y, 1978, Relating the multipole moments of the heart to activated parts of the epicardium and endocardium,
Ann. Biomed. Eng., 6:492505.
Schilling, R. J., Kadish, A. H., Peters, N. S., Goldberger, J., Wyn Davies, D., 2000, Endocardial mapping of
atrial fibrillation in the human right atrium using a noncontact catheter, European Heart Journal.
550564.
Schmitt, Zrenner, B., Schneider, M., Karch, M., Ndrepepa, G., Deisenhofer, Weyerbrock, S., Schreieck,
J., and Schoemig, A., 1999, Clinical experience with a novel multielectrode basket catheter in right atrial
tachycardias, Circulation. 99:24142422.
Schmitt, O. H., 1969, Biological information processing using the concept of interpenetrating domains, in Infor
mation Processing in the Nervous System, (Leibovic, K. N. ed.), SpingerVerlag, New York.
Shahidi, A. Savard, P., and Nadeau, R., 1994, Forward and inverse problems of electrocardiography: modeling
and recovery of epicardial potentials in humans, IEEE Trans. Biomed. Eng. 41:249256.
Strickberger, S. A., Knight, B. P., Michaud, G. E, Pelosi, E, and Morady, E, 2000, Mapping and ablation of
ventricular tachycardia guided by virtual electrograms using a noncontact, computerized mapping system. 1.
Am. Col. Cardiol. 35:414421.
Taccardi, B., Arisi, G., Macchi, E., Baruffi, S., and Spaggiari, S., 1987, A new intracavitary probe for
detecting the site of origin of ectopic ventricular beats during one cardiac cycle, Circulation. 75:
272281.
Throne, and Olsen, L., 1994, Ageneralized eigensystemapproach to the inverse problem of electrocardiography,
IEEE Trans. Biomed. Eng. 41:592600.
Throne, R. D., Olsen, L. G., 2000, A comparision of spatial regularization with zero and first order Tikhonov
regularization for the inverse problem of electrocardiography, Computers in Cardiology. 27:493496.
Tikhonov, A., and Arsenin, 1977, Solutions of IllPosed Problems, John Wiley and Sons, New York.
Tilg, B., Wach, E, SippensGroenwegen, A., Fischer, G., Modre, R., Roithinger, E Mlynash, M., Reddyuu, G.,
Roberts, T., Lesh, M., and Steiner, P., 1999, Closedchest validation of source imaging from human ECG
and MCG mapping data, in: Proceedings of the 21st Annual International Conference of the IEEE EMBS,
October 19991First Joint BMESIEMBS Conference, IEEE Press.
Tilg, B., Fischer, G., Modre, R., Hanser, E, Messnarz, B., Wach, P., Pachinger, Hintringer, E, Berger, T.,
AbouHarb, M., Schoke, M., Kremser, C., and Roithinger, E, 2001, Feasibility of activation time imaging
within the human atria and ventricles in the catheter laboratory, Biomedizinishe Technik 46:213215.
Tuch, D. S., Wedeen, V.J., Dale, A. M., and Belliveau, J. w., 1997, Conductivity maps of white matter fibertracts
using magnetic resonance diffusion tensor imaging, Proc. Third int. conf. On Fundamental Mapping of the
Human Brain, Neuroimage. 5:s44.
Twomey, S., 1963, On the numerical solution of Fredholm integral equations of the first kind by the inversion of
the linear system produced by quadrature, J. ACM, 10:97101.
Ueno, S., and Iriguchi, N., 1998, Impedance magnetic resonance imaging: a method for imaging of impedance
distribution based on magnetic resonance imaging, J. Appl. Phys. 83:64506452.
160 F. Greensite
van Oosterom, A., 1987, Computing the depolarization sequence at the ventricular surface from body surface
potentials, in: Pediatric and Fundamental Electrocardiography, (1. Liebman, Plonsey, and Y.Rudy, eds.),
Martinus Nijhoff, Zoetermeer, The Netherlands, pp. 7589.
van Oosterom, A., 1999, The use of the spatial covariance in computing pericardial potentials. IEEE Trans. Biomed.
Eng. 46:778787.
Velipasaoglu, E. P., Sun, H., Zhang, E, Berrier, K. and Khoury, D. S., 2000, Spatial regulariation of the
electrocardiographic inverse problem and its application to endocardial mapping, IEEE Trans. Biomed. Eng.
47:327337.
Wach, P., Modre, R., Tilg, B., Fischer, G., 2001, An iterative linearized optimization technique for nonlinear
illposed problems applied to cardiac activation time imaging, COMPEL 20:676688.
Waller, A., 1889, On the electromotive changes connected with the beat of the mammalian heart, and of the human
heart in particular, Phil. Trans. R. Soc. Lond. B. 180: 169194.
Waller, A., 1911, quoted in Cooper 1. K., 1987, Electrocardiography 100 years ago: origins, pioneers, and con
tributors, NEJM. 315:461464.
Wahba, G., 1977, Practical approximated solutions to linear operator equations when the data are noisy, SIAM 1.
Numer. Anal. 14:651667.
Wilson, EN., Macleod, A. G., and Barker, P. S., 1933, The distribution of the action currents produced by heart
muscle and other excitable tissues immersed in extensive conducting media, 1. Gen. Physiol. 16:423456.
Wilson, EN., Johnston, E D., and Kossmann, C. E., 1947, The substitution of the tetrahedron for the Einthoven
triangle. Am. Heart J., 33:594603.
Yamashita, Y., and Geselowitz, D., 1985, Sourcefield relationships for cardiac generators on the heart surface
based on their transfer coefficients, IEEE Trans. Biomed. Eng. BME32:964970.
Zablow, 1966, An equivalent cardiac generator which preserves topolgraphy, Biophys. 1. 6:535536.
5
THREEDIMENSIONAL
ELECTROCARDIOGRAPHIC
TOMOGRAPHIC IMAGING
BinHe*
University of Illinois at Chicago
5.1 INTRODUCTION
Cardiac electrical activity is distributed over the three dimensional (3D) myocardium.
is of significance to noninvasively image distributed cardiac electrical activity throughout
the 3Dvolume of the myocardium. Such knowledge of the source distribution would play an
important role in our effort to relate the electrocardiographic inverse solutions with regional
cardiac activity.
Historically, attempts to noninvasively obtain spatial information regarding cardiac
electrical activity started from body surface potential mapping by using a larger number
of recording leads covering the entire surface of the body (Taccardi, 1962). From such
measurements, instantaneous equipotential contour maps on the body surface have been
obtained and shown to provide additional information when compared to a conventional
electrocardiogram (See Flowers Horan, 1995 for review). Since body surface potential
maps (BSPMs) are manifestation of cardiac electrical sources on the body surface, efforts
have been made to solve the electrocardiography inverse problemto seek the generators
of BSPMs.
Equivalent dipole solutions have been investigated with the aim of extracting use
ful information regarding cardiac electrical activity. Such efforts included (l) Moving
Dipole Solutions (Mirvis et al., 1977; Savard et al., 1980; Okamoto et al., 1983; Gulrajani
et al., 1984), in which one or more current dipoles are estimated at the location(s)
that best describe the body surface recorded electrocardiograms; and (2) Fixed Dipoles
Present address for correspondence: University of Minnesota, Department of Biomedical Engineering, 7105
BSBE, 312 Church Street, Minneapolis, MN 55455 Email: binhe@umn.edu
161
162 B.He
Solutions (Barber Fischman, 1961; Bellman et aI., 1964; He Wu, 2001), in which an
array of dipoles are arranged at fixed locations where their moments are determined by
minimizing the difference between the modelgenerated and the measured body surface
electrocardiograms.
It has been demonstrated that the single moving dipole solution can provide a good
representation of welllocalized cardiac electrical activity (Savard et aI., 1980). Efforts have
also been made to estimate two moving dipole solutions, although technical challenges exist
when the number of equivalent dipoles increases from one to two (Okamoto et aI., 1983;
Gulrajani et al., 1984). Due to the illposedness of the inverse problem, currently there is no
well established method to estimate three or more moving dipoles. In addition to equiva
lent dipole approach, equivalent multipole models have also been investigated (Geselowitz
1960; Hlavin & Plonsey, 1963; Pilkington & Morrow, 1982) during the early stage of elec
trocardiography inverse solutions, in an attempt to obtain equivalent 3D information on
cardiac electrical activity. The limitation of the multipole approach, however, is its inability
of localizing cardiac electrical activity.
In the past decade, most research on the electrocardiography inverse problem has been
carried out in the line of heart surface inverse solutions. As reviewed in Chapter 4, these
research efforts are mainly related to epicardial potential inverse solutions or heart surface
activation imaging.
Three dimensional electrocardiographic tomographic imaging has received much at
tention since 2000. He and Wu reported their effort on electrocardiographic tomography in
their presentation at the World Congress on Medical Physics and Biomedical Engineering
held in Chicago in 2000. In this work, He Wu demonstrated the feasibility in a computer
simulation study to image the 3D distribution of cardiac dipole source distribution from
noninvasive body surface electrograms by using the Laplacian weighted minimum norm
approach (He Wu, 2000, 2001). In subsequent work, He and coworkers developed a
heartmodel based 3D activation imaging approach (He Li, 2002; He et al., 2002), and a
3D transmembrane potential (TMP) imaging approach (He et al., 2003), which were in
troduced in a presentation at the 4th International Conference of Bioelectromagnetism in
2002 (He Li, 2002). Ohyu et al. (2002) have developed an approach to estimate the ac
tivation time and approximate amplitude of the TMP from magnetocardiograms using the
Wiener estimation technique. Skipa et al. presented their effort to estimate transmembrane
potentials from body surface electrocardiograms at the 4th International Conference on
Bioelectromagnetism (2002).
In this Chapter, we review the principles and methods of performing 3D electrocardio
graphic tomographic imaging, with a focus on introducing the recently developed distributed
3D electrocardiographic tomographic imaging techniques.
5.2 THREEDIMENSIONAL MYOCARDIAL DIPOLE SOURCE IMAGING
5.2.1 EQUIVALENT MOVING DIPOLE MODEL
Equivalent dipole inverse solutions were among the earliest efforts ofobtaining electro
cardiography inverse solutions. Early efforts were made on moving dipole inverse solutions
ThreeDimensional Electrocardiographic Tomographic Imaging 163
where one or two equivalent dipoles were used to represent cardiac electrical activity in the
sense that the dipolegenerated body surface potential maps (BSPMs) matches the measured
BSPMs well (Mirvis et aI., 1977; Savard et al., 1980; Okamoto et al., 1983; Gulrajani et aI.,
1984). In the moving dipole model, the locations of the equivalent dipoles vary from time
to time which provide information on the centers of gravity of electrical activity within the
heart. Such location information offers an important capability for moving dipole solutions
to localize the regions of myocardial tissues which are most responsible for the measured
BSPMs.
A limitation of this approach, however, is that the inverse solution is sensitive to
measurement noise and thus limiting the number of moving dipoles that can be reliably
estimated from the measured BSPMs. For this reason, the moving dipole inverse solution
may be useful in localizing a focal cardiac source during the initial phase ofcardiac activation
for a single activity. For general cardiac activation, the moving dipole inverse solution fails
to represent the complex cardiac electrical activity. A detailed review on an equivalent
moving dipole solution can be found in reference (Gulrajani et al., 1984).
5.2.2 EQUIVALENT DIPOLE DISTRIBUTION MODEL
As early as the 1960's, Barber and Fischman suggested the possibility of modeling
cardiac electrical activity using an array of current dipoles located at fixed locations within
the myocardium (Barber Fischman 1961; Bellman et aI., 1964). In this model, the dipoles
are not moving butfixedover time, while its moments remain variable. Yet this model did not
receive much attention in the field of electrocardiography inverse problem in the past three
decades, partially due to the dominance of the epicardial potential (Barr et aI., 1977; Frazone
et aI., 1978; Shahidi et aI., 1994; Throne Olson, 1994, 1997; Johnston Gulrajani, 1997;
Oster et aI., 1997; He Wu, 1997; Greensite Huiskamp, 1998; Burnes et al., 2000) and
the heartsurface activation time (Cuppen and van Oosterom 1984; Huiskamp Greensite
1997; Greensite 2001; Modre et al., 2001; Pullan et al., 2001) inverse solutions developed
during the same period.
Recently, the fixed dipole array model has been expanded into a volume distribution
of current dipoles for the purpose of tomographic dipole source imaging (He & Wu, 2000,
2001). In this work, He Wu modeled cardiac electrical sources by means of a large number
of current dipoles over the 3D volume of the ventricles. Each of the dipoles was located
at a particular position, representing the local electrical activity, while the moments of the
dipoles were varied over time. The magnitude function of the regional dipoles provided
a spatial distribution of current strength with the 3D myocardial volume. Estimation of
this current dipole moment provides a means of imaging the spatial distribution of current
sources.
5.2.3 INVERSE ESTIMATION OF 3D DIPOLE DISTRIBUTION
The key hypothesis under the 3Ddipole distribution imaging is based on the assumption
that the electrical sources located at a small region of myocardial tissue are coherent and
164 B.He
can be approximated by a current dipole. By assigning one such current dipole to each
"small" region of the myocardium, the following mathematical model, which relates the
current dipole distribution inside the myocardium to the body surface ECG measurements,
can be obtained:
V=AX (5.1)
where is the vector consisting of m body surfacerecorded ECG signals, X is the un
known vector consisting of the moments of the current dipoles, which are located at n sites
covering the entire myocardial volume, and A is the transfer matrix. The measurement at
each electrode sensor is produced by a linear combination of all dipole components, with
columns in A serve as weighting factors. By solving (5.1), one obtains an estimation of
3D current dipole source distribution corresponding to each measured BSPM. Since the
number of measurement electrodes is always far less than the dimension of the unknown
dipole source vector X, this problem is an underdetermined inverse problem and a proper
regularization strategy is necessary for obtaining a reasonable solution to (5.1).
The minimum norm (MN) solution is one of the feasible solutions (Hamalainen
Ilmoniemi, 1984)
(5.2)
where (*)+ denotes the MoorePenrose inverse. As the minimum norm solution is intrinsi
cally biased towards the superficial position, the weighted minimum norm solution (Jeffs
et aI., 1987) and the Laplacian weighted minimumnorm solution (LWMN) (PascualMarqui
et aI., 1994) has been proposed to solve the linear inverse problem. He Wu investigated
3D electrocardiography dipole source imaging using the principles of LWMN (He Wu,
2000,2001). LWMN utilizes a weighting operator LW, where L is a Laplacian operator and
W is a diagonal 3n by 3n matrix with Wii Ai and Ai is the zth column of the transfer
matrix Assuming the weighting factor is nonsingular, then
(5.3)
and the LWMN solution of (5.1) becomes
(5.4)
When using LWMN the resulting solution tends to be oversmoothed due to the con
straints of minimizing the Laplacian of the signal. For wellfocused cardiac sources, such
as the sites of origins of cardiac arrhythmias, a recursive weighting strategy, which was
previously developed for improving the performance of MN MEG imaging (Gorodnitsky
et aI., 1995), has been used to search for focal sources in the heart from initial LWMN
estimates. This algorithm recursively enhances the values of some of the initial solution
elements, while decreasing the rest of the elements until they become zero. In the end, only
a small number of winning elements remain nonzero, yielding the desired type of localized
energy distribution of the solution.
ThreeDimensional Electrocardiographic Tomographic Imaging
5.2.4 NUMERICAL EXAMPLE OF 3D MYOCARDIAL DIPOLE
SOURCE IMAGING
165
Computer simulation results for 3D myocardial dipole source imaging have recently
been reported (He Wu, 2001). A 3D hearttorso inhomogeneous volume conductor model
(Wu et aI., 1999) was used in the simulation. Considering the low conductivity of the lungs ,
the conductivity ratio of torso to lungs was set to 1:0.2; and the conductivity for myocardial
muscle was assumed to be the same as the torso (Mulmivuo Plonsey, 1995; Gulrajani ,
1998). The ventricles were divided into an equidistant lattice structure of 1,124 nodes
with a resolution of 6.7 mm. A regional current dipole was assigned on each of the nodes,
resulting in 1,124 regional dipoles.
Fig. 5.1illustrates an example of3D cardiac dipole source imaging. Twodipole sources ,
oriented from the waist towards the neck, were used to approxiate two localized cardiac
sources, which were located close to the endocardium at the right ventricle and the epi
cardium at the left ventricle (Fig. 5.l(a)). Gaussian white noise of 5% was added to the
body surface potentials calculated from assumed cardiac dipole sources to simulate noise
contamin ated body surface ECG measurements. The inverse imaging algorithm described
in Section 5.2.3 was used to attempt to reconstruct the source distribut ion within the my
ocardium, without a priori knowledge of the number of primary current dipoles.
The LWMN solution is illustrat ed in Fig. 5.1(b), where the red and yellow color s
illustrate the strength of the equivalent dipole source distribution throughout the ventricles.
Fig. 5.1(b) shows that the LWMN solution reached maxima in both the right ventricle and
the left ventricle, overlying with the locations of the source dipoles. The LWMN solution
showed a stronger source distribution over the left ventricle probably because the dipole
in the left ventricle is located closer to the chest when compared with the dipole located
in the right ventricle. In addition, Fig. 5.1(b) shows that there is another major area of
activity appearing over the posterior ventricular wall in the LWMN solution. The recursively
weighted LWMN solution is illustrated in Fig. 5.l (c) where after 20 iterations the source
strength distribution is well focused at two locations. One of the source localization results
was consistent with the "true" dipole at the right ventricle , while the other was located at
the left ventricle, but shifted about 1 em towards the direction of the endocardium from the
"true" diople position.
Previous studies have shown that the equivalent dipole solution suffers from existing
experimental noise if the number of the moving dipoles increases to two or more (Okamoto
et al., 1983). In a clinical setting, however, it is necessary to localize and image sites of
origins of arrhythmias even without knowing in advance how many dipoles should be used.
Hence, there is a need to develop a technique, which can localize and image sites of origins of
cardiac arrhythmias without a priori constraints on the number of equivalent moving dipoles
(such as one dipole). The LWMN approach which He Wu (2000, 2001) have applied to
cardiac dipole source imaging, on the other hand, does not attempt to make assumptions
on the number of focal cardiac sources. Some of the a priori information being taken into
account in the 3D cardiac dipole source imaging approach is that the myocardial electrical
activation is smooth over a reasonably small region. With such constraints , the estimated
inverse solution provides a smoothed distribution of current density over a large area of
myocardium (Fig. 5.1(b)). For focal sources, as illustrated in Fig. 5.1(a), additional strategy
S
i
m
u
l
a
t
e
d
S
o
u
r
c
e
s
(
a
)
L
W
M
N
S
o
l
u
t
i
o
n
(
b
)
R
e
c
u
r
s
i
v
e
S
o
l
u
t
i
o
n
(
c
)
F
I
G
U
R
E
A
n
u
m
e
r
i
c
a
l
e
x
a
m
p
l
e
o
f
3
D
c
a
r
d
i
a
c
d
i
p
o
l
e
s
o
u
r
c
e
i
m
a
g
i
n
g
.
(
a
)
T
w
o
d
i
p
o
l
e
s
o
u
r
c
e
s
(
r
e
d
d
o
t
s
)
,
o
r
i
e
n
t
e
d
f
r
o
m
t
h
e
w
a
i
s
t
t
o
w
a
r
d
s
t
h
e
n
e
c
k
,
w
e
r
e
u
s
e
d
t
o
a
p
p
r
o
x
i
a
t
e
t
w
o
l
o
c
a
l
i
z
e
d
c
a
r
d
i
a
c
s
o
u
r
c
e
s
,
l
o
c
a
t
e
d
c
l
o
s
e
t
o
t
h
e
e
n
d
o
c
a
r
d
i
u
m
a
t
t
h
e
r
i
g
h
t
v
e
n
t
r
i
c
l
e
a
n
d
t
h
e
e
p
i
c
a
r
d
i
u
m
a
t
t
h
e
l
e
f
t
v
e
n
t
r
i
c
l
e
.
G
a
u
s
s
i
a
n
w
h
i
t
e
n
o
i
s
e
o
f
5
%
w
a
s
a
d
d
e
d
t
o
t
h
e
b
o
d
y
s
u
r
f
a
c
e
p
o
t
e
n
t
i
a
l
s
t
o
s
i
m
u
l
a
t
e
n
o
i
s
e

c
o
n
t
a
m
i
n
a
t
e
d
b
o
d
y
s
u
r
f
a
c
e
E
C
G
m
e
a
s
u
r
e
m
e
n
t
s
.
(
b
)
T
h
e
L
W
M
N
s
o
l
u
t
i
o
n
s
h
o
w
s
s
m
o
o
t
h
e
d
d
i
s
t
r
u
b
u
t
i
o
n
o
f
c
u
r
r
e
n
t
d
e
n
s
i
t
y
d
i
s
t
r
i
b
u
t
i
o
n
w
i
t
h
i
n
t
h
e
m
y
o
c
a
r
d
i
u
m
.
w
h
e
r
e
t
h
e
r
e
d
a
n
d
y
e
l
l
o
w
c
o
l
o
r
s
i
l
l
u
s
t
r
a
t
e
t
h
e
s
t
r
e
n
g
t
h
o
f
t
h
e
e
q
u
i
v
a
l
e
n
t
d
i
p
o
l
e
s
o
u
r
c
e
d
i
s
t
r
i
b
u
t
i
o
n
.
(
c
)
T
h
e
r
e
c
u
r
s
i
v
e
l
y
w
e
i
g
h
t
e
d
L
W
M
N
s
o
l
u
t
i
o
n
s
h
o
w
s
w
e
l
l
f
o
c
u
s
e
d
s
o
u
r
c
e
s
t
r
e
n
g
t
h
d
i
s
t
r
i
b
u
t
i
o
n
w
h
i
c
h
c
o
r
r
e
s
p
o
n
d
w
e
l
l
w
i
t
h
t
h
e
t
w
o
"
t
r
u
e
"
d
i
p
o
l
e
s
o
u
r
c
e
s
a
s
s
h
o
w
n
i
n
(
a
)
.
S
e
e
t
h
e
a
t
t
a
c
h
e
d
C
D
f
o
r
c
o
l
o
r
f
i
g
u
r
e
.
(
f
r
o
m
H
e
W
u
,
2
0
0
1
w
i
t
h
p
e
r
m
i
s
s
i
o
n
)
@
I
E
E
E
ThreeDimensional Electrocardiographic Tomographic Imaging 167
such as recursive focusing is needed to obtain the inverse solution, in which it is as
sumed that the sites of origins of cardiac arrhythmias are localized over small regions
inside the myocardium. With this additional constraint, the inverse dipole source distri
bution shows localized distribution of current density close to the original "true" dipole
sources.
Note, however, that the LWMN inverse solution with a recursive weighting strat
egy still shows certain shift towards the "interior" of the myocardium from the "true"
solution positions (Fig. 5.1). Both the LWMN algorithm and the recursive weighting algo
rithm may contribute to such "shift." Although the work reported by He Wu suggests
the promise of imaging cardiac electrical activity using LWMN approach, a systematic
study should be conducted to evaluate the reconstruction results for a number of source
configurations including sources located in various regions of the heart with various orien
tations.
5.3 THREEDIMENSIONAL MYOCARDIAL ACTIVATIONIMAGING
Myocardial activation imaging received much attention in recent years, in which local
activation time over the heart surface is estimated from BSPMs (Cuppen and van Oosterom
1984, Huiskamp and Greensite 1997, Greensite 2001, Modre et al., 2001, Pullan et al
2001). As reviewed in Chapter 4, this approach is based on the bidomain theory, which
allows direct linking of the heart surface activation time with body surface potentials under
the assumption of the electrical isotropy (or "equal anisotropy") within the myocardium
(Greensite 2001).
Recently, the concept of myocardial activation imaging has been extended from 2D
heart surface to 3D myocardial volume (He et al., 2002; He Li, 2002; Ohyu et al., 2002).
In these approaches, the activation time throughout the 3D myocardium is estimated from
body surface electrograms by means of a heartexcitationmodel (He et al., 2002; He Li,
2002) or a Wiener inverse filter (Ohyu et al., 2002). In this Section, the heartmodelbased
3D activation imaging approach (He et al., 2002) is presented.
5.3.1 OUTLlNE OF THE HEARTMODEL BASED 3D ACTlVATION TIME
IMAGING APPROACH
The 3D distribution of activation time throughout the ventricles has been estimated
with the aid of a heartmodel based approach, in which the a priori knowledge on cardiac
electrophysiology is embedded. Fig. 5.2 illustrates a schematic diagram of this approach. A
realistic geometry computer hearttorso model is used to represent the relationship between
3D activation sequences within the myocardium with BSPMs. The a priori knowledge on
cardiac electrophysiology and the detailed anatomic information on the heart and torso are
embedded in this hearttorso model. The 3D myocardial activation sequence is estimated
as the parameter of the heartmodel and obtained by means of a nonlinear estimation
procedure.
168 B.He
FIGURE 5.2. Schematic diagram of 3D electrocardiography tomographic imaging. See attached CD for color
figure (From He et aI., Phys Med BioI, 2002 with permission)
A preliminary classification system (PCS) is employed to determine the cardiac sta
tus based on the a priori knowledge of the cardiac electrophysiology and the measured
BSPM by means of an artificial neural network (ANN) (Li He, 2001). The output of the
ANN based PCS provides the initial estimate of heart model parameters to be used later
in a nonlinear optimization system. Using these initial parameters as the result of PCS, the
optimization system then minimizes objective functions that assess the dissimilarity be
tween the measured and hearttorsomodelcalculated BSPMs. The heart model parameters
corresponding with the calculated BSPM are employed to produce 3D myocardial activa
tion sequence if the measured BSPM and the hearttorsomodelcalculated BSPM matches
well. Before the objective functions satisfy the given convergent criteria, the heart model
parameters are adjusted with the aid of the optimization algorithms and the optimization
procedure proceeds.
5.3.2 COMPUTER HEART EXCITATION MODEL
Numerous efforts have been made to develop computer heart models that can simulate
cardiac electrophysiological processes as well as the relationship between cardiac activities
and BSPMs (See Chapter 2 for review). Although the more detailed information incorpo
rated the better, the cellular automaton heart excitation model (Aoki et al., 1987; Lu et al.,
ThreeDimensional Electrocardiographic Tomographic Imaging 169
1993) has been used in the 3D activation time imaging research due to its capability of
simulating cardiac activation, BSPMs, and computational efficiency.
In this work (He et aI., 2002), we used a cellular automaton ventricle model that was
constructed as a 3D array of approximately 42,000 myocardial cell units with a spatial
resolution of 1.5 mm. The ventricles consisted of 50 layers with interlayer distance of
being 1.5 mm, and were divided into 53 myocardial segments. Each segment is comprised
of approximately the same number of myocardial cell units. The action potential of each of
heart units was already determined according to the cardiac action potential experimentally
observed and stored in the action potential data file. From the epicardiumto the endocardium,
the refractory period of the action potential of cardiac cellular units gradually increased for
the Twave simulation. The primary current dipole sources are proportional to the gradient
of the transmembrane potentials at adjacent cardiac units (Miller Geselowitz, 1978).
The anisotropic propagation of excitation in the ventricular myocardium was incorpo
rated into this heart model (He et aI., 2003) in order to obtain more accurate simulation of
the body surface ECG and myocardial activation sequence (Nenonen et al., 1991, Lorange
and Gulrajani 1993; Wei et al., 1995; Franzone et al., 1998; Huiskamp, 1998; Fischer et aI.,
2000). Ventricular myocardium was divided into different layers with thickness of 1.5 mm
from epicardium to endocardium. The myocardial fiber orientations were rotated counter
clockwise over 120
0
from the outermost layer (epicardium, 60°) to the innermost layer
(endocardium, +60°) (Streeter et aI., 1969) with identical increment between the consecu
tive layers. All units on a myocardial layer of ventricles from epicardial layer to endocardial
layer had identical fiber orientation. For each myocardial unit, a fiber direction vector, which
is located on its local tangential plane, was determined by its fiber angle. The fiber orien
tations of all myocardial units of ventricles were determined, and put in the realistically
shaped inhomogeneous torso model for calculating the body surface ECG. Excitation con
duction velocity of myocardial units was set to 0.6 m/s and 0.2 m/s along the longitudinal
and transverse fiber direction, respectively. Electrical conductivity of myocardial units was
set to 1.5 mS/cm along the longitudinal fiber direction and 0.5 mS/cm along the transverse
fiber direction (Nenonen et aI., 1991).
Fig. 5.3 shows the realistic geometry inhomogeneous hearttorso model (a), an example
of simulated sinus rhythm (b), and an example of paced activity (c). Fig. 5.3(b) shows the
activation sequence corresponding to sinus rhythm (left), and an example of the anterior
BSPM and a chest ECG lead simulated during sinus rhythm (right). Fig. 5.3(c) shows an
example of the simulated BSPM on anterior chest (middlebottom) at 30 ms following
pacing the anterior wall of the ventricle, and the ventricular excitation sequence over the
epicardium (middletop) corresponding to the pacing site (left), and a chest ECG lead (right).
The pacing site is shown on the left.
5.3.3 PRELIMINARY CLASSIFICATION SYSTEM
There are a large number of parameters associated with the cellular automaton heart
model that need to be determined in order to estimate the activation sequence. A Preliminary
Classification System (PCS) is used to approximately classify, from BSPMs, cardiac status
with the a priori knowledge on cardiac electrophysiology and the mapping relationship
between cardiac activation and resulting BSPMs. An ANN has been used to serve as a PCS.
In this implementation (Li He, 2001), a threelayer feedforward ANN was used, with the
170
Sin
~
Pa in
B.He
( a)
( c)
FI GURE 5.3. Illustration of computer hearttorso modeling and simulation. (a) Realistic geometry heart torso
model. (b) Simulation of sinus rhythm: Left panelIntracardiac activation sequence over three slices within
the ventricles; Right panelAn example of simulated anterior BSPM and chest ECG lead. (c) Simulation of
epicardial pacing: Left panel Heart model and a pacing site at middle anterior epicardium of ventricle. Middle
panels Anterior view of simulated epicardial isochrone (top), and an example of the BSPM over the anterior
chest following pacing (bottom). Righ panelA simulated chest ECG lead. See attached CD for color figure.
(Modified from He et aI., Phys Med BioI, 2002 with permission)
number of neurons in the input layer being set to the number of body surface electrodes, and
the number of neurons in the output layer being set to the number of myocardial segments
being studied. Gaussian white noise (GWN) was added to the BSPMs to simulate noise
cont aminat ed body surface ECG measurements . The BSPM maps during 25 to 50 ms after
initial activation were used as input s to train the ANN.
5.3.4 NONLINEAR OPTIMIZATION SYSTEM
The heart model parameters associated with myocardial activation sequence are esti
mated by minimizing dissimilarity between the measured and hearttorsomodelgenerated
(referred to as "simulated" below) BSPMs. The dissimilarity between the measured and
simulated BSPMs is described by appropri ate characteristic parameters extracted from the
BSPMs. The following three objecti ve functions (Li He, 2001) have been used to re
flect the dissimilarity between the measured and simulated BSPMs for paced ventricular
activation:
(a) which was constructed with the average correlation coefficient (CC) between
the measured and simulated BSPMs from instant T to instant T2 of the cardiac excitat ion
after detection of initial activation, is defined as:
z
L
2
 T
J
) (5.5)
ThreeDimensional Electrocardiographic Tomographic Imaging
where t) is the CC between the measured and simulated BSPMs at instant t. x is a
parameter vector of the spatial location of initial activation in the computer heartexcitation
model.
(b) which was constructed with the deviation of the positions of minima of
the measured and simulated BSPMs from instant to instant
2
, is defined as:
T2
L r:»,
t=T,
(5.6)
where and represent the positions of the minima in the measured and
simulated BSPMs at instant t, respectively. The definition of x is the same as that in Eq.
(5.5).
(c) which was constructed with the relative error of the number of body
surface recording leads, at which the potentials are less than a certain negative threshold,
in the measured and simulated BSPMs from instant to instant is defined as:
(5.7)
where i» and are the
numbers of recording leads, at which the potentials are less than a given threshold
0), in the measured and simulated BSPMs at instant respectively. and are
the ithlead measured and simulated potentials at instant t, respectively. u(e) is the unitstep
function, which gives a unity output if the potential at a lead is less than the preset threshold.
N is the number of body surface recording leads. The definition of x is the same as that in
Eq. (5.5).
Combining the above objective functions, the mathematical model of the optimization
for this heartmodelbased electrocardiographic imaging can be represented as the following
minimization problem:
£minp,
XEX
(5.8)
where X is the probable value region of the parameters in the computer heartexcitation
model. is a vector of heart model parameters. is the optimal value of the objective
function and are the allowable errors of the objective function
and respectively. Eq. (5.8) was solved by means of the Simplex Method.
5.3.5 COMPUTER SIMULATION
The feasibility of 3D myocardial activation imaging has been suggested in a computer
simulation study (He et al., 2002), and presented in this section. In this simulation study,
pacing protocols were used to simulate paced cardiac activation. By setting pacing sites
in different myocardial regions of the heart excitation model, sequential pace maps were
obtained by solving the forward problem using the hearttorso computer model. Two pacing
protocols, singlesite pacing and dualsite pacing, were used to evaluate the performance of
172 B. He
the 3Dmyocardial activation imaging approach. Gaussian white noise (GWN) of 10 !.LV was
added to the BSPMs at each time instant after the onset of pacing, to simulate the noise
contaminated body surface potential measurements. The maximum value of the BSPM
during the QRS compl ex was set to 3 mV.
The performance of activation time imaging was tested by singlesite pacing in 24
different sites throughout the ventricles. The CC and RE between the vector of simulated
activation time and the vector of estimated activation time were calculated for each of the
24 pacing sites. The vector of activation times consists of the activation time of each voxel
within the ventricles. Averaged over all 24 sites, the RE and CC between the "true" and
estimated activation times are 0.07 0.03 and 0.9989 0.0008, respectively, suggesting
the high degree of fidelity of the inverse estimation of activation time in the ventricles.
Fig. 5.4 shows two typical simulation examples. The top rows show the simulated
"true" activation sequence, and the bottom rows show the inversely estimated activation
sequence. Each row shows the activation sequence in 5 longitudinal sections «b)'""' (f))
and I transverse section of ventricles (a). Five horizontal lines in the transverse section
True
(A)
Estimated
(8) (b) (d) (e)
True
(B)
Estimated
(8) (b) (c) (d) (e)
FIGURE 5.4. Two examples of activation time imaging results during dualsite ventricular pacing. The activa
tion sequence within the ventricles was inversely estimated from the BSPM with V Gaussian white noise
being added. First row shows the simulated "true" activation sequence , and second row the activation sequence
correspondi ng to the inversely est imated result. Each row shows the isochrone in 5 longitudi nal sections «b) (f»
and I transverse section of the ventricles (a). Five horizontal black lines in the transverse section of the ventricles
from top to bottom respectively indicate the positions of 5 longitudinal sections from (b) to (f). The unit of color
bar is millisecond. The distance between both neighboring sections is 4.5 mm. (A) One pacing site at septum
endocardium of left ventricle, and another at intramural of leftanterior wall (marked by yellow dots). (B) Both
pacing sites at leftposterior intramural adjacent to endocardium (marked by yellow dots). See attached CD for
color figure. (From He et aI., Phys Med Bioi, 2002 with permissi on)
ThreeDimensional Electrocardiographic Tomographic Imaging
TABLE5.1. Effects of heart and torso geometry uncertainty on the
activation time imaging
Pacing Region BA BRW BLW BS BP Mean ± SD
NM 0.08 0.03 0.06 0.04 0.05 0.05 ± 0.02
LSX 0.08 0.03 0.06 0.10 0.06 0.06 ± 0.03
RSX 0.06 0.03 0.06 0.09 0.07 0.06 ± 0.02
FSY 0.08 0.03 0.07 0.10 0.06 0.05 ± 0.02
BSY 0.08 0.06 0.07 0.12 0.05 0.07 ± 0.03
NTM+ 10% 0.09 0.11 0.08 0.05 0.07 0.08 ± 0.02
NTM  10% 0.09 0.10 0.06 0.10 0.07 0.08 ± 0.02
Note:
NM: Normal Model.
LSXlRSX: Left/Right shift along the xdirection.
FSYIBSY: Front/Back shift along the ydirection.
NTM IO%/NMT 10%: 10% enlargement and reduction of the normal torso model.
173
of ventricles from top to bottom respectively indicate the positions of the 5 longitudinal
sections from (b) to (f). In panel (A), one pacing site is located at the left ventricular septal
endocardium, and another pacing site is located at the left anterior intramural wall. In panel
(B), both pacing sites are located at leftposterior intramural adjacent to endocardium. In
both cases, 10 f.LV GWN was added to the BSPMs to simulate noisecontaminated body
surface ECG recordings. Fig. 5.4 suggests that, for dual site pacing at two separate locations
(A) or adjacent locations (B), the 3D myocardial activation imaging can reconstruct well
the activation sequence, although the estimated activation sequence showed little delayed
activation as compared with the "true" activation sequence.
Effects of hearttorso geometry uncertainties were tested by selecting five pacing sites,
in five different regions adjacent to the AVring(BA: basalanterior; BRW: basalrightwall;
BP: basalposterior; BLW: basalleftwall; BS: basalseptum). The modified (enlarge or
reduce by 10%) torso models or positionshifted heart models (in 4 directions) were used
in the forward BSPM simulation, and 10 f.LV GWN was added to the simulated BSPMs. By
using the modified hearttorso models in the forward simulations while the standard model in
the inverse calculation, the effect of intersubject geometry variation was initially evaluated.
Table 5.1 shows the RE between the simulated "true" ventricular activation sequence and
the estimated activation sequence following a singlesite pacing. NM refers to normal case,
in which only measurement noise is introduced without geometry uncertainty. Note that the
hearttorso geometry uncertainty showed little effect on the activation sequence estimation
as determined by the RE measure. For example, 2% increase in RE was obtained for the
backward shift of the heart along the ydirection (BSY), as compared with the NM case. The
estimation errors associated with the 10% enlarged or reduced torso models have averaged
REof8%.
Effects of conduction velocity of ventricular activation in the heart model were assessed
by varying the conduction velocity in the forward heart model. The BSPMs were simulated
by using this altered ventricle model and noise was added to simulate noisecontaminated
BSPM measurements. The standard heart model, in which the average conduction velocity
was used, was then used to estimate the inverse solutions. Fig. 5.5 shows a typical simulation
example, with the same format as in Fig. 5.4. GWN of 10 f.LV was added to the BSPMs
B.He
(0 (e) (d) (c) (0) (a)
Troe
FIGURE 5.5. An example of activation time imaging results with variation in the conduction velocity. Same
format of display as in Fig. 5.5. The top rows showthe simulate d "true" activation sequence with altered conduction
velocity in the forward heart model, and the bottom rows show the inversely estimated activation sequence. The
results correspond to 10% increase in conduction velocity following a singlesite pacing. See attached CD for
color figure. (From He et aI., Phys Mcd Bioi, 2002 with permission)
to simulate noisecontaminated body surface ECG recordings. The top row shows the sim
ulated "true" activation sequence with altered conduction velocity in the forward heart
model, and the bottom row shows the inversely estimated activation sequence, following a
singlesite pacing when the conduction velocity of the forward heart model was increased
by 10%. When the average conduction velocity is different in the forward heart model when
compared with that in the heart model used in the inverse procedure, the estimated activation
time at specific regions within the ventricles differs from the original activation time dis
tribution in the forward solution. In particular, the early activation moved down toward the
apex direction. Nevertheless, the overall distributions of the activation time are not affected
substantially. The RE and CC between the "true" and estimated activation sequences within
the ventricles are 0.0916/0.106 and 0.996/0.998, respectively, corresponding to 5%110%
increase in the conduction velocity.
5.3.6 DISCUSSION
In this Section, a new approach for noninvasive 3D cardi ac activation time imaging
by means of a heartexcitationmodel is reviewed. This approach is based on the observa
tion that a priori information regarding cardiac electrophysiology should be incorporated
into the cardiac inverse solutions in order to obtain useful information on the 3D cardiac
activation from the twodimensional electrical measurements over the body surface . In this
approach, the a priori information on cardiac electrophysiology is incorporated into the
heartexcitationmodel, which is not an equivalent physical source model but an equivalent
physiological source model. By linking this physiological source model with body surface
ECG measurements, physiological parameters of interest are estimated from body surface
ECG recordings. A unique feature of such an approach is that the rich knowledge we have
gained in the forward whole heart modeling (See Chapters 2 and 3) can be directly applied
to the 3D cardiac imaging. Furthermore, the anisotropic nature of myocardial propagation
can also be incorporated in the 3D myocardial activation imaging, as shown in this Sec
tion. Such a priori electrophysiological information serves as constraints when solving the
inverse problem, leading to robust 3D inverse solutions.
Three Dimensional Electrocardiographic Tomographic Imaging 175
Although a cellularautomaton heartexcitationmodel has been used for the 3D activa
tion time imaging, it is anticipated that more sophisticated 3D heartexcitationmodels (e.g.
Gulrajani et al., 2001) can be realized in 3D activation time imaging in the future, providing
needed spatial resolution for more complicated cardiac activation sequences. With rapid
advancement in computer technology, this goal seems to be more feasible than expected.
In parallel to the heartexcitationmodel based 3D activation imaging approach, Ohyu
et al. (2002) has reported another 3D activation imaging approach in which the action
potentials are approximated by a step function with the initial activation being varied from
site to site within isotropic ventricles. The distribution of activation time was then connected
to the BSPMs and estimated by means of Wiener inverse filter. Both this activation imaging
approach (Ohyu et al., 2002) and the dipole source distribution imaging approach (He Wu,
2001) reviewed in Section 5.2 use linear systems connecting the inverse solution parameters
directly with the BSPMs. It would be of interest to evaluate the heartmodelbased activation
imaging reported by He et al. (2002), with linear system based activation imaging reported
by Ohyu et al. (2002).
5.4 THREEDIMENSIONAL MYOCARDIAL TRANSMEMBRANE
POTENTIAL IMAGING
In association with activation time, the transmembrane potential (TMP) reflects im
portant electrophysiological properties on the local myocardial tissues. The TMP has been
estimated over the heart surface from magnetocardiograms (Wach et al., 1997). Recently,
efforts have been extended from the heart surface to the 3D myocardium. He and coworkers
reported (2002, 2003) their effort to estimate TMP distribution within the 3D ventricles
from body surface electrocardiograms by means of a heartmodel based imaging approach.
Ohyu et al. has developed an approach to estimate the activation time and approximate
amplitude of the TMP from magnetocardiograms using Wiener inverse filter (2002). Skipa
et al. reported their initial results on estimation of TMP distribution within the heart from
BSPMs (2002).
In this Section, we present the heartmodebased 3D TMP imaging approach and its
applications to imaging TMP distributions associated with paced ventricular activity, and
acute myocardial infarction. The whole procedure of the heartmodelbased TMP imaging
approach may also be illustrated in the schematic diagram in Fig. 5.2, except that the
reconstructed 3D cardiac sources are spatiotemporal distribution of TMP.
Similar to the 3D heartmodel based activation imaging, the following procedures are
used. A realistic geometry 3D hearttorsomodel is constructed based on the knowledge
of cardiac electrophysiology and geometric measurements via CTIMRI. The anisotropic
nature of myocardium can be incorporated into this computer heart model. The BSPMs are
linked with the 3D TMP distribution by means of the hearttorsomodel. To reduce the di
mensionality of the parameter space, a preliminary classification system (PCS) is employed
to classify cardiac status based on the a priori knowledge of cardiac electrophysiology
and the BSPM, by means of an ANN. The output of the PCS provides the initial estimate
of heart model parameters which are employed later in a nonlinear optimization system.
The nonlinear optimization system then minimizes the objective functions that assess the
dissimilarity between the "measured" and modelgenerated BSPMs.
176 B.He
the "measured" BSPM and the hearttorsomodelgenerated BSPM match well, the
3D distribution of transmembrane potentials is determined from the heart model parame
ters corresponding with the resulting BSPM. the results do not match, the heart model
parameters are adjusted with the aid of the optimization algorithms and the optimiza
tion procedure proceeds until the objective functions satisfy the given convergent criteria.
When the procedure converges, the TMP distribution throughout the 3D myocardium is
determined.
The feasibility of imaging 3DTMP distribution has been suggested in computer sim
ulation in paced activities (He et al., 2003). The performance of the above TMP imaging
approach was tested by singlesite pacing in 24 different sites throughout the ventricles.
GWN of 10 J.LV was added to the BSPMs, and GWN of 10 mm was added to the body
surface electrode positions, to simulate noisecontaminated body surface ECG recordings.
Fig. 5.6 shows the TMP amplitude distributions (ab) of ventricular depolarization follow
ing a singlesite pacing at the septum in 5 longitudinal sections within the ventricles (c).
The TMP distribution in each longitudinal section at 8 typical instances (from 6 ms to 48
ms with a time step of 6ms) after the onset of pacing is shownin one row. The 5 longitudinal
sections within the ventricles arc illustrated in Fig. 5.6(c) by 5 horizontal black lines in the
Layel'3
6ms 12ms 18ms 24ms 3Jms
•
35ms 42ms 48ms
FI GURE 5.6. An example of TMP imaging results duri ng singlesite ventricular pacing. (a) and (b) illustrate
the forward and inverse solution of the TMP distributions in 5 longitudinal sections within the ventricles, duri ng
ventricular depolarization following a singlesite pacing at the septum. The TMP distri bution in each longitudinal
section at 8 typical instances (from 6 ms to 48 ms with a time step of 6 ms) after the onset of pacing is shown
in one row. The 5 longitudinal sections within the ventricles are illustrated in (c) by 5 horizontal black lines in
the transverse section of the ventricles from top to bottom indicating their positions. The gray regions in the
longitudinal sections indicate the resting cell units. The Max and Min of color bars correspond to the maximum
and minimum values of the TMP amplitude during the first 60 ms from the onset of activation. See attached CD
for color figure.
ThreeDimensional Electrocardiographic Tomographic Imaging 177
La 3
6nu
uyu.1S
FIGURE 5.6.
transverse section of the ventricles from top to bottom indicating their positions. The gray
regions in the longitudinal sections indicate the resting cell units . The Max and Min of color
bars correspond to the maximum and minimum values of the TMP amplitude during the first
60 ms from the onset of activation. Fig. 5.6 suggests that the inverse TMP distribution cap
tures well the overall spatiotemporal patterns of the forward TMP distribution following a
single site pacing at the septum, but with a slight shift of the area of initial activation towards
the base (as observed in Layer 3 in the inverse TMP distribution). Averaged over 24 sites
for singlepacing, the RE and CC between the "true" and estimated TMP distributions are
0.1266 ± 0.0326 and 0.9915 ± 0.0041 , respectively, indicating that the 3D TMP imaging
approach can reconstruct well the TMP distributions within the ventricles corresponding to
a welllocalized ventricular activati on.
178
(a)
(b)
B.He
FIGURE 5.7. A numerical example of myocardial infarction imaging. (a) Green shows preset acute myocardial
infarction. (b) Red shows estimated infarcted area over the same layer in the heart model. See attached CD for
color figure.
The effects of hearttorso geometry uncertainties on the performance of the 3D TMP
imaging approach has also been evaluated with torso size, heart position uncertainties being
considered. The 3D TMP imaging approach was found to be robust for up to 10% torso size
variation and up to 10 mm heart position shift. See reference (He et al., 2003) for detailed
descriptions of the results for TMP imaging of paced activities.
The 3D TMP imaging approach has also been applied to image acute myocardial
infarction (MI) (Li He, in press). Fig. 5.7 illustrates an example of myocardial infarction
imaging. In this case, GWN of5 f..L Vwas added to the BSPMs to simulate noisecontaminated
body surface ECG recordings. Fig. 5.7(a) shows the "true" MI that is marked by green color,
and Fig. 5.7(b) shows the inversely estimated MI that is marked by red color. A "true" MI
is located in the middle left wall (MLW) of ventricle and distributes in 7 myocardial layers
(not shown here). The estimated MI is located close to the "true" MI site, and has similar
shape with the "true" MI, except that some small MI areas occurred in layers adjacent to
the 7 "true" MI layers (not shown here). Fig. 5.7 suggests the feasibility of applying the 3D
TMP imaging approach to imaging the spatial location and extent of the acute MI in the
ventricles .
5.5 DISCUSSION
In this Chapter, we reviewed the recent progress in the development of noninva
sive 3D electrocardiography tomographic imaging approaches, including 3D dipole source
distribution imaging, 3D activation imaging, and 3D transmembrane potential imaging.
These activities may be classified as two general approaches in terms of methodology. One
is to solve the system equations connecting electrophysiological characteristics (such as
current density, activation time and TMP) to BSPMs. This approach involves solving the
system equations using inverse techniques such as weighted minimum norm (He Wu,
2000,200I ; Skipa et al., 2002) and Weiner technique (Ohyu et al., 2002). Another approach
is to solve the electrocardiography tomographic imaging problem indirectly, by means of
a heartmodel. In this heartmodel based approach, we have developed a localization ap
proach to localize the site of origin of activation from body surface ECG recordings
He, 2001), an activation imaging approach to image the activation time distribution (He et
al., 2002) , and an TMP imaging approach to image distribution of transmembrane potential s
ThreeDimensional Electrocardiographic Tomographic Imaging 179
throughout the ventricles from BSPMs (He et al., 2003). The 3D TMP imaging approach
has been applied to image dynamic spatiotemporal patterns of activation induced by pacing
(He et al., 2003), and to image and localize the site and size of acute myocardial infarction
(Li He, in press).
The heartmodel based approaches are based on our observation that a priori informa
tion regarding the distributed cardiac electrophysiological process should be incorporated
into the cardiac inverse solutions in order to obtain useful information on the distributed
3D cardiac electrical activity from the twodimensional BSPMs. In the present approach,
the a priori information on cardiac electrophysiology is incorporated into the distributed
heartmodel, which is not an equivalent physical source model but an electrophysiological
source model, in which knowledge of electrophysiology and pathophysiology is imbed
ded. The distributed electrophysiological process within the heart is represented by cellular
automata, on each of which the site of origin of activation, activation time, or transmem
brane potential are determined based on the knowledge of cardiac electrophysiology. In
such approaches, since substantial electrophysiology a priori information is incorporated
into the inverse solutions, more accurate inverse solutions are anticipated as compared with
other approaches without taking this information into account. Such electrophysiology
priori information not only includes more accurate forward solution at each time point, but
also a more realistic timevarying dynamics as set by the heart electrophysiology model.
Therefore, it is not surprising that good matches between "true" cardiac electrical activity
and estimated inverse solutions are obtained by means of the heart model based approaches.
On the other hand, the system equation approach has the benefit that there is no need to
limit the search space for heart model parameters, as currently being practiced in the heart
model based approaches. The inverse solutions are obtained directly by solving the system
equations that link the electrophysiological properties with BSPMs via biophysical rela
tionships. It would be of interest to compare the performance of these two approaches for
3D electrocardiography tomographic imaging.
The inverse problem of electrocardiography has been solved by means of equivalent
point sources (dipole localization), distributed twodimensional heart surface imaging meth
ods (epicardial potential imaging, and heart surface activation imaging), and 3D distributed
source imaging approaches. While the 3D distributed source imaging, as reviewed in this
chapter, represents an important advancement in the field of electrocardiography inverse
problem, all 3D electrocardiography tomographic imaging approaches have only been eval
uated, up to date, in computer simulations. It is of ultimate importance and significance to
experimentally validate the 3D distributed source imaging approaches, in order to establish
electrocardiography tomographic imaging as a useful means for imaging noninvasive three
dimensional distribution of cardiac electrical activity, for aiding clinical diagnosis and man
agement of a variety of cardiac diseases, and for guiding radiofrequency catheter ablative
interventions.
ACKNOWLEDGEMENT
The author wishes to thank his postdoctoral associates and graduate students, Dr.
Guanglin Li, Dr. Dongsheng Wu, and Xin Zhang, with whom this work was conducted.
This work was supported in part by NSF BES0201939, a grant from the American Heart
Association #0140132N, and NSF CAREER Award BES9875344.
180
REFERENCES
B.Re
Aoki, M., Okamoto, Y.,Musha, T., and Harumi, K.: Threedimensional simulation of the ventricular depolarization
and repolarization processes and body surface potentials: normal heart and bundle branch block. IEEE Trans.
Biomed. Eng., 34: 454462, 1987.
Barber, M.R., Fischman, E.1.: Heart dipole regions and the measurement of dipole moment. Nature, 192: 141142,
1961.
Barr, R.C., Ramsey, M., Spach, M.S.: Relating epicardial to body surface potential distributions by means of
transfer coefficients based on geometry measurements. IEEE Trans. Biomed. Eng., 24: III, 1977.
Bellman, R., Collier, C, Kagiwada, H., Kalaba, R., Selvester, R.: Estimation of heart parameters using skin
potential measurements. Comm. ACM, 7: 666668,1964.
Burnes, J.E., Taccardi, B., MacLeod, R.S., Rudy, Y.: Noninvasive ECG imaging of electrophysiologcially abnormal
substrates in infarcted hearts: a model study. Circulation, IOJ: 533540, 2000.
Cuppen, J.1.M., Van Oosterom, A.: Model studies with inversely calculated isochrones of ventricular depolariza
tion. IEEE Trans. Biomed. Eng., 31: 652659,1984.
de Guise, J., Gulrajani, R.M., Savard, P., Guardo, R., Roberge, EA.: Inverse recovery of two moving dipoles from
simulated surface potential distributions on a realistic human torso mode!. IEEE Trans. Biomed. Eng., 32:
126135, 1985.
Fischer, G., Tilg, B., Modre, R., Huiskamp, G.1., Fetzer, J., Rucker, Wach, P.: A bidomain model
based BEMFEM coupling formulation for anisotropic cardiac tissue. Ann Biomed Eng., 28: 12291243,
2000.
Flowers, N.C., Horan, L.G.: Body surface potential mapping. In: Cardiac Electrophysiology, edited by Zipes, D.P.,
Jalife, J. 2nd Edition. W.B. Saunders Company, pp. 10491067, 1995.
Frazone, P.c., Taccardi, B., Viganotti, c.: An approach to the inverse calculation of epicardial potentials from
body surface maps. Adv. Cardio!., 21: 5054, 1978.
Franzone, P.C., Guerri, L., Pennacchio, M., Taccardi, B.: Spread of excitation in 3D models of the anisotropic
cardiac tissue. III. Effects of ventricular geometry and fiber structure on the potential distribution. Math.
Biosci., 151: 5198,1998.
Geselowitz, D.B.: Multipole representation for an equivalent cardiac generator. Proc. IRE., 48: 7579, 1960.
Gorodnitsky, I.E, George, 1.S., Rao, B.D.: Neuromagnetic source imaging with FOCUS: a recursive weighted
minimum norm algorithm. Electroenceph. din. Neurophysio!., 95: 231251,1995.
Greensite, E, Huiskamp, G.: An improved method for estimating epicardial potentials from the body surface.
IEEE Trans. Biomed. Eng., 45: 17, 1998.
Greensite, E: Myocardial Activation Imaging. In: Computational Inverse Problems in Electrocardiography, edited
by Johnston, P. WIT Press Brisol, 143190,2001.
Gulrajani, R., Roberge, EA., Savard, P.: Moving dipole inverse ECG and EEG solutions. IEEE Trans. Biomed.
Eng., 31: 903910,1984.
Gulrajani, R.M., Trudel, M.C., Leon, L.1.: A membranebased computer heart model employing parallel process
ing. Biomedizinische Technik, Brand 46, Erganzungsband, 2: 2022, 2001.
Hamalainen, M., Ilmoniemi, R.: Interpreting measured magnetic fields of the brain: estimates of current distribu
tions. Helsinki University of Technology Report, TKKF A559, 1984.
He, 8., Wu, D.: A bioelectric inverse imaging technique based on surface Laplacians. IEEE Trans. Biomed. Eng.,
44:529538,1997.
He, B., Wu, D.: Threedimensional source imaging of cardiac electric activity. Proc. of World Congress on Medical
Physics and Biomedical Engineering, CDROM, 2000.
He, 8., Wu, D.: Imaging and visualization of 3D cardiac electric activity. IEEE Trans. Inf. Techno!. Biomed., 5:
181186,2001.
He, 8., Li, G., Zhang, X.: Noninvasive Threedimensional Activation Time Imaging of Ventricular Excitation by
Means of a HeartExcitationMode!' Physics in Medicine and Biology, 47: 40634078, 2002.
He, B., Li, G.: Noninvasive threedimensional myocardial activation time imaging by means of a heartexcitation
mode!. Int. J. of Bioelectromagnetism, 4(2): 8788, 2002.
He, B., Li, G., Zhang, X.: Noninvasive Imaging of Ventricular Transmembrane Potentials within Threedimensional
Myocardium by Means of a Realistic Geometry Anisotropic Heart Mode!' IEEE Trans. Biomed. Eng., 50( I0):
11901202,2003.
ThreeDimensional Electrocardiographic Tomographic Imaging 181
Hlavin, J.M., Plonsey, R.: An experimental determination of a multipole representation of a turtle heart. IEEE
Trans. Bimed. Eng., 10: 98, 1963.
Huiskamp, G., Greensite, E: A new method for myocardial activation imaging. IEEE Trans. Biomed. Eng., 44:
433446,1997.
Huiskamp, G.: Simulation of depolarization in a membraneequationsbased model of the anisotropic ventricle.
IEEE Trans. Biomed. Eng., 45: 847855, 1998.
Jeffs, B., Leahy, R., Singh, M.: An evaluation of methods for neuromagnetic image reconstruction. IEEE Trans.
Biomed. Eng., 34: 713723,1987.
Johnston, P.R., Gulrajani, R.M.: A new method for regularization parameter determination in the inverse problem
of electrocardiography. IEEE Trans. Biomed. Eng., 44: 1939, 1997.
Li, G., He, 8.: Localization of the site of origin of cardiac activation by means of a heartmodelbased
electrocardiographic imaging approach. IEEE Trans. Biomed. Eng., 48: 660669, 2001.
Li, G., He, B.: Noninvasive Estimation of Myocardial Infarction by Means of a HeartModelBased Imaging
Approach.: simulation study Med. Bio!' Eng. Comput., In press.
Lorange, M., Gulrajani, R.M.: A computer heart model incorporating anisotropic propagation. Model construc
tion and simulation of normal activation. 1. Electrocardiol., 26: 245261,1993.
Lu, W, Xu, Z., Fu, Y Microcomputerbased cardiac field simulation model. Med. Bio. Eng. Comput., 31: 384387,
1993.
Malmivuo, J., Plonsey, R.: Bioelectromagnetism. Oxford University Press, 1995.
Gulrajani, R.M.: Bioelectricity and Biomagnetism. John Wiley Sons, 1998.
Martin, RO., Cox, J.W, Keller, EW, Terry, EH., Brody, D.A.: Equivalent cardiac generators; Two moving dipoles
and moving dipole and quadripole. Ann. Biomed. Eng., 2: 164183, 1974.
Miller, W.T., Geselowitz, D.B.: Simulation studies of the electrocardiogram The normal heart. Circ. Res., 43:
301323,1978.
Mirvis, D.M., Keller, EW., Ideker, R.E., Cox, J.W, Dowdie, RJ., Zettergren, D.G.: Detection and localization
of multiple epicardial electrical generators by a twodipole ranging technique. Circ. Res., 41: 551557,
1977.
Modre, R., Tilg, B., Fischer, G., Wach, P.: An iterative algorithm for myocardial activation time imaging. Computer
Methods and Programs in Biomedicine, 64: 17,2001.
Nenonen, 1., Edens, J., Leon, LJ., Horacek, B.M.: Computer model of propagated excitation in the anisotropic
human heart: 1. Implementation and algorithms. In: Compter in Cardiology, 545548,1991.
Okamoto, Y, Teramachi, Y, Musha, T.: Limitation of the inverse problem in body surface potential Mapping.
IEEE Trans. Biomed. Eng., 30: 749754,1983.
Ohyu, S., Okamoto, Y, Kuriki, S.: Use of the ventricular propagated excitation model in the magnetocardiographic
inverse problem for reconstruction of electrophysiological properties. IEEE Trans. Biomed. Eng., 49: 509
519,2002.
Oster, H.S., Taccardi, B., Lux, R.L., Ershler, P.R, Rudy, Y: Noninvasive electrocardiographic imaging: recon
struction of epicardial potentials, electrograms, and isochrones and localization of single and multiple elec
trocardiac events. Circulation, 96: 10121024, 1997.
PascualMarqui, R.D., Michel, C.M., Lehmann, D.: Low resolution electromagnetic tomography: a new method
for localizing electrical activity in the brain. Int. J. Psychophysiol., 18: 4965, 1994.
Pilkington, T.c., Morrow, M.N.: The usefulness of multi poles in electrocardiography. CRC Crit. Rev. Biomed.
Eng., 7:175,1982.
Pullan, AJ., Cheng, L.K., Nash, M.P., Bradley, CP, Paterson, DJ.: Noninvasive electrical imaging of the heart:
theory and model development. Ann. Biomed. Eng., 29: 81736,2001.
Savard, P., Roberge, EA., Perry, J., Nadeau, R.A.: Representation of cardiac electrical activity by a moving dipole
for normal and ectopic beats in the intact dog. Circ. Res., 46: 415425, 1980.
Selvester, RH.S.: Recommendation for nomenclature of myocardial subdivisions. J. Electrocardil., 25: 161162,
1992.
Shahidi, A.Y.,Savard, P., Nadeau, R: Forward and inverse problems of electrocardiography: modeling and recovery
of epicardial potentials in humans. IEEE Trans. Biomed. Eng., 41: 249256, 1994.
Skipa, 0., Sachse, N.E, Werner, Dossel, 0.: Transmembrane potential reconstruction in anisotropic heart
mode!. Proc. of International Conference on Bioelectromagnetism, 1718, 2002.
Streeter, Jr. D.D., Spotnitz, H.M., Patel, Jr. D.P., Ross, J., Sonnenblick, E.H.: Fiber orientation in the canine left
ventricle during diastole and systole. Circ. Res., 24: 339347,1969.
182 B.Ue
Taccardi, B. Distribution of heart potent ial on the thoracic surface of normal human subjects. Circ. Res., 12:
341352, 1963.
Throne, R.D., Olson, L.G.: A genera lized eigensystem approach to the inverse problem of electrocardiography.
IEEE Trans. Biomed. Eng., 41: 592600,1994.
Throne, RD., Olson, L.G. : General ized eigensys tem techniques for the inverse problem of electrocardiography
applied to a realistic hearttorso geometry. IEEE Trans. Biomed. Eng., 44: 447454,1 997.
Wach, P. Tilg, B., Lafer, G., Rucker, W.: Magnetic source imaging in the human heart: estimating cardiac electrical
sources from simulated and measured magnetocardi ogram data. Med. BioI. Eng. Comput. , 35: 157166,
1997.
Wei, D., Okazaki, 0 ., Harurni, K., Harasawa, E., Hosaka, H.: Compar ative simulation of excitation and body
surface electrocardiogram with isot ropic and anisotropic computer heart model. IEEE Trans. Biomed. Eng.,
42: 343 357, 1995.
Wu, D., Tsai, H.C., He, B.: On the Estimation of the Laplacian Electrocardiogram during Ventricular Activation.
Ann. Biomed. Eng., 27: 731745 ,1999.
BODYSURFACE LAPLACIAN
MAPPING OFBIOELECTRIC
SOURCES
Bin He* and lie Lian
Department of Bioengineering, University of Illinois at Chicago
6.1 INTRODUCTION
6.1.1 HIGHRESOLUTION ECG AND EEG
Targeting two of the most lifecritical organs, the heart and brain, the electrocardiogram
(ECG) and the electroencephalogram (EEG) are the two important bioelectric recordings
to study the cardiac and neural activity.
Conventional ECG and EEG have many advantages. First, they are noninvasive mea
surement. Second, they are very convenient for application and have relatively low cost.
More importantly, they have unsurpassed millisecondscale temporal resolution, which is
essential for revealing rapid change of dynamic patterns of heart and brain activities. How
ever, the major limitation of the conventional ECG and EEG is their relatively low spatial
resolution as compared to some other imaging modalities, such as the computed tomography
(CT) or the magnetic resonance imaging (MRI).
One reason contributing to the low spatial resolution is the limited spatial sampling.
Conventional EEG uses the standard international 1020 system, which has about 20 elec
trodes over the scalp, with corresponding interelectrode distance of about 6 em (Nunez
et al., 1994). For the ECG measurement, the most commonly used configuration in a clinical
setting is the 12lead ECG. Despite its great success in many clinical applications, it has a
major limitation in that it contains very little spatial information, and doctors have to infer
the cardiac status mainly based on temporal analysis of the ECG waveforms. Therefore, one
way to enhance the spatial resolution of ECG and EEG is to increase the spatial sampling,
by using larger number of surface electrodes in ECG and EEG measurement.
Address all correspondence to: Bin He, Ph.D. University of Minnesota, Department of Biomedical Engineering,
7105 BSBE, 312 Church Street, Minneapolis, MN 55455. Email: binhe@umn.edu
183
184 B. He and J. Lian
However, even with very highdensity spatial sampling, the spatial resolution of the
EEG and ECG is still limited, because of the volume conduction effect. In other words, the
electrical signals will get smeared as they pass through the media between the bioelectric
sources and the body surface sensors. For the brain, it's the head volume conductor, partic
ularly the skull layer, which has low conductivity (Nunez, 1981, 1995). For the heart, it's
the torso volume conductor, including the effects of lungs, the ribs and other tissues (Mirvis
et al., 1977; Spach et al., 1977; Rudy Plonsey, 1980).
Therefore, advanced techniques are desired in order to compensate for the volume
conduction effect and enhance the spatial resolution of the ECG and EEG. As reviewed in
Chapters 4 and 5 with applications to the heart, one of such methods is to solve the socalled
inverse problem, which attempts to estimate the bioelectric sources from the body surface
potential measurements. Another method is the surface Laplacian, which will be thoroughly
discussed in this chapter.
6.1.2 BIOPHYSICAL BACKGROUND OF THE SURFACE LAPLACIAN
The concept of the Laplacian originated centuries ago, and the Laplacian operator has
been widely used in digital image processing as a spatial enhancement method. Similarly,
the Laplacian technique can also be used for highresolution bioelectric mapping.
By definition, the surface Laplacian (SL) is defined as the 2nd order spatial derivative of
the surface potential. Due to its intrinsic spatial highpass filtering characteristics, the SLcan
reduce the volume conduction effect by enhancing the highfrequency spatial components,
therefore can achieve higher spatial resolution than the surface potentials (Figure 61).
Consider the nonorthogonal curvilinear coordinate system on a general surface Q,
u= x, v= y, and z = f (u, v), where f(u, v) is a continuous function whose 2 order
partial derivatives exist. is the analytical surface potential function (whose 2
order partial derivatives exist) on Q, the SL of can be written in tensorial formulation
(Courant Hilbert, 1966; Babiloni et al., 1996):
{a [ (av av)] a [ (av av)]}
Vs =.y'g au gIla;; g12a;; av g21a;; g22a;;
where the components of the metric tensor are given by (Babiloni et al., 1996):
af af

au av
gl2 g21
1 + ( ~ r
g22 =
(61)
(62a)
(62b)
(62c)
(62d)
Body Surface Laplacian Mapping of Bioelectric Sources 185
FIGURE 61. Schematic illustration of the SL as a spatial enhancement method. The cardiac activity located at
the anterior apex (black circle) is sensed by potential measurement over the larger area on the chest (light grey)
but by Laplacian measurement over the smaller area on the chest (dark grey). (From Tsai et aI., Electromagnetics,
2001 with permission)
For the plane model where z j(u, v) 0, u x and v y, the SL is reduced to:
(63)
For the sphere model, assume z .Jl  u v x and v y, the SL is then given
by (Perrin et al., 1987a):
(64)
The Laplacian electrogram (we refer electrogram to either ECG or EEG when the
heart or the brain is concerned) shall be defined as the negative SL of the surface potential
electrogram (He, 1999; He Wu, 1999), to facilitate the interpretation of the Laplacian
maps in comparison to the potential maps. As stated in equation (63), assuming a planar
surface in the vicinity of the observation point, a reasonable approximation of the local
area of the body surface would be the tangential plane at the point of interest, over which a
local Cartesian coordinate system y, z) can be considered. Assuming z to be normal to
the tangential plane, the Laplacian ECGIEEG at the observation point becomes (He, 1999;
He Wu, 1999):
where denotes the current density and Jeq is an equivalent current source (He Cohen,
1992a, 1995; He, 1997, 1998a, 1999; He Wu, 1999).
Unlike the ECG and EEG inverse problems, the SL approach does not attempt to lo
cate the bioelectric sources inside the heart and brain. Instead, the Laplacian ECGIEEG
186 B. He and J. Lian
can be viewed as a twodimensional (2D) projection of the threedimensional (3D) bio
electric source onto the body surface. Therefore, as shown in equation (65), the Laplacian
ECG/EEG can be interpreted as an equivalent current source density on the body surface,
which has the similar physical units as the primary bioelectric source density. On the other
hand, compared to the ECG and EEG inverse approaches, the SL approach does not require
exact knowledge about the conductivity distribution inside the torso and head volume con
ductors and has unique advantage of referenceindependence as compared with the potential
measurement.
6.2 SURFACE LAPLACIANESTIMATION TECHNIQUES
6.2.1 LOCAL LAPLACIANESTIMATES
If the body surface in the vicinity of the measurement point can be approximately
represented by a planar surface, the SL can be calculated using equation (63). In practice,
the second order derivatives can be approximated by means of finite difference (Hjoth,
1975).
Consider a grid of unipolar electrodes with equal interelectrode distance on the body
surface (Figure 62A), the regular Laplacian electrogram at each nonboundary electrode
(the crosshatched circle) can be estimated by using the regular finite difference represen
tation (Hjoth, 1975; He Cohen, 1992a; Wu et al., 1999; Lian et al., 2002):
LR(i, ~ ~ {V(i, ~ [V(i  1, V(i 1, v«, j  1) V(i, j I)]}
(66)
where V(i, and LR(i, represent the potential and the regular Laplacian electrogram at
the electrode respectively. For each nonboundary electrode, equation (66) uses the
potential measurement at five electrodes (the crosshatched circle and its four neighboring
open circles in Figure 62A) to estimate the Laplacian electrogram at the center electrode
(the cross hatched circle). Similarly, the Laplacian electrogramat electrode can also be
estimated from the potential recorded from this electrode and those recorded from its other
four neighboring electrodes in the diagonal direction (neighboring black circles surrounding
the crosshatched circle in Figure 62A). Denote the distance from the center electrode to its
diagonal neighboring electrodes as d (for uniform grid, d .fib), the diagonal Laplacian
electrogram can also be estimated by (Wu et al., 1999; Lian et al., 2002):
LD(i, ~ :2 {V(i, j)  ~ [ V ( i  1, j  V(i  1, j 1)
+V(i+l,jl)+V(i+l,j+I)]}
where L D(i, represents the diagonal Laplacian electrogram at the electrode (i, j).
Body Surface Laplacian Mapping of Bioelectric Sources 187
••
.......'
• •
{
• •
• •
. /
.
•••••••••••••••
•••••••••••••••
•• 0 ••••••••••••
•••••••••••••••
•••••••••••••••
•••••••••••••••
•••••••••••••••
•••••••••••••••
A
FIGURE 62. Schematic illustration of the local Laplacian estimates. (A) Regular or diagonal 5point local
Laplacian estimation. (B) Circular finite difference local Laplacian estimation.
A more general form of the finite difference reprentation of the SL utilizes the po
tential information from more local electrodes to realize the circular Laplacian electrode
(He Cohen, 1992a). As illustrated in Figure 62B, to estimate the Laplacian electro
gram at the center electrode, the unipolar potential data are obtained from this elec
trode as well as from electrodes located along a small circle (with radius surrond
ing it, and the finite difference representation of the Laplacian electrogram is given by
(Le et aI., 1994; Wei et al., 1995; He, 1997, 1998a, 1999; Wei Mashima, 1999; Wei,
2001):
4 ( 1 n )
u, ~ 2" V
o
  LVi
r n
(68)
where Vo and La represent the potential and circular Laplacian electrogram at the cen
ter electrode, respectively, and 1, 2, ... , represents the potential at one of the
surrounding electrodes.
Another local Laplacian estimate uses bipolar concentric electrode that consists of two
parts: a conductive disk at the center and a surrounding conductive ring (Fattorusso et al.,
1949; He Cohen, 1992a). In the bipolar approach, the Laplacian electrogram may be
estimated as (He & Cohen, 1992a, 1995; He, 1997):
Lo ~ i (Vo __ 1. Vdl)
r 2nr
(69)
where the integral is taken around a circle of radius
In addition, some other localbased Laplacian algorithms were proposed in order to
achieve more acurate numerical estimates, for instance, by local modeling of the scalp
and the potential distribution (Le et aI., 1994), or by means of the local polynomial fitting
(Wang Begleiter, 1999).
188
6.2.2 GLOBAL LAPLACIAN ESTIMATES
He and J. Lian
From equations (61) and (62), it can be seen that both the analytical model s of the
potential distribution V(u, and the surface geometry f eu, are required to calculate the
SL. However, in real applications, the only data available are the limited number of potential
recordings and body surface geometri c coordinate samplings. Thus, V(u, and f eu,
must be interpolated or approximated, and the most widely used interpolation scheme is
the spline interpolation. Among the investigations on the spline SL, of noteworthy is the
spherical spline SL (Perrin et al., 1987a,b, 1989), the ellipsoidal spline SL (Law et al.,
1993), and the realistic geometry spline SL (Babiloni et al., 1996, 1998; He et al., 2001,
2002; Zhao He, 2001).
Considering the nonplanar shape of the body surface, a global SL estimation using
spline technique will be more accurate than the localbased SL estimates. Furthermore, as
a major advantage over the localbased SL estimation, the spline SL has been shown to
provide a more robust characteristic against noise (Perrin et al., 1987a; Law et aI., 1993;
Babiloni et al., 1996; He et al., 2001, 2002).
In the following, a recently developed realistic geometry spline Laplacian estimation
algorithm is presented (He et al., 2001, 2002). Estimation of the parameters associated
with the spline Laplacian is formulated by seeking the general inverse of a transfer matrix.
The number of spline parameters, which need to be determined through regularization, is
reduced to one in the present approach , thus enabling easy implementation of the realistic
geometry spline Laplacian estimator.
6.2.2.1 Spline interpolation ofthe surface geometry
Given body surface sampling point s (Xi, Yi, z.) . 1, ... , where is the number of
sampling points for coordinate measurement, the mathematical model of the body surface
geometry Z can be described by 2D thin plate spline (Harder & Desmarais, 1972;
Perrin et al., 1987a,b; Babiloni et al., 1996; He et al., 2001, 2002):
2(m l) ( 2 2) dk k
log d
i
+W Y
i=1 i=l d=Ok=O
where m (spline order) is set to 2 (Perrin et aI., 1987a,b; Babiloni et aI., 1996; He et aI.,
2001,2002), d; (x  Xi )2 Yi)2, Km I and are basis function and osculat
ing function , respectively, and is a constant which accounts for effective radius of the
recording sensor (Harder Desmarais, 1972; Perrin et al., 1987a). The coefficients Pi and
qdk are the solutions of following matrix equations (Duchon, 1976; Perrin et aI., 1987a,b;
He et al., 2002):
KP +EQ=Z
ETp =0
(611a)
(6 11b)
Body Surface Laplacian Mapping of Bioelectric Sources 189
where Q, and Z are the vectors containing and Zi respectively, the matrices and
E are composed of elements of basis function and sampling coordinates, respectively.
6.2.2.2 Spline interpolation ofthe surface potential distribution
Similarly, given body surface potential recordings Vi at positions Yi, z. ), i
1, 2, . . . , where is the number of recording electrodes, the body surface potential
distribution over the 3D space at an arbitrary point z) can be modeled by the 3D spline
(Babiloni et al ., 1996, 1998; He et aI., 2001, 2002):
~ ~ (2m 3)/2 ~ ~ ~
= ~ t i = ~ t i r i ~ ~ ~ r d k g X Y
i=1 i =1
(612)
where m (spline order) is set to 3 (Law et aI., 1993; Babiloni et aI., 1996, 1998; He et aI.,
2001, 2002), rl (x  Xi)2 + Yi)2 + (z  Zi)2, H
m

l
and R ml are basis and osculat
ing functions, respectively, and the coefficients and can be determined by solving the
matrix equations (Law et aI., 1993; Babiloni et aI., 1996; He et al., 2002):
(613a)
(613b)
where T, R, and V are the vectors containing t. , and respectively, the matrices H
and are composed of elements of basi s function and electrode coordinates, respectively.
6.2.2.3 Determination ofthe spline parameters
In an attempt to overcome the illposeness of the systems, approximation instead of
interpolation of the surface geometry and potential distribution are used by introducing cor
rection terms in equations (611a) and (613a), which are respectively changed to (Babiloni
et aI., 1996, 1998; He et aI., 2001, 2002):
(K wI)P EQ Z (614a)
(614b)
where 1is the identity matrix, parameters wand are used to improve the numerical stability
of the systems. The optimal values of these two parameters need be determined separately
by either "tuning procedure" or other regularization techniques (Babiloni et al., 1996, 1998).
Instead of searching the optimal parameters in two dimensions, the above equations
can be reformulated by combining equations (611a,b) and (613a,b) into one linear system
equation (He et aI., 2001, 2002):
AX=B (615)
190 B. He and J. Lian
where
0
~ ]
0 0
(616a) A
 0
0 H
0 0 F
T
x= T Rf
(616b)
B = 0
O]T
(616c)
Then the problem becomes seeking the solution of equation (615). Applying the
concept of the general inverse, we have (He et al., 2001, 2002):
(617)
where A# is the pseudoinverse of A. Matrix A is illposed, thus regularization methods
must be used to improve the stability of the system. Notably, after reformulating the matrix
equations into one unified linear system in equation (615), only one single regularization
parameter needs to be determined when seeking the general inverse A#. Therefore, the
present method not only can significantly reduce the computation effort and improve the
efficiency and stability of the spline SL algorithm, but also can be combined with many
regularization techniques which have been extensively studied to determine the param
eter in seeking the general inverse A# (for details on the regularization techniques, see
Chapter 4).
6.2.3 SURFACE LAPLACIANBASED INVERSE PROBLEM
The SLbased ECG or EEG inverse problem has also been explored to achieve high
resolution heart or brain electric source imaging. One of the approaches is to estimate the
epicardial potentials from the body surface Laplacian ECG (He, 1994; Wu et al., 1995, 1998;
He & Wu, 1997, 1999; Johnston, 1997; Throne & Olson, 2000), or estimate the cortical
potentials from the scalp Laplacian EEG (He, 1998; Babiloni et al., 2000; Bradshaw
Wikswo, 2001).
As illustrated in Figure 63, if Vis an isotropic homogeneous volume conductor sur
rounded by an outer surface S1 and an inner surface and there is no current source existing
within the potential on the inner surface can be related to the potential or Laplacians on
the outer surface.
Applying Green's second identity to the volume results in (Barr et al., 1977):
u . 1. . 2E..dS
~ ~ ~ ~
(618)
where u(r'lthe electrical potential at the observation point
dQthe solid angle of an infinitesimal surface element ds as seen from
a
authe first derivative of potential u with respect to the outward normal to dS
r;
Body Surface Laplacian Mapping of Bioelectric Sources 191
Sl
FIGURE 63. Schematic illustration of the arbitrarily shaped volume conductor.
By discretizing the surfaces Sl and S2 into triangular elements, and taking the limit
of observation point approaching the surface element on and S 2, respectively, from the
inside of the following matrix equations can be obtained:
~ ~
PIIU, P'2U2 GI2r2 =0
~  
P21U P22U2 G22r2
(619)
(620)
where k is the vector consisting of the electrical potentials at every surface element
on Si , and is the vector consisting of the normal derivatives of the electrical poten
tials at every triangle element on Sk but just inside of VI. P P
12,
P21 , P22, G
12
, and
G
22
are coefficient matrices (Barr et aI., 1977). Solving equations (619) and (620)
leads to the followingequation that relates the inner surfacepotential 2 to the outer surface
potential U,:
(62 1)
where T
l2
(P Gl 2G221 P2dl(G12G221 P22  P12 ).
The surface Laplacian of the potential at the position the outer surface Sl can be
written as follows (Wuet aI., 1998; He Wu, 1999):
4* 1 (a
2
1 (a
2
1 au ( a
2
1
Ls(r ) = U • d + U • d an2 + ar. · an2( ,:) dS
(622)
where n is the normal direction of the surfaceSl at Similarly, by discretizing the surfaces
Sl and S2 into triangular elements, the following matrix equation can be obtained:
........ .... >. ......
Ls=AU,+BU (623)
192 B. He and J. Lian
~
where L, is the vector consisting of the surface Laplacians at every surface element on S"
and and C are coefficient matrices (Wu et aI., 1998; He Wu, 1999). From equations
(619) , (6:?l) , and (623), we can relate the inner surface potential to the outer surface
Laplacian by transfer matrix
(624)
where H = A · T
l2
+ B  C · G
22'
. ( P
2
, . Tl2 + P22)
The potentialbased inverse problem seeks the inner surface potentials from the outer
surface potentials by solving equation (621):
(625)
On the other hand, the SLbased inverse problem seeks the inner surface potentials based
on solving equation (624):
(626)
where # denotes the general inverse of the transfer matrix.
In addition, the hybrid potentialLaplacianbased inverse solution can also be solved
by minimizing the error function (He Wu, 1999; Throne Olson, 2000):
(627)
where ex is a weighting coefficient. The resulting inverse solution is given by (He Wu,
1999):
(628)
Equation (628) suggests that the inner surface potentials can be estimated from both the
outer surface Laplacians and out surface potentials.
6.3 SURFACE LAPLACIANIMAGING OF HEART
ELECTRICAL ACTIVITY
6.3.1 HIGHRESOLUTION LAPLACIANECG MAPPING
By applying the SL technique to the potential ECG, body surface Laplacian mapping
(BSLM) was first proposed by He and Cohen (l992a,b). Theoretical and experimental
studies have been carried out, demonstrating the unique feature of BSLM in effectively
reducing the torso volume conduction effect and enhancing the capability of localizing
and mapping multiple simultaneously active myocardial electrical events (He & Cohen,
Body Surface Laplacian Mapping of Bioelectric Sources 193
1992a,b; He et aI., 1993, 1995, 1997,2002; Oostendorp van Oosterom, 1996; Umetani
et aI., 1998; Wei Harasawa, 1999; Wu et aI., 1999; Tsai el aI., 2001; Besio et al., 2001;
Wei et aI., 2001; Li et aI., 2002).
6.3.2 PERFORMANCE EVALUATION OF THE SPLINE LAPLACIAN ECG
Through human experiments and computer simulations, we have systematically eval
uated the signal to noise ratio of the Laplacian ECG, during ventricular depolarization and
repolarization, and demonstrated the feasibility of recording the Laplacian ECG, using the
5point local SL estimator (Wu et aI., 1999; Lian et aI., 2001,2002). Further improvement of
the Laplacian ECG estimation may be achieved by using the globalbased spline Laplacian
technique.
In this section, we present the performance evaluation of the 3D spline SL algorithm
(see Section 6.2.2). Computer simulations were conducted using both a spherical model
and a realistic geometry hearttorso model, and comparison studies were also made with
the 5point local SL estimator (He et aI., 2002).
Given the torso surface geometry coordinates and the potential measurement, the
Laplacian ECG was estimated by using the realistic geometry spline SL algorithmas detailed
in Section 6.2.2. The linear inverse problem in equation (617) was solved by using the
truncated singular value decomposition (TSVD) (Shim Cho, 1981), and the truncation
parameter was determined by means of the discrepancy principle (Morozov, 1984).
6.3.2.1 Effects of noise
We first evaluated the effect s of noise on the SL estimation, by approximating the
torso volume conductor as a homogeneous singlelayer unitradius sphere model , with
normalized interior conductivity of 1.0. A radial dipole or a tangential dipole was used
to represent a localized cardiac electrical source. Three eccentricities (0.5, 0.6, 0.7) were
used to assess the effect of the source depth on the SL estimation. The Gaussian white
noise (GWN) of different noise levels (5%, 7%, 10%) was added to the dipolegenerated
surface potentials sampled from 129 surface electrodes, simulating noisecontaminated
potential ECG measurement. Two cases of geometry noise were also considered (2% ge
ometry noise plus 10% potential noise, and 5% geometry noise plus 5% potential noise).
For each noise level, ten trials of noise were generated and simul ations were conducted.
The correlation coefficient (cq values between the estimated SL and the analytical SL
for all ten trials were averaged and shown in Table 61. The SL was estimated by three
different methods : (1) 5point local SL (5PL), (2) twoparameter spline SL (2SL), and (3)
the recently developed oneparameter spline SL (1SL). The twoparameter spline SL was
estimated optimally by using the tuning procedure (Babiloni et aI., 1996, 1998) to find
the optimal values of (J) and in equation (614) and the oneparameter spline SL was
estimated optimally by searching the optimal truncation parameter in TSVD procedure (He
et aI., 2002). The optimal parameters correspond to the maximum CC between the analytical
SL and the estimated SL. The quant ity w in equation (610) was set to 0.16 in spline SL
estimation.
194 B. He and J. Lian
TABLE 61. The CC values between the analytical and estimated SL under different levels of noise
in onesphere model
GWN
5%PN 7%PN 10%PN 2%GN+IO%PN 5%GN+5%PN
5PL 2SL ISL 5PL 2SL ISL 5PL 2SL ISL 5PL 2SL ISL 5PL 2SL ISL
IRD, r=0.5 0.79 0.98 0.96 0.73 0.97 0.95 0.63 0.96 0.94 0.58 0.93 0.92 0.72 0.91 0.90
IRD, r = 0.6 0.90 0.97 0.98 0.85 0.96 0.97 0.77 0.95 0.95 0.58 0.91 0.93 0.70 0.91 0.92
IRD, r= 0.7 0.96 0.96 0.96 0.93 0.95 0.95 0.89 0.93 0.92 0.76 0.89 0.90 0.80 0.92 0.93
ITD, r = 0.5 0.71 0.98 0.96 0.62 0.97 0.95 0.50 0.96 0.94 0.65 0.90 0.91 0.67 0.88 0.90
ITD, r = 0.6 0.84 0.95 0.98 0.78 0.95 0.95 0.67 0.94 0.95 0.60 0.92 0.92 0.66 0.92 0.92
ITD, r = 0.7 0.94 0.96 0.96 0.91 0.95 0.95 0.85 0.91 0.93 0.71 0.88 0.88 0.79 0.91 0.92
Note: ROradial dipole, TOtangential dipole, PNpotential noise, GNgeometry noise.
Three findings are obvious from Table 61. First, for all three different SL estimators,
the higher the noise level, the smaller the Cc. Second, for all the cases studied, the spline SL
has superior performance (higher CC) than the 5point local SL, while the twoparameter
spline SL and oneparameter spline SL have similar performance. Third, the 5point local
SL has the best performance for superficial sources and under low potential noise level,
but its performance degrades dramatically as the source moves to deeper position or under
higher noise levels. On the other hand, the spline SL generally has good performance over
a broader source depths (from 0.5 to 0.7), and shows more robust characteristics against the
noise in potential measurement. Specifically, for the oneparameter spline SL estimator, the
CC values for all cases studied are greater than 0.92 under 10% potential noise, and equal
or greater than 0.90 under 5% potential noise plus 5% geometry noise.
6.3.2.2 Effects ofnumber ofrecording electrodes
Table 62 shows the effects of number of recording electrodes on the SL estimation.
One or multiple dipoles with varying orientations were placed in the spherical conductor
model (see note under Table 62 for dipole configurations), and 5% GWN was added to the
dipolesgenerated potentials. The CC values between the analytical SL and the estimated
SL with different electrode numbers and different dipole configurations are shown in Table
62. Similarly, three different SL estimators were evaluated and compared. The quantity
in equation (610) was set to 0.16, 0.18, 0.20, and 0.24 corresponding to 129,96,64, and
TABLE 62. The CC values between the analytical and estimated SL corresponding to different
electrode numbers and dipole configurations in onesphere model
Electrode Number
129 96 64 32
5PL 2SL ISL 5PL 2SL ISL 5PL 2SL ISL 5PL 2SL ISL
Config. A 0.84 0.95 0.98 0.78 0.97 0.97 0.88 0.96 0.97 0.72 0.95 0.84
Config. B 0.90 0.97 0.98 0.92 0.97 0.98 0.92 0.97 0.97 0.85 0.96 0.96
Config. C 0.77 0.96 0.99 0.82 0.94 0.97 0.80 0.94 0.96 0.59 0.92 0.88
Config. D 0.87 0.96 0.98 0.88 0.97 0.95 0.88 0.97 0.90 0.49 0.73 0.71
Config.E 0.74 0.94 0.98 0.75 0.89 0.94 0.77 0.92 0.92 0.47 0.67 0.83
Note: Configurations A: ITO at r= 0.6; B: IRO at r = 0.6; C: two +zdirection dipoles at (±0.3, 0.0, 0.5); 0: one lxdirection
dipole at (0.0, 0.0, 0.7) and two lzdirection dipoles at (0.0, ±OA, 0.5); E: 4RO at r = 0,6, each one is ttl] with respect to the
zaxis. ROradial dipole, TOtangential dipole.
Body Surface Laplacian Mapping of Bioelectric Sources 195
32 electrodes, respectively. The spline SL was estimated optimally by seeking the optimal
regularization parameter(s).
Table 62 clearly indicates the correlation between the goodness of SL estimation and
the number of surface electrodes. In general, the more electrodes being used, the higher
CC of the SL estimation. The CC values drop significantly when 32 electrodes are used,
which is consistent with the fact that a minimum sampling in the space domain is needed
to restore the spatial frequency spectrum. Again, Table 62 indicates that twoparameter
spline SL and oneparameter spline SL have comparable performance, and are more robust
against measurement noise than the 5point local SL estimation.
6.3.2.3 Effectsofregularization
In simulations, the optimal SL can be estimated by means of a priori information of
the analytical SL, i.e., by seeking the optimal parameter that maximizes the CC between
the analytical SL and the estimated SL. In real applications, the SL can also be estimated
without using the a priori information of the analytical SL, for example, by using the
discrepancy principle (Morozov, 1984). In Table 63, the upper rows show the CC between
the analytical SL and the optimal estimated SL. The lower rows show the CC between the
analytical SL and the estimated SL obtained by means of the discrepancy principle, without
the a priori information on the analytical SL. In this simulation, multiple dipoles with
varying orientations were placed in the spherical conductor model (see note under Table
63 dipole configurations), and 5% GWN was added to the analytical surface potentials
sampled at 129 recording electrodes. The quantity in equation (617) was set to 0.16 in
spline SL estimation.
Table 63 indicates that the regularization results always have lower CC than the
optimal results (by definition). However, the results obtained via regularization (by using
the discrepancy principle in this case) are comparable to the optimal SL estimates. Out of
four source configurations, the CC values for configurations B and D are almost similar
for these two types of results. For Configuration C, the CC of the regularization is smaller
than the optimal result by 1%. For configuration A, the CC of the regularization result is
smaller than the optimal result by less than 3%. However, the absolute CC is 96% or above,
suggesting the feasibility of the estimation of the SL through regularization.
Figure 64 depicts one typical example of the normalized surface potential maps and
the SL maps corresponding to source Configuration C in Table 63. In this figure, (A) is the
noisecontaminated surface potential map, (B) is the analytical Laplacian ECG map, (C) is
the optimal spline Laplacian ECG map estimated by means of a priori information, (D) is
TABLE63. Comparison of the optimalestimated splineSL and the splineSLestimated by using
the discrepancy principle in onesphere model
Dipole Configuration
Optimal spline SL
Regularized spline SL
0.99
0.96
0.98
0.98
C
0.98
0.97
0.98
0.98
Note: Configurations A: two +zdirection dipoles at (±O.3, 0.0, 0.5); B: one +xdirection dipole at (0.0, 0.0, 0.7) and two +z
direction dipoles at (0.0, ±O.4, 0.5); C: 4RD at r 0.6, each one is with respect to zaxis; D: 4TD at r 0.7, each one is
rr/4 with respect to zaxis. RDradial dipole, TDtangential dipole.
196 B. He and J. Lian
. r ~ . ~ 8 : J _ ~ 8 : J  ~ 8 : J . ~ ~
·1 0 1  0 1 I 0 1· 0 1 ·1 0 1
FIGURE 64. A typical example of the normalized potential map (A) and the Laplacian maps
See text for details. See the attached for color figure. (From He et al., 2002 with permission)
©
the spline Laplacian ECG map estimated by means of the discrepancy principle, and (E)
is the Laplacian ECG map estimated by the 5point SL estimator. Figure 64 indicates that,
from the viewpoint of imaging and mapping, the regularization spline SL estimate is almost
identical to the optimal spline SL estimate, and similar to the analytical SL result, for the
case studied. This is consistent with the high CC values obtained (Table 63) between the
analytical SL, the optimal spline SL, and the regularization spline SL estimates. Also noted,
the 5point local SL is more sensitive to the measurement noise as compared to the spline
SL, especially at the border regions.
6.3.2.4 Simulationin a realistic geometry hearttorso model
The performance of the present 3D spline SL estimator was further examined using
a realistic geometry hearttorso model, where cardiac electric activity was simulated by
pacing one or two sites in the ventricles of the heart model (He et al., 2002) (Figure
65A). The potential ECG induced by ventricular pacing was simulated by means of the
boundary element method (Aoki et al., 1987). The GWN was added to the simulated surface
potentials to simulate noisecontaminated potential BCGmeasurement. The Laplacian ECG
was estimatedfrom the noisecontaminated potential ECG using the 3Dspline SL algorithm,
and comparison was also made with the conventional 5point local SL estimation.
Figure 65B depicts the simulation results when simultaneously pacing two sites (site
#1 at the free wall of right ventricle, site #2 at the ventricular anterior) in the ventricular
base, and the single site pacing results corresponding to these two sites are shown in Figure
65C and Figure 65D, respectively. In these figures, (i) shows the activation sequence
inside the ventricles induced by the pacing. (ii) shows the 5% GWN contaminated body
surface potential map over the anterior chest immediately following the pacing. (iii) and
(iv) respectively show the estimated body surface Laplacian ECG maps over the anterior
chest, by using the 5point local SL estimator and the oneparameter spline SL estimator.
Note that in Figure 65B, the estimated Laplacian ECG maps provide multiple and more
localized areas of activity overlying the two pacing sites, whereas the body surface potential
map does not reveal the spatial details on this source multiplicity, due to the smearing effect
of the torso volume conductor. Also noted, that the spline SL estimate is less noisy and
can separate the two areas of activity more efficiently than the 5point local SL estimate. In
Figures 65CD, both the potential ECG and Laplacian ECG maps corresponding to single
site pacing reveal one pair of negative/positive activity, while the Laplacian ECG maps
Body Surface La placian Mapping of Bioelectric Sources 197
o
IIAX
41
.02
43
Nor
1
IIAX
lUX
O.
IIAX
42
43
0 O. 4. 0 O.
AX
.0 1
.02
43,
0 O.
1)1 .0 . 0 O.
B C
41
) 42
43
(I)
41
)42
43
.01
42
43
4.
FIGURE 65. Computer simulation of the spline Laplacian mapping in a realistic geometry hearttorso model.
(A) Hearttorso model and the locations of two pacing sites. (B) Dualsite pacing example. (C) Single site (#1)
pacing example. (D) Single site (#2) pacing example. See text for details. See the attached CD for color figure.
(From He et aI., IEEETBME, 2002 with permission) IEEE
provide much more localized spatial pattern. The two pairs of negative/posit ive activities
revealed in Figure 65B correspond well to the activities observed in Figures 65CD.
Consistently, the 5point local SL estimates are noisier than the spline SL estimates.
6.3.2.5 Spline Laplacian ECG mapping in Humans
Applying the spline Laplacian algorithm we have developed, body surface Laplacian
mapping has been explored in a group of healthy male subjects during ventricular and atrial
depolarization. Ninetyfive channel body surface potential ECG was recorded simultane
ously over the anterolateral chest in the subjects. The Laplacian ECG was estimated from
198
{  +2. ~ + f r .
1'3
P2
'. 2.55 '. _ .:0.25
B. He and J. Lian
.25
(A) (B) (C) (D)
FIGURE 66. Body surface potential and Laplacian maps of a healthy human Subject around the peak ofRwave.
Time instant is referenced to the onset of QRS and illustrated by a vertical line labeled in the Lead I ECG tracing
(A). The BSPM map is shown in (B). The spline BSLM map is shown in (C) with spatial details denoted by letter
'P' and 'N' followed with a numerical number for positive and negative activities, respectively. The corresponding
BSLM map estimated by the 5point SL estimator is shown in (0). The physical units of the color bars in the
BSPM and the BSLM maps are mVand mv/cmr, respectively. See the attached CO for color figure.
the recorded potentials during QRS complex and the Pwave by means of the oneparameter
spline SL estimator. For all subjects, more spatial details were observed in the SL ECG
maps as compared with the potential ECG maps, with spline SL more robust against noise
than the 5point SL (Li et al., 2003). Figure 66 shows one example of the SL ECG map
over the anterolateral chest of a healthy male subject around the peak of Rwave (Figure
66A). Figure 66B shows the potential map, which shows a pair of positivity and negativity
over the anterolateral chest. The corresponding spline BSLM map is shown in Figure 66C,
illustrating a localized negative activity, N2, located over the central chest, a positive activity
P2 slightly shifted toward the left lateral chest with respect to the position of N2, another
positive activity to the left of P2, and another negative activity N3 appeared in leftsuperior
area. Figure 66D shows the SL ECG map estimated using the 5point local SL estimator.
Note that the local 5point SL estimate (Figure 66D) shows more focused activities as
compared with the potential map (Figure 66B), but failed to reveal the spatial details as
illustrated in the spline SL map (Figure 66C). The negative and positive activities observed
in the group of human subjects have been related to the epicardial events (Li et al., 2003).
Figure 67 shows an example of spline SL mapping in a healthy human subject during
atrial depolarization (Lian et al., 2002b). Compared with the diffused potential map (Figure
67B), the corresponding spline BSLM map (Figure 67C) clearly shows two major positive
activities, PI and P2, representing the local maxima on the right and left anterior chest,
42 ms
A (8)
Nl
(C)
FIGURE 67. Body surface potential and Laplacian maps during the mid P wave in a healthy human subject.
(a) The potential P wave recorded from the left lower anterior chest, and the time instant for constructing the maps.
(b) The instantaneous BSPM map shows smooth pattern of potential distribution (colorbar unit: j..LV). (c) The
instantaneous BSLM map shows two major positive activities PI and P2, associated with three negative activities
NI, N2, and N3 (colorbar unit: j..LV/cm"), See the attached CO for color figure.
Body Surface Lapl acian Mapping of Bioelectric Sources 199
respectively. Correspondingly, three associated negative activities can alsobe observedand
denoted as Nl , N2, and N3, representing the local minima on right, middle and left chest
separated byPI andP2, respectively. Dataanalysis andcomputer simulationstudies suggest
that the positivities PI and P2 may correspond to the activation wavefronts in the right and
left atria, respectively (Lian et al., 2002b).
Compared to the smooth pattems of the BSPMs, more spatial details are revealed in
the BSLMmaps during ventricular and atrial activation, whichcould be correlated with the
underlying multiple myocardial activation wavefronts.
6.3.3 SURFACE LAPLACIAN BASED EPICARDIAL INVERSE PROBLEM
The feasibility on solving the ECGinverse problemby means of the LaplacianECG
(He, 1994; Wu et aI., 1995, 1998; He Wu, 1997, 1999; Johnston, 1997), and a hybrid
approach using both potential ECG and Laplacian ECG for epicaridial inverse problem
(He &Wu, 1999;Throne & Olson, 2000) have also been explored (see Section6.2.3).
As an example, Figure 68 shows the simulation results for testing the feasibility
of the Laplacian ECG based epicardial potential inverse solution in a realistically shaped
hearttorso model. Current dipoles located inside the anterior myocardium pointing from
endocardium to the epicardium were used to simulate anterior sources. To simulate the
noisecontaminatedexperimental recordings, upto20%GWNwasaddedto theheartmodel
generated body surface ECG signals before the epicardial potentials were reconstructed.
Figure 68 plots the relative error (RE) and CCbetween the epicardial potentials calculated
fromtwo anteriordipoles using the boundary element method, and the epicardial potentials
1.0 tt::a:::::::::::;t ~
.......................................................................
: : ~      :
c
C
o
~
0.11
0.6
11.3 0.2 11. 1
11.11 +    .     .  .
0.0
oise Level in Surface Signals
FI GURE 68. RE and CC values between the forward epicardial potent ials and the epicardial potentials recon
structed from the potential ECG and the Laplacian ECG in a realistically shaped heart torso model with two dipoles
located in the anterior ventricular wall. (From He Wu, Crit Rev BME, 1999 with permission from Begell House)
200 B. He and J. Lian
reconstructed from the noisecontaminated potential ECG and surface Laplacian ECG over
the whole torso. Note that the Laplacian ECG based epicardial inverse solutions always
provide smaller RE as compared with the potential ECG based epicardial inverse solutions
for the same noise level. For a larger noise level of 20% in the Laplacian ECG, the Laplacian
ECG based epicardial inverse solutions still show a comparable performance as compared
with the potential ECG based inverse solutions at a lower noise level of about 3%. Due to
increased noise level in Laplacian ECG as compared with the potential ECG, the real merits
of the Laplacian ECG based inverse solution would depend on how accurate one may record
or estimate the Laplacian ECG in an experimental setting. The data comparing the epicardial
inverse solutions obtained from the body surface potentials with those obtained from the
body surface Laplacians, which are estimated from the noisecontaminated potentials, are
currently lacking in the literature.
6.4 SURFACE LAPLACIANIMAGING OF BRAIN
ELECTRICAL ACTIVITY
6.4.1 HIGHRESOLUTION LAPLACIANEEG MAPPING
As a spatial enhancement method, the SL technique has been applied for many years
to highresolution EEG mapping. The SL has been considered an estimate of the local
current density flowing perpendicular to the skull into the scalp, thus it has also been
termed current source density or scalp current density (Perrin et aI., 1987a,b; Nunez et aI.,
1994). In addition, the relationship between the SL and the cortical potentials has also been
investigated (Nunez et aI., 1994; Srinivasan et aI., 1996). Since Hjorth's early exploration on
scalp Laplacian EEG (Hjorth, 1975), a number of efforts has been made to develop reliable
and easytouse SL techniques. Of noteworthy is the development of spherical spline SL
(Perrin et aI., 1987a,b), ellipsoidal spline SL (Law et aI., 1993), and the realistic geometry
spline SL (Babi1oni et al., 1996, 1998; He, 1999; Zhao & He, 2001; He et al., 2001).
6.4.2 PERFORMANCE EVALUATION OF THE SPLINE LAPLACIANEEG
In this section, we present the performance evaluation of the realistic geometry spline
Laplacian estimation algorithm in highresolution EEG mapping (He et al., 2001). The
evaluation was conducted by computer simulations using both a 3concentricsphere head
model (Rush Driscoll, 1969) and a realistic geometry head model. In addition, we ex
amined the performance of the spline SL algorithm in highresolution mapping of neural
sources using experimental visual evoked potential (VEP) data.
The realistic geometry spline SL estimation algorithm is detailed in Section 6.2.2. The
linear inverse problem in equation (617) was solved by the TSVD (Shim Cho, 1981), and
the truncation parameter was determined by means of the discrepancy principle (Morozov,
1984).
6.4.2.1 Effects ofnoise
The analytical SL was used to evaluate the numerical accuracy and reliability of the
spline SL estimation from the scalp potentials. The accuracy of the spline SL estimator was
evaluated by the CC and RE between the estimated and analytic SL distributions.
Body Surface Laplacian Mapping of Bioelectric Sources
1
(,)
.6
(,)
RE
a:
.4
:=:
~ :
.2
201
2 5 1
Hoi
15 2
FIGURE 69. Effects of noise on the spline SL estimation in a 3sphere inhomogeneous head model. (From He
et aI., Clin Neurophy, 2001 with permission)
Figure 69 shows an example of the simulation results for the effects of noise in the
3concentricsphere head model. A radial dipole (Rad) or a tangential dipole (Tag) located
at an eccentricity of 0.6 was used to represent a welllocalized areas of brain electrical
activity. The scalp potential and the scalp SL at 129 electrodes were calculated analytically
(Perrin et al., 1987a). Noise of up to 25% was added to the analytical potentials to simulate
noisecontaminated scalp potential measurements. For each noise level, ten trials of GWN
were generated and simulation conducted. The RE and the CC between the estimated SL
and the analytical SL for all ten trials were averaged and displayed in Figure 69. The SL was
estimated by minimizing the RE between the analytical SL and the SL estimate obtained by
solving equation (617) for a regularization parameter. The parameter in equation (610)
was set to 0.16.
Figure 69 indicates that the higher the noise level the larger the RE (or smaller the
CC). The CC value was greater than 96% for the radial dipole, and greater than 91% for
the tangential dipole, for up to 25% noise level. Figure 69 suggests that the oneparameter
realistic geometry SL estimator is robust against the additive white noise in the scalp
potenti al measurements.
6.4.2.2 Effects ofnumber ofrecording electrodes
Table 64 shows an example of the simulation results with different number of surface
electrodes in the 3concentricsphere head model. The RE and CC between the analytical
SL and the estimated spline SL with different electrode number and different dipole con
figurations are shown in Table 64. One or multiple dipoles with unity strength were placed
in the spherical conductor model (see note under Table 64 for detailed description of the
parameters of the dipole configurations). The GWN of specified noise level was added to
the scalp potentials to simulate noise contaminated EEG measurements. The parameter w
in equation (610) was set to 0.16,0.18,0.20 and 0.24 corresponding to 129,96,64 and 32
electrodes, respectively. The SL was estimated by minimizing the RE between the analytic
SL and the SL estimate.
Table 64 clearly indicates the correlation between the goodness of the SL estimation
and the number of surface electrodes. In general , the more electrodes used, the higher
202 B. He and J. Lian
TABLE64. The RE and CC values between analytical and estimated SL corresponding
to different electrode numbers and dipole configurations in 3sphere model
Electrode Number 129 96 64 32
Configuration A
RE 0.25 0.27 0.29 0.32
CC 0.97 0.96 0.96 0.95
Configuration B
RE 0.30 0.31 0.40 0.51
CC 0.95 0.95 0.92 0.86
Configuration C
RE 0.37 0.40 0.46 0.51
CC 0.93 0.92 0.89 0.86
Configuration D
RE 0.28 0.33 0.36 0.64
CC 0.97 0.94 0.93 0.77
Configuration E
RE 0.20 0.22 0.28 0.42
CC 0.98 0.98 0.96 0.91
Note: Configurations A: IRO at r 0.5, with 15%GWN; B: ITO at r 0.7, with 5% GWN; C: I dipole at (0.0,
0.1,0.75) pointing to +x direction, another dipole at (0.0, 0.1,0.6) pointing to +z direction, with 5%GWN; 0:
2RO at r 0.7, and ITO at r 0.6, each has an angle of rr/6 with respect to zaxis, with 5% GWN; E: 2RO at
xaxis and 2TO at yaxis, all at r 0.65, and each has an angle of rr/6 with respect to zaxis, with 10%. RO: radial
dipole. TO: tangential dipole.
CC (or lower RE) of the SL estimation. This phenomenon is consistent with the fact that a
minimum sampling in the space domain is needed to restore the spatial frequency spectrum.
As can be seen from Table 64, the CC for the SL estimation is 92% or larger for all
cases when 96 or more electrodes were used. The CC for Configuration C was about 89%
when 64 electrodes were used and 86% when 32 electrodes were used. This relatively
low CC values, as compared with other configurations may be explained by the large
eccentricity of the dipoles in the Configuration C. The closer the dipole is located to the
scalp, the sharper the spatial distribution of the scalp Laplacian. Thus higher spatial sampling
rate is desired. This phenomenon is further observed when only 32 electrodes are used, the
CC values dropped lower than 90% for Configurations B, C and D. Of interesting is the
low CC value for Configuration D when 32 electrodes were used. The CC dropped from
93%, when 64 electrodes were used, to 77% when 32 electrodes were used. Table 64
suggests that a highdensity electrode array of 96 or more is desirable for scalp spline SL
mapping.
6.4.2.3 Effects ofregularization
Table 65 shows examples of the simulation results comparing the SL estimate obtained
by means of a priori information of the analytical SL (denoted below as the optimal SL
estimate), and the SL estimate without using the a priori information of the analytical SL.
In this simulation, the discrepancy principle (Morozov, 1984) was used to determine the
truncation parameter of the TSVD procedure. In Table 65, the upper rows show the RE
and CC between the analytical SL and the optimal SL estimate. The lower rows show
the RE and CC between the analytical SL and the estimated SL obtained by means of the
discrepancy principle. In this simulation, 129electrodes were uniformly distributed over the
upper hemisphere, the parameter w in equation (610) was set to 0.16, and GWN of varying
noise level was added to the analytical scalp potentials to simulate noisecontaminated
potential measurements.
Body Surface Laplacian Mapping of Bioelectric Sources 203
TABLE65. Comparison of the optimal estimated spline SL and the spline SL esti
mated by using the discrepancy principle in 3sphere model
Dipole Configuration C D
Optimal RE 0.36 0.34 0.20 0.29
TSVD CC 0.95 0.94 0.98 0.98
Estimated RE 0.38 0.40 0.25 0.44
TSVD CC 0.94 0.93 0.97 0.91
ConfigurationsA: 2RDat r 0.70, each has an angle of with respect to zaxis, with 5%GWN; B:
2RDat r 0.6, and ITDat r 0.7, each has an angle of with respect to zaxis, with 10%GWN; C: 2RD
and 2TD, all at r =0.65, and each has an angle of n16with respect to zaxis, with 10%GWN; D: 4RDat r =
0.75, each has an angle of '!f17 with respect to zaxis, with 5% GWN. RD: radial dipole. TD: tangential dipole.
o
~
o.s
o
A
o.s
o
~
. , ~  ~    : '
B
FIGURE 610. Two examples of the normalized potential EEG and Laplacian EEG maps. See text for details.
See the attached CD for color figure. (From He et a!., CEn Neurophy, 2001 with permission)
Table65 indicatesthat the regularizationresults are always worsethanthe optimal SL
estimates (by definition), sincethe optimal SLestimates are obtainedby minimizingthe RE
betweenthe analyticalandestimatedSL. However, Table65showsthat the resultsobtained
via regularization (by using the discrepancy principle in this case) are comparable to the
optimal SLestimates. Out of four sourceconfigurations, the CCfor Configurations A, Band
Care almost similar for these twotypes of results. For Configuration D, the CCof the regu
larizationis smaller than the optimal SLestimate by about 6.5%. Howeverthe absoluteCC
is above91%, suggestingthe feasibilityof the estimationof the SL through regularization.
Figure610depictstwoexamplesof thenormalizedpotentialandLaplacianEEGdistri
butionscorrespondingtoConfiguration AandConfiguration BinTable65. For each source
configuration, the first panel shows is the noisecontaminatedscalp potential map, the sec
ondpanel showsthe analytical splineLaplacianEEGmap, the thirdpanel showsthe optimal
204
..
B. He and J. Lian
L
FIGURE 611. A realistic geometry head model built from one subject with two simulated dipoles located within
the brain. (From He Lian, Crit Rev BME, 2002 with permission from Begell House)
estimated Laplacian EEG map by means of a priori information, and the last panel shows
the regularization estimated Laplacian EEG map by means of the discrepancy principle.
6.4.2.4 Simulation in a realistic geometry head model
The computer simulation was also conducted by using a realistic geometry head model
built from one healthy subject (male, 34 years old) who was later involved in the VEP
experiment. The CT images of the subjects were obtained, and the BEM models of the scalp,
the skull, and the brain surfaces of the subject were constructed (Figure 611). Two artificial
dipoles were used to simulate two simultaneously active brain electrical sources. One dipole
source is located in right medial temporal lobe with orientation tangential to the cortical
surface, and another dipole is located in right inferior frontal lobe with radial orientation.
Figure 612 shows the result of spline Laplacian imaging of the simulated dipole
sources in realistic geometry head model. The scalp potentials generated by the two artificial
dipoles were contaminated with 5% GWN to simulate the measurement noise, and show
a blurred dipole pattern of distribution with frontal positivity and posterior negativity. The
spline Laplacian EEG map, however, effectively reduces the blurring effect caused by
the head volume conductor, and clearly reveals two localized activities corresponding to
the underlying dipole sources.
6.4.2.5 Surface Laplacian imaging of visual evokedpotential activity
Besides simulations, human VEP experiments were carried out to examine the per
formance of the spline SL estimator. The same above subject who gave written informed
consent was studied in accordance with a protocol approved by the UIC/IRB. Visual stimuli
were generated by the STIM system (Neuro Scan Labs, VA). The 96channel VEP signals
referenced to right earlobe were amplified with a gain of 500 and band pass filtered from 1
Hz to 200 Hz by Synamps (Neuro Scan Labs, VA), and were acquired at a sampling rate of
Body Surface Laplacian Mapping of Bioelectric Sources 205
FIGURE 612. Spline Laplacian mapping of the simulated dipole sources in a realistic geometry head model.
(A) Scalp potential map. (B) Estimated spline Laplacian EEG map. See the attached CD for color figure. (From
He Lian, Crit Rev BME, 2002 with permission from Begel! House)
1,000 Hz by using SCAN 4.1 software (Neuro Scan Labs, VA). The electrodes' locations
were measured using Polhemus Fastrack (Polhemus Inc. , Vermont). Full or half visual field
pattern reversal checkerboards (black and white) with reversal interval of 0.5 sec served as
visual stimuli and 300 reversals were recorded to obtain averaged VEP signals . The display
had a total viewing angle of 14.3
0
by ILl 0 , and the checksize was set to be 175' by 135'
expressed in arc minutes . The SL was estimated at the peak of the P100 component.
Figure 613 shows the recorded scalp potential maps and the estimated spline Laplacian
EEG maps at the PIOO peak time point of the pattern reversal VEP recorded from 96
A c
FIGURE 613. Spline Laplacian mapping of YEP activity in a human subject. (A) Scalp potential map elicited
by the full visual field stimuli. (B) Spline Laplacian EEG map in response to the full visual field stimuli. Scalp
potential map elicited by the left visual field stimuli. (D) Spline Laplacian EEG map in response to the left visual
field stimuli. See the attached CD for color figure. (From He et aI., Clin Neurophy, 2001 with permission)
206 B. He and J. Lian
electrodes over the scalp. As shown in Figure 613A, the scalp potential map elicited
by the full visual field stimuli is characterized by the strong but diffused activity that
distributed symmetrically over the occipital area. The estimated spline Laplacian EEG
map (Figure 613B), on the other hand, greatly improves the spatial resolution and clearly
reveals two dipolelike sources located in the visual cortices in both hemispheres. Notably,
the distribution of the positivity and negativity on each part of the scalp suggests orientation
of cortical current sources , pointing toward the contralateral hemisphere. As shown in Figure
613C, in response to the left visual field stimuli, a dominant positive potential component
was elicited with a widespread distribution on the left scalp. However, the estimated spline
Laplacian EEG map (Figure 613D) shows a dominant dipolelike current source located
in the right visual cortex. Similarly, the positivitynegativity distribution on the right scalp
in the spline Laplacian EEG map suggests orientation of cortical current sources, pointing
toward the contralateral hemisphere.
6.4.3 SURFACE LAPLACIAN BASED CORTICAL IMAGING
Using the procedure as detailed in Section 6.2.3, the application of using the scalp SL
to reconstruct the cortical potentials has also been explored (He, 1998b). Figure 614 shows
(a)
(c)
(b)
(d)
FIGURE 614. An example of cort ical imaging from scalp Laplacian EEG. (a) The "ture" cortical potential
distribution generated by four radial dipoles located at eccentricity of 0.8 in the 3sphere head model. (b)(d)
cortical potential distributions reconstructed from scalp Laplacian EEG with (b) 10%, (c) 30%, and (d) 50%
Gaussion white noise, respectively. See the attached CD for color figure. (From He, IEEEEMB, 1998 with
permission) © IEEE
Body Surface Laplacian Mapping of Bioelectric Sources 207
FIGURE 615. Cortical imaging of the SEP activity over a realistic geometry head model in a human subject,
using standard and the SL prefiltered WMN estimate. See text for details. See the attached CD for Color figure.
(From Babiloni et aI., MBEC, 2000 with permission)
an example of the simulation study based on the 3concentricsphere head model (Rush
Driscoll, 1969). Figure 614(a) shows the "true" cortical potential distribution generated
by four radial dipoles. Figures 614(b)(d) show the inverse cortical potential distributions
estimated from the scalp SL with 10%, 30%, and 50% GWN added to the scalp SL,
respectively. Notice that the higher the noise level in the scalp SL, the higher the background
noise level in the estimated cortical potentials. However, reconstruction of the multiple
extrema of the cortical potential distribution is quite robust, suggesting the feasibility and
unique feature of the cortical potential imaging from the scalp SL. In a separate study,
by investigating the spatial filter characteristics of the sourceLaplacian relationship,
Bradshaw and Wikswo (2001) demonstrated that dramatic improvement is evident in
the SLbased inverse solution, as compared with inverse reconstruction from the raw
data.
In another approach, the SL prefiltered EEG data was used as input for the weighted
minimum norm (WMN) linear inverse estimate of the cortical current sources, in order
to remove subcortically originated EEG potentials from the scalp potential distribution
(Babiloni et aI., 2000). As an example, Figure 615 shows the application of this technique
to the cortical imaging of the somatosensory evoked potentials (SEPs) over the realistic
geometry head model of one subject. For all the SEP components being examined (P20
N20, P22, N30P30), the cortical imaging inverse solutions have enhanced spatial resolution
than the scalp potential maps. Moreover, with respect to the WMN estimate, the SL pre
filtered WMN estimate presented enhanced spatial information content, in that the potential
maxima over the cortical surface were sharper and more localized.
208
DISCUSSION
B. He and J. Lian
The SL, as demonstrated by many investigators, enjoys enhanced spatial resolution and
sensitivity to regional bioelectrical activity located close to the surface recording electrodes,
and has unique advantage of reference independence. Conventionally, the localbased SL
operators have been used to estimate the Laplacian ECG or Laplacian EEG, by approxi
mating planar surface at the recording electrode. More accurate estimation can be achieved
by taking into account the realistic geometry of the body surface using the spline interpola
tion scheme. On the other hand, due to the highpass spatial filtering characteristics of the
local SL operator, amplification of the noise associated with the potential measurements
is unavoidable. Spatial lowpass filters, such as the Gaussian filter et al., 1994) and
Wiener filter (He, 1998a), have been shown to be useful in improving the signaltonoise
ratio of the SL. The spline SL, on the other hand, has been shown to provide an intrinsic
spatiallowpass filtering in addition to its spatialhighpass filtering characteristics (Nunez
et al., 1994; Srinivasan et al., 1998).
Estimation of the spline SL from potentials does not require the information on the
conductivity distribution inside the volume conductor. On the other hand, the spline SL es
timation techniques do need a mathematical model describing the geometry of the surface
over which the SL is to be estimated. The spline SL has been estimated over a spherical,
ellipsoidal, and a realistic geometry surface (Perrin et al., 1987a,b, 1989; Law et al., 1993;
Babiloni et aI., 1996, 1998; He, 1999; Zhao He, 2001; He et al., 2001, 2002). Further
more, the recently developed 3D spline SL algorithm (see Section 6.2.2) eliminates the
need of determining two spline parameters, and provides a rational determination of the
spline parameter through regularization process (He et al., 2001,2002). Such new approach
provides comparable computational accuracy and stability, while substantially reduces the
computational burden of optimizing two independent regularization parameters, as required
in the previously reported approaches.
The performance of the present realistic geometry spline SL estimator has been eval
uated through a series of computer simulations. Based on the onesphere homogeneous
volume conductor model, the simulation results demonstrate that the performance of the
oneparameter spline SL algorithm is comparable with that of the traditional twoparameter
spline SL algorithm (Tables 6163), but with much greater computational efficiency. The
simulation study also demonstrates that the spline SL estimators are more robust against
additive noise in both potential and geometry measurements as compared to the 5point
local SL estimator (Table 61), and consistent results are found for different numbers of
recording electrodes (Tables 62). An interesting finding is that the 5point local SL has
good performance only for shallow sources and under low noise level, while the spline SL
has good performance over a broader source depths and under variant noise levels (Table
61). This can be explained by the highpass spatial filter property of the 5point local SL
estimator, versus the bandpass spatial filter property of the spline SL estimator (Nunez
et al., 1994; Srinivasan et al., 1998). Table 63 and Figure 64 further suggest that the SLcan
be estimated by using the well established regularization techniques, such as the discrepancy
principle (Morozov, 1984), without apriori information of the "true" SL. In addition, based
on the 3sphere inhomogeneous volume conductor model, consistent evidence is shown
that the spline SL estimation algorithm is robust against noise in potential measurements
(Figure 69), provides consistent performance for different number of recording electrodes
(Table 64), and can be estimated by means of the discrepancy principle (Table 65).
Body Surface Laplacian Mapping of Bioelectric Sources 209
The application of the 3D spline SL estimator to the realistic geometry volume con
ductor models further suggests the potential use of the realistic geometry spline Lapla
cian ECGIEEG estimation. Figure 65 indicates that the Laplacian ECG maps provide a
muchlocalized projection onto the body surface in the areas directly overlying the heart,
and are especially useful in identifying and characterizing the source multiplicity as com
pared with the potential ECG maps. Similarly, Figure 612 indicates that the Laplacian EEG
map can effectively reduces the smoothing effect of the head volume conductor, and clearly
localizes the underlying brain electrical sources. The human YEP experiments further sug
gest the usefulness of the Laplacian EEG mapping. The full visual field stimuli elicited
symmetrical potential distribution about the midline of the occipital scalp. The estimated
Laplacian EEG map showed much more localized dipolar current sources in both visual
cortices, with dipolar orientations pointing toward the respective opposite hemisphere. is
widely accepted that the half visual field stimuli activates the visual cortex on the contralat
eral hemisphere of the brain. But paradoxically, the left visual field stimuli elicited stronger
positive potential distribution over the ipsilateral side of the scalp, which might be misin
terpreted as left visual cortex activation. However, by using the 3D spline SL method, the
estimated Laplacian EEG map clearly indicated that the right visual cortex was activated,
and these results are consistent with previous reports (Barrett et aI., 1976; Blumhardt et aI.,
1977; Towle et aI., 1995).
In summary, the SL, as a spatial enhancement method, can enhance the highfrequency
spatial components of the surface EEG and ECG. The 3D spline SL algorithm can take into
consideration of the realistic geometry of the body surface, and is applicable to both brain
and heart electrical source imaging. Only one spline parameter needs to be determined
through regularization procedure in this spline SL algorithm, thus enabling easy imple
mentation of the spline SL in an arbitrarily shaped surface of a volume conductor. Both
computer simulations and preliminary human experiments have demonstrated the excellent
performance of the 3D spline SL in highresolution ECG and EEG mapping, suggesting it
may become an alternative for noninvasive mapping of heart and brain electrical activity.
ACKNOWLEDGEMENT
The authors would like to thank their colleagues Dr. G. Li and Dr. D. Wu for useful
discussions. This work was supported in part by a grant from American Heart Association
#0140132N, NSF CAREER Award BES9875344, and NSF BES0201939.
REFERENCES
Aoki, M., Okamoto, Y, Musha, T.,and Harurni, K.: Threedimensional simulation of the ventricular depolarization
and repolarization processes and body surface potentials: normal heart and bundle branch block. IEEE Trans.
Biomed. Eng., 34: 4 5 ~ 6 2 , 1987.
Babiloni, E, Babiloni, C, Carducci, E, Fattorini, Onorati, P., and Urbano. A.: Spline Laplacian estimate of
EEG potentials over a realistic magnetic resonanceconstructed scalp surface model. Electroenceph. din.
Neurophysiol., 98: 363373,1996.
Babiloni, E, Carducci, E, Babiloni, and Urbano, A.: Improved realistic Laplacian estimate of highlysampled
EEG potentials by regularization techniques. Electroenceph. Clin. Neurophysiol., 106: 336343, 1998.
Babiloni, E, Babiloni, C, Locche, Cincotti, E, Rossini, P.M., and Carducci, E: Highresolution electro
encephalogram: source estimates of Laplaciantransformed somatosensoryevoked potentials using a realistic
210 B. HeandJ. Lian
subject head model constructed from magnetic resonance images. Med. Bioi. Eng. Comput., 38: 512519,
2000.
Barr, R.C., Ramsey, M. III, and Spach, M.S.: Relating epicardial to body surface potential distributions by
means of transfer coefficients based on geometry measurements. IEEE Trans. Biomed. Eng., 24: III,
1977.
Barrett, G., Blumhardt, L., Halliday, A.M., Halliday, E., and Kriss, A.: A paradox in the lateralisation of the visual
evoked response. Nature, 261: 253255,1976.
Besio, WG; Lu, C.C., and Tarjan, P.P.: A feasibility study for body surface cardiac propagation maps of humans
from Laplacian moments of activation. Electromagnetics, 21: 621632, 2001.
Blumhardt, L.D., Barrett, G., and Halliday, A.M.: The asymmetrical visual evoked potential to pattern reversal
in one half field and its significance for the analysis of visual field defects. British Ophthalmology, 61:
454461, 1977.
Bradshaw, L.A., and Wikswo, J.PJr.: Spatial filter approach for evaluation of the surface Laplacian of the elec
troencephalogram and magnetoencephalogram. Ann. Biomed. Eng., 29: 202213, 2001.
Courant, R. and Hilbert, D.: Methods of mathematical physics. New York: Interscience. 1966.
Duchon, J.: Interpolation des fonctions de deux variables suivant Ie principe de la flexion des plaques minces.
R.A.I.R.O. Anal. Num., 10: 512,1976.
Fattorusso, Thaon, M., Tilmant, 1.: Contribution of I'etude de I'electrocardiogramme precordial. Acta Cardi
ologica., 4: 464487,1949.
Harder, R. and Desmarais, R.: Interpolation using surface spline. 1. Aircraft, 9: 189191, 1972.
He, B. and Cohen, RJ.: Body surface Laplacian ECG mapping. IEEE Trans. Biomed. Eng., 39: 11791191,
I992a.
He, B. and Cohen, RJ.: Body surface Laplacian mapping of cardiac electrical activity. Am. J. Cardiol., 70:
16171620, 1992b.
He, B., Kirby, D., Mullen, T., and Cohen, RJ.: Body surface Laplacian mapping of cardiac excitation in intact
pigs. Pacing Clin. Electrophysiol., 16: 10171026, 1993.
He, B.: On the Laplacian inverse electrocardiography. Proc. Ann. Int. Conf. IEEE Eng. Med. BioI. Soc., 145146,
1994.
He, B., Chernyak, Y, and Cohen, RJ.: An equivalent body surface charge model representing three dimensional
bioelectrical activity. IEEE Trans. Biomed. Eng., 42: 637646, 1995.
He, B. and Cohen, RJ.: Body surface Laplacian ECG mappingA review. Crit. Rev. Biomed. Eng., 23: 475510,
1995.
He, B.: Principles and applications ofthe Laplacian electrocardiogram. IEEE Eng. Med. BioI. Mag., 16: 133138,
1997.
He, B. and Wu, D.: A bioelectric inverse imaging technique based on surface Laplacians. IEEE Trans. Biomed.
Eng., 44: 529538, 1997.
He, B., Yu, X., Wu, D., and Mehdi, N.: Body surface Laplacian mapping of bioelectrical activity. Methods In!
Med., 36: 326328, 1997.
He, B.: Theory and applications of bodysurface Laplacian ECG mapping. IEEE Eng. Med. BioI. Mag., 17:
102109,1998a.
He, B.: High resolution source imaging of brain electrical activity. IEEE Eng. Med. BioI. Mag., 17: 123129,
I998b.
He, B.: Brain electrical source imaging: Scalp Laplacian mapping and cortical imaging. Crit. Rev. Biomed. Eng.,
27: 149188, 1999.
He, B. and Wu, D.: Laplacian electrocardiography. Crit. Rev. Biomed. Eng., 27: 285338,1999.
He, B., Lian, 1., and Li, G.: Highresolution EEG: a new realistic geometry spline Laplacian estimation technique.
Clin. Neurophysiol., 112: 845852,2001.
He, B., Li, G., and Lian, J.: A spline Laplacian ECG estimator in a realistic geometry volume conductor. IEEE
Trans. Biomed. Eng., 49: 110117, 2002.
He, B., Lian, J.: Spatioternporal Functional Neuroimaging of Brain Electric Activity. Critical Review ofBiomedical
Engineering, 30: 283306, 2002.
Hjorth, B.: An online transformation of EEG scalp potentials into orthogonal source derivations. Electroenceph.
Clin. Neurophysiol., 39: 526530, 1975.
Johnston, P.R.: The Laplacian inverse problem of electrocardiography: an eccentric spheres study. IEEE Trans.
Biomed. Eng., 44: 53948,1997.
BodySurfaceLaplacianMappingof BioelectricSources 211
Law, S.K., Nunez, P.L., and Wijesinghe, R.S.: Highresolution EEG using spline generated surface on spherical
and ellipsoidal surfaces. IEEE Trans. Biomed. Eng., 40: 145153, 1993.
Le, J., Menon, V, and Gevins, A.: Local estimate of surface Laplacian derivation on a realistically shaped
scalp surface and its performance on noisy data. Electroenceph. Clin. Neurophysiol., 92: 433441,
1994.
Lian, 1., Srinivasan, S., Tsai, H., and He, B.: Comments on "Is accurate recording of the ECG surface Laplacian
feasible?" IEEE Trans. Biomed. Eng., 48: 610613, 2001.
Lian, J., Srinivasan, S., Tsai, H., Wu, D., and He, B.: On the estimation of noise level and signal to noise ratio of
Laplacian ECG during ventricular depolarization and repolarization. Pacing Clin. Electrophysiol., 25(10):
14741487,2002.
Lian, J., Li, G., Cheng, J., Avitall, B., and He, B.: Body surface Laplacian mapping of atrial depolarization in
healthy human subjects. Med. Bio!. Eng. Comput., 40(6): 650659, 2002b.
Li, G., Lian, J., He, B.: On the Spatial Resolution of Body Surface Potential and Laplacian Pace Mapping. Pacing
and Clinical Electrophysiology, 25: 420429, 2002.
Li, G., Lian, Salla, P., Cheng, Ramachandra, 1., Shah, P., Avitall, B., and He, B.: Body surface Laplacian
electrogram of ventricular depolarization in normal human subjects. Cardiovasc. Electrophysiol., 140):
1627,2003.
Mirvis, D.M., Keller, EW., Ideker, RE., Cox, Zettergren, D.G., and Dowdie, RJ.: Values and limitations
of surface isopotential mapping techniques in the detection and localization of multiple discrete epicardial
events. 1. Electrocardiol., 10: 347358,1977.
Morozov, VA.: Methods for solving incorrectly posed problems. Berlin: SpringerVerlag, 1984.
Nunez, PL.: Electric field ofthe brain. London: Oxford University Press, 1981.
Nunez P.L.: Neocortical dynamics and human EEG rhythms. New York: Oxford University Press, 1995.
Nunez, P.L., Silibertein, RB., Cdush, PJ., Wijesinghe, R.S., Westdrop, A.E, and Srinivasan, R.: A theoretical and
experimental study of high resolution EEG based on surface Laplacian and cortical imaging. Electroenceph.
Clin. Neurophysiol., 90: 4057,1994.
Oostendorp, T.E and van Oosterom, A.: The surface Laplacian of the potential: theory and application. IEEE
Trans. Biomed. Eng., 43: 394403, 1996.
Perrin, E, Bertrand, and Pernier, J.: Scalp current density mapping: value and estimation from potential data.
IEEE Trans. Biomed. Eng., 34: 283288, 1987a.
Perrin, Pernier, Bertrand, Giard, M.H., and Echallier, Mapping of scalp potentials by surface spline
interpolation. Electroenceph. Clin. Neurophysiol., 66: 7581, 1987b.
Perrin, Pernier, Bertrand, and Echallier, Spherical splines for scalp potential and current density
mapping. Electroenceph. Clin. Neurophysiol., 72: 184187, 1989.
Rudy, Y and Plonsey, R.: A comparison of volume conductor and source geometry effects on body surface and
epicardial potentials. eire. Res., 46: 283291,1980.
Rush, S. and Driscoll, D.A.: EEG electrode sensitivityan application of reciprocity. IEEE Trans. Biomed. Eng.,
16: 1522, 1969.
Shim, YS. and Cho, Z.H.: SVD pseudoinversion image reconstruction. IEEE Trans. Acoust. Speech. Processing,
29: 904909,1981.
Spach, M.S., Barr, R.C., Lanning, C.E, and Tucek, P.c.: Origin of body surface QRS and Twave potentials from
epicardial potential distributions in the intact chimpanzee. Circulation, 55: 268278, 1977.
Srinivasan, R., Nunez, P.L., Tucker, D.M., Silberstein, RB., Cadusch, PJ.: Spatial sampling and filtering of EEG
with spline Laplacian to estimate cortical potentia!. Brain Topography, 8(4): 355366, 1996.
Srinivasan, R, Nunez, P.L. and Silberstein, R.B.: Spatial filtering and neocortical dynamics: estimates of EEG
coherence. IEEE Trans. Biomed. Eng., 45: 814826,1998.
Throne, R.D. and Olson, L.G.: Fusion of body surface potential and body surface Laplacian signals for electro
cardiographic imaging. IEEE Trans. Biomed. Eng., 47: 452462, 2000.
Towle, VL., Cakmur, R., Cao, Y, Brigell, M., and Parmeggiani, L. Locating VEP equivalent dipoles in magnetic
resonance images. Int. Neurosci., 80: 105116, 1995a.
Tsai, H., Ceccoli, H., Avitall, B., and He, B.: Body surface Laplacian mapping of anterior myocardial infarction
in man. Electromagnetics, 21: 607619, 2001.
Umetani, Okamoto, Y, Mashima, S., Ono, Hosaka, H., and He, B.: Body Surface Laplacian mapping
in patients with left or right ventricular bundle branch block. Pacing Clin. Electrophysiol., 21: 20432054,
1998.
212 B. He andJ. Lian
Wang, and Begleiter,H.: Local polynomial estimate of surface Laplacian. BrainTopogr., 12: 1929, 1999.
Wei,D., Harasawa,E., andHe, B.: Simulatedbodysurfacepotentialand Laplacianmapsduringtheleft ventricular
breakthrough. Proc. ofthe Ann. Int. Can! Of IEEE Eng. In Med. Biol. Soc., 223224, 1995.
Wei, D., Mashima, S.: Predictionof accessorypathwaylocations in WolffParkinsonWhite syndromewith body
surface potential Laplacianmaps: A simulationstudy. Heart 40: 451459,1999.
Wei, D.: Laplacianelectrocardiogramssimulatedusing realisticallyshaped hearttorsomodel during normal and
abnormal ventriculardepolarization. Electromagnetics, 21: 593605, 2001.
Wu, D., Saul, J.P., and He, B.: Epicardial InverseSolutions fromBody Surface LaplacianMaps: A Model Study.
Proc. ofthe Ann. Int. Can! Of IEEE Eng. In Med. Biol. Soc., Montreal, 1995.
Wu, D., Wang, Y.,and He, B.: Reconstructionof epicardial potentialsfrombody surfaceLaplacianmaps by using
a realisticallyshaped hearttorsomodel. Proc. Ann. Int. Can! IEEE Eng. Med. Biol. Soc., 8385, 1998.
Wu, D., Tsai, H.C., andHe, B.: On the estimationof theLaplacianelectrocardiogramduringventricularactivation.
Ann. Biomed. Eng., 27: 731745, 1999.
Zhao, and He, B.: A newalgorithm for estimating scalp Laplacian EEG and its application to visualevoked
potentials. Electromagnetics, 21: 633640, 2001.
7
NEUROMAGNETIC SOURCE
RECONSTRUCTIONANDINVERSE
MODELING
Kensuke Sekihara' and Srikantan S. Nagarajarr'
Departmentof ElectronicSystems and Engineering, Tokyo Metropolitan Instituteof Technology,
Asahigaoka 66, Hino, Tokyo1910065, Japan
2Department of Radiology, University of California, San Francisco, 513 Parnassus Avenue, S362
San Francisco, CA 94143, USA
7.1 INTRODUCTION
The human brain has approximatel y neurons in its cerebral cortex. Their electro
physiological acti vity generates weak but measurable magnetic fields outside the scalp.
Magnetoencephalography (MEG) is a method which measures these neuromagnetic fields
to obtain information about these neural acti vities (Hamalainen et al., 1993; Roberts et al. ,
1998; Lewine et al., 1995). Among the various kinds of funct ional neuroimaging methods,
such a neuroelectromagnetic approach has a major advantage in that it can provide fine
time resolution of millisecond order. Therefore, the goal of neuromagnetic imaging is to vi
sualize neural activit ies with such fine time resolution and to provide functional information
about brain dynamics. To attain this goal, one technical hurdle must be overcome. That is,
an efficient method to reconstruct the spatiotemporal neural activities from neuromagnetic
measurements needs to be developed. Toward this goal, a number of algorithms for recon
structing spatiotemporal source activities have been investigated (Baillet et al., 2001) .
This chapter deals with this neuromagnetic reconstruction problem. However, we do
not provide a general review of various algorithms for this reconstruction problem. Instead,
we describe a particular class of source reconstruction techniques referred to as the spatial
filter, whi ch allows the spat iotemporal reconstruction of neural activities without assuming
any kind of source model. Furthermore, among the spatial filter techniques, we focus on
adapti ve spatial filter techniques. The se techniques were origi nally developed in the fields
of array signal processing, including radar, sonar, and seismic exploration , and have been
Corresponding author: Kensuke Sekihara Ph.D., Tokyo Metropolitan Institute of Technology, Asahigaoka 66,
Hino, Tokyo 1910065, Japan, Tel:81425858642, Fax:81425858642, Email: ksekiha@cc.tmit.ac.jp
213
214 K. Sekihara and S. S. Nagarajan
widely used in such fields (van Veen et al., 1988). Nonetheless, the adaptive spatial filter
techniques are relatively less acknowledged in the MEG/EEG community.
In this chapter, we also formulate reconstruction techniques based on linear least
squares methods (Hamalainen and Ilmoniemi, 1984) as the nonadaptive spatial filter. This
formulation enables us to compare the leastsquaresbased techniques with the adaptive
spatial filter techniques on a common, unified basis. Actually, we compare these two types
of techniques by using the same figure of merit called the resolution kernel, and show that the
adaptive techniques can provide much higher spatial resolution than the leastsquaresbased
methods.
The organization of this chapter is as follows: Following a brief review of the neu
romagnetometer hardware in Section 7.2, we describe the forward modeling and some
basic properties of MEG signals in Section 7.3. Section 7.4 presents the formulation of the
linear leastsquaresbased methods as the nonadaptive spatial filter. Section 7.5 describes
the adaptive spatial filter techniques. In Section 7.6, we present a quantitative compari
son between the adaptive and nonadaptive methods using the resolution kernel criterion.
Section 7.7 presents a series of numerical experiments on the adaptive spatial filter per
formance. In Section 7.8, we demonstrate the effectiveness of the adaptive spatial filter
techniques by applying them to two sets of MEG data.
7.2 BRIEFSUMMARY OFNEUROMAGNETOMETER HARDWARE
is generally believed that the neuromagnetic field is generated by the postsynaptic
ionic current in the pyramidal cells of cortical layer III. Here, neuronal cells are organized
into socalled columnar structure, and the synchronous activity of these cells results in
superimposed magnetic fields strong enough to be measured outside the human head. The
average intensity of this neuromagnetic field, however, is around a few hundred femto
Tesla To measure such an extremely weak magnetic field, a neuromagnetometer uses
a special device called a superconducting quantum interference device (SQUID) (Clarke,
1994; Drung et al., 1991). This device is so sensitive that it can in principle measure a single
quantum of magnetic flux.
When measuring such a weak neuromagnetic field, a major problem arises from the
background environmental magnetic noise. Background magnetic noise is generated by
electronic appliances such as computers, power lines, cars, and elevators. Such noise is
common at a site where the neuromagnetometer is installed, and its average intensity is
usually five to six orders of magnitude greater than the neuromagnetic field. One obvious
way to reduce it is to use a magneticallyshielded room. However, a typical mediumquality
shielded room, most commonly used in MEG measurements, can reduce the background
noise by up to only three orders of magnitude.
To further reduce the background noise, neuromagnetometers are usually equipped
with a special type of detector configuration called a gradiometer (Hamalainen et al., 1993;
Lewine et al., 1995). The firstorder gradiometer consists of two coils of exactly the same
area; they are connected in series, but wound in opposite directions. Therefore, the gra
diometer cancels the electric current induced by the background noise fields because the
One IT is equivalent to I x 10
Neuromagnetic Source Reconstruction and Inverse Modeling 215
sources of such noise fields are generally far from the gradiometer and induce nearly the
same amount of electric current in both coils. The gradiometer can achieve two to three
orders of magnitude reduction in the background noise, and can remove the influence of
the residual noise field within the magneticallyshielded room. The reduction performance,
however, depends on the manufacturing precision of the two coils.
Aside from the gradiometer, several other methods of removing the external noise have
been investigated (Vrba and Robinson , 2001; Adachi et ai., 2001). Many of these methods
use extra sensors that measure only the noise fields. (Such sensors are usually located apart
from the sensor array which measures the neuromagnetic fields.) Quasirealtime electronics
then perform the online subtraction between the outputs from the extra sensors and from
the regular sensors. A method that does not require additional sensor channels has also
been developed. This method applies a technique called signalspace projection, and can
be implemented completely as a postprocessing procedure (Parkkonen et ai., 1999). These
external noise cancellation methods make it possible to use a magnetometer as a sensor
coil instead of the gradiometer. They also permit the measurement of neuromagnetic fields
outside a magneticallyshielded room.
The most remarkable advance in neuromagnetometer hardware over the last ten years
has been the rapid increase in the number of sensors. Since neuromagnetometers with
a 37channel sensor array became commercially available in the late 80s (Lewine et al.,
1995), the number of sensors in commercially available neuromagnetometers has constantly
increased. The latest neuromagnetometers are equipped with 200300 sensor channels, with
wholehead coverage of the sensor array.
7.3 FORWARD MODELING
7.3.1 DEFINITIONS
Let us define the magnetic field measured by the rnth detector coil at time as
and a column vector as a set of measured data where
is the total number of detector coils and the superscript indicates the matrix transpose . A
spatial location is represented by a threedimensional vector = y, The source
current density at and time is defined as a threedimensional column vector t) .
The magnitude of the source current is denoted as and the orientation
of the source is defined as a threedimensional column vector t)
t), t), whose component (where equals y, or z in this chapter),
is equal to the cosine of the angle between the direction of the source moment and the
direction.
Let us define l!n(r) as the rnth sensor output induced by the unitmagnitude source
located at and directed in the direction. The column vector is defined as
Then, we define a matrix which represents the sensi
tivity of the whole sensor array at as The rnth row of
represents the sensitivity at of the rnth sensor. Then, using
the superposition law, the relationship between and s is expressed as
(7.1)
216 K. Sekihara and S. S. Nagarajan
Here, is the noise vector at The sensor sensitivity pattern, represented by the matrix
is customarily called the sensor lead field (Hamalainen 1993; Sarvas, 1987),
and this matrix is called the leadfield matrix. We define, for later use, the leadfield vector
in the sourcemoment direction as which is obtained by using
7.3.2 ESTIMATION OF THE SENSOR LEAD FlEW
The problemof the source reconstruction is the problem of obtaining the best estimate
of from the array measurement is thus apparent that, to solve this recon
struction problem, we need to have a reasonable estimate of the leadfield matrix. In this
subsection, we describe how we can obtain the lead field using the spherically symmetric ho
mogeneous conductor model, which is most commonly used in estimating the MEG sensor
lead field. Also, we briefly mention a realisticallyshaped volumeconductor model, which
can generally provide more accurate lead field estimates, particularly for nonsuperficial
brain regions. Estimation of the leadfield is called the forward problem, because this is
equivalent to estimating the magnetic field from a point source located at a known location;
this problem stands in contrast to the inverse problem in which the source configuration is
estimated from a known magnetic field distribution.
Let us define the electric potential as the magnetic field as and the electric
current density as j. The source current defined in Section 7.3.1 is called the primary
current, (alternatively called the impressed current), which is directly generated from the
neural activities. There is another type of electric current called the return current or volume
current. results from the electric field in the conducting medium, and it is not directly
caused by the neural activities. Defining the conductivity as the return current is expressed
as  pV where  V is equal to the electric field. Thus, the total electric current is
expressed as
(7.2)
The relationship between the total current and the resultant magnetic field is given
by the BiotSavart law,
r  r'
itr': x
4n
(7.3)
where is the magnetic permeability of the free space. In order to derive the analytical
expression for the relationship between the primary source current and the magnetic field,
we first consider the case in which a whole space is filled with a conductor with constant
conductivity In this case, it is easy to show that the following relationship holds:
fir r'
4n
(7.4)
Here the magnetic field is denoted as
o
for later convenience. Note that Eq. (7.4) is similar
to Eq. (7.3). The only difference is that the total current density is replaced by the
primary current density
Neuromagnetic Source Reconstruction and Inverse Modeling 217
We then proceed to deriving a formula for the magnetic field outside a spherically
symmetric homogeneous conductor, B. To do so, we make use of the fact that the radial
component of the magnetic field is equal to the radial component of
o
in Eq. (7.4)
(Cuffin and Cohen, 1977; Sarvas, 1987), i.e.,
o.
(7.5)
where is the unit vector in the radial direction defined as (Note that we set
the coordinate origin at the sphere origin.) The relationship V x 0 holds outside the
volume conductor, because there is no electric current. Thus, can be expressed in terms
of the magnetic scalar potential
This potential function is derived from
00 00
0 0
(7.6)
(7.7)
where we use the relationship in Eq. (7.5). By substituting Eq. (7.4) intoEq. (7.7), we finally
obtain
x
A '
where
The formula for is then obtained by substituting Eq. (7.8) into Eq. (7.6), i.e.,
A
where
(7.8)
(7.9)
2
VA r',
To obtain the component of the leadfield matrix, we first calculate
m
)
(where m is the mth sensor location) by using Eq. (7.9) with where
is the unit vector in the direction. When the sensor coil is a magnetometer coil, (which
measures only the magnetic field component normal to the sensor coil), is calculated
from m) . where I'::il is a unit vector expressing the normal direction of
the mth sensor coil. When the sensor coil is a firstorder axial gradiometer with a baseline
of is calculated from m) . I,::il  m rr. This
218 K. Sekihara and S. S. Nagarajan
represents the sensitivity of the sensor to the primary current density located at and
directed in the ~ direction.
One of the important properties of the lead field obtained using the spherically
symmetric homogeneous conductor model is that if and are parallel, i.e., if the
primary current source is oriented in the radial direction, no magnetic fields are generated
outside the spherical conductor from such a radial source. Also, we can see that when
approaches the center of the sphere, l ~ 0) becomes zero, and no magnetic field is generated
outside the conductor from a source at the origin.
The sphericallysymmetric homogeneous conductor is generally satisfactory in ex
plaining the measured magnetic field when only superficial sources exist , i.e., when all
sources are located relati vely close to the sensor array. This is because the curvature of the
upper half of the brain is well approximated by a sphere. However, for sources located in
lower regions of the brain, the model becomes inaccurate because the curvature of the lower
brain regions significantly differs from a sphere. Such errors caused by misfits of the model
may be reduced by using a spheroidallysymmetric conductor model (Cuffin and Cohen ,
1977) or an eccentric sphere conductor model (Cuffin, 1991).
More fundamental improvements can be obtained by using realisticallyshaped
volumeconductor models. Such conductor models can be constructed by first extracting the
brain boundary surface from the subject's 3D MRI. We denote this surface ~ . We then use
the following Geselowitz formula (Geselowitz, 1970; Sarvas , 1987) to calculate magnetic
fields outside the volume conductor:
J.l o l ' r  r'
x
:E
(7.10)
where the integral on the righthand side indicates the surface integral over ~ ; r' represents
a point on ~ , and!:E is a unit vector perpendicular to ~ at Here, we assume that the
conductivity within the brain is uniform, and denote it The second term on the righthand
side of Eq. (7.10) represents the influence from the volume current, and to calculate this
term, we need to know on ~ , which is obtained by solving (Sarvas, 1987)
where
p p , r r'
:E
r r' ,
= 
(7.11)
(7.12)
and rand r' are points on the boundary surface.
Because the brain boundaries are irregular, the magnetic field can only be ob
tained numerically using the boundary element method (BEM). In this calculation, we
first estimate the electric potential on the brain boundary surface ~ by numerically
solving Eq. (7. 11). We then calculate magnetic fields out side the brain using Eq. (7.10).
The details of these numerical calculations are out of the scope of this chapter, and they
are found in (Barnard et al., 1967; Hamalainen and Sarvas, 1989). The numerical method
Neuromagnetic Source Reconstruction and Inverse Modeling 219
mentioned so far assumes uniform conductivity within the brain boundary, and is called
singlecompartment BEM. It is usually used in estimating MEG sensor lead fields (Fuchs
et al., 1998). It can be extended to the multiplecompartment BEM and such models are
usually used for estimating the EEG sensor lead fields. The BEMbased realisticallyshaped
volume conductor models generally provide significant improvements in the accuracy of
the forward calculation particularly for deep sources (Fuchs et al., 1998; Cuffin, 1996), al
though they are computationally expensive. Improvements in the computational efficiency
of the BEM have been reported (Bradley et al., 2001; van't Ent et al., 2001).
7.3.3 LOWRANKSIGNALS AND THEIR PROPERTIES
Let us consider specific cases where the primary current consists of localized discrete
sources. The number of sources is denoted and we assume that is less than the number
of sensors The locations of these sources are denotedr , .r, ... ,rQ. The sourcemoment
distribution is then expressed as
Q
t)
q=]
(7.13)
where and this integral extends the small region around where the
qth source is confined. This type of localized source is called the equivalent current dipole
with moment sD. Here, sD is called the moment because it has a dimension of current x
distance. The basis underlying the equivalent current dipole is physiologically plausible
(Okada et al., 1987), and sources of neuromagnetic fields are often modeled with the current
dipoles.
Since there is little advantage to explicitly differentiating the current density v t)
and the current moment for simplicity we keep the same notation to
express the current moment. Then, the Qdimensional source magnitude vector is defined
as vet) t), t), ... , t)]T. We define a 3Q x Q matrix that expresses the
orientations of all sources as Wet) such that
[
T J ( r ~ , t)
lJi"(t)=
o
o
o
The composite leadfield matrix for the entire set of Q sources is defined as
(7.14)
Then, substituting Eq. (7.13) into Eq. (7.1), we have the discrete form of the basic relation
ship between bet) and vet) such that
bet) [LclJi"(t)]v(t) net). (7.15)
220 K. Sekihara and S. S. Nagarajan
Let us define the measurement covariance matrix as
b
; i.e.,
b
T where
(.) indicates the ensemble average. (This ensemble average is usually replaced by the time
average over a certain time window.) Let us also define the covariance matrix of the source
moment activity as R
s
; i.e., R, Then, using Eq. (7.15), we get the
relationship between the measurement covariance matrix and the sourceactivity covariance
matrix such that
(7.16)
where the noise in the measured data is assumed to be white Gaussian noise with a variance
of and I is the M x M identity matrix.
Let us define the kth eigenvalue and eigenvector of R
b
as Ak and ei, respectively. Let
us assume, for simplicity, that all sources have fixed orientations. Then, unless some source
activities are perfectly correlated with each other, the rank of R, is equal to the number
of sources Q. Therefore, according to Eq. (7.16), has Q eigenvalues greater than
and M  eigenvalues that are equal to The signal whose covariance matrix has such
properties is referred to as the lowrank signal (Paulraj et al., 1993; Sekihara et al., 2000).
Let us define the matrices and as = and =
The column span of E
s
is the maximumlikelihood estimate of the signal subspace of R
b
,
and the span of is that of the noise subspace (Scharf, 1991). In the lowrank signals, the
measurement covariance matrix R
b
can be decomposed into its signal and noise subspace
components; i.e.,
Here, we define the matrices As and AN as
As diag[AI, ... , AQ] and AN diag[AQ+l, ... ,
(7.17)
(7.18)
where diagl .. ] indicates a diagonal matrix whose diagonal elements are equal to the entries
in the brackets.
The most important property of the lowrank signal is that, at source locations, the
leadfield matrix is orthogonal to the noise subspace of This can be understood by first
considering that
(7.19)
Since L; is a full columnrank matrix, and we assume that R, is a full rank matrix, the
above equation gives
(7.20)
This implies that the leadfield matrices at the true source locations are orthogonal to any
noise level eigenvector; that is, they are orthogonal to the noise subspace (Schmidt, 1981),
Neuromagnetic Source Reconstruction and Inverse Modeling
i.e.,
221
(7.21)
Since the equation above holds, the relationship 0 also holds. These orthog
onality relationships are the basis of the eigenspaceprojection adaptive beamformer de
scribed in Section 7.5.2, as well as the basis of the wellknown MUSIC algorithm (Schmidt,
1986; Schmidt, 1981; Mosher et al. , 1992).
7.4 SPATIAL FILTER FORMULATIONAND NONADAPTIVE SPATIAL
FILTER TECHNIQUES
7.4.1 SPATIAL FILTER FORMULATION
Spatial filter techniques estimate the source current density (or the source moment) by
applying a linear filter to the measured data. Because the source is a threedimensional vector
quantity, there are two ways to implement the spatial filter approach in the neuromagnetic
source reconstruction. One is the scalar spatial filter and the other is the vector spat ial filter.
In the scalar approach, we use a single set weights that characterizes the properties
of the spatial filter, and define the set of the filter weights as a column vector 7]) =
[WI 7]) , 7]) , . . . , Here the weight vector depends both on the location
and the source orientation 7]. This weight vector 7]) should only pass the signal from
a source with a particular location and an orientation 7]. The weight vector rejects not only
the signals from other locations but also the signal from the location if the orientation of
the source at differs from 7]. Then, the magnitude of the source moment is estimated using
a simple linear operation,
(7.22)
where the estimate of the source magnitude is denoted t ). When using the scalartype
beamformer in Eq. (7.22), we need to first determine the beamformer orientation 7] to
estimate the source activity at a specific location However, this 7] is generally unknown,
although several techniques have been developed to obtain the optimum estimate of the
source orientation (Sekihara and Scholz, 1996; Mosher et al., 1992).
The vector spatial filter uses a weight matrix that contains three weight vectors
and which respectively estimate the y, and z components of the
source moment. That is, the source current vector is estimated from
= = (7.23)
where t) is the estimate of the source current vector. The vector spatial filter estimate
the source orientation as well as the source magnitude.
The application of a spatial filter weight artificially focuses the sensitivity of a sensor
array on a specific location r , and this location r is a controllable parameter. Therefore, in
222 Sekihara and S. S. Nagarajan
(7.24)
a postprocessing procedure, we can scan the focused region over a region of interest to
reconstruct the entire source distribution.
7.4.2 RESOLUTION KERNEL
The major problem with spatial filter techniques is how to derive a weight vector with
desirable properties. To develop such weight vectors, we need a criterion which characterizes
how appropriately the weight has been designed. The resolution kernel can play this role.
Combining Eqs. (7.1) and (7.23), we obtain the relationship,
=
where,
= (7.25)
This is called the resolution kernel, which expres ses the relat ionship between the
original and estimated source distributions (Menendez et al, 1997; Menendez et al., 1996;
Lutkenhoner and Menendez, 1997). Therefore, the resoluti on kernel can provide a measure
of the appropriateness of the filter weight. In other words, the weight must be chosen so that
the resoluti on kernel has a desirable shape, which generally satisfies the three propert ies:
(i) peak at the source location, (ii) a small mainlobe width, and (iii) a low sidelobe
ampl itude.
The most important property among them is that the kernel should be peaked at the
source locat ion. Only in this case, the reconstructed source distribution can be interpreted
as a smoothed version of the true source distribution . However, if this condition is not met,
the reconstructed source distribution should contain systematic bias and may be totall y
different from the true source distributi on. The kernel should also have a small mainlobe
width, so that the reconstruction results have high spatial resolution. When the kernel has
a lower sidelobe amplitude, the results have less systematic noise and artifacts.
7.4.3 NONADAPTIVE SPATIAL FILTER
Minimumnormspatialfilter
There are generally two types of spatial filter techniques. One is a nonadaptive method
in which the filter weight is independent of the measurements. The other is an adaptive
method in which the filter weight depends on the measurements. The primary interest of
this chapter is the application of the adaptive spatial filter technique to the neuromagnetic
source reconstruct ion. However, before proceeding to describe the adaptive spatial filter,
we briefly describe the nonadaptive spatial filter in order to clarify the difference bet ween
these two types of spatial filter methods.
The bestknown nonadaptive spatial filter is the minimumnorm estimate (Hamalainen
and Ilmoniemi , 1984; Hamalainen and Ilmoniemi , 1994; Wang et al. , 1992; Graumann,
1991). The filter weight can be obtained by the following minimization:
min!
2
(7.26)
Neuromagnetic Source Reconstruction and Inverse Modeling 223
where is the threedimensional delta function. By making the resolution kernel close
to the deltafunction, the weight is obtained and it is expressed as
The estimated current density is then expressed as
(7.28)
The matrix G is often referred to as the gram matrix. The p and q element of G is given by
calculating the overlap between the lead fields of the pth and qth sensors,
o..
(7.29)
Unfortunately, in biomagnetic instruments, the overlaps between the adjacent sensor
lead fields are very large, as depicted in Fig. 7.1(a). As a result, G
p
•
q
has a moreorless
similar value for various pairs of and Consequently, the matrix G is generally very
poorly conditioned. This fact greatly affects the performance of this nonadaptive spatial
filter method because it requires calculation of the inverse of G, a process which is very
erroneous if G is nearly singular.
This gram matrix G is usually numerically calculated by introducing pixel grids
throughout the source space. Let us denote the locations of the pixel grid points
1, 2, ... , and the composite leadfield matrix for the entire pixel grids
Then, the matrix G is calculated from
(7.30)
(a) (b)
FIGURE 7.1. Schematic views of the sensor lead field. (a) Biomagnetic instrument and (b) Xray computed
tomography.
K. Sekihara and S. S. Nagarajan
However, to avoid the numerical instability when inverting it, the following regularized
version is usually used:
(7.31)
where y is the regularization parameter. The final solution is expressed as
(7.32)
The regularization, however, inevitably introduces considerable amounts of smearing into
the reconstruction results. Besides, the solution obtained using the minimumnorm spatial
filter suffers a geometric bias in that the current estimates are forced to be closer to the
sensor array than their actual locations.
It should, however, emphasized that the poor performance of the minimumnormspatial
filter is not because the method itself has a serious defect but because a mismatch exists
between the method and the biomagnetic instruments. This can be understood byconsidering
a situation for other imaging modalities such as the Xray computed tomography (CT). As
is shown in Fig. 7.1(b), the overlaps between the lead fields of different sensors are very
small for the Xray CT. As a result, the matrix G is close to the identity matrix and the
nonadaptive spatial filter method works quite well. Indeed, the minimumnorm spatial filter
technique is considered identical to the filteredbackprojection algorithm (Herman, 1980)
used for image reconstruction from projections in commercial Xray CT systems. In the
next subsection, we briefly describe investigations into ways of improving the performance
of the minimumnormbased spatial filter.
Leastsquaresbased interpretation ofthe minimumnormmethods
The minimumnorm spatial filter is commonly derived by minimizing the least
squaresbased cost function. Actually, this leastsquaresbased interpretation is much more
popular than the spatialfilterbased interpretation described in Section 7.4.3. Namely, the
solution in Eq. (7.32) minimizes the cost function,
(7.33)
where S is a source vector whose elements consist of the current estimate at the pixel points,
t), ... In Eq. (7.33), the first term on the righthand side is the
leastsquares error term and the second term is the total sum of the current norm. Therefore,
the optimum solution minimizes the total current norm as well as the leastsquares error.
This is why the method is often referred to as the minimumnorm estimate.
The trick to improving the performance of the minimumnorm method is to use a more
general form of the cost function, expressed as
(7.34)
where iJ! represents some kind of weighting applied to the solution vector and Y
represents the weighting applied to the residual of the leastsquares term. The solution
Neuromagnetic Source Reconstruction and Inverse Modeling
derived by minimizing this cost function is expressed as
(7.35)
In this solution, the gram matrix becomes L ~ The inclusion of the
matrices and gives a greater degree of freedom in regularizing G, and by choosing
appropriate forms for these matrices, the numerical instability can be improved without
introducing unwanted side effects such as image blur.
In general, the matrix P is derived from a desired property of the solution. One widely
used example is the minimum weightednorm constraint in which we use P whose
nondiagonal elements are zero, and diagonal terms are given by
3k+l,3k+l 
and (7 36)
3k+2,3k+2 k) 11
This weight can reduce the geometric bias of the minimumnorm solution to some ex
tent, compensating for the variation in the leadfield norm. Low resolution electromagnetic
tomography (LORETA) (PascualMarqui and Michel, 1994; Wagner et al., 1996) is another
popular application of this particular type of seeks the maximally smooth solution by
using P pLpR, where pR is the Laplacian smoothing matrix.
Bayesiantype estimation methods determine P based on prior knowledge of the neural
current distribution (Schmidt et al., 1999; Baillet and Gamero, 1997). Determination of P
by fMRI has been proposed (Liu et al., 1998; Dale et al., 2000). The matrix Y is generally
determined from the noise properties. When the measurements contain nonwhite noise and
we know the noise covariance matrix, is usually set to the inverse of the noise covariance
matrix. The determination of the optimumforms for the matrices and has been an active
research topic, and many investigations have been performed in this direction. However,
we will not digress into the details of these investigations. Instead, in the following section
we describe different approaches known as the adaptive spatial filter, which does not use
the gram matrix of the lead field.
7.4.4 NOISE GAINAND WEIGHT NORMALIZATION
The spatial filter weights determine the gain for the noise in the reconstructed results.
In the scalar spatial filter techniques, the output noise power due to the noise input, Pn, is
given by
(7.37)
where R; is the noise covariance matrix. When the noise is uncorrelated white Gaussian
noise, the output noise power is equal to
(7.38)
where 0'2
T
is the power of the input noise. Therefore, the norm of the filter
weight vector is called the noise power gain or the white noise gain. In vector
226 K. Sekihara and S. S. Nagarajan
spatial filter techniques, the output noise power is expressed as
n
= =
n
When the input noise is uncorrelated white Gaussian noise, this reduces to
(7.39)
(7.40)
Here, the square sum of the norm of the filter weights
Ilwy(r)f
z(r)11
is the noise power gain. A minimumnorm spatial filter with weight
normalization has also been proposed (Dale et al., 2000) The output of this spatial filter is
expressed as
~
t)
tr{W
(7.41)
Because the weight norm is the noise gain, the output of this spatial filter is interpreted as
being equal to the SNR of the minimumnorm filter outputs.
7.5 ADAPTIVE SPATIAL FILTER TECHNIQUES
7.5.1 SCALAR MINIMUM VARIANCEBASED BEAMFORMER TECHNIQUES
The adaptive spatial filter techniques use a weight vector that depends on measure
ment. The bestknown adaptive spatial filter technique is probably the minimumvariance
beamformer. The term "beamformer" has been customarily used in the signalprocessing
community with the same meaning as "spatial filter". In this method, the spatial filter weights
are obtained by solving the constrained optimization problem,
and consequently we get
subject to (7.42)
(7.43)
where is defined as
The idea behind the above optimization is that the filter weight is designed to minimize
the total output signal power while maintaining the signal from the pointing location r.
Therefore, ideally, this weight only passes the signal from a source at the location with
the orientation 17, and suppresses the signals from sources at other locations or orientations.
One difficulty arises when applying it to actual MEGfEEG source localization problems.
That is, when we use the spherically symmetric homogeneous conductor model to calculate
the beamformer output has erroneously large values near the center of the sphere. This
is because becomes very small when approaches the center of the sphere.
Neuromagnetic Source Reconstruction and Inverse Modeling 227
A variant of the minimumvariance beamformer, proposed by Borgiotti and Kaplan
(Borgiotti and Kaplan, 1979), uses the optimization,
subject to (7.44)
The resultant weight vector is expressed as
(7.45)
Because represents the noise power gain, the output of the above bearnformer
directly corresponds to the power of the source activity normalizedby the power of the output
noise. This BorgiottiKaplan beamformer is known to provide a spatial resolution higher
than that of the minimumvariance beamformer (Borgiotti and Kaplan, 1979). Moreover,
it can easily be seen that the output of the beamformer in Eq. (7.45) does not depend on
Thus, related artifacts are avoided.
Another more serious problem with the adaptive beamformer techniques described so
far is that they are very sensitive to errors in the forward modeling or errors in estimating
the data covariance matrix. Since such errors are nearly inevitable in neuromagnetic mea
surements, these techniques generally provide noisy spatiotemporal reconstruction results,
as demonstrated in Section 7.7.
One technique has been developed to overcome such poor performance (Cox et al.,
1987; Carlson, 1988). The technique, referred to as diagonal loading, uses the regularized
inverse of the measurement covariance matrix, instead of its direct matrix inverse. Although
this technique has been applied to the MEG source localization problem(Robinson and Vrba,
1999; Gross and Ioannides, 1999; Gross et al., 2001), it is known that the regularization
leads to a tradeoff between the spatial resolution and the SNR of the beamformer output.
7.5.2 EXTENSION TO EIGENSPACEPROJECTION BEAMFORMER
We here describe the eigenspaceprojection beamformer (van Veen, 1988; Feldman
and Griffiths, 1991), which is tolerant of the abovementioned errors and provides improved
output SNR without sacrificing the spatial resolution in practical lowrank signal situations.
Using Eqs. (7.43) and (7.17), and defining we rewrite the weight
vector for the minimumvariance beamformer as
(7.46)
where
In Eq. (7.46), the second term on the righthand side, should ideally be equal
to zero because the leadfield matrix is orthogonal to at the source locations as
indicated by Eq. (7.21). Various factors, however, prevent this term from being zero, and a
228 K. Sekihara and S. S. Nagarajan
nonzero seriouslydegradesSNRas explainedin the next section. Therefore, the
eigenspacebased beamformeruses only the first termof Eq. (7.46) to calculateits weight
vector i.e.,
= =
(7.47)
Note that is equal to the projectionof onto the signal subspace of
b
• Namely,
the following relationship holds (Feldmanand Griffiths, 1991; Yu and Yeh, 1995):
EsEJ (7.48)
Therefore, the extensionto an eigenspaceprojection beamformeris attainedby projecting
the weight vectors onto the signal subspaceof the measurement covariance matrix.
7.5.3 COMPARISON BETWEENMINIMUM VARIANCE AND EIGENSPACE
BEAMFORMER TECHNIQUES
Althoughtheminimumvariancebeamformerideallyhas exactlythesameSNRas that
of theeigenspacebasedbeamformer, theSNRof theeigenspacebeamformeris significantly
higher inpractical applications. The reasonfor this highSNRcanbe understoodas follows.
Let us assume that a single source with a moment magnitudeequal to exists at We
assume that the estimated lead field, is slightlydifferent from the true lead field
The estimateof is derived and the average
power of the estimatedsource moment is expressed as
(7.49)
where is the averagepower of definedby = For the minimumvariance
beamformer, this
s
is expressedas
~ 2 1 2 ~ T
P
s
b
Ps[1 (7.50)
where we use the orthogonalityrelationship O. For the eigenspaceprojection
beamformer, it is also expressedas
~ ~ T T 2
= Ps[1 (7.51)
The average noise power
n
is obtained using Eqs. (7.38) and (7.46) and, for the
minimumvariance beamformer, it is expressedas
(7.52)
Neuromagnetic Source Reconstruction and Inverse Modeling
For the eigenspaceprojection beamformer,
n
is expressedas
229
(7.53)
Thus, the output SNR of the minimumvariance beamformer, SNR(MV), is expressed as
(Chang and Yeh, 1992; Chang and Yeh, 1993)
(7.54)
The SNRfor the eigenspacebased beamformer, SNR(ES). is thus
(7.55)
The only difference between Eqs. (7.54) and (7.55) is the presence of the second term
I T in the denominatorof the righthandside ofEq. (7.54). It is readilyapparent
that SNR(MV) and are equal if wecanuseanaccuratenoisesubspaceestimateandan
accurateleadfield vector, becausethetermi
T
(r isexactlyequal tozerointhiscase.
It is, however, generally difficultto attain the relationship, I T O. One obvious
reasonfor thisdifficultyis that whencalculating inpractice, psteadof usi'¥ R
b
, the
pIecovariancematrix
b
must be used;
b
is calculatedfrom
b
= 1/ L k=l b(tk )b T (tk)
where is the number of time points.
Anotherfactorthat is specifictoMEGandcausesl tohavea nonzerovalue
is that it is almost impossible to use a perfectlyaccurate leadfield vector. This is because
the conductivity distribution in the brain is usually approximated by using some kind of
conductor modelsuch as the spherically homogeneous conductor modelto calculate
the lead field matrix. Although this error may be reduced to a certain extent by using a
realistic head model, the error cannot be perfectly avoided. Let us define the overall error
in estimating as Assumingthat [1
2»11
/12, we can rewriteEq. (7.54) as
(7.56)
Note that, in the denominatorof the righthandside of this equation, the normof the matrix
e T has an order of magnitudeproportional to e 11 /A whereANrepresentsone of
the noiselevel eigenvalues of Ri : The eigenvalue ANis usually significantly smaller than
the signallevel eigenvalues. Therefore, Equation(7.56) indicates that even when the error
" e " is verysmall, the terme T maynot be negligiblysmall comparedto the first term
in the denominator. Thus, inpracticetheeigenspaceprojection beamformerattainsan SNR
significantly higher than that of the minimumvariancebeamformer.
230
7.5.4 VECTORTYPE ADAPTIVE SPATIAL FILTER
K. Sekihara and S. S. Nagarajan
The scalar beamfonner techniques described in the preceding subsections require
determination of the beamformer orientation to calculate We, here, describe the
extension to vectortype adaptive spatial filter techniques, which does not require the pre
determination of
A naive way of extending to the vector beamformer is to simply use the scalar beam
former weight vector obtained with to estimate the source in the direction (where
= = where uif (r ,
is obtained using Eq. (7.43). The use of such weight vectors, however, generally gives er
roneous results, and the cause of this estimation failure can be explained as follows. Let us
assume that a single source with its orientation equal to 1];f exists at Its
activation time course is assumed to be Then, we can express the measured magnetic
field as =
zs(t)[Z(r).
This can be interpreted as showing
that the magnetic field is generated from three perfectly correlated sources located at the
same location r , with moments equal to yo and
zs(t)f
z.
Let us, for example, consider the case of estimating the x component of the source
moment. The estimated moment, is expressed as
= = 1]v 1];
(7.57)
Since the weight is obtained by imposing the constraint 1,
we have
(7.58)
In this equation, there is no guarantee that the relationships w = 0 and
0 hold. Instead, and generally have fairly
large negative values, resulting in considerable errors.
The above analysis also suggests how we can avoid such errors. Equation (7.57)
indicates that the weight should be derived with the multiple constraints,
(7.59)
That is, we impose the null constraints on the directions orthogonal to the one to be esti
mated. We here omit the notation for the weight expression unless this omission causes
ambiguity. Similarly, to derive Wy and lL!:, the following constraints should be imposed,
=
W =
0,
and
and
= 0
= 1.
(7.60)
(7.61)
Neuromagnetic Source Reconstruction and Inverse Modeling 231
The minimumvariance beamformer with such multiple linear constraints, referred to as
the linearly constrained minimumvariance beamformer (Frost, 1972), is known to have the
following solution:
(7.62)
It is clear from the discussion above that, when estimating one of the three orthogonal
components of the source moment, we need to suppress the other two components. By
doing this, we can avoid the errors caused by the perfectly correlated virtual sources, so the
beamformer can detect the source moment projected in three orthogonal directions. Note
that the set of weight vectors in Eq. (7.62) has been previously reported (van Dronge1en
et al., 1996; van Veen et al., 1997; Spencer et al., 1992). In these reports, however, the
necessity for imposing the null constraints was not fully explained.
Vectorextended BorgiottiKaplan beamformer
The extension of the BorgiottiKaplan beamformer in Eq. (7.45) is performed in the
similar manner. The weight vectors are obtained by using the following constrained mini
mizations,
subject to
T
0, and 0, (7.63) 1,
W
x
subject to and
0
(7.64) mm w
v
0, 1,
w,
subject to 0, 0, and (7.65) min
z
z
z
z
1.
,.
We first derive the expression for .r Let us introduce a scalar constant ~ such that
= ~ where ~ can be determined from the relationship = 1. Then, the con
strained optimization problem in Eq. (7.63) becomes
(7.66)
The solution of this optimization problem is known to have the form
Then, we have
T 2 T
~ n x'
where
n
(7.67)
(7.68)
232 K. Sekihara and S. S. Nagarajan
Thus , we get ( l/jI; nIxfrom the relationship w;W
x
Using exactly the same
derivation, the weights wyand W
z
can be derived, and a set of the weights is expressed as
RbI L(r )[L (r)R
b
L(r )r
.
jI[nh
(7.69)
Extensionto eigenspaceprojection vector beamformer
The extension to the eigenspaceprojection vector beamformer is attained by using
(7.70)
The projection onto the signal subspace , however, cannot preserve the null constraints
imposed on the orthogonal components. This can be understood by considering, for example,
the case of w
x
' The null constraints in this case should be 0 and O.
However, let us consider
w;JY(r) (EsEJ wxfJY(r) w;EsEJF(r),
= (EsEJ wxfZZ(r) = w; EsEJfZ(r).
(7.71)
Because P(r) and fZ(r) are not necessarily in the signal subspace, we generally have
EsEJP(r) P(r) and EsEJZZ(r) fZ(r) , and therefore 0 and O.
It can, however, be shown that the eigenspaceprojection beamformer in Eq. (7.70) can
still detect the three orthogonal components of the source moment even though the null
constraints are not preserved (Sekihara et aI.,
7.6 NUMERICAL EXPERIMENTS: RESOLUTIONKERNEL
COMPARISON BETWEENADAPTIVEAND NONADAPTIVE SPATIAL
FILTERS
7.6.1 RESOLUTION KERNEL FOR THE MINIMUMNORM SPATIAL FILTER
We compare the resolution kernels for the minimumnorm and the minimumvariance
spatial filter techniques. These two methods are typical and basic spatial filter techniques
in each category. In these numerical experiments, we use the coil configuration of the 148
channel Magnes 2500™ neuromagnetometer (4D Neuroimaging, San Diego). The sensor
coils are arranged on a helmetshaped surface whose sensor locations are shown in Fig. 7.2.
The coordinate origin is chosen as the center of the sensor array. The z direction is defined
as the direction perpendicular to the plane of the detector coil located at this center. The x
direction is defined as that from the posterior to the anterior, and the y direction is defined
as that from the left to the right hemisphere. The values of the spatial coordinates y , z)
Neuromagnetic Source Reconstruction and Inverse Modeling 233
o
5
• •
, ~
.'
: ~ . . .

e. ••
e.
~
~
••• • • •
 ~ :
. . • ' e " ••• '.
y
20 20
FIGURE 7.2. ThesensorlocationsandthecoordinatesystemusedforplottingtheresolutionkernelsinSection7.6.
The filledspots indicatethe locationsof the 148sensors, and the hatchedrectangleshows the plane of x = 0 on
whichthe resolutionkernels were plotted.
are expressed in centimeters. The coordinate system is also shown in Fig. 7.2. The origin
of the spherically symmetric homogeneous conductor was set to (0, 0, 11).
To plot the resolution kernel, we assume a vertical plane x 0 located below the center
of the sensor array (Fig. 7.2). The power ofthe resolution kernel II lR11 was calculated using
Eqs. (7.25), (7.27), and (7.31) for the minimumnorm method. A point source was assumed
to exist at (0,0, 6), i.e., r' in Eq. (7.25) was set to (0,0, 6). The kernel was plotted
within a region defined as 5 :s y :s 5 and 9 :s z :s Ion the vertical plane of x O.
The resulting resolution kernels are shown in Fig. 7.3. Here, the results in Fig.7.3(a) show
the kernel obtained from the original minimumnorm method. is well known that the
original minimumnorm method suffers from a strong geometric bias toward the sensors.
The results in Fig. 7.3(a) confirm this fact. The kernels of the minimumnorm method with
the leadfield normalization are shown in Fig.7.3(b) and (c). Here, in Eq. (7.36) was used
and the regularization parameter y was set at O.OOOlA for (b) and O.OOlA for (c). These
results show that the leadfield normalization significantly improves the performance of the
minimumnorm method. However, the resolution is still significantly low, particularly in
the depth direction. Moreover, the peak of the kernel is located a few centimeters shallower
than the assumed location; the depth difference depends on the choice of the regularization
parameter. The results in Fig.7.3(d) show the kernel from the minimumnorm method with
the normalized weight (Eq. (7.41)). The main lobe is significantly sharper than those in the
leadfield normalization cases of (b) and (c). However, the peak is located 2cm deeper than
its original position.
This A1 is the largesteigenvalueof the grammatrix J;
234 K. Sekihara and S. S. Nagarajan
'        ~ ~  ~   ~   '
" ~
~
5
. ()
8
. <)
(b)
3
c 5
. (>
·7
8
. <)
o
(c) y
3
8
~ 5
FIGURE 7.3. Results of plotting the resolution kernels for minimumnormbased spatial filter techniques.
(a) Results for the original minimumnorm method. (b) Results for the minimumnom, method with the nor
malized lead field. The regularization parameter y was set to O.OOO])cj. (c) Results for the minimumnorm
method with the normalized lead field. The regularization parameter y was set to 0.00 lA (d) Results for the
minimumnorm method with weight normalization. The parameter y was set to 0.000 A for the results in
(a) and (d). Here, Al is the largest eigenvalue of the gram matrix L ~ L N .
7.6.2 RESOLUTION KERNEL FOR THE MINIMUM VARIANCE ADAPTIVE
SPATIAL FILTER
Next, the resolution kernel for the minimumvariance vector beamformer technique
was plotted. The kernel was calculated using Eqs. (7.25) and (7.62). Here, the covariance
matrix was calculated from R 1+ wherer ' is setto the sourcelocation
equal to (0,0, The calculated resolution kernels for the four SNR values of J
2
are shown in Fig. 7.4. First of all, the kernel is peaked exactly at the target source location
(0,0, The kernel's sharpness depends on the SNR value. This is one characteristic of
the adaptive spatial filter methods. We encounter the SNR between 8 and 2 in most actual
measurements. In such an SNR range, the kernel sharpness obtained with the minimum
variance spatial filter is significantly higher than that with the nonadaptive minimumnorm
method. These results clearly demonstrate that the minimumvariance filter can provide
more accurate reconstruction results with significantly higher spatial resolution than the
minimumnorm spatial filter.
Neuromagnetic Source Reconstruction and Inverse Modeling 235
2
3 3
4
~ 5 ~ 5
6 6
7 7
8 ·8
9 9
2 2 4 4 2 0 2
(a)
Icrn)
. J
2 2
3 3
4
.s 5
,
Q5
6 6
7 ·7
8 8
9 9
2 0 4 4 2 0 2
(d)
FIGURE 7.4. Results ofplotting the resolution kernels for the minimumvariance adaptive spatial filter technique.
The plots are for SNR values of (a) 8, (b) 4, (c) 2, and (d)
7.7 NUMERICAL EXPERIMENTS: EVALUATIONOF ADAPTIVE
BEAMFORMER PERFORMANCE
7.7.1 DATA GENERATIONAND RECONSTRUCTION CONDITION
We next conducted a series of numerical experiments to test the performance of the
adaptive beamformer technique. Here, we assume the 37sensor array of Magnes" neu
romagnetometer in which the sensor coils are arranged in a uniform, concentric array on
a spherical surface with a radius of 12.2 em. The sensors are configured as firstorder ax
ial gradiometers with a baseline of 5 em. Three signal sources were assumed to exist on
a plane defined as x O. The source configuration is schematically shown in Fig. 7.5(a)
and their time courses are shown Fig. 7.5(b). The magnetic field was generated at a lms
interval from 0 to 400 ms. Gaussian noise was added to the generated magnetic field, and
the SNR, defined as the ratio of the Frobenius norm of the signalmagneticfield data matrix
to that of the noise matrix, was set to 16. The generated magnetic field is shown also in
Fig.7.5(b).
The covariance matrix R
b
was calculated using the data obtained with this time window
between 0 and 400 ms. We express the sourcemoment vector using the two tangential
components ((), and the radial component is assumed to be zero. The reconstruction was
236
5
..
E
15
K. Sekihara and S. S. Nagaraj an
37channel sensor array
three sources
5
sphere for the forward modeling
FIGURE 7.5. (a) The coordinate system and the source configuration used in the numerical experiments in
Section 7.7. The cross section atx °is shown. The square shows the reconstruction region for the experimental
results shown in Figs. 7.67.9. (b) Time courses of the three sources assumed in the numerical experiments. Time
courses fromthe first to the third sources are shown fromthe top to the third row. respectively. The three vertical
broken lines indicate the time instants 220, 268, and 300ms, at which the sourcemoment magnitude is displayed.
The bottom rowshowsthe generated magnetic field.
performed by using
= wI and = w[ (7.72)
The reconstruction region was defined as the area between 4 :::: y :::: 4 and 8 :::: z ::::  3
on the plane x 0, and the reconstruction interval was 1 mm in the y and z directions.
Once and were obtained, cp, an angle representing the mean source
direction in the ¢  eplane was calculated using
cp arctan (
if
0
~ an:tan (
if
0
(7.73)
where (.) indicates the average over the time window with which R
b
was calculated. Then,
the time course expressed in the mean source direction, t ), and that in its orthogonal
Neuromagnetic Source Reconstruction and Inverse Modeling 237
220 268 300
"0
~
. ~
0
c
"0
~
. ~
0
0
c
I
"0
~
.!::
:;
0
0
c
100 200 300 400
late ncy (ms)
"'C
2
0
"E
0
(b)
100 200
latency (ms)
FIGURE 7.5.
300 400
direction, 1 t) were given by
t ) = t ) cos q; t ) sin q; ,
cos q;  sin q; . (7.74)
In the following experiments, we used and t) when displaying the time course
of a source activity.
To display the results of the spatiotemporal reconstruction, three time points at 220,
268, and 300 ms were selected. The amplitude of the second source happened to be zero
at 220 ms, and all the sources had nonzero amplitudes at 268 ms, while only the second
238 K. Sekihara and S. S. Nagarajan
source had a nonzero amplitude at 300 ms. The snapshots of the source magnitude dis
tribution s ~ ( r , s ~ ( r , at these three time points, and the time averaged
reconstruction t )2) are presented in the following experiments.
7.7.2 RESULTS FROM MINIMUMVARIANCE VECTOR BEAMFORMER
The results of the spatiotemporal reconstruction obtained using the minimumvariance
vector beamformer in Eq. (7.62) are shown in Fig. 7.6(a). The estimated time courses at
the pixels nearest to the three source locations are shown in Fig. 7.6(b). These results show
that the reconstruction at each instant in time was fairly noisy: the snapshot at 220 ms
showed some influence from the second source, and the snapshot at 300 ms contained the
activities of the first and third sources. The timeaveraged reconstruction, however, resolved
three active sources. In Fig. 7.6(b), only t) shows the source activity time courses, but
S.L t) contains no significant activities. This confirms the fact that the source orientation
is fixed during the observation period.
We next tested the minimumvariance beamformer with the regularized inverse
y instead of The regularization parameter was set to 0.003).. The re
sults in Fig. 7.7(a) show that a considerable amount of blur was introduced. The es
timated time courses are shown in Fig. 7.7(b). This figure shows that the SNR of the
beamformer output increased considerably in this case, although each time course shows
some influence from neighboring sources. The results demonstrate that the regulariza
tion leads to a tradeoff between the spatial resolution and the SNR of the beamformer
output.
7.7.3 RESULTS FROM THE VECTOREXTENDED BORGIOTT/KAPLAN
BEAMFORMER
The reconstruction results from weight vectors obtained using Eq. (7.69) are shown
in Fig. 7.8. The weight is equivalent to the vectorextended BorgiottiKaplan (BK) beam
former without the eigenspace projection. Comparison between the timeaveraged recon
struction in Fig. 7.6(a) and in Fig. 7.8(a) confirms that the BorgiottiKaplantype beam
former has a spatial resolution much higher than the minimumvariance beamformer. The
spatiotemporal reconstruction, however, is very noisy in both cases.
7.7.4 RESULTS FROM THE EIGENSPACE PROJECTED VECTOREXTENDED
BORGIOTTI KAPLAN BEAMFORMER
We then applied the eigenspaceprojected BK beamformer obtained using Eqs. (7.69)
and (7.70) to the same computergenerated data set. The reconstructed source distributions
are shown in Fig. 7.9(a), and the estimated time courses are shown in Fig. 7.9(b). Comparison
between Figs. 7.8 and 7.9 confirms that the eigenspace projection can improve the SNR
with almost no sacrifice in spatial resolution. Comparing the result s in Fig. 7.9 with the
minimumvariance results in Fig. 7.6, we can clearly see that the eigenspaceprojected BK
beamformer technique significantly improved both spatial resolution and output SNR.
This AI is the largest eigenvalue of
Neuromagnetic SourceReconstruction and InverseModeling 239
E.
6
8
6
8
2
~
~
:
. ~
~
200
latency
300
2
FIGURE7.6. (a) Results of the spatiotemporal reconstruction obtained using the minimumvariancebased
vector beamformer in Eq. (7.62). The upperleft, upperright, and lowerleft maps show the snapshots of the
sourcemoment magnitude at 220 ms, 268 ms, and 300 ms, respectively. The lowerright map shows the time
averaged reconstruction. (b) Estimated time courses from the first to the third sources are shown from the top to
the bottom, respectively. The two time courses in each panel correspond to t) and I). The three vertical
broken lines indicate the time instants at 220, 268, and 300 ms.
240
~
N 6
4
K. Sekihara and S. S. Nagarajan
E
~
6 .
(a)
2 0 2 4 2 0 2 4
Y(em) y (cm)
0
. ~
100 200
latency (ms)
300 400
FIGURE 7.7. (a) Results of the spatiotemporal reconstruction obtained using the minimumvariancebased
vector beamformer in Eq. (7.62) together with the regularized inverse The parameter y was set to
0.003Al' where Al is the largest eigenvalue of Rs; (b) Estimated time courses from the first to the third sources
are shown from the top to the bottom, respectively.
Neuromagnetic SourceReconstruction and Inverse Modeling 241
4
2
N 6
4
2
N 6
o 4
~ .
~ .
0
y
400 300 100
1" '
o
FIGURE 7.8. (a) Results of the spatiotemporal reconstruction with the vectorextended BorgiottiKaplantype
beamformer (Eq. (7.69)). (b) Estimated time courses from the first to the third sources are shown from the top to
the bottom, respectively.
242 K. Sekihara and S. S. Nagarajan
latency (rns)
4
"
6
8
4
'2
Q
6
8
2 0 2 4 2 0 2 4
y y
"
FIGURE 7.9. (a) Results of the spatiotemporal reconstruction obtained using the eigenspaceprojected Borgiotti
Kaplan vector beamformer technique (Eqs. (7.69) and (7.70». (b) Estimated time courses from the first to the third
sources are shown from the top to the bottom, respectively.
Neuromagnetic Source Reconstruction and Inverse Modeling
z
FIGURE 7.10. The x, y, and zcoordinates used to express the reconstructi on results in Section 7.8. The coordinate
origin is defined as the midpoint between the left and right preauricular points. The axis directed away from the
origin toward the left preauricular point is defined as the y axis, and that from the origin to the nasion as
the +x axis. The +z axis is defined as the axis perpendicular to both these axes and is directed from the origin to
the vertex.
7.8 APPLICATIONOF ADAPTIVE SPATIAL FILTER TECHNIQUE
TO MEG DATA
This section describes the application of the adaptive spatial filter technique to actual
MEG data. The MEG data sets were collected using the 37channel Magnes" neuromag
netometer. The first data set is an auditory and somatosensory combined response, which
contains two major source activities. We show that the adaptive spatial filter technique can
reconstruct these two sources and retrieve their time courses. The second data set is the
somatosensory response with very high SNR achieved by averaging 10000 trials. With this
data set, we show that the adaptive technique can separate cortical activities only 0.7cm
apart . Throughout this section, we use the head coordinates shown in Fig. 7.10 to express
the reconstruction results.
7.8.1 APPLICATIONTOAUDITORYSOMATOSENSORY COMBINED RESPONSE
The evoked response was measured by simultaneously presenting an auditory stimulus
and a somatosensory stimulus to a male subject. The auditory stimulus was a 200 ms pure
tone pulse with 1 kHz frequency presented to the subject's right ear, and the somatosensory
stimulus was a 30 ms tactile pulse delivered to the distal segment of the right index finger.
These two stimuli started at the same time. The sensor array was placed above the subject's
left hemisphere with the position adjusted to optimally record the Nlm auditory evoked
field. A total of 256 epochs were measured, and the response averaged over all the epochs
is shown in the upper part of Fig. 7.11.
The adaptive vector beamformer technique was applied to localize sources from this
data set. The covariance matrix R
b
was calculated with the time window between 0 ms
and 300 ms. We calculated, by using Eqs. (7.69) and (7.70), the eigenspaceprojected
BorgiottiKaplan weight matrix containing the two weight vectors and estimated
the source magnitude vector Flr, t ) using Eq. (7.72). The signal subspace dimension Q was
244 K. Sekihara and S. S. Nagarajan
§
200
. ~
c
0
200 0,::
50 0 50 100 150 200 250 300 350
latency (ms)
10
E
5 5 65ms
~
0
5 5
E
5
E
5 38ms
~ ~
0
5 5
10
( ~ }
= 5
~
5 94ms
E
O
5
5 0 5 10 5 0 5 5 0 5 10
FIGURE 7.11. Results of the spatiotemporal reconstruction from the auditorysomatosensory combined re
sponse shown in the upper trace of this figure. The auditorysomatosensory combined response was measured by
simultaneously applying an auditory stimulus and a somatosensory stimulus. A total of256 epochs were averaged.
The contour maps show reconstructed source magnitude distributions at three different latencies (65, 138, and
194 ms). The reconstruction grid spacing was set to 5 mm. The maximumintensity projections onto the axial (left
column), coronal (middle column), and sagittal (right column) directions are shown. The letters Land R indicate
the left and right hemispheres. The circles depicting a human head show the projections of the sphere used for the
forward modeling.
set to two because the eigenvalue spectrum of R
b
showed two distinctly large eigenvalues.
The maximumintensity projections of the reconstructed moment magnitude t)
2
onto
the axial, coronal, and sagittal planes are shown in Fig. 7.11. The source magnitude at three
latencies, (65, 138, and 194 ms), is shown in this figure. The source magnitude map at
138 ms contains a source activity presumably in the primary somatosensory cortex. The
source magnitude map at 194 ms shows a source activity in the primary auditory cortex.
The map at 65 ms contains both of these activities.
The time courses of points in the primary somatosensory and auditory cortices are
shown in Figs. 7.l2(a) and (b), respectively. The coordinates of these cortices were deter
mined from the maximum points in the source magnitude maps at 138 ms and 194 ms.
In Fig. 7.12(a) the P50 peak, which is known to represent the activity of the primary
Neuromagnetic Source Reconstruction and Inverse Modeling
65
o (a) 50
.S 0
;; 0.5
100 150 200 250 300
latency (ms)
...
.5
.e
0.5
__
o 50
(b)
FIGURE 7.12. Time courses of the points nearest to (a) the primary somatosensory cortex and (b) the primary
auditory cortex. The solid and broken plotted lines correspond, respectively, to 511 t) and 5L t). Three vertical
broken lines indicate the time instants at 65, 138, and 194 ms.
somatosensory cortex, is observed at a latency of about 50 ms. In Fig. 7.12(b), the auditory
Nlm peak is observed at a latency of about 100 ms.
7.8.2 APPLICATION TO SOMATOSENSORY RESPONSE: HIGHRESOLUTION
IMAGING EXPERIMENTS
Electrical stimuli with 0.2 ms duration were delivered to the right posterior tibial
nerve at the ankle with a repetition rate of 4 Hz. The MEG recordings were taken from the
vertex centering at Cz of the intemationall020 system. An epoch of 60 ms duration was
digitized at a 4000 Hz sampling frequency and 10000 epochs were averaged. The upper part
of Fig. 7.13 shows the MEG signals, recorded over the foot somatosensory region in the left
hemisphere. The eigenspaceprojected BorgiottiKaplan beamformer was applied to this
MEG recording. The covariance matrix Rb was calculated with a time window between
20 and 45 ms containing 100 time samples. The maximumintensity projections of the
reconstructed source magnitude t) 11 onto the axial, coronal, and sagittal planes are
shown in Fig. 7.13.
The source magnitude maps revealed initial activation in the anterior part of the S1
foot area at 33.1 ms, followed by coactivation of the posterior part of S1 cortex at 36.2 ms.
The posterior activation became dominant at 37.2 ms and the initial anterior activation
completely disappeared at 39.1 ms. Fig. 7.14 shows the source magnitude map at 36.9 ms
overlaid, with proper thresholding, onto the subject's MRI. Here, the anterior source was
probably in area 3b and the posterior source was in an area near the marginal sulcus. The
246
§
"0
0
c..
20 25 30 35 40 45
latency (ms)
K. Sekihara and S. S. Nagarajan
50 . 12[:J] " . ....
E . ' . E E
0 '.'
4 .' ..' 4 . .' .
5 ....
. ..
E . , .
0 . . . .
4 : 4  
5 . ... . . .. ...•. . .
500 ' 
0 · · · .... 
: :: 4 · · : 4 v , •• . . : •
5 .    .. '  .
5a" ...
0 . .: . 8 · ' 8 .. .. . .: . ..
5 · · .. . .  .
33.lms
36.2ms
37.2ms
39. 1ms
o 5 10 10 5 0 5 5 o 5
FIGURE 7.13. Results of the spatiotemporal reconstruction from the somatosensory response shown in the upper
trace of this figure. The somatosensory response was measured using the right posterior tibial nerve stimulation.
The contour maps show reconstructed source magnitude distributions at four different latencies. The reconstruction
grid spacing was set to I mm. The maximumintensity projections onto the axial (left column), coronal (middle
column), and sagittal (right column) directions are shown. The letters Land R indicate the left and right hemispheres.
The circles depicting a human head show the projections of the sphere used for the forward modeling.
separation of the two sources was approximately 7 mm, demonstrating the highresolution
imaging capability of the adaptive spatial filter techniques. Details of this investigation
have been reported (Hashimoto et al., 2001a), and the results of applying the adaptive
beamformer technique to the response from the median nerve stimulation have also been
reported (Hashimoto et al., 200Ib).
ACKNOWLEDGMENTS
The author would like to thank Dr. D. Poeppel, Dr. Marantz, and Dr. Roberts for
providing the auditory data. We are also grateful to Dr. Hashimoto, and Dr. Sakuma for
Neuromagnetic Source Reconstruction and Inverse Modeling 247
FIGURE 7.14. The source magnitude reconstruction results at a latency of 36.9 ms. The source magnitude map
was properly thresholded and overlaid onto the sagittal cross section of the subject's MRI. The colors represent the
relative intensity of the source magnitude; the relationship between the colors and relative intensities is indicated
by the color bar. The anterior source was probably in area 3b and the posterior source was in an area near the
marginal sulcus. The separation of the two sources was approximately 7 mm in this case. See the attached CD for
color figure.
providing the somatosensory data and for useful discussion regarding the interpretation of
the reconstructed results. This work has been supported by GrantsinAid from the Kayamori
Foundation of Informational Science Advancement; GrantsinAid from the Suzuki Foun
dation; and GrantsinAid from the Ministry of Education, Science, Culture and Sports in
Japan (C13680948). This work has also been supported by the Whitaker Foundation, and
by National Institute of Health. (P4IRRI255303 and ROIDC0048550IAI).
REFERENCES
Adachi, Y, Shimogawara, M., Higuchi, M., Haruta, Y, and Ochiai, M., 2001, Reduction of nonperiodical extra
mural magnetic noise in MEG measurement by continuously adjusted least squares method, in Proceedings
of12th International Conferences on Biomagnetism, (R. Hari et al., eds.), Helsinki University of Technology,
pp. 899902.
A. M. Dale, A. M., Liu, Fischl, B. R., Buckner, R. Belliveau, J. Lewine, J. D., and Halgren, E., 2000,
Dynamic statistical parametric mapping: Combining fMRI and MEG for highresolution imaging of cortical
activity, Neuron, 26, pp. 5567.
Baillet S. and Garnero, 1997, A Bayesian approach to introducing anatomofunctional priors in the EEGIMEG
inverse problem, IEEE Trans. Biomed. Eng., pp. 374385.
Baillet, S., Mosher, J. and Leahy, R. M., 2001, Electromagnetic brain mapping, IEEE Signal Processing
Magazine, 18, pp. 1430.
Barnard, A., Duck, I., Lynn, M., and Timlake, W., 1967, The application of electromagnetic theory to electrocar
diography II. Numerical solution of the integral equations, Biophys. 1.,7, pp. 433462.
Borgiotti G. and Kaplan, J., 1979, Superresolution of uncorrelated interference sources by using adaptive array
technique, IEEE Trans. Antenn. and Propagat., 27, pp. 842845.
248 K. Sekihara and S. S. Nagarajan
Bradley, C. P., Harris, G. M., and Pillan, A. J., 2001, The computational performance of a highorder coupled
FEMJBEM procedure in electropotential problems, IEEE Trans. Biomed. Eng., 48, pp. 12381250.
Carlson, D., 1988, Covariance matrix estimation errors and diagonal loading in adaptive arrays, IEEE Trans.
Aerospace and Electronic Systems, 24, pp. 397401.
Chang, and Yeh, C. C.• 1992, Performance of DMI and eigenspacebased beamformers, IEEE Trans. Antenn.
Propagat., 40. pp. 13361347.
Chang, and Yeh, C. C., 1993, Effect of pointingerrors on the performance of the projection beamformer, IEEE
Trans. Antenn. Propagat., 41, pp. 10451056.
Clarke, J., 1994, SQUIDs, Scientific American, 271, pp. 3643.
Cox, H., Zeskind, R. M., and Owen, M. M., 1987, Robust adaptivebeamforming, IEEE Trans. Signal Process.,
35, pp. 13651376.
Cuffin and Cohen D., 1977, Magnetic fields of a dipole in special volume conductor shapes, IEEE Trans.
Biomed. Eng., 24, pp. 372381, 1977.
CuffinB. 1991, Eccentricspheres models of the head, IEEE Trans. Biomed. Eng., 38, pp. 871878.
Cuffin, B. 1996, EEGlocalizationaccuracy improvements usingrealisticallyshaped head models, IEEE Trans.
Biomed. Eng., 43, pp. 299303, 1996.
de Peralta Menendez, R. G., Gonzalez Andino, S., and Lutkenhoner, B.• 1996, Figures of merit to compare
distributedlinear inversesolutions. Brain Topography, 9, pp. 117124,1996.
de Peralta Menendez, R. G.• Hauk, 0 ., Gonzalez Andino, S., Vogt, H., and Michel, C., 1997, Linear inverse
solutions withoptimal resolution kernels applied to electromagnetic tomography, Human Brain Mapping.S,
pp. 454467, 1997.
Drung, D., Cantor, R., Peters, M., Ryhanen, P., and Koch, H., 1991, IntegratedDC SQUID magnetometer with
highdv/db, IEEE Trans. Magn., 27, pp. 30013004.
Feldman, D. D. and Griffiths, J., 1991, Aconstrained projection approachfor robust adaptivebeamforming, in
Proc. Int. Conf. Acoust ., Speech, Signal Process. , Toronto,May, pp. 13571360.
Frost, O. T., 1972, An algorithmfor linearlyconstrainedadaptivearray processing, Proc. IEEE, 60, pp. 926935.
Fuchs, M., Drenckhahn, R., Wischmann, H.A., and Wagner, M., 1998, An improved boundary element method
for realistic volumeconductor modeling, IEEE Trans. Biomed. Eng., 45, pp. 980997.
Geselowitz, D. B., 1970, On the magnetic fieldgeneratedoutside an inhomogeneous volumeconductor by internal
current sources, IEEE Trans. Biomed. Eng., 2, pp. 346347.
Graumann, R., 1991, The reconstruction of current densities, Tech. Rep. TKKFA689, Helsinki University of
Technology.
Gross J. and loannides, A. A., 1999, Linear transformations of data space in MEG, Phys. Med. BioI., 44, pp.
2081 2097.
Gross, J., Kujara, J., Harnalainen, M. S., Timmermann, Schnitzler,A., andR. Salmelin, 200I, Dynamic imaging
of coherent sources: Studying neural interactions in the human brain, Proceedings of National Academy of
Science, 98, pp. 694699.
Harnalainen, M. S. and I1moniemi, R. J., 1984, Interpreting measured magnetic fields of the brain: Estimates of
current distributions, Tech. Rep. TKKFA559, Helsinki University of Technology.
Harnalainen, M. S. andSarvas, J., 1989, Realisticconductivitygeometrymodelof thehumanheadforinterpretation
of neuromagneticdata, IEEE Trans. Biomed. Eng., 36, pp. 165171.
Harnalainen, M. S., Hari. RxIlmoniemi.R.J., Knuutila, J., andLounasmaa, O. V., 1993,Magnetoencephalography
theory, instrumentation, andapplications to noninvasivestudies of the workinghumanbrain, Rev. Mod. Phys.,
65, pp. 413497.
Hamalainen, M. S. and I1moniemi, R. J., 1994, Interpreting magnetic fieldsof the brain: minimumnormestimates,
Med. BioI. Eng. Comput., 32, pp. 3542.
Hashimoto, Sakuma, K., Kimura, T., Iguchi, Y., and Sekihara, K., 200l a, Serial activation of distinct cy
toarchitectonic areas of the human SI cortex after posterior tibial nerve stimulation, NeuroReport , 12, pp.
18571862.
Hashimoto, Kimura, T., Iguchi, Y., Takino, R., and K. Sekihara, K., 200lb, Dynamic activation of distinct
cytoarchitectonic areas of the human SI cortex after median nerve stimulation, NeuroReport, 12, pp. 1891
1897.
Herman, G. T., 1980. Image Reconstruction from projec tions, Academic Press. NewYork, USA.
J. D. Lewine J. D. and Orrison Jr., W. w., 1995, Magnetoencephalography and magnetic source imaging, in
Functional Brain Imaging, (w. W. OrrisonJr. et al., eds.), pp. 369417. MosbyYear Book, Inc.
Neuromagnetic Source Reconstruction and Inverse Modeling
Liu, A. K.•Belliveau. 1. w..and Dale. A. M.•1998. Spatiotemporal imagingof humanbrainactivity usingfunctional
MRI constrained magnetoencephalography data: Monte Carlo simulations. Proc. Natl. Acad. Sci., 95, pp.
89458950.
B. and de Peralta Menendez. R. G.• 1997. The resolution fieldconcept. Electroenceph. Clin. Neuro
physiol.• 102, pp. 326334.
Mosher. J. C; Lewis. P.S.•and Leahy. R. M.• 1992. Multipledipole modelingand localizationfromspatiotemporal
MEG data. IEEETrans. Biomed. Eng., 39. pp. 541 557.
Okada, Y., Lauritzen. M.• and Nicholson, C., 1987, MEGsource models and physiology, Phys. Med. Bioi, 32, pp.
4351.
Parkkonen, T.. Simola, J. T.• Tuorinierni, J. T.• and Ahonen. A. 1999. An interference suppression system
for multichannel magnetic field detector arrays. in Recent Advances in Biomagnetism, (T. Yoshimoto et al.•
eds.), Tohoku University Press, Sendai, pp. 1316.
PascualMarqui, R. D. and Michel, C. M., 1994, Lowresolution electromagnetic tomography: A newmethod for
localizing electrical activity in the brain. Int. J. Psychophysiol., 18. pp. 4965.
Paulraj, A.. Ottersten, B.•Roy. R.•Swindlehurst. A.•Xu. G., and Kailath, T.•1993.Subspace methods for directions
ofarrivalestimation. in Handbook of Statistics. (N. K. Boseand C. Rao, eds.), Elsevier SciencePublishers,
Netherlands. pp. 693739.
Roberts, T. P. Poeppel, D., and Rowley, H. A.• 1998. Magnetoencephalography and magnetic source imaging,
Neuropsychiatry, Neuropsychology. and Behavioral Neurology. 11, pp. 4964 .
Robinson, S. E. and Vrba, J., [999, Functional neuroimaging by synthetic aperture magnetometry (SAM), in
Recent Advances in Biomagnetism, (T. Yoshimoto et aI., eds.), Tohoku University Press. Sendai, pp. 302
305.
Sarvas. L, 1987, Basic mathematical and electromagnetic concepts of the biomagnetic inverse problem. Phys.
Med. BioI., 32, pp. 11 22.
Scharf, 1991. Statistical Signal Processing: detection. estimation. and time series analysis, AddisonWesley
Publishing Company. New York.
Schmidt. D. M.•George, J. S., and Wood, C. c.. 1999. Bayesian inference applied to the electromagnetic inverse
problem, Human Brain Mapping, 7, pp. 195212.
Schmidt. 1981.A signal subspace approach tomultipleemitter locationandspectral estimation, PhDthesis,
Stanford University, Stanford, CA.
Schmidt, 1986, Multiple emitter location and signal parameter estimation, IEEE Trans. Antenn. Propagat.,
34. pp. 276280.
Sekihara, K.•Poeppel, D.. Marantz. A., and Miyashita, Y., 20()(). Neuromagnetic inverse modeling: applications of
eigenstructurebasedapproaches to extracting cortical activities fromMEGdata. in Image, Language, Brain.
(Alec Marantz et al.•eds.), The MIT Press. Cambridge. pp. 197231.
Sekihara, and Scholz, B.• 1996. Generalized Wiener estimation of threedimensional current distribution from
biomagnetic measurements. in Biomag 96: Proceedings of the Tenth International Conference on Biomag
netism, (C. J. Aine et al.•eds.), SpringerVerlag, New York. pp. 338341.
Sekihara, K., Nagarajan, S. S.•Poeppel, D.•Marantz, A., and Miyashita, Y.. 2001, Reconstructing spatiotemporal
activities of neural sources using an MEG vector beamforrnertechnique, IEEE Trans. Biomed. Eng.• 48. pp.
760771.
Spencer, M. E.•Leahy. R. M.• Mosher. J. c.. and Lewis, P. S., 1992. Adaptive filters for monitoring localized brain
activity from surface potential time series, in Conference Recordf or 26th Annual Asilomer Conference on
Signals. Systems, and Computers, November, pp. 156161.
van Drongelen, W.• Yuchtrnan, M., van Veen, B. D., and van Huffelen, A. 1996. A spatial filtering technique
to detect and localize multiple sources in the brain, Brain Topography. 9, pp. 3949.
van Veen, B. D. and Buckley, M., 1988. Beamforrning: A versatile approach to spatial filtering, IEEE ASSP
Magazine. 5. pp. 424, April.
van Veen, D., 1988, Eigenstructure based partially adaptive array design, IEEE Trans. Antenn. Propagat., 36,
pp. 357362.
van Veen, B. D.• van Drongelen, W.• Yuchtrnan, W. and Suzuki, A.• 1997, Localization of brain electrical ac
tivity via linearly constrained minimum variance spatial filtering. IEEE Trans. Biomed. Eng.• 44. pp. 867
880.
van't Ent, D., de Munck, 1. C.•and Kaas, A. 2001, A fast method to derive realistic BEM models for ElMEG
source reconstruction. IEEETrans. Biomed. Eng., 48, pp. 14341443.
250 K. SekiharaandS. S. Nagarajan
Vrba, J. and Robinson, S., 2001, The effect of environmental noise on magnetometer and gardiometerbased
MEG systems, in Proceedings of 12th International Conferences on Biomagnetism, (R. Hari et al., eds.),
Helsinki University of Technology, pp. 953956.
Wagner, M., Fuchs, M., Wischmann, H.A., Drenckharn, R., and Kohler, T., 1996, Smooth reconstruction of
cortical sources from EEG or MEG recordings, Neurolmage, 3, pp. S168.
Wang, J. Z., Williamson, S. J., and Kaufman, 1992, Magnetic source images determined by a lead
field analysis: The unique minimumnorm leastsquares estimation, IEEE Trans. Biomed. Eng.,
pp.565575.
Yu, J. and Yeh, C C, 1995, Generalized eigenspacebased beamformers, IEEE Trans. Signal Process., pp.
24532461.
8
MULTIMODAL IMAGING FROM
NEUROELECTROMAGNETICAND
FUNCTIONAL MAGNETIC
RESONANCE RECORDINGS
Fabio Babiloni and Febo Cincotti
Dipartimento di Fisiologia Umana e Farmacologia, Universita di Roma "La Sapienza", Roma, Italy
8.1 INTRODUCTION
Human neocortical processes involve temporal and spatial scales spanning several orders of
magnitude, from the rapidly shifting somatosensory processes characterized by a temporal
scale of milliseconds and a spatial scales of few square millimeters to the memory processes,
involving time periods of seconds and spatial scale of square centimeters. Information about
the brain activity can be obtained by measuring different physical variables arising from the
brain processes, such as the increase in consumption ofoxygen by the neural tissues or a vari
ation of the electric potential over the scalp surface. All these variables are connected in direct
or indirect way to the neural ongoing processes, and each variable has its own spatial and
temporal resolution. The different neuroimaging techniques are then confined to the spatio
temporal resolution offered by the monitored variables. For instance, it is known from physi
ology that the temporal resolution of the hemodynamic deoxyhemoglobin increase/decrease
lies in the range of 12 seconds, while its spatial resolution is generally observable with the
current imaging techniques at few mm scale. Today, no neuroimaging method allows a spa
tial resolution on a mm scale and a temporal resolution on a msec scale. Hence, it is of interest
to study the possibility to integrate the information offered by the different physiological
variables in a unique mathematical context. This operation is called the "multimodal inte
gration" of variable X and Y, when the X variable has typically particular appealing spatial
resolution property (mm scale) and the Y variable has particular attractive temporal prop
erties (on a ms scale). Nevertheless, the issue of several temporal and spatial domains is
Corresponding author: Dr. Fabio Babiloni, Dipartimento di Fisiologia Umana e Farmacologia, Universita di
Roma "La Sapienza", P.le A. 5, 00185 Roma, Italy, Tel: +390649910317, Fax: +390649910917, Email:
fabio.babiloniteuniromal.it
251
252 F. Babiioni and F. Cincotti
critical in the study of the brain functions, since different properties could become observ
able, depending on the spatiotemporal scales at which the brain processes are measured.
Electroencephalography (EEG) and magnetoencephalography (MEG) are two inter
esting techniques that present a high temporal resolution, on the millisecond scale, adequate
to follow brain activity. Unlikely, both techniques have a relatively modest spatial resolution,
beyond the centimeter. In spite ofa lack of spatial resolution, neural sources can be localized
from EEG or MEG data by making a priori hypotheses on their number and extension. A
more detailed description of the techniques involved in the high resolution EEG and MEG
recordings and imaging can be found in Chapter 7 and Chapter 8. Here, we briefly recall
that the so called high resolution EEG methods included: subject's multicompartment
head model (scalp, skull, dura mater, cortex) constructed from magnetic resonance im
ages, the sampling of EEG potentials from 64128 electrodes; (ii) the computation of
the surface Laplacian from the scalp potential recordings and/or the use of multidipole
source model for characterizing the neural active sources. However, the spatial resolution
of the EEG/MEG techniques is fundamentally limited by the intersensor distances and by
the fundamental laws of electromagnetism (Nunez, 1981). On the other hand, the use of
aprioriinformation from other neuroimaging techniques like functional magnetic resonance
imaging (fMRI) with high spatial resolution could improve the localization of sources from
EEG/MEG data.
This chapter deals with the multimodal integration of electrical, magnetic and hemody
namic data to locate neural sources responsible for the recorded EEG/MEG activity. The ra
tionale of the multimodal approach based on fMRI, MEG and EEG data to locate brain activ
ity is that neural activity generating EEG potentials or MEG fields increases glucose and oxy
gen demands (Magistretti et aI., 1999). This results in an increase in the local hemodynamic
response that can be measured by fMRI (Grinvald et al., 1986; Puce et al., 1997). On the
whole, such a correlation between electrical and hemodynamic concomitants provides the
basis for a spatial correspondence between fMRI responses and EEG/MEG source activity.
The chapter is organized as follows: first, a brief introduction on the principles at the basis of
the fMRI recordings will be presented; then a recall ofthe principal techniques used for EEG
and MEG for locating neural sources will be presented, with special emphasis on cortical
imaging and linear distributed solutions. This last technique will be employed to show both
the mathematical principle and the practical applications of the multimodal integration of
EEG, MEG and fMRI for the localization of sources responsible for intentional movements.
8.2 GENERALITIES ON FUNCTIONAL MAGNETIC
RESONANCE IMAGING
A brain imaging method, known as fMRI, has gained favor among neuroscientists over
the last few years. Functional MRI reflects oxygen consumption and, as oxygen consump
tion is tied to processing or neural activation, can give a map of functional activity. When
neurons fire, they consume oxygen and this causes the local oxygen levels to briefly decrease
and then actually increase above the resting level as nearby capillaries dilate to let more oxy
genated blood flowinto the active area. The most used acquisition paradigm is the socalled
Blood Oxygen Level Dependence (BOLD), in which the fMRI scanner works by imaging
blood oxygenation. The BOLD paradigm relies on the brain mechanisms, which overcom
pensate for oxygen usage (activation causes an influx of oxygenated blood in excess of that
Multimodal Imaging from Neuroelectromagnetic
B
253
FI GURE 8.1. Physiologic principle at the base of the generation of fMRI signals. A) Neurons increase their
firing rates, increasing also the oxygen consumption. B) Hemodynamic response in a second scale increases the
diameter of the vessel close to the activated neurons. The induced increase in blood flow overcomes the need for
oxygen supply. As a consequence, the percentage of deoxyhemoglobin in the blood flow decreases in the vessel
with respect to the figure A). See the attached CD for color figure.
used and therefore the local oxyhemoglobinconcentration increases). Oxygen is carried to
the brain in the hemoglobin molecules of blood red cells. Fig. 8.1 shows the physiologic
principle at the base of the generation of signals. In this figure it is shown how the
hemodynamic responses elicited by an increased neuronal activity (A) produces a decrease
of the deoxyhemoglobincontent of the blood flow in the same neuronal district after few
seconds (B). The magnetic properties of hemoglobin differ when it is saturated with oxy
gen compared to when it has given up oxygen. Technically, deoxygenated hemoglobin is
"paramagnetic" and therefore has a short T2 relaxation time. As the ratio of oxygenated
to deoxygenatedhemoglobin increases, so does the signal recorded by the MRI. Deoxy
hemoglobinincreases the rate of depolarizationof hydrogen nuclei creating the MR signal
thus decreases the intensityof the T2 image. The bottomline is that the intensity of images
increases withthe increaseof brain activation. The problemis that at the standardintensity
used for the staticmagneticfield (1.5Tesla) this increase is small (usually less than 2%) and
easily obscured by noise and different artifacts. By increasing the static field of the fMRI
scanner, the signal to noise ratio increases to more convenient values. Static fieldvalues of
3 Tesla are nowcommonly used for research on humans, while recently fMRI scanner at 7
Teslawas employed to map hemodynamic responses in the human brain (Bonmassar et al.,
2001). At such high field value, the possibility to detect the initial increase of deoxyhe
moglobin(the initial "dip") increase. The interest in the detection of the dip is based on the
fact that this hemodynamic response happens on timescale of 500 ms (as revealedby hemo
dynamic optic measures; Malonek and Grinvald, 1996), compared to 12 seconds needed
for theresponse of the vascular systemto theoxygendemand. Furthermore, inthe latter case
the response has a temporal extension well beyond the activation occurred(10 seconds). As
a last point, the spatial distribution of the initial dip (as describedby using the optical dyes;
F. Babiloni and F. Cincotti
Malonek and Grinvald, 1996) is sharper than those related to the vascular response of the
oxygenated hemoglobin. Recentl y, with high field strength MR scanners at 7 or even 9.4
Tesla (on animals), a resolution down to cortical column level has been achieved (Kimet al.,
2000). However, at the standard field intensity commonly used in fMRI studies (1.5 or 3
Tesla), the identification of such initial transient increase of deoxyhemoglobin is controver
sial. Compared to positron emitted tomography (PET) or single photon emitted tomography
(SPECT), fMRI does not require the injection of radiolabeled substances, and its images
have a higher resolut ion (reviewed in Rosen et al., 1998). PET, however, is still the most in
formati ve technique for directly imaging metabolic processes and neurotransmitter turnover.
8.2.1 BLOCKDESIGNAND EVENTRELATEDfMRI
Though dynamic fMRI experiments were early recognized to be fundamentally differ
ent from previous hemodynamically based functional imaging methods (like, for instance,
Positron Emitted Tomography; PET), early studies in fMRI typically used experimental
paradigms that could have been easily performed by using previous nuclear technologies.
Spec ifically, most experiments were performed by using extended periods of "on" versus
"off" activations, in a way called block designs paradigm. Such paradigm had been used
in dozens of functional studies of sensory and higher cortical function using PET and
single photon emission computed tomography (SPECT) for more than a decade. Neverthe
less, although such block designs are a necessity when imagi ng hemodynami cs by using
techniques that requir e a quasiequilibrium physiological state for periods up to I min,
they were clearl y not requi red for fMRI experiments, where activity was detectable within
seconds from stimulus onset. Movement away from block designs was gradual and aided
by a number of studies exploring fMRI signal responses to brief stimulus events (as long as
2 seconds or less; Blami re et al., 1992). A detectable signal change in fMRI was shown to
be produced by 2 s or shorter stimulus (Blamire et al., 1992; Bandett ini, 1993). Moreover,
it was also shown that visual stimulation as brief as 34 msec in duration could elicit small,
but clearly detectable, signal changes (Savoy et al., 1995). All these data suggest that tMRI
is sensitive to transient phenomena and can provide at least some degree of quantitative
information on the underlying neuronal behavior. Together, these results thus suggest that
it should be possible to interpret transient fMRI signal changes in ways directly analogous
to electrophysiologic evoked potenti als. A first step in this direction was made by Dale
and Buckner (reviewed in Rosen et al., 1998) who showed that visual stimuli lateralized
to one hemifield could be detected within intermixed trial paradigms. By using methods
similar to those applied in the field of evoked response potenti al research, the trials were
selectively averaged to reveal the predicted pattern of contralateral visual cortex activation.
Taken together with the above observations, these collective data demonstrate convincingly
that fMRI is capable of detecting changes related to singletask events and brief epochs
of stimulation. Hence, the paradigm in which the fMRI information was collected on a
trialbytrial basis is called "eventrelated fMRI" .
8.3 INVERSE TECHNIQUES
The ultimate goal of any EEG, MEG and fMRI recordings is to produce information
about the brain activity of a subject during a particular sensorimotor or cogniti ve task.
Multimodal Imaging from Neuroelectromagnetic 255
The mathematical procedures that allow us to recover information about the activity of the
neural sources from the noninvasive EEG/MEG recordings are called inverse techniques.
The inverse techniques have been systematically treated and reviewed, with application to
ECG (Chapter 4), and MEG (Chapter 7). Mathematical models for the head as volume
conductor and for the neural sources are employed by linear and nonlinear minimization
procedures to localize putative sources of EEG data. Several studies have indicated the
adequacy of the equivalent current dipole as a model for the cortical sources (Nunez, 1981,
1995), while the importance of realistic geometry head volume conductor models for the
localization of cortical activity has been stressed more recently (Gevins, 1989; Gevins
et aI., 1991, 1999, Nunez, 1995). Results of previous intracranial EEG studies have led
support to the idea that high resolution EEG techniques (including head/source models and
proper regularization inverse procedures) might model with an acceptable approximation
the strengths and extension of cortical sources of surface EEG data, at least in certain
conditions (Le and Gevins, 1993; Gevins et aI., 1994; He et aI., 2002). We briefly present
a survey on the principal inverse techniques, with a particular emphasis on the socalled
"distributed" solutions, which we will use to demonstrate the multi modal integration of
EEG/MEG and fMRI.
8.3.1 ACQUISITIONOF VOLUME CONDUCTOR GEOMETRY
A key point of highresolution EEG and MEG technologies is the availability of the
accurate model of the head as a volume conductor by using anatomical MRIs. These images
are obtained by using the MRI facilities largely available in all research and clinical institu
tions worldwide. Reference landmarks such as nasion, inion, vertex, and preauricular points
may be labeled using vitamin E pills as markers. Tlweighted MR images are typically used
since they present maximal contrast between the structures of interest.
Contouring algorithms allow the segmentation of the principal tissues (scalp, skull,
dura mater) from the MR images (Dale et aI., 1999). Separate surfaces of scalp, skull, dura
mater and cortical envelopes are extracted for each experimental subject, yielding a closed
triangulated mesh. This procedure produces an initial description of the anatomical structure
that uses several hundred thousands pointsquite too much for subsequent mathematical
procedures. These structures are thus downsampled and triangulated to produce scalp, skull
and dura mater geometrical models with about 10001300 triangles for each surface. These
triangulations were found adequate to model the spatial structures of these head tissues.
A different number of triangles are used in the modeling of the cortical surface, since
its envelope is more convoluted than the scalp, skull and dura mater structures. A number
of triangles variable from 5000 to 6000 may be used to model the cortical envelope for the
purpose of following the spatial shape of the cerebral cortex. In order to allow coregistration
with other geometrical information, the coordinates of the triangulated structures are referred
to an orthogonal coordinate system (x, y, z) based on the positions of nasion and preauricular
points extracted from the MR images. For instance, the midpoint of the line connecting
the preauricular points can be set as the origin of the coordinate system and, with the
y axis going through the right preauricular point, the x axis lying on the plane determined
by nasion and preauricular points (directed anteriorly) and the z axis normal to this plane
(directed upward). Once the model of scalp surface has been generated, the integration of the
electrodes' positions are accomplished by using the information about the sensor locations
256 F. Babiioni and F. Cincotti
FIGURE 8.2. Realistic MRIconstructed head of a human subject. Electrode positions (128) are shown on the
MRIconstructed scalp surface and on the underlying cortex surface.
produced by the 3D digitizer. The sensors positions on the scalp model are determined by
using a nonlinear fitting technique. Fig. 8.2 shows the results of the integration between
the EEG scalp electrodes position and a realistic head model.
8.3.2 DIPOLELOCALIZATION TECHNIQUES
Dipole localization techniques produced estimates of the position and moment of one
or several equivalent current dipoles localized in a head model from the noninvasive EEG
and/or MEG recordings. From the position of the localized current dipoles in the head model,
inferences about the neural sources in the real brain are inferred. So far, two approaches to
dipole localization have become popular in neuroscience, and both rely on the solution of
nonlinear minimization algorithms.
The first approach is the socalled "moving dipole" method (Cohen et al., 1990).
Dipoles are found at a succession of discrete times, with no a priori assumption about
the relation of the localized dipoles at different time instants. In general, it is difficult to
locate more than two dipoles for each potential/magnetic field recorded due to the numerical
instability of the inverse procedure. However, also with this limitation, this procedure is
very popular and allows locating sources mainly in the primary sensory cortical areas.
Fig. 8.3 shows the localization of a current dipole (the red arrow) indicating the restricted
cortical areas responsible for the generation of the characteristic magnetic field distribution
recorded by the magnetic sensors 20 ms after a stimulus delivery at right wrist in humans.
Multimodal Imaging from Neuroelectromagnetic 257
FIGURE 8.3. Localization of the equivalent current dipole (the red arrow) indicating the restricted cortical areas
responsible for the generation of the characteristic magnetic field distribution occurring over the magnetic sensors
20 ms after a stimulus delivered at the right wrist of a subject (N201P20). The position of the dipole is integrated
into a realistic head model built by segmenting sequential magnetic resonance images of the subject. See the
attached CD for color figure.
The position of the localized dipole was then integrated into a realistic head model built
according to the procedures described above.
The second approach to the dipole localization combines both the spatial and temporal
properties of scalp potentials/fields, to increase the ratio of available data to degrees of
freedom for the minimization procedures. This results in an increase of the number of
dipoles that may be reliably localized from EEG and MEG recordings. Different constraints
are applied to find the best inverse solutions, for example setting the position of the dipoles
and estimating the time/series of the dipole moments, or determining the orientation of the
dipoles and setting their positions. This last approach is called multiple source analysis
(MSA; Scherg and von Cramon, 1984; Ebersole 1997, 1999; Scherg et aI., 1999).
8.3.3 CORTICAL IMAGING
The possibility to model the complex head geometry with the finite element technique
allowed Alan Gevins and colleagues to derive a method, they called deblurring, that esti
mates potential distribution on the dura mater surface by using non invasive EEG recordings
(Le and Gevins, 1993; Gevins et aI., 1994). This method still uses nonlinear minimization
techniques but did not use any explicit model of the neural sources. In fact, by just applying
Poisson's equation, Gevins and coworkers were able to move back from the scalp potential
distribution to the dura mater potential distribution. This method was also validated by
258
3) Generated cortical
potentials from the
estimated dipole
strenghts
F. Babiloni and F. Cincott i
~ · I
I) EEG signa s
2) Estimated cortical
strenghts from EEG
siznals
FIGURE 8.4. Possible representation of the corti cal imaging technique. The acquired EEGscalp potentials are
used to estimate the cortical dipole strengths at the dipole layer level, here represented with the realistic cortical
surface. The estimated cortical dipole strengths (2) are then used to generate the potential distribution over a
dura mater surface (3) using the basic laws of electromagnetism. Such distribution can be generated at any other
modeled head structure.
using epicortical recording s, and the deblurred dura mater potential distributions showed a
clear improvement with respect to the examination of the raw potential distribution s over
the scalp. It is worth of note that the mathematical model supporting the deblurring method
is not suitable to accommodate fMRI or PET information. In fact, mathematical frame
work that allow integration between electromagnetic and metabolic modalities, require the
sources of currents in the brain to be explicitly modeled.
Another technique useful to recover improved images of cortical distributions from
EEG scalp recordings is known as cortical imaging. In this technique, an explicit model
of the neural sources, i.e. the current dipole, is used. In general, a layer of current dipoles
simulates the cortical surface, and the retrieved dipole strengths are then used to generate
potential distributions over a surface of the head model simulating the dura mater. Fig. 8.4
shows the idea at the base of the cortical imaging technique. A three shell (scalp, inner and
outer surface of the skull) realistic head volume conductor is represented, together with
the cortical dipole layer. It has been proven that even the use of homogeneous spherical
volume conductor for the head and a realistic cortical surface for the dipole layer provided
more focused and detailed information than the raw scalp potential s (Sidman et al., 1992,
Srebro et aI., 1993, Srebro and Oguz, 1997). However, it must be noted that the conductivity
ratio between skull and scalp is far than 1 as assumed in homogeneous models. The value
adopted for such ratio by all the researchers in the field in these last 30 years is 1:80
(Rush and Driscoll , 1968), or even I :15 as stated more recentl y (Oostendorp et al., 2000) .
According to this observation, several researchers (He et aI., 1996; Babiloni et aI., 1997;
He, 1999; He et al., 1999) developed cortical imaging techniques that took into account
the inhomogeneity of the head as volume conductor by using realisti c head model s and
boundary element mathematics. By regularization, dura mater potentials obtained both
Multimodal Imaging from Neuroelectromagnetic 259
from simulation and real recordings presented improved spatial characteristics with respect
to the use of raw scalp potentials. is worth of note that the mathematical framework of the
cortical imaging technique allows in principle the integration of tMRI priors. In fact, since
the cortical imaging method is a linear inverse technique (He, 1999), all the multimodal
integration we will present in the following paragraph for the EEG/MEG and tMRI data for
the distributed linear inverse solution can be in theory applied also for the cortical imaging.
Beside to the use of Green's second identity, another approach to the imaging of the
cortical potential distribution from noninvasive recordings was made by using the spherical
harmonic functions (Nunez et al., 1994; Edlinger et al., 1998). However, the mathematical
framework developed for the estimation of cortical potentials do not allow easily the inte
gration of tMRI information. In fact, since this method recovers the "deblurred" cortical
potential distribution but not the cortical current strengths, it is difficult to integrate the
information about the activation of patches of cortical tissues obtained by tMRI.
8.3.4 DISTRIBUTED LINEAR INVERSE ESTIMATION
As seen before, when the EEG activity is mainly generated by circumscribed cortical
sources (i.e. shortlatency evoked potentials/magnetic fields), the location and strength of
these sources can be reliably estimated by the dipole localization technique (Scherg et al.,
1984, Salmelin et al., 1995). In contrast, when EEG activity is generated by extended cortical
sources (i.e. eventrelated potentials/magnetic fields), the underlying cortical sources can
be described by using a distributed source model with spherical or realistic head models
(Grave de Peralta et al., 1997; PascualMarqui, 1995; Dale and Sereno, 1993). With this
approach, typically thousands of equivalent current dipoles covering the cortical surface
modeled and located at the triangle center were used, and their strength was estimated
by using linear and non linear inverse procedures (Dale and Sereno, 1993; Uutela et al.,
1999). Taking into account the measurement noise supposed to be normally distributed,
an estimate of the dipole source configuration that generated a measured potential b can be
obtained by solving the linear system:
Axl n e b (8.1)
where A is a m x n matrix with number of rows equal to the number of sensors and number
of columns equal to the number of modeled sources. We denote with A the potential
distribution over the m sensors due to each unitary jth cortical dipole. The collection of
all the mdimensional vectors A 1, ... , describes how each dipole generates the
potential distribution over the head model, and this collection is called the lead field matrix
A This is a strongly underdetermined linear system, in which the number of unknowns,
dimension of the vector x, is greater than the number of measurements b of about one
order of magnitude. In this case from the linear algebra we know that infinite solutions for
the x dipole strength vector are available, explaining in the same way the data vector b.
Furthermore, the linear system is illconditioned as results of the substantial equivalence
of several columns of the electromagnetic lead field matrix In fact, we know that each
column of the lead field matrix arose from the potential distribution generated by the dipolar
sources that are located in similar positions and have orientations along the cortical model
used. Regularization of the inverse problem consists in attenuating the oscillatory modes
260 F. Babiloni and F. Cincotti
generated by vectors that are associated with the smallest singular values of the lead field
matrix A, introducing supplementary and aprioriinformation on the sources to be estimated.
In the following, we characterize with the term "source space" the vector space in which
the "best" current strength solution x will be found. "Data space" is the vector space in
which the vector b of the measured data is considered. The electrical lead field matrix A
and the data vector b must be referenced consistently. Before we proceed to the derivation
of a possible solution for the problem drawn in (8.1) we recall few definitions of algebra
useful for the following. A more complete introduction to the theory of vector spaces is
out of the scope of this chapter, and the interested readers could refer to related textbooks
(Spiegel, 1978; Rao and Mitra, 1977). In a vector space provided with a definition of a inner
product (,.), it is possible to associate a value or modulus to a vector b by using the notation
(b, b) [b]. The notion of length of a vector can be generalized even in a vector space in
which the space axes are not orthogonal. Any symmetric positive definite matrix M is said
a metric for the vector space furnished with the inner product (.,.) and the squared modulus
of a vector b in a space equipped with the norm M is described by
(8.2)
With these recalls in mind, we now face the problem to derive a general solution of the
problem described in Eq. 8.1 under the assumption of the existence of two distinct metrics
Nand M for the source and the data space, respectively. Since the system is undetermined,
infinite solutions exist. However, we are looking for a particular vector solution ethat
has the following properties: 1) it has the minimum residual in fitting the data vector b
under the norm M in the data space 2) it has the minimum strength in the source space
under the norm N. To take into account these properties, we have to solve the problem
utilizing the Lagrange multiplier and minimizing the following functional that express
the desired properties for the sources x (Tikhonov and Arsenin, 1977; Dale and Sereno,
1993; Menke, 1989; Grave de Peralta and Gonzalez Andino, 1998; Liu, 2000) :
b l l ~ +)..1 I I x l I ~ )
(8.3)
The solution of the variational problem depends on adequacy of the data and source space
metrics. Under the hypothesis ofM and N positive definite, the solution of Eq. 3 is given by
taking the derivatives of the functional and setting it to zero. After few straightforward
computations the solution is
(8.4)
where G is called the pseudoinverse matrix, or the inverse operator, that maps the measured
data b onto the source space e.Note that the requirements of positive definite matrices for
the metric Nand M allow to consider their inverses . Last equation stated that the inverse
operator G depends on the matrices M and N that describe the norm of the measurements
and the source space, respectively. The metric M, characterizing the idea of closeness in
the data space, can be particularized by taking into account the sensors noise level by
using the Mahalanobis distance (Grave de Peralta and Gonzalez Andino, 1998). If no
a priori information is available for the solution of linear inverse problem, the matrices M
MultimodaI Imaging from Neuroelectromagnetic 261
and N are set to the identity, and the minimum norm estimation is obtained (Hamalainen
and Ilmoniemi, 1984). However, it was recognized that in this particular application the
solutions obtained with the minimum norm constraints are biased toward those sources
that are located nearest to the sensors. In fact, there is a dependence of the distance on the
law of potential (and magnetic field) generation and this dependence tends to increase the
activity of the more superficial sources while depresses the activity of the sources far from
the sensors. The solution to this bias was obtained by taking into account a compensation
factor for each dipole that equalizes the "visibility" of the dipole from the sensors. Such
technique, called column norm normalization by Lawson and Hanson in 1974, was used
in the linear inverse problem by PascualMarqui, 1985 and then adopted largely by the
scientists in this field. With the column norm normalization the inverse of the resulting
source metric is
(8.5)
in which is the ith element of the inverse of the diagonal matrix and is
the L2 norm of the ith column of the lead field matrix A. In this way, dipoles close to
the sensors, and hence with a large will be depressed in the solution of the inverse
problem, since their activations are not convenient from the point of view of the functional
cost. The use of this definition of matrix N in the source estimation is known as weighted
minimum norm solution (PascualMarqui, 1995; Grave de Peralta et aI., 1997).
The described mathematical framework is able to accommodate the information com
ing from EEG, MEG and tMRI data, as we will demonstrate in the following paragraphs.
8.4 MULTIMODAL INTEGRATION OF EEG, MEG AND FMRI DATA
Before we describe how it can be possible to implement methods that fuse data from
all modalities, some remarks are necessary about the neural sources that mayor may not
be retrieved by multimodal EEGMEGtMRI integration. In the following paragraphs, we
will present possible techniques for multimodal integration of EEG, MEG and tMRI data
by using a particularization of the metrics of the data and the source space, in the context of
the distributed linear inverse problem. In particular, we will show that the metric of the data
space M can be characterized to take into account the EEG and MEG data. Furthermore,
we will demonstrate how the source metric N can be particularized by taking into account
the information from the hemodynamic responses of the brain voxels.
8.4.1 VISIBLE ANDINVISIBLE SOURCES
Any neuroimaging technique has its own visible and invisible sources. The visible
sources for a particular neuroimaging technique are those neuronal pools whose spatio
temporal activity can be at least in part detected. In contrast, invisible sources are those
neural assemblies that produce a pattern of the spatiotemporal activity not detectable by
the analyzed neuroimaging technique. In the case ofEEG(or MEG) technique, it is clear that
the visible sources are generally located at the cortical level, since the cortical assemblies
are close to the recording sensors, and the morphology of the cortical layers allows the
262 F. Babiloni and F. Cincotti
generation of open (rather than closed) electromagnetic fields. On the other hand, it is often
poorly understood that the invisible sources for the EEG (or MEG) are all those cortical
assemblies that do not fire synchronously together. In fact, in a dipole layer composed by
coherent sources and Nincoherent ones, the potentials due to individual coherent sources are
combined by linear superposition, while the combination of the incoherent sources is only
due to statistical fluctuations. The ratio between the contributions of coherent to incoherent
source can be expressed by M1v'N(Nunez, 1995). Hence, if N is very large, say about 10
million of incoherent neurons that fire continuously, and is a small percentage of such
neurons (say 1%; about 100,000 neurons) that instead fire synchronously, we obtain that the
potential measured at the scalp level will be determined by 10 1M , with a net result of
about 30. Hence, only 1% of the active sources produce a potential larger than the other 99%
by a factor of 30 just because of the synchronicity property. This means that a cortical patch
may generate an EEG signal with no modification of its metabolic consumption, simply
by increasing the firing coherence of a small percentage of neurons. As a consequence of
that, neuroimaging techniques based on imaging of the metabolic/hemodynamic request of
the neural assemblies may detect no activity change with respect to the baseline condition.
However, there are other situations in which the visible sources for metabolic techniques
such as fMRI and PET can be invisible for EEG or MEG techniques. Stellate cells are
neurons present in the human cerebral cortex, and represent 15% of the neural population
of the neocortex (Braitenberg and Schuz, 1991). These cells occupy a spherical volume
within the cortex, thus generating essentially a closedfield electromagnetic pattern. Such
a field cannot be recorded at the scalp level by electrical or magnetic sensors, although the
actual firing rate of such stellate neurons is rather high with respect to the other cortical
neurons. This means that these neuronal populations present high metabolism requirements
that can be detected by the fMRI or PET techniques, while at the same time they are
"invisible sources" for the EEG and MEG techniques. Other example of invisible sources
for the EEG and MEG techniques are represented by the neural assemblies located at the
thalamic level, since they are also arranged in such a way to produce closed electromagnetic
field, while having high metabolic requirements.
8.4.2 EXPERIMENTAL DESIGNAND COREGISTRATION ISSUES
8.4.2.a Experimental design
Experimental setups that take into account both the electrical and the hemodynamic
responses as dependent variables have to be designed with particular attention. There are
two main important classes of setups that can be considered in a study of this type, depending
whether simultaneous EEG and fMRI measurements or just separate EEG/MEG and fMRI
recordings are scheduled. In the first case, many issues related to the coregistration of
the head can be easily overcome. However, in both cases the differences between the
hemodynamic and electric behavior have to be taken into account. In fact, consideration
about the signal to noise ratio (SNR) can limit the use of similar paradigms for EEG
and fMRI recordings. For instance, EEG/MEG response to very brief stimuli (such as
Somatosensory Evoked Potentials; SEPs, i.e. short electrical shock) can be recorded with a
high SNR, while the hemodynamic responses decrease its SNR by decreasing the stimulation
length. Furthermore, it has also been demonstrated that while EEG amplitudes decrease
Multimodal Imaging from Neuroelectromagnetic 263
by increasing the stimulation rate, the opposite it is true for the hemodynamic response
amplitudes (Wikstrom et al., 1996; Kampe et al., 2000). Experimental design for either
separate or simultaneous collection of electrophysiological and hemodynamic variables can
be easier when eventrelated fMRI technique is used, in contrast to the blockdesign fMRI.
In these experimental paradigms, the availability of the time behavior of the hemodynamic
response can be useful to design similar stimulation setup for both modalities.
S.4.2.b Coregistration
In the multimodal integration ofEEGIMEGdata with the fMRI, a common geometrical
framework has to be derived in order to locate appropriately the voxels whose EEG responses
is high and voxels whose hemodynamic response is increased/decreased during the task
performance. The issue of deriving a common geometrical framework for the data obtained
by different imaging modalities is called the "coregistration" problem (van den Elsen
et al., 1993). Several techniques can be used to produce an optimal match between the
realistic head reconstruction obtained in the high resolution EEGIMEG by the MRIs of
the experimental subject and the fMRI image coordinates. The first body of techniques
is based on the presence of landmarks on the both images used for the coregistration.
Corresponding landmarks have to be determined in both modalities (Fuchs et al., 1995).
A second body of techniques is based instead on the matching of surfaces belong to the
same head structure, as obtained by the different image modalities. In these techniques a
prerequisite is the segmentationof the structures whose surfaces have to be matched (Wagner
and Fuchs, 2001). With the volumebased registration technique no additional information
as landmarks or surface detection is necessary (Wells et al., 1997). In the case in which the
multimodal EEG and fMRI is performed simultaneously, the setup of a common geometrical
framework becomes simpler. In this case registration can be performed based on a scanner
coordinate system. As additional advantage, simultaneous measurement of EEG and fMRI
also allows an accurate coregistration of the electrode positions, a problem that in the other
cases have to be solved by using nonlinear minimization techniques.
8.4.3 INTEGRATION OF EEG AND MEG DATA
As mentioned before, electroencephalography (EEG) and magnetoencephalography
(MEG) are useful tools for the study of brain dynamics and functional cortical connectivity
due to their high temporal resolution (in the range of milliseconds). While the EEG reflects
the activity of neural generators oriented both tangentially or radially with respect to the
surfaces of electrodes, the MEG is more sensitive to cortical generators oriented tangentially
to the surfaces of sensors. However, the recorded EEG is a distorted copy of the cortical
potential distribution due to the poor conductivity of the skull, while the MEG is insensitive
to the different head tissues conductivities. In this framework an important question arises,
namely the importance of using one (EEG, MEG) or both modalities (EEG and MEG)
for increasing the accuracy of the estimated neural activity. Simulation studies aimed at
integrating data from MEG and EEG sensors with phantoms demonstrated an improvement
of spatial accuracy of the reconstruction methods when MEG and EEG data are fused
together (Phillips et al., 1997, Baillet et al., 1997, 1999). These simulation studies suggest the
possibility to practically integrate data from both EEG and MEG modalities in the solution
264 F. Babiloni and F. Cincotti
of some neurophysiological problems also in the case of distributed source problem. was
also demonstrated that the use of combined EEG and MEG data increase the stability and
the accuracy of the source activity estimate from primary sensory cortical areas of man with
respect to using modalities separately (Stock et aI., 1987; Fuchs et aI., 1998). In this context,
the question if the use ofcombined EEG and MEG measurements lead to a better estimate of
the distributed cortical activity with respect to the use of EEG or MEG separately, has been
recently addressed (Babiloni et aI., 2001). Here, we would like to expand in the following
paragraph concepts about the possibility to integrate EEG and MEG data in the context of
distributed linear inverse solutions.
At a first look, the attempt to integrate the EEG and MEG data in order to increase the
quality of the source reconstruction fails when we consider that the units of the potential and
magnetic field differ. How can one fuse such data together? In order to combine the different
measures of electric and magnetic data, both have to be converted to a common basis. This
conversion was performed by normalizing the measured signals to their individual noise
amplitudes, yielding unitfree measures for both electric and magnetic modalities. Such
normalization procedure was accomplished by using the covariance matrix of the electric
and magnetic noise as metric in the data space for the solution of the linear inverse problem.
The estimation of the noise covariance matrices requires the recording of several single
sweeps of EEG and MEG data, and the possibility to determine a segment of the recorded
data in which no taskrelated activity is present. Then, on all the sweeps recorded and for
the time period of interest, the maximum likelihood estimate for the covariance matrices of
the electrical N, and magnetic N noise matrices have to be computed. With the use of these
matrices we can produce the block covariance matrix of the electromagnetic measurement,
by posing S N, N i.e horizontally adjoining the two matrices. The forward solution
specifying the potential scalp field due to an arbitrary dipole source configuration is solved
on the basis of the linear system
(8.6)
where (i) E is the electric lead field matrix obtained by the boundary element technique for
the realistic MRIconstructed head model; (ii) B is the magnetic lead field matrix obtained
for the same head model; (iii) x is the array of the unknown cortical dipole strengths; (iv) v
is the array of the recorded potential values; and (v) m is the array of magnetic field values.
The lead field matrix E and the array v were referenced consistently. In order to scale EEG
and MEG, the rows of the lead field matrix E and B were first normalized by the norm of
rows (Phillips et aI., 1997). This scaling was equally applied on the electrical and magnetic
measurements arrays, v and m.
As noted before, the inverse operator G is expressed in terms of the matrices M and
N that regulates the metric in the measurement and source space, respectively. Here, M is
now equal to the inverse of the covariance matrix S of the noise of the normalized EEG
and MEG sensors, while N is the matrix that regulates how each EEG or MEG sensor is
influenced by dipoles located at different depths of the source model. The covariance matrix
S was derived from the normalized EEG (v) and MEG (m) data by maximum likelihood
estimation as described before. The matrix N is a diagonal matrix in which the ith element
is equal to the norm of the ith column of the normalized lead field matrix
Multimodallmaging from Neuroelectromagnetic 265
RIGHT MOVEMENT
EEG
t 0 +1
T I M E ~
MEG
t 0 +1
T1ME<ec)
FIGURE 8.5. Disposition of the electric (up) and magnetic (bottom) sensors for the recording of EEGand MEG
data related to unilateral voluntaryfingermovements (separate recording sessions). Averaged MEGand EEGtime
series (waveforms) recorded fromtwo selected magnetic (Ml and M2) andelectric (El and E2) sensors are shown
on the right of the figure. These sensors overlay the primarysensorimotor cortex contralateral to the movement.
See the attached CDfor color figure.
In the following , the above methodology is applied to the EEG and MEG data related
to the preparation and the execution of the voluntary movement of the right index finger.
Fig. 8.5 presents the disposition of the electric (up) and magnetic (bottom) sensors for the
recording of EEG and MEG data (separate recording sessions). Averaged MEG and EEG
time series (waveforms) recorded from two selected magneti c (MI and M2) and electric
(EI and E2) sensors in a normal healthy subject execut ing the movement are shown on the
right of Fig. 8.5. These sensors overlay the primary sensorimotor cortex contralateral to
the movement , which is known to be active both during preparation and execution of the
movement. Fig. 8.6 shows linear inverse estimates from EEG, MEG, and combined EEG
MEG data recorded fromthe subject about 110 ms after the onset ofEMG activity associated
with selfpaced right finger movement. Raw EEG and MEG distributions present large and
distant negative and positive maxima preponderant in the side contralateral to the movement,
the electric field being tilted of 90· with respect to the magnetic field. In contrast, linear
inverse estimates were characterized by circumscribed zone of negativity and positivity in
both sides. Linear inverse estimate of EEG data (MovementRelated Response 1; MRR 1)
shows negative maxima in the mesialfrontal and contralater al frontal areas, and a zone
of minor negativity in the ipsilateral frontal area. In addition, there is a reversed parietal
266 F. Babiloni and F. Cincotti
FIGURE8.6. Amplitude gray scale 3D maps showing linear inverse estimates of electroencephalographic (EEG),
magnetoencephalographic (MEG), and combined EEGMEG data recorded (12850 channels, respectively) from
a subject about 110 ms after the onset of the electromyographic response accompanying a voluntary brisk right
middle finger extension. Percent gray scale is normalized with reference to the maximum amplitude calculated
for each map. Maximum negativity is coded in white and maximum positivity (+1 in black.
positivity in both sides. Compared to this distribution, linear inverse estimate of MEG data
(Movement Evoked Field l; MEF 1) present closer bilateral frontal negativity and parietal
positivity as well as no negativity in the mesialfrontal area. The linear inverse estimate of
the combined EEGMEG data show a high spatial resolution content integrating features of
linear inverse estimate of EEG and MEG data considered separately. This result does not
depend on the increase of the total number of sensors (EEG electrodes MEG coils) since
it was also reproducible by comparing purely EEG data and data obtained substituting some
of the EEG channels with MEG channels (thus preserving the total number of channels in
the two datasets, Babiloni et al., 2001).
is worth of note that the presented results are obtained with separate EEG and MEG
recordings. An important source of variance in the linear inverse source analysis of combined
EEG MEG data might be caused by the nonsimultaneous recording of these data sets, when
attentional, learning and emotional variables are unpaired across the recording blocks.
Regarding the experiments presented above, it must be stressed that movementrelated
potentials/fields are very stable across experimental sessions performed in different days.
Furthermore, the participant subjects were preliminary trained to stabilize a simple motor
Multimodal Imaging from Neuroelectromagnetic 267
performance in which learning, attentional attentive and emotional concomitants seem to be
negligible. Finally, the possibility of combining MEG and EEG data recorded in different
experimental sessions offers the opportunity of scientific and clinical cooperation between
centers that own only one of the imaging systems. In contrast, the use of simultaneous
multimodal EEGMEG recordings is nowadays possible only in very few laboratories.
In conclusion, the present results for the multimodal integration of EEG and MEG sug
gest that the linear inverse source estimates of the combined EEGMEG data improve with
respect to those of EEG or MEG data considered separately. The methodological approach
includes MRbased realistic head and cortical source modeling, high spatial sampling of
real EEG and MEG data, and regularized linear inverse estimate mathematics (weighted
minimum norm source estimate). The application of this technology supports the hypothe
sis that in humans the preparation and execution of one of the simple unilateral volitional
digit acts is subserved by a distributed cortical system including SMA and contralaterally
preponderant Ml and S1.
8.4.4 FUNCTIONAL HEMODYNAMIC COUPLING AND INVERSE ESTIMATION
OF SOURCE ACTIVITY
Here we describe how it is possible to take into account the information from the
hemodynamic coupling between the cortical areas in the estimation of the timevarying
source strengths by solving the electromagnetic inverse problem.
8.4.4.a Multimodal integration ofEEG/MEG andfMRI data with dipole
localization techniques
As described previously, the dipole localization techniques locate neural activity by
using few current equivalent dipoles from noninvasive EEG and/or MEG recordings. The
use of spatial information from hemodynamic methods as a constraint to the electromagnetic
inverse problem necessitates the assumption that the brain areas that appear active with
different methods are to some extent the same. Dipole localization techniques could use the
tMRI information essentially in two way: 1) by setting the foci of the tMRI hotspots in
the brain as initial locations of the equivalent current dipoles localization procedure. Since
the minimization procedure is non linear, it is hence dependent by the initial position of the
current dipole(s); 2) by setting the position of the current dipoles at the tMRI hotspots,
allowing to the localization procedure to rotate freely only the direction of the dipoles to
fit the EEG or MEG data. Methods I) and 2) can also be combined. In order to take into
account possible electrical sources that were not detected as active fMRI spots, some current
dipoles that are not constrained to fMRI spots' positions may be added to the source model.
Recently, a simulation study suggested that EEG dipole fits could benefit from fMRI
constraints (Wagner and Fuchs, 2001) . A reconstructed dipole in the vicinity of each fMRI
hotspot yields the corresponding source time course. However, spatially unconstrained
dipoles are then necessary to account for remaining simulated EEG activity, to appropriately
locate also sources that are invisible to the fMRI hotspots . Moving from the simulation
to the application studies, the spatial correspondence between EEG/MEG and tMRI has
been mostly investigated in the context of motor and somatosensory evoked activity. In
the localization of the primary sensory and motor areas using equivalent current dipole
268 F. Babiloni and F. Cincotti
modeling, typical spatial differences between the dipole location and the center of tMRI
activation have been between 10 to 16 millimeters (Sanders et al., 1996; Beistener et al.,
1997). In light of the current limited knowledge, the localization results seem to agree
reasonably well when locating responses from primary sensory areas. Extending comparison
to more complex cortical networks has indicated mostly converging activation patterns
(Ahlfors et al., 1999; Korvenoja et al., 1999). In the last study of Korvenoja and coworkers,
five subjects participated in both a somatosensory evoked field (SEFs) recording with a
Neuromag 122 and a tMRI recording. The goal was to maximize the response amplitude
in both imaging modalities, in order to achieve the best possible signaltonoise ratio. The
SEFs were first modeled independently of tMRI results. Thereafter a multidipole model
was constructed by placing equivalent current dipoles (ECDs) to tMRI activation centroids.
MEG data indicated activity in some of the eight source areas but no tMRI activation was
seen, then ECD location from independent MEG data analysis was used. The eight dipoles
were spatially fixed, but were allowed to change their orientation and amplitude to explain
the data (rotating dipoles). The ECD orientations were remarkably stable over the whole
analysis period (0 to 400 ms post stimulus). The timecourses of activation in the model
were found to agree with data, which have been obtained with invasive electrophysiological
methods (Allison et al., 1989; 1996). In another study performed by the same research
group, the temporal dynamics of visual motion areas was analyzed by combining tMRI and
MEG data (Ahlfors et al., 1999). Also in that case, results indicated that while complete
overlap of activation patterns determined independently from MEG and tMRI did not exist
on individual level, the activation patterns did converge at the group level. In a study on
the sources of movementrelated magnetic fields (Baillet et al., 2001), the locations of the
contralateral and SMA sources found with MEG dipole localizations were found rather
close to the maximum of the closest tMRI clusters (12 mm and 4 mm, respectively).
Cortical remapping of the focal parametric source model was performed and the cortical
clusters corresponding to the contralateral and central sources indicated a good match in
location with the tMRI regions. Other studies have combined the analysis of hemodynamic
and electrophysiological data that were collected separately, by using both the Positron
emission tomography (PET) and EEG (Heinze et al., 1994; Snyder et al., 1995; Heinze
et al., 1998) or the tMRI and EEGIMEG (Belliveau 1993; Morioka et al., 1995; George
et al., 1995; Menon et al., 1997; Opitz et al., 1999). Example of integration of hemodynamic
or metabolic techniques and invasive cortical recordings can be found in the study of Luck
and colleagues with tMRI and EEG invasive recordings (Luck, 1999) or in that of Lamusuo
and coworkers, that integrates PET, MEG and invasive recordings (Lamusuo et aI., 1999).
Simultaneous acquisition of EEG and tMRI data is necessary when the activity of
interest cannot be easily reproduced, as it happens in the case of the epilepsy studies. In
this case the epileptiform activity between seizures could be produced by different cortical
generators. Simultaneous EEG/tMRI recordings (HuangHellinger et al., 1995; Warach
et aI., 1996; Seeck et aI., 1998; Krakow et al., 1999) have been used to measure such
activity, by hypothesizing that epileptiform discharges are likely to produce neural activity
measurable by tMRI (Ives et aI., 1993). These studies recorded interleaved EEG and tMRI
to monitor for the presence of interictal activity without, however, localizing the EEG
activity. In a recent study of comparison of spiketriggered tMRI activation hotspots and
EEG dipole model in six epileptic patients, an average of 3 em of mismatch between the
tMRI hotspots and EEG localized current dipoles was found (Lemieux et al., 2001). In this
Multimodal Imaging from Neuroelectromagnetic 269
FIGURE 8.7. Comparisons between foci of interictal epileptic activity as obtained by MEG dipole localization
technique, over threshold fMRI clusters and site of brain lesion in a patient evaluated before surgery. See the
attached CD for color figure.
case the authors concluded that the combination of EEG and tMRI techniques offers the
possibility of advancing the study of the generators of epileptiform electrical activity. Such
results are in line with those found by the group of Romani and coworkers, in which they
found some centimeters of displacement between tMRI hotspots, localized dipoles on the
base of MEG recordings and brain l ~ s i o n s before surgical operation in epileptic patients.
Fig. 8.7 presents such displacements between the various hotspots and the brain lesion.
There are occasions where disagreement in spatial activation patterns could exist for
the integration of EEG/MEG and tMRI data. is not clear, for example, whether very
shortlasting synchronous firing, which can be detected in EEG and MEG, will produce
a detectable hemodynamic change. Eventrelated synchronization and resynchronization
are phenomena that possibly remain undetected by observing hemodynamic changes. For
example, in the study by Ahlfors et al. (1999) MEG indicated activity over the frontal cortex
bilaterally while tMRI did not demonstrate any activity in similar areas. This is a typical
example of the coexistence of visible and invisible sources in a same behavioral task for
the MEG and tMRI techniques.
Taken together, the results above indicate that while it may not be possible to simply
restrict the source model solutions to areas where tMRI shows activation, it still seems to
be a valuable aid in the validation of the source model. Converging lines of evidence from
multiple methods will increase the likelihood of correct solution. The ultimate way to vali
date the inverse solution would be the invasive recordings. However, it is worth of note that
270 F. Babiloni and Cincotti
fMRI priors used in conjunction with the dipole localization techniques require a significant
manual intervention, since one must decide in which regions of the fMRI activation a dipole
is to be "placed". This could provide at least a difficulty in the replication of findings be
tween different researchers, since the dependence of the results by experimenter's choice. In
this respect, it is of interest to note that the combined fMRI, EEG and MEG procedure that
utilizes the mathematical framework of distributed linear inverse solutions do not requires
such "manual" intervention. Such procedure will be presented in the following paragraphs.
8.4.4.b Multimodal integration ofEEG/MEG andfMRI data with distributed model by
using diagonal source metric
Here, we present two characterizations of the source metric N that can provide the ba
sis for the inclusion of the information about the statistical hemodynamic activation of ith
cortical voxel into the linear inverse estimation of the cortical source activity. In the fMRI
analysis, several methods to quantify the brain hemodynamic response to a particular task
have been developed. However, in the following we analyze the case in which a particular
fMRI quantification technique has been used, called Percent Change (PC) technique. This
measure quantifies the percentage increase of the fMRI signal during the task performance
with respect the rest state (Kim et aI., 1993). The visualization of the voxels' distribution
in the brain space that is statistically increased during the task condition with respect to the
rest is called the PC map. The difference between the mean rest and movementrelated
signal intensity is generally calculated voxelbyvoxeI. The restrelated fMRI signal inten
sity is obtained by averaging the premovement and recovery fMRI. Bonferronicorrected
Student's ttest is also used to minimize alpha inflation effects due to multiple statistical
voxelbyvoxel comparisons (Type I error; p 0.05). The introduction of fMRI priors into
the linear inverse estimation produces a bias in the estimation of the current density strength
of the modeled cortical dipoles. Statistically significantly activated fMRI voxels, which are
returned by the percentage change approach (Kim et aI., 1993), are weighted to account for
the EEG measured potentials.
In fact, a reasonable hypothesis is that there is a positive correlation between local
electric or magnetic activity and local hemodynamic response over time. This correlation
can be expressed as a decrease of the cost in the functional of Eq. 8.3 for the sources Xj in
which fMRI activation can be observed. This increases the probability for those particular
sources Xj to be present in the solution of the electromagnetic problem. Such thoughts can
be formalized by particularizing the source metric N, to take into account the information
coming from the fMRI. The inverse of the resulting metric is then proposed as follows
(Babiloni et aI., 2000):
(8.7)
in which )ii and has the same meaning described above. is a function of the
statistically significant percentage increase of the fMRI signal assigned to the ith dipole of
the modeled source space. This function is expressed as
2=1+(K1)
(8.8)
Multimodal Imaging from Neuroelectromagnetic 271
where is the percent age increase of the tMRI signal during the task state for the ith voxel
and the factor tunes tMRI constraints in the source space. Fixing 1 let us disregard
tMRI priors , thus returning to a purely electrical solution; a value for 1 allows only the
sources associated with tMRI active voxels to participate in the solution. It was shown that
a value for K in the order of 10 (90% of constraints for the tMRI information) is useful to
avoid mislocalization due to over constrained solutions (Liu et aI., 1998; Dale et aI., 2000;
Liu, 2000). In the following the estimation of the cortical activity obtained with this metric
will be denoted as diagtMRI, since the previous definition of the source metric N results
in a matri x in which the offdiagonal elements are zero.
8.4.4.c Multimodal integration of EEG/MEG andfMRIdata with distributed model
by usingfull source metric
In the previous paragraphs, we observed that incorporating a priori information for
each cortical voxel about geometrical orientation with respect to the cortical surface and
about hemodynamic response we obtain an estimate of the cortical activity that improves
the reconstruction generated without such constraints. However, it must be noted that
all the formulations presented in literature on the integration of EEG, MEG with tMRI
did not take into account informati on about the functional coupling of the neural sources.
In fact, these formulation s only use the information about the presence or absence of a par
ticular source located at the voxel level in the set of those whose hemodynamic responses
have been elicited by the considered task. However, the theoretical possibility to include
this source of information in the linear inverse problem was already mentioned in previous
articles (Dale and Sereno, 1993; Liu et aI., 1998; Dale et aI., 2000; Liu, 2000).
In this paragraph we present an extension of the linear inverse problem aimed to
taking also into account information about the functional coupling of the cortical sources,
as provided experimentally by the hemodynamic responses returned by the eventrelated
tMRI. In particular, we estimate the hemodynamic correlation of the neural sources by
using the crosscorrelation technique on the hemodynamic waveforms obtained during the
performance of the task under the tMRI scanner. These correlation values are then used as
additional constraints in the solution of the electromagnetic linear inverse problem
together with the cortical orientation constraints and the presence of statistically significant
activation of the hemodynamic response. We take advantage of the offdiagonal elements of
the matrix N to insert the information about the functional coupling of the cortical sources.
In particular we set the generic entry of the inverse of matrix N as in the following
(8.9)
where and have the same meaning described above and is the degree
of functional coupling between source i and source during the particular task analyzed .
Information on coupling is revealed by the correlation of their hemodynamic responses
obtained by the eventrelated tMRI data. In the following the estimation of the cortical
activity obtained with this metric will be denoted as corrtMRI. It is of interest that in
the case of uncorrelated sources tcorn, == 0, i i= j; corn, == 1), the corrtMRI formulation
leads back to the diagtMRI one. Fig. 8.8 summarizes the different approaches pursued here
272 F. Babiloni and F. Cincotti
.. .....
~ 102
~
.J::, 101 . .. . : ' . / . ~ :  . ~ ....! ;: ..•... . y...... .. .... • sa
~ 'W.
123 • 56 7 a
Time[sl
=
corr  fMRl
=
8 , diag  fMRI
FIGURE 8.8. Upper part: Estimat e of the hemodynamic coupling between two generic cortical sources (ith and
j th) as obtained by the computat ion of the crosscorrelati on between the waveforms of the IMRI responses. These
waveforms Sj ) are obtained during a simple voluntary movement (right middle finger extension). Lower part :
mathematical formulation of the inverse of the source metric N to be used in the solution of the linear inverse
problem. Corr(Sj, Sj ) is the zerolag correlation between the two hemodynamic waveforms Sj, and Sj, and <l ij is
the Kroneker symbol.
in order to insert the hemodynamic constraints in the solution of the linear inverse problem
for the estimation of the cortical sources of the recorded EEG in a unique mathematical
formulation.
8.4.4.d Application ofthe multimodal EEGfMRI integration techniques to the
estimation ofsources ofselfpaced movements
In this section we will provide a practical example of the application of multimodal
integration techniques of EEG, MEG and fMRI (as theoretically described in the previous
sections) to the problem of detection of neural sources subserving unilateral selfpaced
movements in humans. The high resolution EEG recordings (128 scalp electrodes) were
performed on normal healthy subjects by using the facilities available at the laboratory
of the Department of Human Physiology, University of Rome "La Sapienza". Realistic
head models were used, each one provided with a cortical surface reconstruction tessellated
with 3,000 current dipoles . Separate block design and eventrelated fMRI recordings of
the same subject s were performed by using the facilities of the Istituto Tecnologie Avan
zate Biomediche (ITAB) of Chiety, Italy, leaded by prof. Gian Luca Romani. Distributed
linear inverse solutions by using hemodynamic constraints were obtained according to the
methodology presented above.
An example of multimodal integration between EEG and fMRI related to a simple vol
untary movement task by using only the hemodynamic information relative to the strength
Multimodal Imaging from Neuroelectromagnetic 273
FIGURE 8.9. Amplitude gray scale 3D maps showing linear inverse estimates from high resolution electroen
cephalographic (HREEG) and combined functional magnetic resonance image (fMRI)HREEG data computed
from a subject about 20 ms after the onset of the electromyographic activity associated with selfpaced right middle
finger movements (motor potential peak, MPp). Percent gray scale of HREEG and combined fMRIHREEG data
is normalized with reference to the maximum amplitude calculated for each map. Maximum negativity ( 00%)
is coded in white and maximum positivity (+I00%) in black.
of fMRI data (according to Eq. 8.7) is provided in Fig. 8.9. This figure shows amplitude
gray scale maps of linear source inverse estimates from EEG and combined fMRIEEG
data, computed about 20 ms after the onset of the electromyographic response to voluntary
right finger movements (Motor Potential peak; MPp). fMRI data indicate maximum acti
vated voxels clustered in bilateral primary motor (Ml), primary somatosensory (Sl), and
supplementary motor (SMA) areas, the fMRI signal intensity being much more higher on
the contralateral (left) side. The linear inverse estimate of neural activity for the HREEG
and combined fMRIHREEG data were mapped over the cortical compartment of a realistic
MRIconstructed subject's head model. The MPp map presents maximum responses in the
contralateral Ml and Sl and in the modeled SMA. With respect to the HREEG solutions
(left), the fMRIHREEG solutions present more circumscribed Ml, S1, and SMA responses.
In addition, the contralateral Ml and Sl responses have similar intensity and are spatially
dissociated.
An example of the multimodal integration between EEG and fMRI data by using both
block and event related experimental designs is depicted in Fig. 8.10. In this figure, the upper
row illustrates the topographic map of readiness potential distribution recorded at the scalp
about 200 ms before a right middle finger extension for another subject analyzed. Note the
extension of the maximum of the negative scalp potential distribution, roughly overlying
274 F. Babiloni and F. Cincotti
Right finger movement
Readiness Potential tMRI Response
notMRI
Linear Inverse solutions
diagtMRI
corrtMRI
100% +100%
FIGURE 8.10. Top left: scalp potential distribution recorded about 200 ms before the movement onset (128
recording channels) in a separate session. This distribution is representative of the socalled readiness potential.
Percent color scale in which maximum negativity is coded in red and maximum positivity is coded in black. Top
right: cortical tMRI response related to the movement. Only the tMRI voxels close to the cortical surface contribute
to source weighting. The intensity of yellow codes the percentage of the increase of the tMRI response. Bottom
row: Cortical distributions of the current density estimated with a linear inverse approach from the readiness
potential shown in the top row. Linear inverse estimates are obtained with no tMRI constraints (left, notMRI)
and two kinds oftMRI constraints, one based on the strengths of the cortical tMRI responses (center, diagtMRI)
and the other on the correlation between tMRI responsive cortical areas (right, corrtMRI). Percent color scale:
maximum negativity is coded in red and maximum positivity is coded in black. See the attached CD for color
figure.
frontal and centroparietal areas contralateral to the movement. The percent values of the
tMRI response during the movement in a separate experimental session are also illustrated.
The maximum values of the tMRI responses are located in the voxels roughly correspond
ing to the primary somatosensory and motor areas (hand representation) contralateral to
the movement. In fact, during the selfpaced unilateral finger extension, somatosensory
reafference inputs from finger joints as well as cutaneous nerves are directed to the primary
somatosensory area, while centrifugal commands from the primary motor area are directed
Multimodal Imaging from Neuroelectromagnetic 275
towardthe spinal cord via the pyramidal system. The lower rowof Fig. 8.10 illustrates the
cortical distribution of the current density estimatedwith linear inverseapproachfrom the
potentialdistributionof the upperrow. LinearinverseprocedureusednotMRI constraint as
well as two types offMRI constraints, i.e. one based on blockdesign (diagfMRI, Eq. 8.7)
and the other on eventrelated design (corrfMRI, Eq. 8.9). The cortical distributions are
represented on the realistic subject's head volume conductor model. Linear inverse solu
tions obtainedwith the fMRI priors (diag and corrfMRI)present more localizedspots of
activations withrespect to thoseobtainedwiththe no fMRI priors. Remarkably, the spotsof
activation are localizedin the hand regionof the primarysomatosensory (postcentral) and
motor (precentral) areas contralateralto the movement. In addition, spots of minor activa
tionwereobservedin the frontocentral medial areas (includingsupplementary motor area)
and in the primary somatosensory and motor areas of the ipsilateral hemisphere. Similar
resultswereobtainedin the other maincomponentsof the movementrelatedpotentials(i.e.
motor potentialsand movementevoked potential).
8.5 DISCUSSION
Thischapterrevieweda mathematical framework for theintegration ofEEG, MEGand
fMRIdata. Besides, advantages anddrawbacks relatedto thelocalizationtechniquesincon
junctionwithfMRI andPETdata havealsobeenreviewed. In general, thereis a rather large
consensus about the need and utility of the multimodal integration of metabolic, hemody
namicandneuroelectrical data. Resultsreviewedin literatureas well as thosepresentedhere
suggesta real improvement in the spatialdetailsof theestimatedneural sourcesbyperform
ingmultimodal integration. However, whilefor themultimodalintegration ofEEG andMEG
data a precise electromagnetic theoryexists, a clear mathematical and physiologiclink be
tweenmetabolic demandsandfiring ratesof theneuronsis still lacking.It is out of doubtthat
whenthislinkis furtherclarified, themodelingof the interaction betweenhemodynamic and
neural firing rate can be further refined. This will lead us to a more proper characterization
of the issues of visible and invisiblesource that at the moment represent the major concern
about the applicability of the multimodal integration techniques (Nunez et aI., 2000).
The results for the multimodal integrationof EEG/MEGand fMRI presented in this
chapter are in line with those regarding the coupling between cortical electrical activity
and hemodynamic measure as indicated by a direct comparison of maps obtained using
voltagesensitive dyes (which reflect depolarization of neuronal membranes in superficial
corticallayers) andmapsderivedfromintrinsicopticalsignals(whichreflectchangesinlight
absorptiondue to changesin bloodvolumeand oxygenconsumption, Shohamet aI., 1999).
Furthermore, previous studies on animals have also shown a strong correlation between
local field potentials,spikingactivity, and voltagesensitive dye signals(Arieli et aI., 1996).
Finally, studies in humans comparing the localization of functional activity by invasive
electrical recordings and fMRI have providedevidenceof a correlation between the local
electrophysiological andhemodynamic responses(Puceet aI., 1997). It is worthof notethat
recentlya studyaimedat investigating this link has beenproduced(Logothetis et aI., 2001).
In this study, intracortical recordingsof neural signalsand simultaneousfMRI signals were
acquired in monkeys. The comparisons were made between the local field potentials, the
multiunit spiking activityand BOLDsignals in the visual cortex. The study supports the
276 F. Babiloni and F. Cincotti
link between the local field potentials and BOLD mechanism, which is at the base of the
procedure of the multimodal integration of EEGIMEG with fMRI described above. This
may suggest that the local fMRI responses can be reliablyused to bias the estimationof the
electrical activity in the regions showinga prominent hemodynamicresponse.
It may be argued that combined EEGfMRI responses could be less reliable for the
modeling of cortical activation in the case of a spatial mismatch between electrical and
hemodynamicresponses. However, previousstudies havesuggestedthat by using the fMRI
data as a partial constraint in the liner inverse procedure, it is possible to obtain accurate
sourceestimatesof electricalactivityeveninthepresenceof somespatialmismatchbetween
the generators ofEEG data and the fMRI signals (Liuet al., 1998; Liu, 2000). Furthermore,
it is questionable whether the level of bias for the hemodynamic constraints in the linear
inverse estimation can be the same with the diagfMRI and corrfMRI approaches. This
issue seems to deserve a specific simulation study, using the literature indexes capable of
assessingthe qualityof the linear inversesolutions(PascualMarqui, 1995;Gravede Peralta
et al., 1996; Grave de Peralta and Gonzalez Andino, 1998, Babiloni et al., 2001).
The multimodal integrationof fMRI, MEGandEEGdata constitutesan unsurpassable
noninvasive technologyfor the analysis of humanhigher brainfunctions at a hightemporal
and a good spatial resolution.
ACKNOWLEDGMENTS
The Authors express their gratitude to the followingcolleagues that have participated
in the researches described above: dr. Claudio Babiloni, dr. Filippo Carducci, prof. Gian
Luca Romani, dr. Cosimo Del Gratta, dr. VittorioPizzella, prof. Paolo Rossini.
REFERENCES
Ahlfors, S.P., Simpson, G.V., Dale, A.M., Belliveau J.w., Liu, A.K., Korvenoja, A., Virtanen, J., Huotilainen,
M., Tootell, R.B., Aronen, H.I., and Ilmoniemi, R.I., 1999, Spatiotemporal activity of a cortical networkfor
processing visual motion revealed by MEG and Journal of Neurophysiology,
Allison, T., McCarthy, G., Wood, e.e.,Darcey, T.M., Spencer, D.O., and Williamson, P.D., 1989, Humancortical
potentials evoked by stimulation of the median nerve. 1. Cytoarchitectonic areas generating shortlatency
activity, Journal ofNeurophy siology
Allison, T., McCarthy, G., Luby, M., Puce, A., and Spencer, D.O., 1996, Localization of functional regions of
human mesial cortex by somatosensory evoked potential recording and by cortical stimulation, Electroen
cephalography Clinical Neurophysiology
Arieli, A., Sterkin, A., Grinvald, A., Aertsen, A.D., 1996, Dynamics of ongoing activity: Explanationof the large
variability in evokedcortical responses, Science 273: I86871.
Babiloni, E. Babiloni, C., Carducci, E, Fattorini L. et al., 1997, A high resolution EEG: a new modeldependent
spatial deblurring method using a realistically shaped MRconstructed subjects head model, Electroenceph.
d in. Neurophysiol.
Babiloni, E, Carducci, E, Cincotti, E, Del Gratta, Roberti, G.M, Romani, G.L., Rossini, P.M., and Babiloni,
C., 2000, Integration of High Resolution EEG and Functional Magnetic Resonance in the Study of Human
MovementRelated Potentials, Methods of Inf ormation in Medicine 39(2):17982.
Babiloni, E, Carducci, E, Cincotti, E, Del Gratta, C., Pizzella, Romani, G.L., Rossini, P.M., Tecchio E, and
Babiloni, C; 2001, Linear inverse source estimate of combined EEG and MEG data related to voluntary
movements, Human Brain Mapping,
Multimodal Imaging fromNeuroelectromagnetic 277
Baillet, S. and Gamero, 1997, A bayesian framework to introducing anatomofunctional priors in the EEG/MEG
inverse problern., IEEE Trans. Biom. Eng. 44:37485.
Baillet, S., Garnero, Marin, G., and Hugonin, P., 1999, J Combined MEG and EEG source imaging by
minimization of mutual information, IEEE Trans. Biom. Eng. 46:52234.
Baillet, S., Leahy, R., Singh, M., Shattuck D., and Mosher, J., 2001, Supplementary Motor Area Activation
Preceding Voluntary Finger Movements as Evidenced by Magnetoencephalography and fMRI, International
Journal of Bioelectromagnetism, 1(3).
Bandettini, P. A. (1993) Functional MRI ofthe Brain, Soc. Magnetic Resonance in Medicine, Berkeley, CA
Blamire, A. M., Ogawa, S., Ugurbil, K., Rothman, D., McCarthy, G., Ellerman, 1. M., Hyder, E, Rattner, Z., and
Shulman, R. G., 1992, Proc. Natl. Acad. Sci. USA 89:1106973.
Braitemberg, V.and Schuz, A., 1991. Anatomy of the cortex. Statistics and Geometry. New York: SpringerVerlag.
Belliveau J. 1993. MRI techniques for functional mapping of the human brain: integration with PET, EEG/MEG
and infrared spectroscopy. In: Quantification ofBrain Function. Elsevier Science Publishers (Excerpta Med
ica), Amsterdam. 63967.
Beisteiner, R., Erdler, M., Teichtrneister, C., Diemling, M., Moser, E., Edward, v., and Deecke 1997, Magne
toencephalography may help to improve functional MRI brain mapping, European Journal ofNeuroscience
9(5): 10727.
Bonrnassar, G., Van Der Moortele, P, Purdon, P, Jaaskelainen, 1., Ives, 1., Vaughan, T., Ugurbil K., and Belliveau
J., 2001, 7 Tesla interleaved EEG and fMRI recordings: BOLD measurements, NeuroImage 13(6):S6.
Cohen, D., Cuffin B., Yunokuchi, K., Manieski, R., Purcell, C, Cosgrove, G.R., Ives, J., Kennedy, 1., and Schomer
D., 1990, MEG versus EEG localization test using implanted sources in the human brain. Ann. Neurol.,
28:811817.
Dale, AM., and Sereno, M., 1993, Improved localization of cortical activity by combining EEG anf MEG with
MRI cortical surface reconstruction: a linear approach, J. Cognitive Neuroscience, 5:16276.
Dale, A M., Fischl, B., Sereno, M. 1.,1999, Cortical surfacebased analysis. 1. Segmentation and surface recon
struction, Neuroimage 9(2): 17994.
Dale, A, Liu, A, Fischl, B., Buckner, R., Belliveau, J. W, Lewine, 1., and Halgren, E., 2000, Dynamic Statistical
Parametric Mapping: Combining fMRI and MEG for HighResolution Imaging of Cortical Activity, Neuron
26:5567.
Ebersole, J. Defining epileptogenic foci: past, present, future., 1997, Journal ofClinical Neurophysiology, 14:470
483.
Ebersole, 1., 1999, The last word. Journal of Clinical Neurophysiology, 16:297302
Edlinger, G., Wach, P, and Pfurtscheller, G., 1998, On the realization of an analytic highresolution EEG, IEEE
Trans. Biomed. Eng. 45:73645.
Fuchs, M., Wischmann, H.A., Wagner, M., and Kruger, 1., 1995, Coordinate System Matching for Neuromagnetic
and Morphological Reconstruction Overlay, IEEE Transactions on Biomedical Engineering. 42:416420.
Fuchs M., Wagner M., Wischmann H.A., Kohler T., Theissen A., Drenckhahn R., Buchner H., 1998, Improving
source reconstruction by combining bioelectrical and biomagnetic data. Electroenceph din Neurophysiol
107:6980.
George J. Aine, J., Mosher, J. Schmidt D. M., Ranken D. M., Schlitt H. A, Wood, Lewine J. D.,
Sanders, J. A., and Belliveau, J. W., 1995, Mapping function in the human brain with MEG, anatomical MRI
and functional MRI. J. Clin. Neurophysiol. 12:406431.
Gevins, A, 1989, Dynamic functional topography of cognitive task, Brain Topogr., 2:3756.
Gevins, A, Brickett, P., Reutter, B., and Desmond, J., 1991, Seeing through the skull: advanced EEGs use MRIs
to accurately measure cortical activity from the scalp, Brain Topogr. 4:125131.
Gevins, A., Le, J., Leong, H., McEvoy, LX, Smith, M.E., 1999, Deblurring, J Clin Neurophysiol, 16(3):20413.
Gevins, A, Le, J., Martin, N., Brickett, P., Desmond, J., and Reutter, B., 1994, High resolution EEG: 124channel
recording, spatial deblurring and MRI integration methods, Electroenceph. din. Neurophysiol. 39:337358.
Grave de Peralta Menendez, R., Gonzalez Andino S., and Lutkenhoner B., 1996, Figures of merit to compare
linear distributed inverse solutions, Brain Topograph 9(2): 11724.
Grave de Peralta, R., Hauk, Gonzalez Andino, S., Vogt, H., and Michel, C.M., 1997, Linear inverse solution
with optimal resolution kernels applied to the electromagnetic tomography, Human Brain Mapping 5, 45467.
Grave de Peralta Menendez, R., and Gonzalez Andino S.L, 1998, Distributed source models: standard solutions
and new developments. In: Uhl, C. (ed): Analysis of neurophysiological brain functioning. Springer Verlag,
pp.176201.
278 F. Babiloni and F. Cincotti
Grinvald, A., Lieke, E., Frostig, R.D., Gilbert, C.D., and Wiesel, TN., 1986, Functional architecture of cortex
revealed by optical imaging of intrinsic signals, Nature 324(6095):3614.
Hamalainen, M., and Ilmoniemi, R., 1984, Interpreting measured magnetic field of the brain: Estimates of the
current distributions. Technical report TKKFA559, Helsinki University of Technology.
He, B., Wang, Pak, S., and Ling, 1996, Cortical source imaging from scalp electroencephalograms, Med.
BioI. Eng. Comput, 34 SuppI, part 2, 2578.
He, B., 1999, Brain Electrical Source Imaging: Scalp Laplacian mapping and cortical imaging, Critical Reviews
in Biomedical Engineering 27,149188.
He B., Wang Wu D., 1999, Estimating cortical potentials from scalp EEG's in a realistically shaped inhomo
geneous head model by means of the boundary element method. IEEE Trans Biomed Eng 46:12648.
He, B., Lian, J., G., 200 Highresolution EEG: a new realistic geometry spline Laplacian estimation technique,
Clinical Neurophysiology 112(5):84552.
He, B., Zhang, Lian, J., Sasaki, H., Wu, D., Towle, Y.L. 2002, Boundary Element Method Based Cortical Potential
Imaging of Somatosensory Evoked Potentials Using Subjects' Magnetic Resonance Images, Neuroimage, in
press.
Heinze, HJ., Mangun,G.R., Burchert, w., Hinrichs, H., Scholz, M., Munte, TE, Gos, A., Scherg, M., Johannes, S.,
and Hundeshagen, H., 1994, Combined spatial and temporal imaging of brain activity during visual selective
attention in humans, Nature 372:54346.
Heinze HJ., Hinrichs H., Scholz M., Burchert W., and Mangun G.R., 1998, Neural mechanisms of global and
local processing. A combined PET and ERP study. 1. Cogn. Neurosci. 10:48598.
HuangHellinger ER., Breiter H.C., McCormak G., Cohen M.S., Kwong K.K., Sutton J.P., Savoy R.L., Weisskoff
R.M., Davis TL., Baker J.R., Belliveau J.W., and Rosen B.R. 1995. Simultaneous functional magnetic
resonance imaging and electrophysiological recording. Hum. Brain Map. 3:1323.
Ives, J. R., Warach, S., Schmitt, E, Edelman, R.R., and Schomer, D.L., 1993, Monitoring the patient's EEG during
echoplanar MRI. Electroenceph. Clin. Neurophysiol. 87:417420.
Kampe, K.K., Jones, R.A., and Auer, D.P., 2000, Frequency dependence of the functional MRI response after
electrical median nerve stimulation, Human Brain Mapping 9(2): 10614
Kim, S., Ashe, 1., Hendrich, K., Ellermann, 1., Merkle, H., Ugurbil, K., and Georgopulos, A. 1993, Func
tional magnetic resonance imaging of motor cortex: hemispheric asymmetry and handedness, Science 261:
6157.
Kim, D.S., Duong TQ., Kim S.G., 2000, Highresolution mapping of isoorientation columns by fMRI, Nature
Neuroscience 3(2): 1649
Krakow, K., Woermann, EG., Symms, M.R., Allen, PJ., Barker, GJ., Duncan, J.S., and Fish, D.R., 1999, EEG
triggered functional MRI of intertictal epileptiform activity in patients with partial seizures, Brain 122: 1679
88.
Korvenoja, A., Huttunen, J., Salli, E., Pohjonen, H., Martinkauppi, S., Paiva, J.M., Lauronen, L., Virtanen, J.,
Ilmoniemi, RJ., and Aronen, HJ., 1999, Activation of multiple cortical areas in response to somatosensory
stimulation: combined magnetoencephalographic and functional magnetic resonance imaging, Human Brain
Mapping 8(1): 1327.
Lamusuo, S., Forss, N., Ruottinen, H.M., Bergman, J., Makela, J.P., Mervaala, E., Solin, Rinne, J.K., Ruot
salainen, U., Ylinen, A., Vapalahti, M., Hari, R., and Rinne, J.O., 1999, eSFjFDGPET and wholescalp
MEG localization of epileptogenic cortex, Epilepsia 40:92130.
Lawson, C.L., and Hanson, R., J. 1974, Solving least squares problems. Prentice Hall, Englewood Cliff, New
Jersey.
Le, J., and Gevins, A., 1993, A method to reduce blur distortion from EEG's using a realistic head model. IEEE
Trans. Biomed. Eng. 40:517528.
Lemieux, L., Krakow, K., Fish, D.R., 200 I, Comparison of spiketriggered functional MRI BOLD activation and
EEG dipole model localization, Neuroimage, 14(5):1097104.
Liu, A.K., Belliveau, J.w., and Dale, A.M., 1998, Spatiotemporal imaging of human brain activity using func
tional MRI constrained magnetoencephalography data: Monte Carlo simulations, Proc. Nat. Acad. Sc.,
95(15):894550.
Liu, A.K., 2000, Spatiotemporal brain imaging, PhD dissertation, Massachusetts Institute of Technology, Cam
bridge, MA.
Logothetis N.K., Pauls J., Augath M., Trinath T, Oeltermann A., 2001, Neurophysiological investigation of the
basis of the fMRI signal. Nature. 412(6843): 1507.
Multimodal Imaging fromNeuroelectromagnetic 279
Luck S. J. 1999. Direct and indirect integration of eventrelated potentials, functional magnetic resonance images,
and singleunit recordings. Hum. Brain Map. 8:115201.
Magistretti, PJ., Pellerin, L., Rothman, D.L., and Shulman, RG., 1999, Energy on demand, Science
283(5401):4967.
Malonek D., Grinvald A., 1996, Interactions between electrical activity and cortical microcirculation revealed by
imaging spectroscopy: implications for functional brain mapping. Science, 272(5261):5514.
Menke W. Geophysical Data Analysis: Discrete Inverse Theory. San Diego, CA Academic Press, 1989.
Menon, Y, Ford, J.M., Lim, KO., Glover, G.H., and Pfefferbaum, A, 1997, Combined EventRelated fMRI
and EEG Evidence For TemporalParietal Cortex Activation During Target Detection, NeuroReport 8:
302937.
Morioka, T., Mizushima, A, Yamamoto, T., Tobimatsu, S., Matsumoto, S., Hasuo, K., Fujii, K., and Fukui, M.,
1995, Functional mapping of the sensorimotor cortex: combined use of magnetoencephalography, functional
MRI, and motor evoked potentials, Neuroradiology 37:52630.
Nunez, EL., Silberstein, R., 2000, On the relationship of synaptic activity to macroscopic measurements: does
coregistration of EEG with fMRI make sense? Brain Topogr. 13(2):7996.
Nunez, E Electric fields of the brain. Oxford University Press, New York, 1981.
Nunez, P. L., 1995, Neocortical dynamics and human EEG rhythms, Oxford University Press, New York.
Opitz, B., Mecklinger, A., VonCramon, D.Y., and Kruggel, E, 1999, Combining electrophysiological and hemo
dynamic measures of the auditory oddball. Psychophysiology 36: 1427.
Oostendorp, T.E, Delbeke, J., Stegeman, D.E, 2000, The conductivity of the human skull: results of in vivo and
in vitro measurements. IEEE Trans Biomed Eng. 47(11): 148792.
PascualMarqui, RD. (1995) Reply to comments by Hamalainen, Ilmoniemi and Nunez. In ISBET Newsletter
N.6, December 1995. Ed: Skrandies., 1628.
Phillips, J.W., Leahy, R., and Mosher, J.e., 1997, MEGbased imaging of focal neuronal current sources, IEEE
Trans. Med. Imag., vol. 16., n.3, pp. 338348.
Puce, A., Allison, T., Spencer, S.S., Spencer, D.D., and McCarthy, G., 1997, Comparison of cortical activation
evoked by faces measured by intracranial field potentials and functional MRI: two case studies, Hum Brain
Mapp 5(4):298305.
Rao, e.R, and Mitra, S.K., Generalized inverse of matrices and its applications. Wiley, New York, 1977.
Rosen, B., Buckner, R., and Dale, A., 1998, Eventrelated fMRI: past, present and future. PNAS, 95:773780.
Rush S., and Driscoll, D.A., 1968, Current distribution in the brain from surface electrodes, Anesthesia Analgesia,
47:71723.
Salmelin, R., Forss, N., Knuutila, J., and Hari, R., 1995, Bilateral activation of the human somatomotor cortex by
distal hand movements, Electroenceph Clin NeurophysioI95:44452.
Sanders, J.A, Lewine, J.D., Orrison, W.w., 1996, Comparison of primary motor localization using functional
magnetic resonance imaging and magnetoencephalography. Human Brain Mapping 4:4757.
Savoy, R.L., Bandettini, P.A., O'Craven, KM., Kwong, K.K, Davis, T.L., Baker, J.R., Weisskoff, R.M., and Rosen,
B.R., 1995, Proc. Soc. Magn. Reson. Med. Third Sci. Meeting Exhib. 2:450.
Scherg, M., von Cramon, D., and Elton, M., 1984, Brainstem auditoryevoked potentials in postcomatose patients
after severe closed head trauma, J NeuroI231(1):15.
Scherg, M., Bast T., and Berg, P., 1999, Multiple source analysis of interictal spikes: goals, requirements, and
clinical value. Journal of Clinical Neurophysiology, 16:214224.
Seeck, M., Lazeyras, E, Michel, Blamke, Gericke, e.A., Ives, J., Delavelle, J., Golay, X., Haenggeli,
e.A., De Tribolet, N., and Landis, T., 1998, Noninvasive epileptic focus localization using EEGtriggered
functional MRI and electromagnetic tomography, Electroenceph. and Clin. Neurophysiol. 106:50812.
Shoham, D., Glaser, D.E., Arieli, A., Kenet, T., Wijnbergen, Toledo, Y., Hildesheim, R., and Grinvald, A.,
1999, Imaging cortical dynamics at high spatial and temporal resolution with novel blue voltagesensitive
dyes, Neuron 24:791802.
Sidman, R., Vincent, D., Smith, D., and Lu, L., 1992, Experimental tests of the cortical imaging technique
applications to the response to median nerve stimulation and the localization of epileptiform discharges,
IEEE Trans. Biomed. Eng. 39:437444.
Spiegel, M. Theory and problems of vector analysis and an introduction to tensor analysis. Me Graw Hill, New
York, 1978.
Srebro, R., Oguz, RM., Hughlett, K, and Purdy, P.D., 1993, Estimating regional brain activity from evoked
potential field on the scalp, IEEE Trans. Biomed. Eng.; 40:509516.
280 F. Babiloni and F. Cincotti
Srebro, R, and Oguz, RM., 1997, Estimating cortical activity from VEPS with the shrinking ellipsoid inverse,
Electroenceph. din. Neurophysi.; 102:343355.
Snyder, A. Z., Abdullaev, Y. G., Posner, M. and Raichle, M. E., 1995, Scalp electrical potentials reflect regional
cerebral blood flow responses during processing of written words, Proc. Natl. Acad. Sci. USA. 92: 168993.
Stok, CJ., Meijs, J.W, and Peters MJ., 1987, Inverse solutions based on MEG and EEG applied to volume
conductor analysis. Phys Med Bioi 32:99104.
Tikhonov, A.N., and Arsenin, Y.Y., Solutions of illposedproblems. Washington D.C., Winston, 1977
Uutela, K., Hamalainen, M., and Somersalo, E., 1999, Visualization of magnetoencephalographic data using
minimum current estimates, Neuroimage, 10(2): 17380.
van den Elsen, P.A., Pol, EJ., Viergever M., 1993, Medical image matching  A review with classification, IEEE
Engineeringin Medicineand Biology, 12:2639.
Wagner, M., and Fuchs, M. 2001, Integration of Functional MR!, Structural MR!, EEG, and MEG, International
Journal of Bioelectromagnetism, 1(3).
Warach, S., Ives, J.R., Schlaug, G., Patel, M.R, Darby, D.G., Thangaraj, Y., Edelman, R.R., and Schomer, D.L.,
1996, EEGtriggered echoplanar functional MRI in epilepsy. Neurology47:8993.
Wells WM., Viola P., Atsumi H., Nakajima S., Kikinis R., 1997, Multimodal volume registration by maximization
of mutual information, Medical ImageAnalysis 1:3551.
Wikstrom H., Huttunen J., Korvenoja A., Virtanen J., Salonen Aronen H., Ilmoniemi RJ. 1996, Effects of inter
stimulus interval on somatosensory evoked magnetic fields (SEFs): a hypothesis concerning SEF generation
at the primary sensorimotor cortex. Electroencephalography and Clinical Neurophysiology 100(6):47987.
THE ELECTRICAL CONDUCTIVITY
OFLIVING TISSUE: A PARAMETER IN
THE BIOELECTRICAL INVERSE
PROBLEM
Maria J. Peters, Jeroen G. Stinstra, and Ibolya Leveles
Facultyof AppliedPhysics, LowTemperature Division, University of Twente
9.1 INTRODUCTION
Electrically active cells within the human body generate currents in the tissues surrounding
these cells. These currents are called volume currents. The volume currents in turn give rise
to potential differences between electrodes attached to the body. When these electrodes are
attached to the torso, electrical potential differences generated by the heart are recorded.
The recording of these electrical potential differences as a function of time is called an
electrocardiogram(ECG). ECG measurements can be used to compute the generators within
the heart. This is called the solution of the ECG inverse problem. This solution may be of
interest for diagnostic purposes. For instance, it can be used to localize an extra conducting
pathway between atria and ventricles. This pathway can then subsequently be removed by
radiofrequent ablation through a catheter. When the active cells are situated within the
brain and the electrodes are attached to the scalp, the recording of the potential difference
measured between two electrodes as a function of time is called an electroencephalogram
(EEG). The EEG inverse problem can, for example, be used to localize an epileptic focus
as part of the presurgical evaluation. The frequencies involved in electrocardiograms and
electroencephalograms are in the range of I1000Hz. Therefore, the Maxwell equations
can be used in a quasistatic approximation, implicating that capacitive and inductive effects
and wave phenomena are ignored as argued by Plonsey and Heppner (1967).
To solve the inverse problem a model is needed of the source and the surrounding
tissues, i.e. the volume conductor. Customarily, the source is modeled by a current dipole
or a current dipole layer and the volume conductor is described by a compartment model,
Corresponding author: Prof. Dr. M. J. Peters, Faculty of Applied Physics, University of Twente, P.O. Box 217,
7500 AE Enschede, The Netherlands, Tel. 31534893138, Fax 31534891099, Email m.j.peterstetn.utwente.nl
281
282 M. J. Peters, J. G. Stinstra, and Leveles
where all compartments are considered to be homogeneous. The head may have a scalp, a
skull, a cerebrospinal fluid and a brain compartment. The torso compartment model may
include the ventricular cavities, the lungs and the surrounding homogeneous medium. The
shape may be a rough approximation of the real geometry, or the surfaces of the various
compartments may have a realistic shape that is obtained from magnetic resonance images.
An electrical conductivity is assigned to each compartment. the inverse solution is used to
localize the sources of the measured potentials, then only the ratio between the conductivities
assigned to the various compartments is of importance. the inverse solution is also used
to estimate the strengths of the sources, then the absolute values of these conductivities are
of importance.
In general, the conductivities of the various human tissues are among other things
dependent on the blood content and temperature, they are a function of the frequency
and strength of the applied current, they show an interindividual variability, and they
are inhomogeneous and anisotropic (Robillard and Poussart, 1977; Rosell et aZ., 1988;
Law, 1993). Moreover, the conductivity may be dependent on the health of the subject,
for instance, edema will change the conductivity, so does the presence of scar tissue or
tumors. The conductivity is called inhomogeneous when the conductivity differs from
place to place. The conductivity is called anisotropic when the conductivity is different in
different directions. For low current densities, the current density is linear with the applied
electric field, in other words the law of Ohm is valid in this case. The averaged Ohmic
conductivity that is assigned to a compartment is called the effective conductivity. The
effective conductivity of an inhomogeneous tissue is the conductivity of a hypothetical
homogeneous medium, which mimics the potential distribution that is found outside the
inhomogeneous tissue. For instance, in case of EEG, the effective conductivities assigned to
the various tissues in the head have to give approximately the same potential distribution at
the scalp as the real inhomogeneous tissues. The problem addressed in the present chapter
is: Which value should be assigned to a certain compartment?
9.1.1 SCOPE OF THISCHAPTER
In order to estimate the effective conductivity of a certain tissue, two approaches are
possible. First, the conductivity of a tissue can be measured in vitro or in vivo by applying
a potential difference by means of a set electrodes and measuring the resulting current.
A second approach is applying knowledge of the chemical composition and biological
morphology of the tissue in order to compute the effective conductivity. As measured
values of the effective conductivities of tissue reported in the literature vary widely, the
second approach may be of help to restrict the uncertainties in the conductivities involved
in the volume conduction problem. In section 9.1.2, the concept that the conductivity is not
a straightforward property of the material is discussed. The effective conductivity from an
experimental point of view will be discussed in section 9.1.3.
In section 9.2 the tissue is modeled as a suspension ofcells in an aqueous surrounding. If
only one type of cells is present and the suspension is a dilute one, the effective conductivity
is expressed by Maxwell's mixture equation. This equation will be discussed in section 9.2.3.
If more than one type of cell is present and the cells are densely packed or clustered, the
effective conductivity is expressed by Archie's law. This will be the subject of section 9.2.4.
The applicability of Maxwell's mixture equation and Archie's law will be illustrated for
various tissues, such as blood, fat, liver, and skeletal muscles .
The Electrical Conductivity of Living Tissue
In section 9.3, it is taken into account that most tissues have a layered structure, each
layer having a different conductivity. The effective conductivity of an entire compartment
will be discussed in section 9.4. The conductivity of a composite medium, like human
tissue, cannot have any value but will be within certain limits. section 9.5 of this chapter is
dedicated to these limits.
9.1.2 AMBIGUITY OF THE EFFECTIVE CONDUCTIVITY
Electrical currents in the human body are due to the movements of ions. As a result
of an electric force acting on an ion that is dissolved in fluid, the ion will develop a mean
drift velocity between collisions. Despite the fact that the average spacing between cells
may be no more than 20 nm, the mean free path of an ion in the extracellular space is
only about 0.01 nm. This represents the distance between collisions with other molecules.
Almost all these collisions take place with water. An ion rarely encounters a cell membrane
and behaves most of the time as though it were in a continuum. The drift velocity in Ohmic
conductors is directly proportional to the electric field. The current density reads:
where (T is the conductivity and E the electric field.
The electrical conductivity cannot be determined unambiguously, because it depends
on the direction of the currents, the extent of the current source, and the positions of the
measuring electrodes. This is illustrated in Fig. where two different situations are com
pared. In the first situation, a uniform current is applied between two parallel flat electrodes.
In the second situation the current is generated by a point source. The expressions for the
conductivity are different for the two situations showing that the effective conductivity is
not merely dependent on the material, but also on the configuration given by the orientation
and location of the source and the measurement points.
a medium is piecewise homogeneous and isotropic then the proper boundary con
ditions are that the normal component of the current density is continuous and the normal
conductivity between
flat elect rodes
IL
source
co nductivity of a
poi nt source
p
FIGURE 9.1. This figure illustrates that the conductivity measured in a homogeneous medium of infinite extent
depends on the electrodes used.
284 M. J. Peters, J. G. Stinstra, and I. Leveles
component of the dielectric displacement makes a step that is equal to the surface charge
density. Hence, all interfaces (including the outerfaces of the cells) will have a continuous
distribution of surface charge representing genuine accumulations of charge. the material
is nonuniform or if the conductivity is anisotropic, we get accumulations of charge within
the material as well as on the interfaces.
The microscopic electrical conductivity is the conductivity that characterizes a part of
tissue that is comparable in size with the dimensions of the cells. The macroscopic effective
conductivity characterizes a part of the tissue that is large compared to the dimensions of the
cells. Several levels of inhomogeneities can be distinguished. We will restrict ourselves to
three levels of inhomogeneities, the microscopic level with typical dimensions of microns,
the millimeter level and the macroscopic level, e.g. compartments with dimensions of several
centimeters. This is illustrated in Fig. 9.2. In the microscopic point of view, forms and
dimensions of cells and the interstitial fluid are taken into account. Near a cell, the electrical
field may change in direction or amplitude due to the charge density on the surface and the
presence of counterions near the surface. For macroscopic purposes one has to consider the
field averaged over regions large enough to contain many thousands of cells or fibers so
that microscopic fluctuations are smoothed over, the 'graininess' of the material is blurred
by distance. At a millimeter level, the layered structure or columnar structure of an organ is
taken into account. For instance, the skin is composed of three layers, namely the epidermis
(the outer nonsensitive and nonvascular layer of the skin that overlies the dermis), the
dermis and subcutis. The epidermis is composed of stratum corneum, stratum lucidum,
stratum granulosum and stratum germinativum. These layers differ in composition and
morphology and consequently in conductivity. The conductivity of the various layers is
averaged and the material acts as a continuum, the averaged conductivity being the effective
conductivity.
9.1.3 MEASURING THE EFFECTIVE CONDUCTIVITY
The uncertainty in measured conductivity values is high, because the measurements are
very complicated. The reasons why they are so complicated will be shortly discussed in this
section. Usually, the Ohmic lowfrequency conductivity of a piece of tissue is determined
from the currentvoltage relation using a two or four points method. The currents used
for the measurements have to be low (about lmA) in order not to trigger an activation of
cells. The field near one cell is very much influenced by the presence of the cell. In order
to have an effective conductivity thousands of cells have to be present within the piece of
tissue measured. Therefore the effective conductivity has to be measured with electrodes
with dimensions of millimeters, these electrodes have to be millimeters apart. The sources
in the brain and heart of EEG and ECG are due to at least 10
5
cells that are active in
synchrony else the EEG or ECG would not be measurable. Consequently, the condition
that the source has dimensions of millimeters is met if brain or heart activity is used for
conductivity measurements.
the measurements are carried outin vitro (i.e. outside the body), the accuracy may be
low because tissue properties change after death. The conductivity will initially drop after
circulatory arrest due to emptying blood vessels and drainage of fluids. For instance, it was
found that the conductivity of frog muscle at 10Hz decreased a factor two after 2.5 hours of
death and that of chicken muscle decreased 70 percent in the first 60 minutes (Zheng et al.,
The Electrical Conductivity of Living Tissue 285
FIGURE 9.2. Illustration of the three levels of inhomogeneities discussed in this chapter. at the left) the head that
is usually modeled by three compartments; in the middle) the layered gray matter; at the right) a suspension of
cells.
1984). However, the conductivity of dog muscle increased immediately after death (Zheng
et al., 1984). This may be caused by a change in the osmotic pressure causing some cells
to swell and burst. Gielen (1983) measured in vivo the lowfrequency conductivity of the
muscles ofa rabbit that were prepared free, the blood supply was unimpaired. He found after
finishing his experiments that the outermost muscle layer was damaged. The cross sections
of the fibers were much larger and rounder than normal. This damage was related to an
increase conductivity. This phenomenon was blamed to osmotic proces ses. The increase
286 M. J. Peters, J. G. Stinstra, and Leveles
will grow in time (Schwan, 1985), because cell membranes after death allowcurrents to pass
more easily. In other words, all in vitro conductivity measurement s should be completed
within a very limited time.
If the measurements are carried out in vivo, commonly animal tissue instead of human
tissue is used. However, it is not clear whether the animal tissue has the same electrical
properties as human tissue although sometimes they seem to be comparable. Moreover, the
in vivo measurements depend on the surrounding tissues. When electrodes are impl anted in
living tissue, the currents applied by these electrodes will not be confined to the tissue that
is between the electrodes, but will spread out through all surrounding tissues. Hence, it is
difficult to estimate which part of the current will flow through the tissue of interest. Some
investigators try to avoid this problem by measuring the conductivity with two electrodes
spaced closely together. However, this raises the question whether on such a scale the in
homogeneities in the structure of the tissue do not disable the measurement of an effective
macroscopic conductivity because the electric field within the tissue near cells has a com
plicated pattern. one happens to be near a cell, the field may be small or point in a totally
different direction. Another probl em experienced is a relatively large extra capacity between
electrode and tissue at low frequencies. Moreover, the electrodeel ectrolyte interface can
produce large errors that depend on the pressure between electrode and organ tissue.
Thence, it is not surprisi ng that the measured lowfrequency conductivities reported in
the literature vary over a wide range. As example some values found for the lowfrequency
conductivity of skeletal muscle tissue at 37°C are given in table 9. 1. No attempt to give
a complete overview has been made. The conductivity in muscle tissue is anisotropic, the
conductivity along the fibers a"is higher than the conductivity perpendicular to the fibers
Five degrees misalignment from true parallel or perpendicul ar orientation during the
measurement would result in an 18 percent overestimate of and a 0.4 underestimate of
(Epstein and Forster, 1983). Consequentl y, misalignment errors will be smallest in and
the anisotropy factor will be easily underestimated. Out of theoretical studies, there might
develop insight into the nature of volume conduction that would permit a proper choice
from the values that are reported.
TABLE 9.1. Some values found in the literature of the conductivity of skeletal
muscle at 37°C in the frequency range of 01000 Hz
anisotropy
species factor reference
cow 0.41 0. 15 2.7 Burger and van Dongen (1961)
rabbi t 0.8 0.06 13
dog 0.67 0.04 17 Rush et al. ( 1963)
dog 0.43 0.21 2.0 Burger and van Mi laan ( 1943)
frog(2I
CC
) 0.09 0.05 1.8 Hart et al. ( 1999)
dog 0.70 0.06 II Zheng et al. (1984)
rabbit 0.75 0.04 17
monkey 0.8 1 0.06 13
dog 0.52 0.08 6.5 Epstein and Foster (1983)
rabbit 0.5 0.08 6. 3 Gielen et
rat 0.5 0.Q7 6. 1 McCrae and Esrick (1993)
The Electrical Conductivity of Living Tissue
TABLE9.2. Examples of thetemperaturedependenceof tissues
287
material conductivity
bile (cowpig) 1.66
amniotic fluid (sheep) 1.54
2.04
Urine (cowpig) 3.33
2.56
skeletal muscle (rat)
parallel 0.42
0.66
perpendicular 0.08
0.20
9.1.4 TEMPERATURE DEPENDENCE
temperature
37
20
25
37.5
37
20
36.7
44.5 (after 30 minutes)
36.6
44.5 (after 30 minutes)
The electrical conduction in living tissue depends on the temperature. Me Rae and
Esrick (1993) heated freshly excised rat skeletal muscle from 39.5 to 50°C. Above 44°C
healthy tissue will be damaged irreversibly. Initiall y. a rapid increase of the lowfrequency
conductivity (with about a factor two) was observed followed by a much slower increase
during which the lowfrequency conducti vity graduall y approached the highfrequency
values. The initial rapid change was associated with microscopically observed fiber rounding
and shrinkage in the radial direction and increasing edema. The subsequent slow change
was associated with disintegration of the tissue. After the cell membranes are destroyed.
the conductivity is no longer frequency dependent and the lowfrequency conductivity has
the same value previously measured at high frequencies. The temperature dependency of the
conductivity of tissue is also caused by the temperaturedependent regulation of the vessel
diameter (vasodilatation). In other words. the blood supply is temperature dependent. The
temperature dependence of the conductivity is illustrated in Fig. 9.3 and table 9.2.
9.1.5 FREQUENCY DEPENDENCE
The lowfrequency conductivity of most tissues like heart muscle. skin. liver. lung. fat.
and uterus is not strongly dependent on frequency. although measurements show that the
conductivity increases with frequency. also at lowfrequencie s. Commonly. this phenomenon
is not taken into account and the conductivity is presumed to be frequency independent for
frequencies lower than 1000Hz Schwan and Foster. 1980). Nevertheless. the con
ductivity does increase with the frequency regardless the kind of tissue. Nicholson (1965)
measured the conductivity of cerebral white matter of a cat. He found an enhancement of
the conductivity in the direction normal to the fibers from 0.11 to when the
frequency changed from 20 to 200Hz. No enhancement was found for the conductivity
parallel to the fibers. Comparison between the experimental data presented by Gabriel et al.
(1996
a
) . and corresponding data previously reported in the literature show good agreement
(Stuchly and Stuchl y, 1980). The conductivity of the gray matter as measured by Gabriel
et al. (1996 in the frequency range from 10Hz to 20GHz is given in Fig. 9.4.
288
>
...
c
"0
C
0
o
36
M. J. Peters,J. G. Stinstra,and Leveles
42 43
40 4 1
39 00 '
FIGURE9.3. Variation of the conductivity with the temperature and time for skeletal muscle, based om mea
surements reported by Gersing (1998).
..§
"C
c
0
0
01L_ ....L.._ l__
40 60
(Hz)
FIGURE9.4. The frequencydependent conductivity of gray matter based on the parameters given by Gabriel
et al. (l996
The Electrical Conductivity of Living Tissue 289
9.1.5.1 Impact ofthe frequency dependence on the EEG
The frequency dependence of the conductivity may influence the EEG. In order to
estimate the impact on EEG, the effects of a frequencydependent conductivity were sim
ulated. The volume conductor model consisted of three concentric spheres representing
the brain, skull and scalp, with radii of 87, 92 and 100mm. This model is often used to
describe the head as a volume conductor. The source was modeled as a current dipole.
Simulations were carried out for three current dipoles, i.e. a central dipole, and a radial
and a tangential dipole at a radius of 80mm. As observation point, a point at the surface
was taken where the potential, found for the lowest frequency, was maximal. The transfer
function was calculated. The transfer function is the relationship between the strength of
the current dipole and the potential in a point at the outermost surface as a function of the
frequency. To enable a comparison between the different cases, all transfer functions were
normalized. there is no frequency dependence, the transfer function would have the value
one for all frequencies.
Two cases were studied. First, the gray matter conductivity was taken from Fig. 9.4
in which the conductivity increases by a factor four; the scalp conductivity was 0.33S/m
and that of the skull 0.0042S/m. Second, the conductivity of all compartments increased
20% in the interval from 1Hz to 100Hz. At 1Hz the conductivity of the brain and scalp
was 0.33S/m and that of the skull 0.0042S/m. The results of these simulations are given
in Fig. 9.5. As shown the volume conductor acts as a lowpass filter. The potential may
drop by approximately a factor two when the frequency is increased from 1 to 100Hz. The
transfer function depends on the depth of the source.
9.2 MODELS OF HUMAN TISSUE
Tissues are composed of cells. The interstitial space between the cells contains fluid.
So, the effective conductivity of a tissue depends on the conductivity of the cells, the volume
fraction occupied by the cells, and the conductivity of the extracellular medium.
9.2.1 COMPOSITES OF HUMAN TISSUE
This section starts with a brief description of tissue at a cellular level. Next, the con
ductivity of a cell and that of the extracellular fluid will be discussed.
9.2.1.1 Cells
All human cells stem from the roundshaped fertilized egg cell. There is no typical cell
shape. Cells come in all shapes: cubes (cells lining sweat ducts), spheres (white blood cells
of the immune system), Bismarck doughnuts (red blood cells), columnar cells, balloonlike
cells (cells lining the urinary bladder), needle shaped ellipsoids or rods (skeletal muscle
cells) and pancakes (cells on the surface of the skin) as illustrated in Fig. 9.6.
Cells vary also considerably in size, and function. For instance, the diameter of a red
blood cell is 7.5f.1m, the diameter of a human egg cell is 140fJ..m, a smooth muscle cell has
a length of 20 to 500f.1m, while a skeletal muscle cell may have a length of 30cm. All cells
290 M. J. Peters, J. G. Stinstra, and Leveles
100 90 80 70 40 50 60
frequency (Hz)
30 20
central dipole
radial dipole at z =80mm
tangential dipole at z 80mrn
all dipole when all conductivities increase by 20%
10
 0_ .. _.
0.8
0.9
c
.2
0.7 e
<2
~
0.6
t:
"0
0.5
~
c
0.4
::
c
0.3
0.2
0.1
0
FIGURE 9.5. Transfer functions for EEG for three dipoles using a threesphere model of the head. Two cases
of frequencydependent behaviour are considered. In the first case the conductivity of the brain compartment is
chosen according to Fig. 9.4. In the second case the conductivity of all three compartments increases linearly.
•
0 
r     ~   r = '
FIGURE 9.6. Cells vary considerably in shape.
The Electrical Conductivity of LivingTissue
TABLE9.3. Values of acell asfoundintheliterature(adaptedfromKotnik
et al., 1997)
291
cell values
extracellular mediumconductivity a
e
membraneconductivity am
cytoplasmicconductivity
cell radius R
Average membranethickness: t 5 nm
lower limit
5.0 IO
4Sm

1
1.0 IO 
2.0 x IO
2Sm 1
1 11
m
higher limit
2.0 Sm
1.2 x IO
6Sm 1
1.0Sm
10011m
are surrounded by a membrane 5 to lOnm thick, visible only with the electron microscope.
The electrical properties of the cell components are different. The highest and lowest values
reported in the literature for the electrical properties of biological cells are given in table 9.3.
The intracellular fluid accounts for about 70 percent of the inner cell volume. The
membranes maintain the integrity of the cell, breaking down after death. Membranes are
highly selective permeability barriers and consist mainly of lipids and proteins. Lipid bilayer
membranes have a very low permeability for ions and most polar molecules. An exception
is water that easily crosses such membranes, and creates a balance in the whole organism.
An ion such as Na+ crosses membranes very slowly because the removal of its shell of water
molecules is highly unfavorable energetically. Membranes contain specific channels and
pumps that regulate the molecular and ionic composition of the intercellular compartment.
A nerve impulse, or action potential is mediated by transient changes in Na+ and K+
permeability. action potential is generated when the membrane potential is depolarized
beyond a critical threshold value. When the transmembrane potential is not exceeding the
threshold value, the relationship between the potential and the current is approximately
linear, so it obeys Ohm's law. For low frequencies, this is the case when the current density
is smaller than 0.5J.LNcm
9.2.1.2 Volume fraction occupied by cells
One can analyse blood and determine the hematocrit content (the volume fraction
occupied by the red blood cells) by measuring the electrical conductivity of blood. The
electrical conductivity is a function of the hematocrit. This function is given in Fig. 9.7.
9.2.1.3 The extracellularfluid
Mammals have a water content of 65 to 70 percent. the early fetal period of humans,
approximately 95 percent of the fetus is water. The proportion of total body weight that is
water decreases throughout the fetal period to reach 75 percent at term. The water content of
various tissues ofadults is given in table 9.4. Biological tissues are inhomogeneous materials
with discrete domains: the extracellular and the intracellular space. Although the cells are of
microscopic size, still they are much larger than the ions in the extracellular space so that the
extracellular fluid can be considered as a continuum. The extracellular and intracellular fluids
are electrically neutral, however, the ion concentrations are quite different in each. Thus,
the electrical properties of the intracellular fluid and the extracellular fluid are different.
292 M. J. Peters, J. G. Stinstra, and Leveles
TABLE 9.4. Water content of various organs (Pehtig
and Kell, 1987; Foster and Schwan, 1986)
tissue type volume fraction of water
gray matter (brain)
white matter (brain)
skeletal muscle
Fat
Liver
Spleen
0.84
0.74
0.795
0.09
0.795
0.795
10.
16.  .   ,.  .   ,...  ,   ....,
E 10
~
~
8
c:
8 6
, , ~
..
..
..
. ..
70 60 30 40 50 20
L
10
FIGURE 9.7. The conductiv ity of blood as a function of the percentage of red blood cells the hematocri t).
The dashed line is the measured curve, the solid line is calculated using Maxwell' s mixture equation, the line with
point s and dashes is calculated using Archie' s law.
9.2.2 CONDUCTIVITIES OF COMPOSITES OF HUMAN TISSUE
A composite cell includes a nucleus, cytoplasm, and a cell membrane. The nucleus
is enclosed by a thin cell membrane. The cytoplasm is a mass of fluid that surrounds the
nucleus and is encircled by the plasma membrane. Within the cytoplasm are specialized
structures called cytoplasmic organelles. In other words, the structure of a cell is such that
the conductivity cannot be expected to be homogeneous. The effective conductivity is the
The Electrical Conductivity of Living Tissue 293
E
FIGURE 9.8. Model used to calculate the effective conductivity of a spherical cell.
conductivity ascribed to an entire cell. In the next section this effective conductivity is
estimated using a simple model for the cell.
9.2.2.1 Effectiveconductivity ofa sphericalcell
The effective conductivity of a cell at low frequencies and for low current densities
will be estimated for a spherical model of a cell. In the model, a sphere of radius rj and
conductivity is surrounded by a thin concentric shell of thickness t and conductivity
representing the membrane. The cell is suspended in a medium with conductivity The
model is depicted in Fig. 9.8. Typically, tis Snm for a cell and its radius R r, tis 10/Lm,
thence t «
The law of conservation of charge in quasistatic approximation reads:
(9.2)
Inserting J Eand E VV yields for a homogeneous region where a is constant
Laplace's equation
(9.3)
We assume that a homogeneous electric field E is applied along the zaxis. The electrical
potential outside the twolayered particle can be calculated by solving Laplace's equation
with the proper boundary conditions. The boundary conditions are that the potential and
the normal component of the current density are continuous across the boundary. On the
other hand we can calculate the potential outside a homogeneous sphere of conductivity
suspended in a medium of conductivity The effective conductivity of the cell is
the conductivity of a uniform sphere that gives the same electrical potential outside the
cell as the twolayered one. Equating the two solutions for the potential and neglecting
the higher order terms of tIR in both the denominator and numerator as t R, yields
294 M. J. Peters, J. G. Stinstra, and Leveles
FIGURE 9.9. Thecurrentdensityarounda spherical cell.The applied fieldwas initiallyuniform. Theintracellular
current density is too small to be depicted.
(Takashima, 1989)
(9.4)
Inserting the values given in table 9.3 shows that cells can be described as nonconducting
particles because the effective conductivity of a cell is about 10
5
times that of the sur
rounding fluid. In good approximation, at low frequencies the currents flowaround the cells
rather than through them as shown in Fig. 9.9. Thence, at low frequencies, the conductivity
is dominated by the conductivity of the extracellular space.
9.2.2.2 Effective conductivity ofa cylindrical cell
Similar calculations can be carried out to deter mine the effective conductivity of a
cylindrical cell. We assume that a current is applied perpendicular to the axis. The effective
conductivity reads
(9.5)
leading to the conclusion that in this case the cylindrical cell like the spherical one can be
considered as a nonconducting particle.
However, if the field is applied parallel to the axis it is a different situation. The cell
has a resistance for a current parallel with the axis. The resistance of an element of material
The Electrical Conductivity of Living Tissue
of length L and cross section A is
295
(9.6)
The cylindrical shell of thickness t, that represents the membrane, is connected in parallel
with the inner cylinder of radius (r  t). The part of the membrane at the top and that at the
bottom of the cylinder are connected in series with the intracellular fluid of the cell. So the
resistance reads
Jl' t)2
L  2t 2t
am
(9.7)
where the effective conductivity is defined as the conductivity of an element of homogeneous
material with the same dimensions, that has the same resistance as the inhomogeneous
element. In good approximation expression (9.7) leads to the expression
(9.8)
Expression (9.8) is only true when the cells are intact, i.e. the membranes at the bottom and
top of the cylinder are not damaged.
9.2.2.3 Conductivity ofextracellularfluid
Since the intracellular space is not playing a role, as the effective conductivity of a
cell is in general zero for low frequencies, the intracellular fluid is not considered here
further. Materials with the highest conductivity are fluids with a low concentration of
cells like urine, amniotic fluid, cerebrospinal fluid and plasma, which will be discussed
below.
Using the ionic concentrations of the different cations and anions in the extracellular
fluid, the conductivity can be approached by
(9.9)
where Ai are the ionic conductivities at 37°C and are the molar concentrations of the
different ions. In table 9.5 the ionic conductivities of various ions at a temperature of 37°C
are given as well as the ionic concentrations in the interstitial fluid, the cerebrospinal fluid
and the blood plasma. Using these values, the conductivity of these fluids can be estimated.
In table 9.6 the computed values are compared with the values cited in literature. From these
values it becomes apparent that the computed values are about 15 to 25 percent higher than
the measured values. This overestimation may be explained by the presence of proteins in
the actual solution and the presence of counterions surrounding its cells.
The extracellular concentration of ions such as K+ undergoes frequent small fluctua
tions, particularly after meals or bouts of exercise. An exception is the brain; if the brain
296 M. J. Peters, J. G. Stinstra, and I. Leveles
TABLE9.5. The ionic concentration mmoIll of differentfluids encountered
in the body and the specific conductivity of each ion in a dilute solution at
37°C.The valueswereobtainedfromAseyev(1998). "Thesevalueswereonly
available for 25°Cand havebeencorrectedby 2 percent/tC
Ion blood plasma interstitial fluid cerebrospinal fluid Ai 1O
4S
m2 jmol
Na+ 142 145 153 63.9
K+
4 4 2 95.7
Mg
2+
2 3 142.4
Ca2+
5 4 3 154.9
Cl
102 116 123 95.4
HC0
3
26 29 55*
P O ~  2 2 296*
protein 17.0 0.0
Other 6.0 6.7
TABLE9.6. Calculated and measured values of the conductivity
of body fluids at 37°C. The measuredvalues are taken from liter
ature: "Geddes and Baker (1967), "Schwan and Takashima (1993),
cBaumann et al. (1997)
conductivity of body fluids interstitial cerebrospinal
(at 37°C in S/m) blood plasma fluid fluid
computed 2.08 2.22 2.12
measured 1.58" 2.0
b
1.79
c
were exposed to such fluctuations the result might be uncontrolled nervous activity, because
K+ ions influence the threshold for the firing of nerve cells. The conductivity of electrolytes
is dependent on the temperature. The conductivity of the amniotic fluid was measured by De
Luca et al. (1996) at a temperature of 20°C. The mayor contributors to the conductivity are
the Nat and CI ions. The ionic conductivity of the former increases by 2.1 percent per °C
and that of the latter by 1.9 percent per "C. overall increase of 17 x 2 34 percent for
the amniotic fluid heated from 20 to 37°C is thus expected. The same rate of increase was
also obtained by Baumann et al. (1997), who measured the conductivity of cerebrospinal
fluid at 25 and 37°C.
9.2.3 MAXWELL'S MIXTURE EQUATION
Spherical particles suspended in a solvent are considered to be a relevant model of
biological tissues. Maxwell (1891) derived an equation for the effective conductivity of
dilute suspensions in aqueous media of spherical particles (that do not have a permanent
electric moment). The derivation of this equation is given in short below. Next, Maxwell's
equation is extended for ellipsoidal particles, as many biological cells are better described
by ellipsoids instead of spheres. The applicability of Maxwell's equation will be illustrated
for blood.
The Electrical Conductivity of Living Tissue 297
FIGURE 9.10. A spherical particle of conductivity surrounded by a solvent with conductivity placed in
an original uniform field E.
9.2.3.1 A dilute solution ofspheres
First, let us consider a spherical particle of radius R and conductivity embedded in
a homogeneous solvent with conductivity a; when a uniform electric field Eo is applied. An
expression for the potential outside the sphere is found by solving the Laplace equation with
the proper boundary conditions using spherical coordinates. The origin of the coordinates
(r, e, rp) is at the center of the sphere . This yields for the potential V outside the sphere of
radius R:
(
ae 
3
)
V(r, 1 3 Eorcose
2ae
(9.10)
Within the sphere, the field is parallel and uniform. Outside the sphere, the field is the
original field plus the field of a dipole at the center of the sphere (see Fig. 9.10).
Second, the suspension is modeled by N small spherical particles of radius R and
conductivity that are surrounded by a large spherical boundary of radius R' as depicted
in Fig. 9.11. Two assumptions are made. First, the volume fraction occupied by the cells is
assumed to be low, the average distance between them being larger than their dimensions
and as a consequence the spheres do not influence each other although they act as dipoles .
The potential is the sum of the potential due to Nsmall homogeneously distributed particles,
yielding:
(
ae 
3
)
V(r, 1 N , 3 Eorcose.
2a
e
(9.11)
Instead of the microscopic point of view, we can look at the sphere from a macroscopic
point of view. We have a spherical shaped medium consisting of an aqueous solution of
spherical particles. The effective conductivity of this sphere is per definition the conductivity
that gives a potential outside of the sphere that is expressed by (9.10). Hence, the potential at
298 M. J. Peters, J. G. Stinstra, and I. Leveles
p
FIGURE 9.11. The model used for the derivation of the Maxwell mixture equation.
a distance r R' from the center of the sphere describing the solution reads:
(
ae  aeff R/
3
)
V(r, 1 3 EorcosB
2ae aeff
(9.12)
where aeff is the effective conductivity of the large sphere containing N particles.
Formulas (9.11) and (9.12) are equivalent expressions for the potential outside the
sphere with radius R'. Noting that NR R' P is the volume fraction occupied by the
particles that are suspended in the large spherical boundary we obtain by equating expres
sions (9.11) and (9.12) the socalled Maxwell's mixture equation:
(9.13)
As argued before at low frequencies and low current densities, the currents will in general
be only in the extracellular space, in that case one can insert apart =O.
A spherical model is a good approximation for many colloidal particles, including
biological cells. However, biological cells often have a complex geometry. Many cells are
better described by ellipsoids; a, b and c are the semiaxes of the ellipsoid. Maxwell's
mixture equation has been derived for ellipsoidal particles that are orientated in parallel by
Sillars (1937). When p is very low and the field is applied along the aaxis, it reads:
a,part__a_e_)_]
aeff =ae
(apart  ae)L
a
(9.14)
where La is the depolarization factor of the ellipsoid in the direction of the aaxis, (Boyle,
1985).
abc ['JO ds
La
0
(a
2
+ s)J(a
2
+ s)(b
2
+ s)(c
2
+ s);
La + L, + L, 1 (9.15)
The Electrical Conductivity of Living Tissue
0.9
0.8
0.7
0.6
~ 0.5
0.4
0.3
0.2
0.1
0
0 2 3 5 6 7
alb
299
FIGURE 9.12. Thedepolarization factor inthedirectionof theaaxisfor anellipsoidof revolution withsemiaxes
a, b, andc, whereb c.
There are three cases of considerably interest for us: spheres, needleshaped ellipsoids
and disklike particles. For a needleshaped ellipsoid or a particle with the shape of a long
cylinder with a b « c, La,L, and L, tend to 1/2, 1/2 and 0, respectively. For a sphere with
a =b =c, the depolarization factor La = 1/3. For a diskshaped ellipsoid with a =b » c,
La, t., and t, tend to 0, 0 and 1 respectively (Fricke, 1953). In Fig. 9.12 the depolarization
factor La is depicted for a spheroid with semiaxes a, b c.
Maxwell's mixture equation has been experimentally tested by Cole et al. (1969) for
variously shaped objects. The particles could be spheres, cubes, or cylinders arranged in
a cubic or hexagonal array or randomly distributed plates. They found that the Maxwell
equation was surprisingly accurate within one percent at concentrations from 30 percent or
less to 90 percent.
9.2.3.1.a. Blood
To illustrate the applicability of Maxwell's mixture equation, the conductivity of blood
is discussed. Blood of normal subjects consists predominantly of red blood cells (erythro
cytes) in plasma. The mature erythrocyte is a cell surrounded by a deformable membrane
well adapted to the need to transverse narrow capillaries. The red cells are biconcave disks,
each with a diameter of about Sum, a thickness of 2fJ..m at its edge, and a volume of about
94fJ..m
3
. In normal adults, the red cells occupy on the average about 48 percent of the volume
of blood of males and about 42 percent in the blood of females. The percentage of the volume
of blood made up by erythrocytes is defined as the hematocrit. Erythrocytes are essentially
300 M. J. Peters, J. G. Stinstra, and Leveles
TABLE9.7. Parametervaluesusedtoestimate theconductivityperpendicular
to the fibers and that parallel to thefibers. ("Kobayashi and Yonemura, 1967)
parameter symbol minimum maximum
effective intracellular conductivity a j 0.55 0.80
extracellular conductivity a. 2.0 2.4
membrane conductivity am
1.0 1.2
6
volume fraction occupied by fibers p 0.85 0.9"
fiber length 5 mm 30cm
membrane thickness t
nonconducting in fields with frequencies up to 100kHz (Trautman and Newbower, 1983).
A typical value for the lowfrequency conductivity of human blood at body temperature and
normal hematocrit is in the range 0.43Q.76S/m (Geddes and Baker, 1967). The orientation
of the ellipsoids with respect to the electric field depends on the blood flow. There is a
periodic noise component in continuous measurements of conductivity, which is presumed
to result from the cyclic reorientations due to the flow. The magnitude of the change with
flow is of the order of three percent of the baseline conductivity. Since the shape of red
blood cells of most species is nonspherical and is subject to changes in aggregation and
orientation with flow, variations in the effective conductivity are understandable.
Geddes and Sadler (1973) measured the conductivity for human, canine, bovine and
equine blood at 25kHz and 37°C, having a hematocrit range extending from 0 to 70 percent.
For human blood they found an exponential fit between the measured data and an exponential
curve for p 53.2eo. 022H with a correlation coefficient of 0.99. Blood can be modeled as a
dilute solution of nonconducting spherical particles within a conducting medium. In Fig. 9.7
both the experimental measurements and the result s using Maxwell ' s mixture equation for
spherical particles are displayed. From this figure one can conclude that the description of
Maxwell is rather accurate. With increasing concentration of red blood cells (p is larger)
the effective conductivity is reduced. The relationship between p and the conductivity can
be used to estimate the hematocrit content (Ulgen and Sezdi , 1998).
9.2.4 ARCHIE'S LAW
Measurements of the conductivity of the cortex do not fit so well with the theory
of Maxwell, they fit better with Archie's law (Archie, 1942; Nicholson and Rice, 1986).
Moreover, Maxwell's theory can only be applied if there is only one type of cell present
and most human tissues are constituted of various types of cells . Archie's law that will
be discussed in the next section is appli cable if there are various types of denselypacked
nonconducting cells suspended in a conducting solution.
To derive Archie' s law the following assumptions are made. a) Tissue is modeled
by cells of various shapes that are suspended in an aqueous conducting medium; b) The
effective conductivity is calculated for a region containing thousands of particles; c) The
particles are assumed to be homogeneously distributed; d) Currents are Ohmic and are
only present in the extracellular space; e) The solvent is homogeneous and isotropic ; An
electric field is applied which is initially uniform; g) The particles do not have a permanent
electric dipole moment.
The Electrical Conductivity of Living Tissue 301
9.2.4.1 Ellipsoidal particles withthe sameorientation
The derivation of an equation for the effective conductivity for higher values of p is
given by Hanai (1960). starts with equation (9.14). Subsequently, the number of particles
is increased. Due to this increase, the effective conductivity increases with a fraction
~ a e f f ' In the next stage the previous mixture acts as a host, and so on. An expression for the
effective conductivity of a densely packed suspension of ellipsoidal particles is obtained by
integration yielding:
'  =lp
Inserting in this equation 0 yields Archie's law:
(9.16)
(9.17)
where a
e
is the conductivity of the fluid surrounding the nonconducting cells, p is the volume
fraction occupied by the cells, and m is the socalled cementation factor that depends on the
shape and orientation of the particles, but not on their sizes. According to Archie's law, the
effective conductivity is proportional to the extracellular conductivity o.: The value of m
depends on the shape of the cells. For instance, its value is 3/2 for spherical particles, m is
for long cylinders with the axis perpendicular to the external field, and it is 1 for cylinders
with the axis parallel to the field. Archie's law also holds for ellipsoidal particles. If the field
is applied along the aaxis of the ellipsoids and the ellipsoids have the same orientation,
(9.18)
9.2.4.1.a. Fat
Since fat tissue consists of about 9 percent water (Foster and Schwan, 1989) and
the interior of the cells is almost completely filled with fat, the interstitial fluid occupies
about 9 percent of the volume (upper limit). Histology shows that fat cells are spherical
shaped particles. According to equation (9.17) an upper limit for the conductivity is 1.9 x
(0.09)3/2 ~ 0.05
9.2.4.1.b. Skeletal muscles
A skeletal muscle fiber represents a single cell of muscle. Each skeletal muscle fiber is a
thin elongated cylinder with rounded ends that are attached to connective tissues associated
with the muscle. Just beneath its cell membrane, the cytoplasm of the fiber contains many
small nuclei and mitochondria. The cytoplasm also contains numerous threadlike myofibrils
that lie parallel to the cylinder axis. The diameter may vary within muscles, between muscles
in the same animals, and between species. Muscle fibers increase in diameter from birth to
maturity and also in response to exercise. The length of a fiber may vary between millimeters
and several tens of centimeters. The amount of connective tissue relative to muscle fibers is
much greater in some muscles than in others and may range from 3 to 30 percent. Connective
tissue is composed of collagen fibers, reticular fibers, elastic fibers and several varieties of
cells, such as fat cells. There is an increase in elastic tissue with aging. Apart from the
302 M. J. Peters, J. G. Stinstra, and Leveles
fibers in the tongue, the fibers have no branches. In long muscles , as in the longest muscle
in the human body the sartorius (52cm long) , the fibers are arranged in parallel. Covering
the surface of each muscle fiber is a thin membrane of about O.I/Lm thickness. The model
used to describe muscle tissue consists of homogeneously distributed cylinders. So far as
connective tissue, blood vessels, and nerve tissue cannot be described by parallel cylinders,
their influence will be neglected.
The effective conductivity of one fiber when the field is applied parallel to the axis is ac
cording to formula (9.9) dependent on the length ofthe fiber. Inserting the values ofTable 9.7
in formula (9.8) yields 0.25 x 1O
5S/m
(short fiber) 15 x 1O
2S/m
(long fiber).
When the fiber is damaged (forinstance, Mc Rae and Esrick (1993) trimmed the fibers) then
may be as high as the effective intracellular conductivity. Currents parallel to the fibers
will be in both domains, the intracellular and extracellular space are connected in parallel.
The effective conductivity of the muscle tissue will be given by (1  p) aextracell paeff.
When the current is applied perpendicular to the fibers, the cell can be described as a
nonconducting cylinder. The effective conductivity perpendicular to the fibers at according
to equation (9.17) will be within the limits 2.0 x (0.1)2 S aI,eff ::: 2.4 x (0.15)2. Thus the
effective conductivity perpendicular to the fibers will be in the range of 0.020.05S/m.
9.2.4.1.c. Cardiac tissue
Cardiac cells can be described as cylinders with a diameter of about 15/Lm and a length
of about 100/Lm. The currents perpendicular to the fibers will circumvent the fibers. The
effective conductivity in the direction perpendicular to the fibers can be calculated using
expression (9.17) with m 2. The volume fraction occupied by the fibers reported by
Clerc ( 1976) is p 0.7. The conductivity of the extracellular space is assumed to be that of
Ringer 69Qcm) yielding a conductivity of I.4S/m. So the conductivity perpendicular
to the fibers is estimated to be 1.4 x (0.3)2 O.13S/m. Rush et ai. (1963) measured the
conductivity of cardiac tissue and found for the conductivity perpendicular to the axis
0.18S/m.
Currents parallel to the fibers will be present both in the extracellular and in the
intracellular space. Cardiac cells are joined at their ends by intercalated disks and each
cell is connected to its neighbors by gap junctions passing through these disks. Cardiac
muscle fibers can be modeled by cylinders that are interconnected by junctions as depicted
in Fig. 9.13. All cylinders are homogeneously distributed and are arranged in parallel. The
current is applied parallel to the axes, such that the volume between the electrodes used to
apply the current contains 10
4
fibers or more. Each cylinder has a length of lOO/Lm and
a radius of 7.5/Lm. All parameters are chosen conform the parameters chosen by Plonsey
and Barr ( 1986) from literature. These values were measured in mammalian ventricular
tissue at a temperature of about 20°e. The resistance of a junction is Rj 0.85 x lO
6Q
.
The resistivity of intracellular fluid is 282Qcm (Chapman and Frye, 1978). Consequently,
the resistance of a cylinder with a length of lO2cm and a diameter of 15/Lmof intracellular
fluid is
The junction and the intracellular space are connected in series , so the resistance of a single
The Electrical Conductivity of Living Tissue 303
e
2 S / ~
=0.7
1.p 0.3
FIGURE 9.13. Model of cardiac tissue.
cylindrical cell is
Rcell R, + R (0.85 + 1.60) x 10 Q .
A homogeneous cylinder with the same dimensions will have the same resistance Rcell,
when an effective conductivity aeff is ascribed to the entire cylinder.
aeff = Ij(RcellO), where 1= 104m, O = x 10
6
)2, yielding aeff =
The two composites of the cardiac tissue, namely the intra and extracellular medium are
connected in parallel. Thus, the effective conductivity along the fibers for cardiac tissue is
(1  p)aextracell + paeff = 0.3 x 1.4 + 0.7 x 0.23 =
Rush et al. (1963) found for the conductivity parallel to the axis of cardiac fibers a value of
O.4S/m.
9.2.4.2 Randomlyorientatedellipsoidal particles
The architecture of human tissue may be such that fibers and cells are oriented along
each other. For instance, skeletal muscle fibers are more or less parallel. However, the
assumption that the cells or fibers have the same orientation is often not plausible. For
example, the red blood cells in blood outside of the body can roughly be described by oblate
E
M. J. Peters, J. G. Stinstra, and Leveles
4 r r  r r  .. ,  .... "T ,
3.5
3
2.5
FIGURE 9.14. The cementat ion factor for a solution of randomly orientated spheroids with semiaxes a. b. and
c, where b c.
spheroids (ellipsoids with a b c) that are randomly oriented. Boned and Peyrelasse
(1983) derived an expression for a solution of randomly orientated ellipsoidal particles.
Every orientation of the particle has the same probability and therefore the Hanai procedure
is performed such that infinitesimal amounts of particles are added such that 1/3 of the
ellipsoids have their aaxes in the direction of the electric field, 1/3 their baxes and 1/3
their caxes. For nonconducting particles again Archie's law is found with a cementation
factor
(9.19)
The lowfrequency conductivity of suspensions of ellipsoidal particles and that of doublet
shaped (budding yeast cells) or biconcave particles (erythrocytes) differs only a few percent
as can be concluded from simulation studies using as numerical method the boundary
element method (Sekine, 2000).
In Fig. 9.14 the cementation factor is depicted for an ellipsoid of revolution with semi
axes a, b and c, where b =c. As can be seen m is minimal for La = 1/3, the depolarization
factor for a sphere, in that case CTeff is maximal.
9.2.4.2.a. Blood
Archie's law can be applicated for blood. Fitting the measured data for human blood
of Geddes and Sadler (1973) to Archie's law leads to a fit for m =1.46 (see Fig. 9.7). For
blood of different animals, a slightly different cementation factor is found .
The Electrical Conductivity of Living Tissue 305
9.2.4.3 Cells of different shape
Human tissue can often be described by particles of different shapes that are immersed
in an aqueous medium. The cortex, for instance, has pyramidal cells and glial cells . The
sizes of the particles do not play a role in the theory derived. A different shaped ellipsoidal
particle is one that has a different ratio of the radii along the three axes. In order to obtain an
expression for a suspension of part icles of different shapes, we applied the Hanai procedure
by adding successively infinitesimal numbers of particles in proportion to their relative
volume fractions , leading again to Archie' s law
WIt =
p
and m,
J=a J
(9.20)
where n is the number of types of nonconducting ellipsoids, the volume fraction of
a particle of type i, p the sum of all volume fractions occupied by particles and Lj the
depolarization factor in the direction of one of the axes (Peters et al., 2001).
9.2.4.3.a. Gray matter
As an example, the effective conductivity of the superficial cortex is estimated, using
parameters found in the literature. The principal neuron types in superficial cortex are
pyramidal cells in layer II and stellate cells in layer III. Axons and dendrites from deeper
layers and axons from subcortical layers occupy a substantial portion of the tissue. The glial
cells of superficial cortex are mostly starshaped cells, i.e. astrocytes and a lesser fraction of
oligondendrocytes. The latter resemble astrocytes, but are smaller and have fewer processes.
A branching astrocyte process has a length of about 40 microns and an average diameter
of about one micron. The shape of the glial cell body in vitro is an intermediate between a
flat disk and a sphere. The pyramidal cells have a cell body and dendrites. The cell body
has a roughly pyramidal shape. The branching properties of dendrites of neurons show a
mean branching angle of about 60 degrees. The diameter of motoneurons is reported as
7 to 14f.lmfor the proximal segment with decreasing values of about 5, 3, and 2f.lm for
successive branches. The fact that solutes of high molecular weight readily pass between
cerebrospinal fluid (CSF) and interstitial fluid supports the supposition that no significant
concentration gradients exist between the two compartments in the steady state. Hence the
CSF reflects the ionic composition of the interstitial fluid and the conductivity of the CSF
will be identical to that of the interstitial fluid i.e.1.8S/m (Baumann et
The gray matter is composed of glial cells that are modeled as spheres, occupying a
volume fraction of 38 percent and pyramidal cells that are modeled as cylinders, occupying
a volume fraction of 46 percent (Havstad, 1967). Both neurons and glial cells are described
by nonconducting particles, i.e. a particle O. The effective conductivity is computed using
equation (9.20) yielding 0.097S/m. A measurement that used a uniform current to
determine the conductivity of gray matter has been carried out by Ludt and Hermann (1973),
who reported a value of 0.10S/m for the rabbit. However, the measurements were carried
out in vitro 15 minutes after death and at room temperature. As found by van Harreveld
and Ochs (1956), the conductivity drops after circulatory arrest by 30 to 35 percent
due to the emptying of blood vessels and drainage of fluid. Thus Ludt and Hermann's
306 M. J. Peters, J. G. Stinstra, and Leveles
measurements indicate that the effective conductivity of cortical tissue in vivo will be about
0.15S/m. Archie's law led to a value of about 0.10S/m. However, the values used for the
calculation may be not optimal. The volume fraction occupied by the extracellular fluid
is in the range 17 to 28 percent (van Harreveld et aI., 1965; Nicholson and Rice, 1986).
The interstitial space and the vascular volume together constitute the extracellular space.
The vascular space has a volume of about 1 to 3 percent. So the value of p) in practice
will be somewhat higher than the value handled by us in the computation of the effective
conductivity.
9.2.4.4 Clustered cells
When spherical particles are clustered in a chain or in a closepacked lattice, the
conductivity is such that these clusters behave as randomly orientated ellipsoids and thus
Archie's law will be obeyed (Grandqvist and Hunderi, 1978).
9.2.4.4.a. Blood
Blood with a high hematocrit content may aggregate. Pftitzner (1987) measured the
conductivity of blood samples containing red blood cells of varying diameter between
100Hz and 100kHz. The conductivity was essentially independent of the diameter of the
cells and the frequency. He found that at a hematocrit of60 percent or more that the dielectric
constant shows a distinct decrease. This was explained by the aggregation of blood cells.
An aggregation in a single cell chain, for instance, would lead to effective depolarization
factors La =0.13; L =L, =0.435, leading according relation (9.19) to m = 1.6. In such
case, the effective conductivity at 80 percent hematocrit would practically be the same as
without aggregation (2 percent difference).
9.2.4.4.b. Liver
The liver consists of different types of cells. These can be divided in hepatocytes and
nonhepatocytes (Raicu et al., 1998
a
) . But since the volume fraction of nonhepatocytes
(0.06) is relatively small to that of the hepatocytes (0.72) a model is used consisting of
one celltype. The hepatic cells are clustered forming plates of one layer in thickness (see
Fig. 9.15). The closepacked clusters of cells are assumed to act as oblate spheroids with axes
FIGURE 9.15. A schematic drawing of a part of a lobule of the liver.
The Electrical Conductivity of Living Tissue 307
FIGURE 9.16. A schematic drawing of the layered human skin.
a, b, c, where a b » c. Choosing the effective depolarization factors La, L and L, to
be 0.1,0.1 and O.S, respectively and applying Archie's law with p 0.72 and as 0.65S/m
(the conductivity of blood) results in O'eff 0.03S/m. This finding is in accordance with
findings of Gabriel et al. (1996
a
) who reported values in the range of 0.02o.03S/m.
Choosing La, L and L, to be 0.2, 0.2 and 0.6 leads to an effective conductivity of
O'eff O.OSS/m, which is more in accordance with the values given by Geddes and Baker
(1967) and Raicu et al. (199S
9.3 LAYERED STRUCTURES
Many tissues are organized in layers, such as the gastrointestinal tract, the retina, or
the gray matter. An example of a layered structure is shown in Fig. 9.16, where the scalp is
depicted. At an interface current lines will change direction. Corning from a medium with
higher conductivity to a medium with lower conductivity, current lines will bend in the
direction of the normal. Coming from a medium with lower conductivity to a medium with
higher conductivity, current lines will bend from the normal. As a consequence, currents
tend to cross the skull in a direction perpendicular to its surface. In other words , the layers
of the skull are approximately traversed in series. Currents in the scalp tend to be parallel
to its surface and the layers of the scalp can be approximated by a parallel connection of
resistances. The simulations discussed in sections 3a and 3b show that this is indeed the
case.
9.3.1 THE SCALP
The outer layer of the scalp is the epidermis of about 0.2mm thickness. This layer is non
sensitive and nonvascular and overlies the dermis. The epidermis is a poorly conductive
308 M. J. Peters, J. G. Stinstra, and Leveles
layer; this is ascribed to the dead nature of one of the layers within the epidermis, the
stratum corneum. The conductivity of the epidermis is approximately 0.026S/m (Sernrov
et al., 1997). The dermis is connective tissue with a thickness of 2rnrn and its conductivity
is estimated to be 0.22S/m (Yamanoto and Yamanoto, 1976). The basis of the dermis is a
supporting matrix with a remarkable capacity for holding water. The dermis has a very rich
blood supply. At the side of the head, subcutaneous fat is found of about 3mm thickness.
Sernrov et al. (1997) used a conductivity value of 0.08S/m for this layer.
To assess the effective conductivity of the scalp, a simulation is performed with a
spherical volume conductor. Two models are used. The first model consists of five shells
representing the brain, skull, fat layer (3mm), dermis (2rnrn) and the epidermis (0.2rnrn).
The conductivities of the three layers are mentioned above. The radii and conductivity of the
brain and skull are 78 and 83mm, and 0.33 and 0.0042S/m, respectively. The other model
has 3 shells: the brain, skull and the skin. The radii and conductivities of the brain and skull
are equal to those of the first model. The thickness of the skin layer is chosen to be the sum
of the thickness of the fat layer, dermis and epidermis. The effective conductivity of the
entire skin layer is calculated (assuming that the currents are crossing the scalp such that
the layers are approximately connected in parallel) by means of
(9.21)
where and a denote the thickness and conductivity of layer i. This yields a value of
0.13S/m for the effective conductivity of the skin. A current dipole on the zaxis is used as
source. The potential is calculated at the outer sphere by means of an analytical expression
(Burik, 1999). The potential using the threeshell model is compared with the potential
using the other fiveshell model. The differences are expressed by the relative difference
measure (RDM) defined as
RDM= (9.22)
where is the calculated value of the potential for the threeshell model, is the
calculated value of the potential for the fiveshell model and N is the number of points
were the values are calculated, The calculations are repeated varying the eccentricities and
orientations of the dipole, and the ratio of the thickness of the three layers of the skin. Some
results are shown in Fig. 9.17. It can be seen that equation (9.21) can be used to calculate
the effective conductivity of the scalp when the thickness of the various layers is known.
If these thicknesses are not known then a 0.13S/m will be an appropriate choice for the
effective conductivity of the scalp.
9.3.2 THE SKULL
Another example of a layered structure is the skull that can also be subdivided into
three layers. The upper and lower layers are structures made out of bone, which are bad
conductors. The middle layer is a relatively good conductor, because it is spongy and
The Electrical Conductivity of Living Tissue 309
c
~
+...,....r , ,,  ,.,...,....... '
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
FIGURE 9.17. The RDM between the scalp potential calculated using an effective conductivity and the one
calculated for a skin composed of 3 layers, namely the epidermis, the dermis and the fat.
contains blood. The total thickness varies across the skull between 4 and 11mm, the mean is
around 6mm. The skull has a conductivity in the direction perpendicular to the surface of a
factor 10 smaller than in the direction parallel to the surface. Usually, a value of 0.0042S/m
is assigned to the conductivity of the entire skull, although recent measurements of the
conductivity of the skull using bone that was temporarily removed during epilepsy surgery
led to values that were a factor ten higher (Hoekema et al., 2001).
Simulations are performed with a spherical volume conductor. The volume conductor
consisting of the brain, skull and scalp is modelled by either a threeshell model in which an
effective conductivity is assigned to the skull or a fiveshell model. The radii and conductivity
of the brain and scalp are 78 and 89.2 mm, and 0.33 and O.l3S/m, respectively. The skull
is described by three layers of conductivity of 0.0029,0.029 and 0.0029S/m, respectively.
These values are chosen such that for a mean skull thickness of 6mm consisting of layers
of 2mm each, the effective conductivity is 0.0042S/m.
The other model has 3 shells: the brain, skull and the scalp. The radii and conductivities
of the brain and scalp are equal to those of the first model. The thickness of the skull is
varied between 2 and 10mm. The ratio between the thicknesses of the three layers describ
ing the skull is varied as well. A current dipole is taken as the source, whose position and
orientation are varied. As the conductivity of the skull is much lower than that of the sur
rounding tissues one may expect that the currents will cross the skull perpendicularly. This
implies that the three layers of the skull are crossed in series. If the effective conductivity is
given by
(9.23)
then the RDM is well below 2 percent. Hence, we may conclude that the effective
310 M. J. Peters, J. G. Stinstra, and Leveles
conductivityof the skull is approximated quite well by expression (9.23). When realistically
shaped models are used in the inverse solution, both the thickness and the conductivity
should vary.
9.3.3 A LAYER OF SKELETAL MUSCLE
In the ECG inverse problem the muscle layer may be considered as one of the com
partments. Some simulation studies suggest that this anisotropic layer is the only thoracic
inhomogeneity with a significant effect on the relationship between epicardial and torso po
tentials (Stanley et al., 1991). The muscles underneath the skin of the torso are essentially
directed parallel to the body surface but otherwise almost uniformly distributed over all
angles(Rush, 1967). This implies that the conductivity parallel to the body surface differs
from that perpendicular to it. Half of the fibers that are parallel to the body surface are
perpendicular to the other half. The conductivity parallel to the body surface will be about
am (at ah) /2. The conductivity perpendicular to the body surface is al. We choose a
coordinate system such that ax a
y
am and a
z
aj.
Expression (9.2) reads in this case:
(9.24)
If the muscle layer is modeled as layer of infinite extent with thickness t, then the proper
boundary conditions are that at the surface the normal component of the current density is
continuous and the potential is continuous.
Equation (9.24) can be solved by a coordinate transformation from (x,y,z) to (x' ,y' ,z'),
where
x' x;y' y and z' zJam/al (9.25)
The scale transformation is such that the currents and the potential are chosen to be invariant.
In other words they are the same at corresponding points of the primed and unprimed system.
The components of the current density being a current divided by a surface transform as:
and
J
z
Jz
(9.26)
The electric field being the gradient of the potential transforms as:
As a consequence, in the primed system:
n2lv, 0 and J7
1
~ a a E ~
= =
(9.27)
(9.28)
Hence, after transformation we have to solve Laplace's equation with the proper boundary
condition (i.e. the zcomponent of the current density is continuous), which is invariant
under the coordinate transformation. The primed system is isotropic with conductivity
The Electrical Conductivity of Living Tissue
Jamal. In other words, the potential of the anisotropic muscle layer is equivalent to the
potential in a homogeneous isotropic medium with an effective conductivity Jamal, if the
thickness is enhanced with a factor Jam/a,.
Inserting 0.05S/m, 0.5S/m for skeletal muscle and taking a layer thickness
t lcm yields for the effective conductivity of the muscle layer O.IIS/m and an
effective thickness of the muscle layer of 5cm.
Stanley et at. (1986) showed that the agreement between calculated and measured
torso potentials significantly improved if the anisotropic nature of the muscle layer was
taken into account. Their results were based on a canine study.
9.4 COMPARTMENTS
Customarily, the volume conductor is described by a compartment model. Each com
partment is considered to be homogeneous. The number of compartments may vary. To
solve the inverse problem for EEG, the head is often described by three compartments,
the brain, skull and scalp. Usually, the conductivity assigned to the brain and to the scalp
compartment are identical. The geometry of these compartments may be a sphere covered
by two spherical shells or the compartments may have a realistic shape. Especially in the
case that a threesphere model of the head is used with standard radii of the spheres, the
interfaces between the compartments will not coincide with the actual interfaces between
the different tissues. Consequently, a compartment will consist of more than one type of
tissue. For instance, the muscles that are present in the occipital and orbitofrontal areas of
the scalp may be partly assigned to the scalp compartment and partly to the skull compart
ment. The cerebrospinal fluid may be assigned mainly to the brain compartment, but also
partly to the skull compartment. Consequently, the effective conductivities do not represent
the actual conductivities of brain matter, skull bone or skin. The effective conductivities are
those values that minimize the differences between the actual EEG and the calculated EEG.
The effective conductivity of a compartment can be estimated in three ways. First, im
planted electrodes can be used to act as a source. Measurements of the potential or magnetic
field distribution generated by the source can be used to estimate the source. The actual
source can be compared with the calculated one. By varying the effective conductivities
used in the models the differences between the actual source and the calculated one can be
minimized. The conductivities that give the smallest differences are taken as the effective
ones. Second, the effective conductivities can be determined by fitting evoked magnetic field
measurements with those of the electrical potential in case that the potentials and magnetic
fields are due to the same source. The third method is based on impedance tomography.
9.4.1 USING IMPLANTED ELECTRODES
Sometimes electrodes are inserted in the brain through trephine holes during presur
gical evaluation of epileptic patients (Veelen et al., 1990). Normally, these electrodes are
used for the measurement of the potential, especially during long term seizure monitoring.
However, when a current is applied to a pair of electrodes, it will act as an artificial source. Its
location and orientation is known from Xray or magnetic resonance images. The artificial
source can be localized from potential measurements on the scalp. The true and calculated
312 M. J. Peters, J. G. Stinstra, and Leveles
location should coincide. This can be established by varying parameters of the model used
in the inverse solution, such as the ratio between the conductivities in a threecompartment
model of the head. Homma et al. (1994) used a realistically shaped model and found that
the inverse solution was the best when the ratio between the conductivities was 1 : 1/80 : 1.
9.4.2 COMBINING MEASUREMENTS OF THE POTENTIAL AND THE
MAGNETIC FlEW
Cohen and Cuffin (1983) measured magnetic fields and electrical potentials that were
evoked by the same stimulus. Both types of measurements were used to localize a dipole
within a standard threesphere model of the head. The locations coincided when using
the ratio 1 : 1/80 : 1 for the equivalent conductivity of the brain, the skull, and the scalp
compartment. Goncalves et (2001) repeated these measurements in four subjects. They
reported values for the conductivity ratio between scalp and skull that varied between 43
and 85. Part of the observed variability may be ascribed to errors in the volume conductor
model, numerical errors, or errors in the measurement. The effect of these errors on the
potential distribution will differ from that on the magnetic field distribution. Part will be
due to the differences between the heads ofthe four individuals. In each head the distribution
of inhomogeneities within compartments will be different.
9.4.3 ESTIMATION OF THE EQUIVALENT CONDUCTIVITY USING
IMPEDANCE TOMOGRAPHY
When a current is applied to the scalp surface, a potential distribution develops across
the head. The relation between the applied currents and the resulting potential depends on
the conductivity distribution within the head. The distribution of the internal conductivity
can be estimated from measurements of this scalp potential. This type of research is known
as electrical impedance tomography. Goncalves et al. (2001) used electrical impedance
tomography to estimate the ratio between the effective conductivity of the skull and the
brain compartment for five subjects under the condition that the brain has the same ef
fective conductivity as the scalp. They used both a spherical and a realisticallyshaped
threecompartment model. For the spherical model they found for the ratio of the brain and
scalp conductivities values that varied between 31 and 124. They ascribed the observed vari
ability to geometrical errors. Differences between the actual head geometry and the three
compartment model are compensated by adjusting the values of the electrical conductivity
of the compartments. These errors will be much smaller in case realisticallyshaped mod
els are used. And indeed for their realisticallyshaped models the spread is decreased. In
the latter case the ratio of the effective conductivity of brain and skull varied between 17
and 65. They suggest that electrical impedance tomography should be a part of the EEG
inverse problem in order to take the individual differences in effective conductivities into
account. Oostendorp et al. (2000) used this method to estimate the effective conductivity
of the skull and scalp compartment for two subjects. Fitting their potential measurements
to the potentials computed by means of the boundary element method yielded a skull con
ductivity of O.013S/m and a brain conductivity of 0.20S/m. The value found for the ratio of
braintoskull conductivity was 15 : 1.
The Electrical Conductivity of Living Tissue
Impedance tomography can also be used to estimate the effective conductivities used
in a compartment model of the torso. Eynboglu et al. (1994) used this method to estimate
the effective conductivities of the heart, lungs and body in a dog. However, the accuracy of
the estimated effective conductivities was only about 40 percent, because the method is not
sensitive for changes in conductivities in the various compartments.
9.5 UPPER AND LOWER BOUNDS
In most cases, it is very difficult to actually calculate the effective conductivity of a
composite material because oflack of detailed information on the microgeometric structure.
Based on information available, lower and upper bounds for the effective conductivity can
be given. The range between these bounds decreases with increasing knowledge about the
constituents of the composite medium.
The conductivity of a composite medium cannot be higher than that of the best con
ducting phase and cannot not be lower than the worst conducting phase. Thus, for a material
consisting of n phases, denoted by the index i 1,2,3, ... , n, with increasing with i,
the limits are
(9.29)
Stricter bounds are found when apart from the conductivities also the volume fractions
of all phases are known. These bounds read (Hashin and Shtrikman, 1962)
(9.30)
The upper bound is, for instance, attained for a twophase composite medium consisting
of a suspension of needleshaped particles with the applied field in the direction of their
principal axes. The lower bound is, for instance, attained for a suspension of thin disks that
are stacked and the field is applied perpendicular to the disks.
In case it is known that the distribution of the constituents is homogeneous and the
medium is isotropic, more rigorous limits are applicable (Hashin and Shtrikman, 1962). In
this case, the upper and lower bound for a twophase material is given by:
(9.31)
p]
   + ~
P2
   + ~
the composite constitutes of homogeneously distributed and randomly orientated
particles within a conducting medium, and the particles are nonconducting spheroids
(L, L
c
) , the upper and lower bounds can be obtained from Archie's law. The depo
larization factors are restrained to La L, 1 and La 0, L, 0, L, O.
According to expression (9.19), the cementation factor m in Archie's law reads in this
case:
(9.32)
314
The maximum effective conductivity is found when
M. J. Peters, J. G. Stinstra, and Leveles
a ~ f f m am
 =O"solv(l  p) In(l  p) =0
a ~ a ~
yielding an upper bound for La = =L, = 1/3
3/2
O"crr .:::: O"sol v p)
(9.33)
(9.34)
In other words , the upper bound is attained when the particles are spheres. Apparently, such
a suspension has the least impact on the flow of currents through the material. The lower
bound O"cff 0 is attained when the particles are thin disks (La = 1, L, =L, = 0). Often
the ratio between the axes of the spheroidal particles is not known, but it is known that a
b = c. In that case the lower bound is found for cylinders, where La = 0, L, = L, = 1/2,
leading to
O"sol v P)5/3 .:::: O"cff (9.35)
In summary, for elongated, homogeneously distributed and randomly orientated non
conducting spheroids in suspension the effective conductivity is limited by the following
bounds
O"sol v(l  P)5/ 3 .:::: O"cff .:::: O"solv P)3/2 (9.36)
In Fig. 9.18 the various upper and lower bounds are plotted for a twophase composite
medium consisting of a nonconducting phase embedded in a conducting medium as a
function of the volume fraction occupied by the nonconducting phase.
9.5.1 WHITE MATTER
White matter appears white because it contains fiber groups that possess myelin sheaths
that consist of layers of membranes. These membranes, composed largely of a lipoprotein
called myelin, have a higher proportion of lipid than other surface membranes. The oligo
dendrocytes (i.e. starshaped cells) are arranged in rows parallel to the myelinated nerve
fibers, with long processes running in the same direction. The white matter can be modeled
as a suspension of elongated particles that may have any direct ion. The extracellular space
measured by van Harreveld and Ochs (1956) in the cerebellum of mice varied between 18.1
and 25.5 percent with a mean of 23.6 percent. The conductivity of the interstitial fluid is
that of the cerebrospinal fluid, i.e.l .8S/m. Thus according to equation (9.36), the effective
conductivity will be between the limits O.lOS/m < O"cff < 0.23S/m.
9.5.2 THE FETUS
In order to simulate the fetal electrocardiogram, a single compartment may describe
the fetus. As no measured values for the conductivity of the human fetus are available,
its conductivity is estimated. In order to be able to use the theory presented above, the
The Electrical Conductivity of Living Tissue 315
x 0 2
......,...r,,
0.2
0.1
0.8
:f O.7[
~ 0.6
c:
80.5
'U 0.4
~
0.3
0.2 0.4 0.6 0.8
FIGURE 9.18. Upper and lower bounds for a material consisting of two composites where one phase is non
conducting as a function of the volume fraction occupied by the nonconducting phase. The conductivity of the
conducting phase is oz.
a) Upper bound if the only information that is available is the value of (f 2 ; the lower bound (l b) coincides with
the xaxis.
(2a) upper bound and (2b) lower bound if the only information that is available is the value of (f2 and the volume
fraction p.
(3a) upper bound, (3b) lower bound if the phases are homogeneously distributed and (f 2 and p are known.
(4a) upper bound and (4b) lower bound if the nonconducting phase consists of spheroidal, homogeneously
distributed particles and the values of (f2 and p are known.
(Sa) upper bound and (5b) lower bound if the nonconducting phase consists of elongated particles that are
homogeneously distributed and randomly orientated and the values of ozand p are known.
fetus is assumed to be a homogeneous conductor. is assumed that the cells in the fetus
are homogeneously distribut ed and randomly orientated and have a shape somewhere be
tween a sphere and a cylinder. Looking at the histology of the fetus most tissues consist of
elongated spheroids or spheres. Disclike cells are less commonly encountered. Based on
these assumptions the conductivity of the fetus can be estimated to be between the limits
oAI  p)5/3 :s afetus :s a
e
p)3/2. The volume fraction of the extracellular space at the
end of gestation is about 40 percent of the total body volume (Brace, 1998; Costarino and
Brans, 1998). The extracellular space includes besides the interstitial fluid the fluids in the
body cavities like the cerebrospinal fluid and the blood plasma. The blood plasma is about
18 percent of the total extracellular water content. As the fetus is considered as one single
entity, there is no objection in taking all the extracellular fluid in account in estimating
the conductivity, as all contribute to the conductivity. In compari son the extracellular fluid
fraction in an adult is about 20percent. Thence, the fetus is at least a factor two more con
ducting than the maternal abdomen. Assuming the conductivity of the extracellular space
in both fetus and adult to comparable at a value of equation (9.36) predicts a range
316 M. J. Peters, J. G. Stinstra, and I. Leveles
of 1.9 (004)5/3 ~ afetus ::::: 1.9 (0.6)3/2 ~ 0.88S/m. Assuming that the vol
ume fraction will be somewhere between 40 and 60 percent and will approach 40 percent
at the end of gestation, a value of O.SS/mis a reasonable choice for the conductivity of the
fetus in the third trimester of pregnancy. This value is used for the solution of the inverse
problem for fetal ECG and leads to reasonable results (Stinstra, 2001).
DISCUSSION
The accuracy of measurements will be limited because the measurements are very
complicated. The accuracy of the computation is limited because the cells vary in shape, they
are not homogeneously distributed, blood supply plays a role, etcetera. Since the model used
to describe a tissue in this chapter is a simplification the results are only an approximation.
However, the results are useful in clarifying the relation between the conductivity and the
structure of the tissue. The results can be used to predict the effects of changes due to,
for instance, temperature, illnesses or age. Anyhow, it makes no sense to use values of the
effective conductivity that suggest an accuracy higher than ten percent by giving the values
with too many digits.
The effective electrical conductivity is a macroscopic parameter that represents the
electrical conductivity of the tissue averaged in space over many cells. Many of the tissues
in the body such as lung, liver, fat, and blood have cells structures that macroscopically show
no preferred direction. Even the heart, which is muscular, has its muscle strands wound in
such a complicated fashion that, overall, no preferred direction can be readily discerned.
Baynhamand Knisley (1999) measured the effective epicardial resistance of rabbit ventricles
and found that in contrast to isolated fibers the ventricular epicardium exhibits an isotropic
effective resistance due to transmural rotation of fibers. Only skeletal muscle cells have a
definite preferred direction when many cells are averaged (Rush et al., 1984). Most cells
have an elongated shape. Thence, the bounds given in section 9.5 can be used to estimate
the effective conductivity. These bounds are not so far apart, so they will help to restrict the
uncertainties in the effective conductivity to be used in the bioelectrical inverse problem.
An exception form long skeletal muscle and heart tissue, as the conductivity parallel to the
fibers will take place both in the extracellular and the intracellular space.
REFERENCES
Archie, G.E., 1942, The electrical resistivity log as an aid in determining some reservoir characteristics, Trans.
Am. Institut. Min. Metal. Eng., 146: 5562.
Aseyev, 1998, Electrolytes. Interparticle interactions. Theory, calculation methods and experimental data, Begell
House inc., New York.
Baumann, S.B., Wozny, D.R., Kelly, S.K., and Meno, EM., 1997, The electrical conductivity of human cere
brospinal fluid at body temperature, IEEE BioMed. Eng., 44: 220223.
Baynham, C.']", Knisley, S.B., 1999, Effective resistance of rabbit ventricles, Ann. of Biomed. Eng., 27:
96102.
Boned, C., and Peyrelasse, J., 1983, Etude de la permittivite complexe d'ellipsoides disperses dans un milieu
continuo Analyses theorique et numerique, Colloid Polym. Sci., 261:600612.
Boyle, M. H., 1985, The electrical properties of heterogeneous mixtures containing an oriented spheroidal dispersed
phase, Colloid Polym. Sci., 263:5157.
Th e Electrical Conductivity or Living Tissue 317
Brace, R.A., 1998, Fluid distribution in the fetus and neonate, in: Fetal and neonatal Physiology. (R. A. Polin,
and W.w. Fox, eds.), Saunders Comp., Philadelphia, pp. 1703 1713.
Burger, H. c., and Dongen, R. van, 1961, Specific electric resistance of body tissues, Phys. Med. Biol., 5:
431437.
Burger, H. C., and Milaan, J. B. van. 1943, Measurement of the specific resistance of the human body to direct
current, Act. Med. Scand., 114:585607.
Burik, M. J. van, 1999, Physical aspects of EEG, PhD thesis, University of Twente, the Netherlands.
Chapman, R.A., and Frye, C.H., 1978. An analysis of the cable properti es of frog ventricular myocardium, J.
Physiol., 283:263283.
Clerc, 1976, Directional differences of impulse spread in trabecular muscle from mammali an heart, Ibid.
255:335 346.
Cohen, D., and Cuffin, B.N., 1983, Demonstration of useful differences between magnetoencephalogram and
electroencephalogram, Electroen. din. Neuro., 56:3851.
Cole, S., Li, C., and Bak, A. E. 1969. Electrical analogues for tissues. Exp. Neurol., 24:459473.
Costarino, A.T.. and Brans, w., 1998, Fetal and neonatal body fluid composi tion with reference to growth
and development, in: Fetal and neonatal Physiology, (R. A. Polin, and W.w. Fox, eds.), Saunders Comp.,
Philadelphia, pp. 17131721.
De Luca, E, Cametti, C, Zimat ore, G., Maraviglia, B., and Pachi, A., 1996, Use of lowfrequency electrical
impedance measurements to determine phospholipid content in amniotic fluid, Phys. Med. BioI., 41:1863
1869.
Epstein. B. R., and Foster, R., 1983, Anisotropy in the dielect ric properties of skeletal muscle, Med. BioI. Eng.
Comput., 21:5155.
Eyuboglu, B. M., Pilkington, T. C; and Wolf, P. D., 1994, Estimation of tissue resistivities from multipleelectrode
measurements, Phys. Med. BioI. 39: 117.
Foster, K. R., and Schwan, H. P., 1989, Dielectric properties of issues and biological materi als: a critical review,
Crit. Rev. Biomed. Eng., 17:25 104.
Foster, K. R., and Schwan, H. P., 1986, Dielectric permittivity and electrical conductivity of biological materials,
in: Handbook of Biological Effects of Electromagnetic Fields, (C. Polk, and E. Postow, eds.), CRC Press,
Inc., Boca Raton, pp. 27.
Fricke, H., 1953, The Maxwel lWagner dispersion in a suspension of ellipsoids , Phys. Chem., 57:934937.
Gabriel, S., Lau, R. w. , and Gabriel, C., 1996
a
, The dielectric properties of tissue: II. Measurements in the
frequency range 10Hz to 20GHz, Phys Med BioI., 41:2251 2269.
Gabriel, S., Lau, R. w. , and Gabriel, C., 1996 The dielectric properties of biological tissues: III. Parametric
models for the dielectric spectrum of tissues, Phys. Med. Bioi., 41:22712293.
Geddes, A., and Baker, E., 1967, The specific resistance of biological materialA compendium of data for
the biomedical engineer and physiologist, Med. BioI. Eng., 5:271 293.
Geddes, A., and Sadler, 1973, The specific resistance of blood at body temperature, Med. BioI. Eng.,
11:336339.
Gcrsing, E., 1998, Monitoring temperature induced changes in tissue during hyperthermia by impedance methods,
Proc. of the X.ICEBI, Universitat Politecnica de Cataluya.
Gielcn, E, 1983, Electrical conductivity and histological structure of skeletal muscle. PhD Thesis, University of
Twente, the Netherlands.
Gielen, E H., Wallingade Jonge, w., and Boon, K. 1984, Electrical conductivity of skeletal tissue: experi
mental results from different muscles in vivo, Med. BioI. Eng. Comput ., 22:569577.
Goncalves, S., Munck, J. C. de, Heethaar, R. M., Lopes da Silva E H., and Dij k, B. W. van, 2000, The application
of electrical impedance tomography to reduce systematic errors in the EEG inverse problema simulation
study, Physiol. Meas., 21:379393.
Grandqvist, C. G., and Hunderi, 1978, Conductivity of inhomogeneous materials: effective medium theory
with dipoledipole interaction, Phys. Rev. B, 18:15541561.
Hanai, T., 1960, Theory of the dielectric dispersion due to the interfacial polarization and its application to
emulsions, KolloidZ., 171:23 31.
Harreveld, van, Crowell , J., Malhtotra S.A., 1965, A study of extracellular space in central nervous tissue by
freezesubstitution, J. Cell Bioi., 25: 1171 37.
Harreveld, A. van, and Ochs, S., 1956, Cerebral impedance charges after circulatory arrest, Am. Physiol.,
187:203207.
318 M. J. Peters, J. G. Stinstra, and Leveles
Hart, E X., Berner N. J., and McMillen R. 1999, Modelling the anistropic electrical properties of skeletal
muscle, Phys. Med Bioi., 44:413421.
Hashin, Z., and Shtrikman S., 1962, A variational approach to the theory of the effective magnetic permeability
of multiphase materials, Appl. Phys., 33:31253131.
Havstad, J. W., 1967, Electrical impedance of cerebral cortex: an experimental and theoretical investigation, PhD
Thesis, Stanford University.
Hoekema, R., Huiskamp, G. J. M., Wieneke, G. H., Leijten, E S. S., van Veelen, C. W. M., van Rijen, C., and
van Huffelen, A. 2001, Measurement of the conductivity of the skull, tempoarily removed during epilepsy
surgery, Biomed Tech., 46:103105.
Homma, S., Musha, Nakajima, Y, Okamoto, Y, Blom, S., Flink, R., Hagbach, E., and Mostrom, U.,
1994, Location of electric current sources in the human brain estimated by the dipole tracing method of the
scalpskullbrain (SSB) head model, Electroen. Clin. Neuro., 91:374382.
Kobayashi, N., and Yonemura, 1967, The extracellular space in red and white muscles of the rat,
Physiol., 17:698707.
Kotnik, T, Bobanovic, E, and Miklavcic, D., 1997, Sensitivity of transmembrane voltage induced by applied
electric fieldsa theoretical analysis, Bioelectroch. Bioener., 43:285291.
Law, S. K., 1993, Thickness and resistivity variations over the upper surface of the human skull, Brain Topogr.,
6:99109.
Ludt, H., and Hermann, H. D., 1973, In vitro measurement of tissue impedance over a wide frequency range,
Biophys. 10:337345.
Maxwell, J. C., 1891, A treatise on electricity and magnetism, volume 1, Arts. 311314, Dover Publ., New York.
Me Rae, D. A., and Esrick, M. A., 1993, Changes in electrical impedance of skeletal muscle measured during
hyperthermia, Int. Hyperthermia, 9:247261.
Nicholson, C., and Rice, M. E., 1986, The migration of substances in the neural microenvironment, Ann. New
York Academy of Sciences, 481:5571.
Nicholson, P. w., 1965, Specific impedance of cerebral white matter, Exp. Neurol., 13:386401.
Oostendorp, T E, Delbeke, J., and Stegeman, D. E, 2000, The conductivity of the human skull; Results of in vivo
and in vitro measurements, IEEE BioMed. Eng., 47:14871492.
Peters, M. J., Hendriks, M., and Stinstra, J. G., 2001, The passive DC conductivity of human tissue described by
cells in solution, Bioelectroch., 53:155160.
Pehtig, R., and Kell, D. B., 1987, The passive electrical properties of biological systems: their significance in
physiology, biophysics and biotechnology, Phys. Med. Bioi., 32:933970.
Pfiitzner, H., 1984, Dielectric analysis of blood by means of a rasterelectrode technique, Med. BioI. Eng. Comput.,
22:142146.
Plonsey, R, and Barr, R.C., 1986, Effect of microscopic and macroscopic discontinuities on the response of cardiac
tissue to defibrilating (stimulating) currents, Med. BioI. Eng. Comput., 24:130136.
Plonsey, and Heppner, D.B., 1967, Considerations of quasistationarity in electrophysiological systems, Bulletin
of mathematical Biophysics, 29:657664.
Raicu, Y, Saibara, T, and Irimajiri, A., 1998
a,
Dielectric properties of rat liver in vivo: a noninvasive
approach using an openended coaxial probe at audiolradio frequencies, Bioelectroch. Bioener., 47:
325332.
Raicu, Y, Saibara, T, Enzan H., and Irimajiri, A., 1998
b
, Dielectric properties of rat liver in vivo: analysis by
modeling hepatocytes in the tissue architecture, Bioelectroch. Bioener., 47:333342.
Robillard, P. N., and Poussart Y, 1977, Specificimpedance measurements of brain tissues, Med. BioI. Eng.
Comput., 15:438445.
Rosell, J., Colominas, J., Riu, P., PallasAreny, R., and Webster, J. G., 1988, Skin impedance from 1 Hz to 1 MHz,
IEEE BioMed. Eng., 35:649651.
Rush, S., 1967, A principle for solving a class of anisotropic current flow problems and applications to electro
cardiography, IEEE BioMed. Eng., BME14:1822.
Rush, S., Abildskov, lA., and Me Fee, 1963, Resistivity of body tissues at low frequencies, Circ. Res.,
XII:4050.
Rush, S., Mehtar, M., and Baldwin, E, 1984, Normalisation of body impedance data: a theoretical study, Med.
BioI. Eng. Comput., 22:285286.
Schwan, H. P., 1985, Dielectric properties of cells and tissues, in: Interactions between Electromagnetic Fields
and Cells, (A Chiabrera, C. Nicolini, and H. P. Schwan, eds.) NATOASI series, vol. 97, Plenum Press, New
York, pp. 75103.
The Electrical Conductivity of LivingTissue 319
Schwan, H. P., and Foster, K. R., 1980, RFField interactions with biological systems: Electrical properties and
biophysical mechanisms, Proc. ofthe IEEE, 68: 104113.
Schwan, H. P., and Takashima, S., 1993, Electrical conduction and dielectric behaviour in biological systems,
Encyclopedia ofApplied Physics, 5:177199.
Sekine, K., 2000, Application of boundary element method to calculation of the complex permittivity of suspensions
of cells in shape ofDoch symmetry, Electroch., 52:17.
Semrov, D., Karba, R, and Valencic, 1997, DC Electrical stimulation for chronic wound healing enhancement.
Part 2. Parameter determination by numerical modelling, Bioelectroch. Bioener., 43:271277.
Sillars, W., 1937, The properties of a dielectric containing semiconducting particles of various shapes, Ins.
Electrical Eng., 80:378394.
Stanley, P. C., Pilkington, T. C., and Morrow, M. N., 1986, The effects of thoracic inhomogeneities on the
relationship between epicardial and torso potentials. IEEE BioMed. Eng., BME33:273284.
Stanley, P. C., Pilkington, T. Morrow, M. N., and ldeker, R. E., 1991, An assessment of variable thickness and
fiber orientation of the skeletal muscle layer on electrocardiographic calculations, IEEE BioMed. Eng.,
38:10691076.
Stinstra, J. G., 2001, Reliability offetal magnetocardiography, PhD thesis, University of Twente, the Netherlands.
Stuchly, M. A., and Stuchly, S. S., 1980, Dielectric properties of biological substancesTabulated, Microwave
Power, 15:1926.
Takashima, S., 1989, Electrical properties of biopolymers and membranes, lOP Publishing Ltd, Bristol.
Trautman, E. D., and Newbower, S., 1983, A practical analysis of the electrical conductivity of blood, IEEE
BioMed. Eng., BME30:141153.
Ulgen, and Sezdi, M., 1998, Electrical parameters of human blood, Proceedings 20th Ann. Int. Conference,
IEEEIEMBS, Hongkong, 29832986.
Veelen, C. van, Debets, R, Huffelen, A. van, Emde Boas, W. van, Binnie, C., Storm van Leeuwen, W., Velis, D. N.,
and Dieren, A. van, 1990, Combined use of subdural and intracerebral electrodes in preoperative evaluation
of epilepsy, Neurosurgery, 26:93101.
Yamamoto, T., and Yamamoto, Y., 1976, Electrical properties of epidermal stratum corneum, Med. Biol. Eng.,
3:151158.
Zheng, E., Shao, S., and Webster, J. G., 1984,Impedance of skeletal musclefrom 1Hz to 1MHz,IEEE BioMed.
Eng., BME·31: 477481.
Action potential 2, 89, 120
Activation time imaging 167
Adaptive spatial filter 226
Anisotropic bidoma in 50
Archie's law 300
Beamformer 226, 230, 231
Bidomain model 124
Bidomain myocardi um 46
BiotSavart law 216
Blockdesign fMRI 254
Body surface isopotential map 107
Body surface Laplacian 144, 183
Body surface Laplacian mapping 192
Body surface potential map 126
BOLD 252
BSLM 192
BSPM 126
Cable theory 17
Cardiac action potential 7, 8, 9
Cardiac arrhythmia 27
Cardiac tissue 302
Cell networks 23
Cells 289
Computer heart model 62
Conductivity tensor 51
Coregistration 262, 263
Cortical imaging 257
Current dipole density 46
DAD 14,25
Defibrillation 72
Dipole distribution imaging 164
Dipole distribution 163, 164
Dipole localization 256, 267
Dipole source imaging 163, 165
Dipole source 44
INDEX
Drug integration 35
EAD 14,25
ECG 183
EEG 183,263
Effective conductivity 47, 283
Electrical conductivity 28 1
Electrocardiographic tomographic imagi ng 161,
168,175
Endocardial potential imaging 129, 147
Endocardial potential 127, 128
Epicardial potential imaging 129, 138
Epicardial potenti al 57
Equivalent conductivi ty 312
Equivalent current density 49
Equivalent dipole 54
Equivalent moving dipole 163
Eventrelated fMRI 254
Extracell ular electrogram 20
Extracell ular fluid 29 1, 295
Fat 301
Fiber orientation 52, 86
Finite difference laplacian 186
Finite difference method 58
Finite element method 58
Finite volume method 60
FitzHughNagumo model 100
fMRI 252
Forward problem 43, 53
Functional magnetic resonance imaging
252
Genetic integration 35
Global Laplacia n estimate 188
Gradiometer 215
Gray matter 305
Green's function 54, 125
321
Modeling and Imaging of Bioelectrical Activity Principles and Applications
BIOELECTRIC ENGINEERING
Series Editor: Bin He
University of Minnesota MinneapoH~ Minnesota
MODELING AND IMAGING OF BIOELECTRICAL ACTIVITY Principles and Applications Edited by Bin He
Modeling and Imaging of Bioelectrical Activity
Principles and Applications
Edited by
Bin He
University of Minnesota Minneapolis, Minnesota
Kluwer Academic/ Plenum Publishers
New York, Boston, Dordrecht, London, Moscow
Library of Congress CataloginginPublicat ion Data Modeling and imaging of bioelectrical activity: principles and applications/edited by Bin He. p. ; cm.  (Bioelectric engineering) Includes bibliographical references and index. ISBN 030648112X 1. HeartElectric propertiesMathematical models. 2. HeartElectric propertiesComputer simulation. 3. BrainElectric propertiesMathematical models. BrainElectric propertiesComputer simulation. I. He, Bin, 1957 II. Series. QP112.5.E46M634 2004
4.
612'.0142T 011 dc22
2003061963
ISBN 030648112X ©2004 Kluwer Academic /Plenum Publishers, New York 233 Spring Street, New York, New York 10013 http://www.wkap.nl/ 10 9 8 7 6 5 4 3 2 1
A C.I.P. record for this book is available from the Library of Congress All rights reserved No part of this book may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, microfilming, recording, or otherwise, without written permission from the Publisher, with the exception of any material supplied specifically for the purpose of being entered and executed on a computer system , for exclusive use by the purchaser of the work. Permissions for books published in Europe: permissions@wkap.nl Permissions for books published in the United States of America: permissions @wkap.com Printed in the United States of America
PREFACE
Bioelectrical activity is associated with living excitable tissue. It has been known, owing to efforts of numerous investigators, that bioelectrical activity is closely related to the mechanisms and functions of excitable membranes in living organs such as the heart and the brain. A better understanding of bioelectrical activity, therefore, will lead to a better understanding of the functions of the heart and the brain as well as the mechanisms underlying the bioelectric phenomena. Bioelectrical activity can be better understood through two common approaches. The first approach is to directly measure bioelectrical activity within the living tissue. A representative example is the direct measurement using microelectrodes or a microelectrode array. In this direct measurement approach, important characteristics of bioelectrical activity, such as transmembrane potentials and ionic currents, have been recorded to study the bioelectricity of living tissue. Recently, direct measurement of bioelectrical activity has also been made using optical techniques. These electrical and optical techniques have played an important role in our investigations of the mechanisms of cellular dynamics in the heart and the brain. The second approach is to noninvasively study bioelectrical activity by means of modeling and imaging. Mathematical and computer models have offered a unique capability of correlating vast experimental observations and exploring the mechanisms underlying experimental data. Modeling also provides a virtual experimental setting, which enables well controlled testing of hypothesis and theory. Based on the modeling of bioelectrical activity, noninvasive imaging approaches have been developed to detect, localize, and image bioelectrical sources that generate clinical measurements such as electrocardiogram (ECG) and electroencephalogram (EEG). Information obtained from imaging allows for elaboration of the mechanisms and functions of organ systems such as the heart and the brain. During the past few decades, significant progress has been made in modeling and imaging of bioelectrical activity in the heart and the brain. Most literature, however, has treated these research efforts in parallel. The similarity arises from the biophysical point of view that membrane excitation in both cardiac cells and neurons can be treated as volume current sources. The clinical observations of ECG and EEG are the results of volume conduction of currents within a body volume conductor. The difference among bioelectrical activity originating from different organ systems is primarily due to the different physiological mechanisms underlying the phenomena. From the methodological point of view,
v
and in particular electrocardiography. It is aimed at serving as a reference book for researchers working in the field of modeling and imaging of bioelectrical activity. Chapter 6 deals with a noninvasive body surface mapping technology . bioengineering.vi Preface therefore. body surface Laplacian mapping has received relatively recent attention in its enhanced capability of identifying and mapping spatially separated multiple activities .surface Laplacian mapping. Although this book focuses on bioelectric activity of the heart and the brain. or medical physics curriculum. Many investigation s have been made in order to inversely estimate and reconstruct potential distribution over the epicardium. and stateoftheart research that are presented in this book should also be applicable to a variety of applications. Chapter 2 discusses the applications of forward theory to whole heart modeling and defibrillation. The subsequent two chapters treat inverse imaging of the brain from neuromagnetic and neuroelectric measurements. and increase our understanding of cardiac pathology in threedimension and whole heart models . Chapter 3 reviews important issues in whole heart modeling and its implementation as well as various applications of whole heart modeling and simulations of cardiac pathologies. Following a review of the theoretical basis of equivalent dipole source models and stateoftheart numerical methods of computing the electrical potential fields. This chapter also illustrates that a noninvasive mapping technique can be applied to imaging of bioelectrical activity originated from different organ systems. and whole organ for bioelectrical activity in the heart. Chapter 5 reviews the recent development in three dimensional electrocardiography tomographic imaging . or equivalent current dipole distribution. It illustrates how modeling can help elucidate the mechanisms of cardiac cells and cell networks. methodology. and as a textbook for graduate students and seniors in a biomedical engineering. Chapter 4 provides a systematic treatment of the methods and applications of heart surface inverse solutions . Inparticular. or activation sequence. Compared with wellestablished body surface potential mapping . as well as functional magnetic resonance imaging (fMRI). as an introduction to investigators who are interested in entering the field or acquiring knowledge about the current state of the field. which illustrates the close relationship between modeling and imaging and the merits of modelbased imaging . Chapter I provides a systematic review of onecell models and cell network models as applied to cardiac electrophysiology. over the heart surface from body surface electrocardiograms. it is possible to estimate threedimensional distributions of electrophysiological characteristics such as activation time and transmembrane potentials. a wholeheartmodel based tomographic imaging approach is introduced. the theory. The purpose of this book is to provide a stateoftheart coverage of basic principles. theories . These approaches and activities are well reviewed in Chapter 4. modeling and imaging of bioelectrical activity can be treated within one theoretical framework. . Chapter 3 also illustrates important clinical applications the modeling approach can offer. and methods of modeling and imaging of bioelectrical activity with applications to cardiac and neural electrical activity. Progress has also been made to estimate endocardial surface potentials and activation sequence from catheter recordings. Recent research shows that. Chapter 2 provides a thorough theoretical treatment of the forward problem of bioelectricity. The first three chapters deal with the modeling of cellular activity. by incorporating a priori information into the inverse solutions. such as the heart and the brain. cell networks. The following two chapters review the theory and methods of inverse imaging with applications to the heart .
NSF BES0218736 and NSF BES020l939. The conductivity parameter is needed in establishing accurate forward models of the body volume conductor and obtaining accurate inverse solutions using modelbased inverse imaging. Chapter 9 deals with tissue conductivity. these four chapters are intended to provide a solid foundation in inverse imaging methods as applied to imaging bioelectrical activity. and multimodal imaging integrating EEG. I am very grateful to them for their contributions during their very busy schedules and their patience during this process. I am indebted to Aaron Johnson Brian Halm.Preface vii Chapter 7 reviews the forward modeling of magnetoencephalogram (MEG). MEG and tMR!. and neuromagnetic source imaging with a focus on spatial filtering approach. Along with Chapters 4 and 5. We hope this book will provide an intellectual resource for your research and/or educational purpose in the fascinating field of modeling and imaging of bioelectrical activity. Chapter 8 provides a general review of tMR!. Chapter 9 systematically addresses this issue for various living tissues. is also greatly appreciated. through grants of NSF CAREER Award BES9875344. and Kevin Sequeira of Kluwer Academic Publisher for their great support during this project. the accuracy of tissue conductivity is crucial in ensuring accurate and robust imaging of bioelectrical activity. Bin He Minneapolis . This book is a collective effort by researchers who specialize in the field of modeling and imaging of bioelectrical activity. Financial support from the National Science Foundation. As most inverse solutions are derived from noninvasive measurements with the assumption of known tissue conductivity distribution. linear inverse solutions for EEG and MEG. Shoshana Sternlicht. an important parameter that is required in bioelectric inverse solutions.
.....2.............2 1.....3.......2 1........... Modeling the Cardiac Action Potential.1 1...4..1 1......4 1..2 Dipole Source Representations 2.....1 Introduction........ 2 43 43 44 44 46 53 53 54 ix ......1 Fundamental Equations 2.. 2.... Shenai Introduction The Onecell Model Voltage Gating Ion Channel Kinetics (HodgkinHuxleyFormalism) ..2 The Bidomain Myocardium.......1 1...2..............3.............5 1...................3 1........2 1.2 1.....1 1.... Vivek Iyer.......................2........ and Mahesh B......1.1.......4 1........ ......3.....4 FROM CELLULAR ELECTROPHYSIOLOGY TO ELECTROCARDIOGRAPHy... Reconstruction of the Local Extracellular Electrogram (Forward Problem) Modeling Pathology in Cellular Networks Modeling Pathology in Threedimensional and Whole Heart Models Myocardial Ischemia Preexcitation Studies Hypertrophic Cardiomyopathy Drug Integration in Threedimensional Whole Heart Models Genetic Integration in Threedimensional Whole Heart Models.......2.3..... Nitish V. Gulrajani 2...........3 Torso Geometry Representations 2. Discussion References 1 1 3 3 7 10 17 17 18 20 23 29 31 31 34 35 35 36 38 THE FORWARD PROBLEM OF ELECTROCARDIOGRAPHY: THEORETICAL UNDERPINNINGS AND APPLICATIONS.........2.... Ramesh M.2...................4 Solution Methodologies for the Forward problem 2. 2...1.......3..3 1..............1 Surface Methods..3 1.... Thakor.............. Modeling Pathologic Action Potentials Network Models Cellcell Coupling and Linear Cable Theory Multidimensional Networks ....CONTENTS 1 1 1................3 1.....
..2..1 4.............3...............4......5....2 4....4...3 119 120 120 123 123 128 129 132 135 138 138 139 141 .....3 4.....2 4.3 2.....6 3..2 3......4........ 82 Hearttorso Geometry Modeling...1..5 2..........4........1 3............. 94 Cardiac Electric Sources and Surface ECG Potentials 100 Computer Simulations and Applications 103 Simulation of the Normal Electrocardiogram 103 Simulation of STT Waves in Pathologic Conditions 107 Simulation of Myocardial Infarction 108 Simulation of Pace Mapping 110 110 Spiral WavesA New Hypothesis of VentricularFibrillation Simulation of Antiarrhythmic Drug Effect 110 Discussion 111 References 114 HEART SURFACEELECTROCARDIOGRAPHIC INVERSE SOLUTIONS Fred Greensite Introduction The Rationale for Imaging Cardiac Electrical Function A Historical Perspective Notation and Conventions The Basic Model and Source Formulations Heart Surface Inverse Problems Methodology Solution Nonuniqueness and Instability Linear Estimation and Regularization Stochastic Processes and Time Series ofInverse Problems Epicardial Potential Imaging Statistical Regularization Tikhonov Regularization and Its Modifications Truncation Schemes 119 4......3.5 3..1 4.....4......1............4 3.........3.3 4...........2..... Computer Heart Models Effects of Torso Conductivity Inhomogeneities Defibrillation. Combination Methods Applications of the Forward Problem..1 3.........3..2 2.......3.....4 4 WHOLE HEART MODELING AND COMPUTER SIMULATION 81 Darning Wei Introduction 81 Methodology in 3D Whole Heart Modeling....x Contents 2.....................3....1 3..6 3..... Future Trends References 58 61 61 62 70 72 75 75 3 3..............2...........2 3.......3..1...5............1 4..5......2....3 2....2......4 4.3.........1 4....3 3..3 3......... 82 Inclusion of Specialized Conduction System 83 Incorporating Rotating Fiber Directions 85 Action Potentials and Electrophysiologic Properties 89 Propagation Models...................3 4.......................2 2...4 3.................3....6 VolumeMethods......3 3.....2 4.....2 3..5 3..2..........2 4.........1 2....
5 4.4.6.3 Surface Laplacian Imaging of Heart Electrical Activity .3 5.1 5.2.4.10 4.7 4.3.6 4.1 4.1 5.9 4.2 Surface Laplacian Estimation Techniques 6.4.3.3 5.4.2 5.6 5.2.2.8 4.3.2 5.2.1 Highresolution ECG and EEG 6.4 5.2.4 5.4.4 4.4.2 Global Laplacian Estimates 6.3 Surface Laplacian Based Inverse Problem 6.1.5 THREEDIMENSIONAL ELECTROCARDIOGRAPHIC TOMOGRAPHIC IMAGING Bin He Introduction '" ThreeDimensional Myocardial Dipole Source Imaging Equivalent Moving Dipole Model Equivalent Dipole Distribution Model Inverse Estimation of 3D Dipole Distribution Numerical Example of 3D Myocardial Dipole Source Imaging ThreeDimensional Myocardial Activation Imaging Outline of the HeartModel based 3D Activation Time Imaging Approach Computer Heart Excitation Model Preliminary Classification System Nonlinear Optimization System Computer Simulation Discussion ThreeDimensional Myocardial Transmembrane Potential Imaging Discussion References 161 161 163 163 163 164 165 167 167 168 169 170 171 174 175 178 180 6 BODY SURFACE LAPLACIAN MAPPING OF BIOELECTRIC SOURCES 183 183 183 184 186 186 188 190 192 Bin He and lie Lian 6.3 5.3.1 Introduction 6.5 4.2 5.Contents xi 4.2 Biophysical Background of the Surface Laplacian 6.1.2 4.1 5.4 5.6 4.6.3.3.1 Local Laplacian Estimates 6.2.7 Specific Constraints in Regularization Nonlinear Regularization Methodology An Augmented Source Formulation Different Methods for Regularization Parameter Selection The Body Surface Laplacian Approach Spatiotemporal Regularization Recent in Vitro and in Vivo Work Endocardial Potential Imaging Imaging Features of the Action Potential Myocardial Activation Imaging Imaging Other Features of the Action Potential Discussion References 142 143 143 143 144 145 146 147 149 149 154 155 156 5 5.5 5.4.2.
4.2 7.3 7.5.4 7.3 6.2 7.3.6 7. Nagarajan Introduction Brief Summary of Neuromagnetometer Hardware Forward Modeling Definitions Estimation of the Sensor Lead Field Lowrank Signals and Their Properties Spatial Filter Formulation and Nonadaptive Spatial Filter Techniques Spatial Filter Formulation Resolution Kernel Nonadaptive Spatial Filter Noise Gain and Weight Normalization Adaptive Spatial Filter Techniques Scalar Minimumvariancebased Beamformer Techniques Extension to Eigenspaceprojection Beamformer Comparison between Minimumvariance and Eigenspace Beamformer Techniques Vectortype Adaptive Spatial Filter Numerical Experiments: Resolution Kernel Comparison between Adaptive and Nonadaptive Spatial Filters Resolution Kernel for the Minimumnorm Spatial Filter Resolution Kernel for the Minimumvariance Adaptive Spatial Filter Numerical Experiments: Evaluation of Adaptive Beamformer Performance Data Generation and Reconstruction Condition Results from Minimumvariance Vector Beamformer Results from the Vectorextended BorgiottiKaplan Beamformer Results from the Eigenspace Projected Vectorextended BorgiottiKaplan Beamformer Application of Adaptive Spatial Filter Technique to MEG Data Application to Auditorysomatosensory Combined Response 213 213 214 215 215 216 219 221 221 222 222 225 226 226 227 228 230 232 232 234 235 235 238 238 238 243 243 .4 7.1 7.1 NEUROMAGNETIC SOURCE RECONSTRUCTION AND INVERSE MODELING Kensuke Sekihara and Srikantan S.5.2 6.5 Highresolution Laplacian ECG Mapping Performance Evaluation of the Spline Laplacian ECG Surface Laplacian Based Epicardial Inverse Problem Surface Laplacian Imaging of Brain Electrical Activity Highresolution Laplacian EEG Mapping Performance Evaluation of the Spline Laplacian EEG Surface Laplacian Based Cortical Imaging Discussion References 192 193 199 200 200 200 206 208 209 7 7.4.1 7.1 7.1 6.3 7.3.4.4 7.3 6.4.3.3 7.7.4.4 7.2 7.1 7.4.5 7.1 7.7 7.1 6.8.7.6.4.2 7.5.3.5.8 7.7.2 7.3 7.xii Contents 6.2 7.6.3.2 6.7.4 6.1 7.3.3 7.
4.1.4.2 .2.8.3 9.3 9.3.2 8.2 9.1 9.2.5 9.2.4.1 8. MEG and tMRI Data Visible and Invisible Sources Experimental Design and Coregistration Issues Integration of EEG and MEG Data Functional Hemodynamic Coupling and Inverse Estimation of Source Activity Discussion References THE ELECTRICAL CONDUCTIVITY OF LIVING TISSUE: A PARAMETER IN THE BIOELECTRICAL INVERSE PROBLEM Maria J.2 9.1 9.1.4 9.1 8.4 8.3.3.1.3.1 9.1 8.3.1 8.2 9.2 8. and Ibolya Leveles Introduction Scope of this Chapter Ambiguity of the Effective Conductivity Measuring the Effective Conductivity Temperature Dependence Frequency Dependence Models of Human Tissue Composites of Human Tissue Conductivities of Composites of Human Tissue Maxwell's Mixture Equation Archie's Law Layered Structures The Scalp The Skull A Layer of Skeletal Muscle Compartments Using Implanted Electrodes Combining Measurements of the Potential and the Magnetic Field 245 247 251 251 252 254 254 255 256 257 259 261 261 262 263 267 275 276 8. Peters.3. Stinstra.3 9.3 8.3 8.4 9.2.4 8. Jeroen G.3.3 8.2.1.4 9.1 9.1 9.5 9 281 281 282 283 284 287 287 289 289 292 296 300 307 307 308 310 311 311 312 9.1.2 9.4.4 8.2 8.2 Application to Somatosensory Response: Highresolution Imaging Experiments References 8 MULTIMODAL IMAGING FROM NEUROELECTROMAGNETIC AND FUNCTIONAL MAGNETIC RESONANCE RECORDINGS Fabio Babiloni and Febo Cincotti Introduction Generalities on Functional Magnetic Resonance Imaging Blockdesign and EventRelated tMRI Inverse Techniques Acquisition of Volume Conductor Geometry Dipole Localization Techniques Cortical Imaging Distributed Linear Inverse Estimation Multimodal Integration of EEG.4.3 9.Contents xiii 7.4.
6 Discussion References INDEX 312 313 314 314 316 316 321 .1 White Matter 9.5 Upper and Lower Bounds 9.3 Estimation of the Equivalent Conductivity using Impedance Tomography 9.4.5.xiv Contents 9.5.2 The Fetus 9.
networks representing tissue diversity and realistic heart geometries can be molded into a whole heart model.1 FROM CELLULAR ELECTROPHYSIOLOGY TO ELECTROCARDIOGRAPHY by Nitish V. and finally. bone and dermis. The bulk electrophysiological signal recorded from these networks is called the local extracellular electrogram.1). Simply put. and Mahesh B. the whole heart model can be placed in a torso model replicating lung. one can reconstruct the salient electric signal (action potential. At each level. 720 Rutland Ave. a linear network (cable). Extending cellular detail to wholeheart electrocardiography requires spanning several levels of analysis (Figure 1. from nucleotide to bedside. electrogram. would prove invaluable to diagnosis and treatment. Shenai t Department of Biomedical Engineering. The onecell model describes an action potential recording from a single cardiac myocyte. The Johns Hopkins University. Thakor. Subsequently. One ultimate goal of the cardiac modeler is to integrate cellular level detail with quantitative properties of the ECG (a property of the whole heart). Fortunately. the rapid explosion in computational power allows the modeler to span the details of each molecular "leaf" to the "forest" of the whole heart.. extrapolation to clinical variables. twodimensional (20) network or threedimensional (3D) network (slab) model of action potential propagation can be constructed. Both the modeler and ranger need to place fundamental elements in the context of a broader landscape. with the recent genome explosion. This magnificent task is not unlike a forest ranger attempting to document each leaf in a massive forest. By connecting an array of these individual myocytes (via gap junctions). ECG) from the cardiac sources by solving the forward problem of electrophysiology (Chapter 2). the modeler needs to examine the "leaves" at even much greater molecular detail. though vigorous reference must be made to numerous . cardiac modeling is equivalent to solving a system of nonlinear differential (or partial differential) equations. But now. cartilage. cardiac modeling is beginning to span the spectrum from DNA to the ECG. Baltimore MD 21205 INTRODUCTION Since many cardiac pathologies manifest themselves at the cellular and molecular levels. such as the electrocardiogram (ECG). Vivek Iyer. Thus.
a ECG ·SG ·100 f~r. and M. Finally. and networks of cells. each cell. heart and torso anatomy are obtained. modeling efforts have primarily focused on accurately reconstructing normal behavior. V. however. B. the model can be made more complex by adding appropriate differential equations to the system. FIGURE 1. Thakor. See the attached CD for color figure. modelers have also begun to revise and extend the quantitative description of these models to include important abnormal behaviors.2 N. Subsequently. as more information about the cellular networks. But with the accumulating experimental history of cardiac disease (such as myocardial ischemia. Network models investigate the connectivity of onecell units organized in arrays. Until recently. V. which are only solved in the time domain.Shenai Cell Network (lD. the onecell model will be expanded to represent multiple dimensions with the incorporation of partial differential equations in space. Levels of Analysis.1. . the appropriate electrical reconstruction is discussed in the context of relevant pathology to emphasize the usefulness of cardiac modeling. At each level of analysis. Thus. The basic electrophysiological recording is the action potential. 2D) Whole Action Potential ~ Electrog ram ~ . tissue structure. a better reconstruction of the ECG becomes possible. These equations provide a quantitative measure of each channel. many patches molded into the shape of a whole heart (in addition to torso variables) gives rise to the ECG. This chapter will first focus on the theoretical onecell equations. An electrical measure of bulk network activity is the extracellular electrogram. longQT syndrome and heart failure). Onecell models include the study of compartments and ion channels and their interactions. laboratory experiments which aim to determine the nature and coefficients of each equation. Iyer. As more experiments are done and data obtained.
the cardiac myocyte contains a prominent intracellular calcium compartmentthe sarcoplasmic reticulum. by simple rate theory.2) The voltage dependence of these ion channels can be understood if these gates are treated as an "energybarrier" model. the ion channel is a multidomain transmembrane protein with "gates" that open and close at certain transmembrane voltages. reviewed below for the cardiac myocyte. [Co]). and its coupling to muscular contraction. 1. many of these membrane channels still follow the same HodgkinHuxley formalism.F.1) where k 1 and L 1 are the forward and reverse rates of the process. presumably due to a larger diversity of ion channels present in the cardiac myocyte. Luo and Rudy 1991.L. 1975). Hodgkin and A.1 THE ONECELL MODEL The origins of the onecell model actually take root from classical neuroscience work conducted by A. Noble and Tsien (McAllister et al. numerous ion channels and intracellular calcium compartment dynamics have been added (DiFrancesco and Noble 1985. this HodgkinHuxley formalism was applied to model the Purkinje fiber action potential by McCallister. Thus. an energy barrier (LlG o) located at a relative barrier position . Na". However. and n open and nclosed are the percentage of open or closed channels (which is proportional to channel "concentration"). Ca2+) must cross the membrane via the transmembrane ion channel. a process symbolically represented by the following equation: (1. Huxley in 1952 (Hodgkin and Huxley 1952). Given the concentration of the charged particle on the inside and outside ([Cil.n open): (1. it was determined that the cardiac action potential is considerably more complex than the neuronal action potential. 1949. This model mathematically formulated the voltagedependent "gating" characteristics of sodium and potassium ion channels in the nerve membrane.1 VOLTAGE GATING ION CHANNEL KINETICS (HODGKINHUXLEY FORMAUSM) At the most fundamental level of electrophysiology. however. Moore and Pearson 1981). In addition. an ion (K+.Vout ). is to characterize the opening and closing of these gates. one would expect the rate of channel opening (dn/dr) to equal (note that nclosed = 1 . Since similar ion channels exist in cardiac cells. Since then. the intercellular connections. respectively. Luo and Rudy 1994). The problem. they were able to derive a quantitative description for current flow across the cell membrane. With the addition of the "slowinward" calcium current in 1976. Typically. Vm (= Vin .From Cellular Electrophysiology to Electrocardiography 3 1. Nevertheless. and the resulting action potential (AP). In famous experiments conducted on the giant axon of the squid.1. described with Eyring Rate Theory (Eyring et al. making the current AP model considerably more complex and robust. Beeler and Reuter (Beeler and Reuter 1976) were able to successfully describe the ventricular action potential with the characteristic "plateau phase" necessary for proper cardiac contraction.
2) circuit. through which ions pass to create an ionic current (lion). The Hodgkin and Huxley model likens the biological membrane to a BatteryResistorCapacitor (BRC model.. R is the gas constant.. thus acting as a capacitor (Cm ).I n ([Clo) .4 N. The V m at which a certain ion is at equilibrium (lion = 0) is termed the Nemst potential (Eion ). Eyring Rate Theory predicts the forward and reverse rates for ion transfer as: kI =K · ( e _~ ) RT • ( e ( ). )t FVm ) RT "G O k_ 1 = K· (e7/T ) . Figure 1. B. Since the membrane confines a large amount of negativelycharged protein within the cell.zF [Cli (1. V. A model that takes timevariance into account was developed by Hodgkin and Huxley in 1952 (Hodgkin and Huxley 1952). across a resistor (or conductance ). and z is the valence of the ion. ( e ~) RT (1.4) Thus . Finally. Ions flow (current) to and from the extra. While the "energybarrier" model predicts voltagedependence. it separates positively and negatively charged compartments.3) where K is a constant. due to its chargeseparating function. The current relates to a transmembrane voltage. from simple circuit analysi s of Figure 1. intracellular [C] i and extracellular [C] o ion concentrations: _ . the solution in Eq. as ions cross the membrane and enter (or leave) the intracellular compartment. Iyer.. the ionic current for a certain ion can be .2. the "battery" which depends on valence . While. V.3) is an extremely simplified version of reality.RT E ion . 1l1 FIGURE 1. The membrane has an inherent capacitance. and a transmembrane voltage (V m).and intracellular domains. and M. electrical repellant charge begins to build that counteracts Vm . it readily suggests that the forward and reverse rates are voltagedependent (thus these rates can be represented as k. Thakor. Shenai K+ Extracel lular Na+ V . V rn (8) along the transmembrane route . ( 1.2. A BatteryResistorCapacitor model of a generic excitable membrane. ( V) and L) ( V» . it does not account for the timevarying features in opening and closing channels. The resistor (1/conductance) represents the ion channel. T is the absolute temperature.
timevarying ion channel conductance. n (the forward and reverse rates.9) An analogous equation can be written for the inward sodium current with the addition of an inactivation mechanism (Figure 1.= {3(V)[l . which has a particular solution under several boundary conditions. increased membrane voltages stochastically increase the probability that these three gates open.n m .From Cellular Electrophysiology to Electrocardiography 5 written as: (1. (1.7). Thus. (1. respectively) : dn(t. Hodgkin and Huxley assumed that the channel was a "gate" as described in Eq. (1..2). Hodgkin and Huxley were able to define regression equations for noo(V) and rm(V). V) dt.. the outward potassium current can be represented as: dV l« = C . t)4 .8) Using an elegant experimental setup that applied a voltageclamp to a giantsquid axon (Cole 1949. n(V.Thus.n r(Vm) (1.a(V)[n] (1.a(Vm) + {3(Vm) (1. which can be rewritten solely in terms of open probability nopen or simply. n(t) follows an inverted exponential time course with the following characteristics: 1 r(v' ) .E K ) 4 (1. except that the inactivation gate closes with increased voltages (Figure lAc). the sodium response to an applied voltage stimulation is biphasic. (1. To determine the dynamics of an individual ion channel. (1.5). they arrived at the open channel probability of n(V.. t) is the voltagedependent. (1. As with the potassium channel. which represent the gating variables for the potassium channel.6) is a firstorder differential equation. t) (V .6) Eq.6) in terms of the quantities derived in Eq. Marmont 1949). First. (1.7) The quantity of noo(Vm) represents the steadystate proportion of open channels after a step voltage has been applied for a nearinfinite amount of time. gives a differential equation that describes the time course of the open probability for a channel: dn dt noo(Vm) . Following the data fitting. Rewriting Eq.8). Following a voltage step LlV(Vm = Vrest + LlV) from the resting membrane potential. allowing . the faster activation gates rapidly open.5) Where g(V.= dt gK . Inactivation follows the same kinetics as Eq. the experimental sodium channel was represented by Hodgkin and Huxley as three voltageactivated gates similar to the potassium activation gates described by Eq. k1(V) and k1(V) are replaced with a(V) and {3(V).3).n] .8). by substituting the open probability into Eq. The variable roo(Vm) characterizes the time the system takes to reach this noo(Vm). To obtain a suitable fit to experimental data.
. The sodium channel is represented by three rapidlyactivating voltagesensitive gates. The lumped probability that all potassium gates will be open is n". .? oo ee so . A) I1aJ(V) 09 sc Volts (mV) B) Illoo(V) ...K+ Open PrOb3bility Na+ Open Prob 3bility n n n n 111 111 111 h Probability aJJ gates are open [ 114 1 I III" I FIGURE 1.?oo eo se Volts (mV) Volts (mV) FIGURE 1. h.. Cl ~ 0 4' 03 1 02 ' 0 \' .. ~ ~ 07 . m. Idealized ion channels..c. ~ ~ 09 · 0& ~ ~ ~ I::l. . . 0$ · ~! §l I::l. ~ 09 06 0 7· 06 0$ 04 03 02 01 C)~(V) ... (B) activation curve for sodium channel. while the probability that the activation and inactivation gates of the sodium channel is m'h. n.4. The potassium channel is generally modeled with four voltageactivation gates.3.. and (C) inactivation curve for sodium channel.. 06 ' ~ C l... with an additional slowly acting voltagesenstive inactivation gate. Activation curves for (A) potassium channels.
However.1. and Iotherchannels is provided via many other channels that vary among celltypes (atrial vs. This simplified approach assumes that the cell membrane contains two distinct types of voltagegated channels (Na+ and K+) that conducting currents in the opposite direction. With the addition of other inward and outward channels (see later sections). Eq. forcing a decrease in the inward current. the slower inactivation gates will close. ventricular cells) and various excitable tissues (heart vs.10) The biphasic nature of the inward sodium current is crucial to the rapid elicitation of an action potential and the characteristic biphasic shape of the action potential. The sodium current can be represented as: (1. 1. Ventricular myocytes express the proper proteome to parlay the electrical excitation into force generating elements that ultimately produce the cardiac output and blood delivery to the rest of the body.8)). with increased voltage. middlemyocardial cells (Mcells). Note that l«. uniform excitation of the ventricular myocytes. and the history of these modeling developments is described below. and other channels are represented by nonlinear terms (i. Although there are slight differences in the quantitative description of the sodium and potassium channels described above. and are both voltage and timedependent. There is no conceptual change in the nature of the current equationthe activation gate n is simply replaced with m and h (though these gates all differ quantitatively. a generalized differential equation can be written: dt dV 1 = C UK M + INa + Iotherchannels + I stim) (1.e.1. while Purkinje fibers represents an efficient conducting system specialized for the fast.11) where I stim represents a stimulation current (provided from a stimulating lead or adjacent cells). in normal and diseased heart function. the methods were quickly adapted to represent the cardiac action potential. Thus. the cardiac myocyte uniquely expresses a diverse set of ion channelswhich give unique electrophysiological properties to different types of heart tissue. represents a system of nonlinear differential equations that must be solved using techniques of numerical integration.2 MODELING THE CARDIAC ACTION POTENTIAL While the model of an action potential was originally described for a neuron. different models exist for transmural orientation (endocardial cells. Additionally.11) coupled with gating equations for each channel (Eq. (1. Models for each type of these cells have been extensively developed and are described in Table 1. the cardiac myocyte also exhibits a considerable inward calcium current that is responsible for the distinguishable "plateau" phasewhich coincides with the muscular contraction in the ventricular myocyte. and epicardial cells). m and n both increase with more positive Vm» while the value of h decreases with more positive Vm). . Within the heart.From Cellular Electrophysiology to Electrocardiography 7 inward current to develop. INa. (1. there exist a variety of cell types that require different considerations when developing a model. Pacemaker cells in the sinoatrial node express channels that allow an autonomous train of action potentials. nervous system). Even within the ventricle. n4 and m 3 h).
This Is current follows HodgkinHuxley formalism. the timeactivated outward IXl current and IK1. ICaL. However. the Is calcium current.IK2 lSi (slowinward Ica) INaCa.T. V.1. In fact.5). IcaL. a timeindependent . (2000) Novelty Squid Axon Purkinje Cell Ventricular Cell Purkinje Cell Ventricular Cell Ventricular Cell Human Ventricular Cell Sinoatrial Nodal cells INa.7 M. After repolarization of the action potential. the flux of [Ca2+]j can be described by: T d[Ca]· = _10. V. Iyer. McCallister. the McCallister. given the vast diversity of cardiac cell types.7 .12) En toto. a generalized plateau and repolarization current. Cabuffering Updated with human data Updated Ca handling 1. Noble and Tsien introduced a prototype numeric model for the rhythmic "pacemaker activity" of cardiac Purkinje cells by using the voltageclamp method to study an outward potassium current.8 N. I K2 is a prominent current in producing the automaticity of pacemaker cells. IK Updated INaCa. INaK. Additionally. At any given state. the initial low level of intracellular calcium.Thakor. As this slow wave of depolarization brings membrane potential towards threshold.IK Ix!. 1975). Is + . Ica. Noble and Tsien (MNT) model reconstructed the entire action potential. [Ca2+]j does not remain constant with the arrival of the transmembrane Is current. This model incorporates an Is component.7 . Shena! TABLE1. and M. However. a slow inward calcium current that is responsible for the slow depolarization and the prominent plateau phase.1 Classical modelsofthe cardiac actionpotential In 1975. the deactivation of outward I K 2 current allows a net inward current to produce a diastolic slow wave of depolarization in between action potentials (Figure 1. in that state variables d (activation) and f (inactivation) describe timevarying conductances of the slow inward current. widely altering the Nemst potential. using a modified HodgkinHuxley sodium conductance for the rapid upstroke phases.T Updated INa.07(10. Thus.[Cali) (1. Classical and Modern Models of Various Cardiac Cell Types Type Classical Models HodgkinHuxley (1952) McCallister. Ica.1. E s. Tsien (1974) BeelerReuter (1977) Modem Models DiFrancescoNoble (1985) LuoRudy Phase I (1991) LuoRudy Phase II (1994) PriebeBeuckelmann (1998) Zhang et al. while using voltageclamp methods to describe an lXI. I K2 (McAllister et al. the prominent plateau phase that is crucial for forceful contraction. this landmark model was able to simultaneously describe characteristic pacemaker activity and rapid conduction velocities associated with Purkinje cells. To this end.2. Beeler and Reuter modeled the intracellular handling of calcium by assuming it flows into the cell and accumulates while being exponentially reduced by an uptake mechanism (in the sarcoplasmic reticulum). the model incorporated four major components: the familiar INa current. unlike other HodgkinHuxley ions. Thus. B. the MNT model could not describe the characteristics of ventricular action potentialsnamely. Noble. Beeler and Reuter developed a numerical model (the BR model) for the ventricular myocyte in 1977 (Beeler and Reuter 1976). the range of [Ca2+]i can range from 1 to 10. INaK.
. .From Cellular Electrophysiology to Electrocardiography 9 McCallister... (OM<') JOO I x . outward potassium current. . au .{. • __ . Id ..t\ J ~<:: .. Noble._:• .1. 1. 8 ( 0 __ 'v..'l /':~ / __' I . I CI "'" FIGURE 1. _ _ .. _ F ._ I.. .1 JSR (CSQ I . o r ". Ts ien (Purkinje Fiber) Beeler and Reu ter (Ventricular Fiber) . .d """ Ia JS:!l "~TlJlNI ~. I • ._ .= ..t.. . A comparison of classical (top) and modern models (bottom).. . .]~CMON) . E f . L 1'41 le. . including determinants of action potential duration.... and oscilliatory behavior in ventricular cells.. I.. With this model. ' . Beeler and Reuter began to predict pathological phenomena.._ II:': ' ''  t ~ .. .I~I I! :' .~ ....2.. a c ~~ I'Ll " lII)'opIMm ~ " "p.. < . • . ... u  . _ _ . ~ •• ..CJ~ ~.. J ~otote I~ I~LCa . .5.' :f'i!..2 Modern models of cardiac action potentials While modem models utilize many of the concepts introduced in the classical models described above. . current models now incorporate a larger repertoire of ion channels.1 t. ':.... <..c •..' ik=§.9) Luo Rudy (membrane) LuoRudy (sarcoplasmic reticulum) A.. ..
Iyer. In addition. Recently. DiFrancesco and Noble described an improved model of the Purkinje action potential (DN model) (DiFrancesco and Noble 1985). and Ca H induced Ca H release.Thakor. Luo and Rudy published an updated version of the DN model that included more recent experimental data for the sodium and potassium currents. Processes not described in the DN model were also added. In 1991. But in 1994. This enhanced model has provided a breakthrough in simulations of excitationcontraction (EC) coupling and reentrant mechanisms of arrhythmogenesis. 1. along with improved assumptions on calcium channels (Ltype and TType) and intracellular calcium handling. the experimental recording technique at the time was rather limited. myocardial ischemia does so by shortening the action potential duration (APD) while LQTS induces arrhythmias by lengthening the APD. In 1985. but also to electrocardiographic detection and appropriate pharmaceutical intervention. While the LuoRudy model describes ventricular action potentials. and could not account for important arrhythmogenic phenomena. However. network sarcoplasmic reticulum. there is a comprehensive understanding of basic ionic mechanisms and their behavior in normal cardiac cells. Zhang et at. such as infarction and/or arrhythmogenesis. improved computational power and numerical techniques can solve hefty systems of differential equations. B.10 N. the sarcolemma Na+ ICaH exchanger. V. Nevertheless. have used existing biophysical data to simulate a family of action potentials recorded in rabbit atria (Lindblad et at. Luo and Rudy published an updated model which comprehensively updated the DN description of the sarcolemma Ltype Calcium channel (lea. the LuoRudy model was updated by substituting animal data in favor of recent human data (Priebe and Beuckelmann 1998). such as the buffering of Ca H in the myoplasm. 1996). have incorporated recent sinoatrial data to formulate a modem model of various sinoatrial nodal cells (central nodal and peripheral nodal cells) (Zhang et at.3 MODELING PATHOLOGIC ACTION POTENTIALS Currently. the sarcoplasmic CaATPase. V. Among many others. are now the focus of intense modeling research. that included the traditional ion channel formulation.d.1. The model consists of three compartmentsthe myoplasm. In 1998. and complex intracellular and sarcoplasmic calcium handling. The various cardiac models listed in Table 1. These studies have contributed not only to a theoretical understanding of the diseases. Though both myocardial ischemia and longQ'I' syndromes can lead to fatal arrhythmias (Wit and Janse 1993. and M. but omitted the BR formation of the inward calcium current (lsi). less is accepted about how impairments of these ionic mechanisms ultimately predict or provoke gross events. ElSherif et al. and the junctional sarcoplasmic reticulum. 1996). . two areas of cardiac pathology. 2000). citing a lack of singlechannel and onecell experimental history (Luo and Rudy 1991). myocardial ischemia and longQ'I' syndromes (LQTS). allowing a more precise description of cellular electrophysiology (onecell) and the interaction of many cells (network models).1 have widely contributed to this theoretical understanding. As a result. the focus of modeling has shifted from describing normal behavior of myocytes to describing pathological phenomena. and a nonspecific Calcium current (Luo and Rudy 1994). Shenai a richer history of experimentation. several other models exist for other cardiac tissues. Lindblad et at. the sarcolemma Na/K pump.
6 summarizes the various cellular phenomena associated with myocyte ion channels. Acidosis 2. Alterations in intracellular ATP ([ATP]i). Rotors 1. These modifications may be achieved by: (1) adding novel channels to the existing repertoire of known membrane channels. LQT3 3.1. can alter the activity of membrane pumps. one must modify existing models of normal behavior. Acidosis 2. Various ion channels have been implicated in pathologic phenomena.1 Myocardial ischemia Myocardial ischemia results from a withdrawal of oxygen from myocardial tissue (due to inefficient or absent perfusion). 1. resulting in disturbances to aerobic respiration and ATP production. Figure 1.3. DADs FIGURE 1. Plateau EADs 1. and thus the distribution of critical ions (Na+ and K+) that are largely responsible for the electrophysiological characteristics of myocardium and proper action potential propagation. Thus. Phase3 EADs 2. DADs 1. Cellular phenomena associated with myocyte ion channel currents. ischemia develops at the cellular level. To study impaired cells.6. when the amount of oxygen (Poz) in . or (3) a combination of new channels and altered channel dynamics.From Cellular Electrophysiology to Electrocardiography 11 1. Phase3 EADs 2. (2) altering the quantitative dynamics of known channelsfor example by altering ionic concentrations or pH.
Shaw and Rudy 1997).Iyer. Weiss et al. Morena et al. 1986. Thus. Thus. First. while large increases in extracellular potassium begin to inactivate the sodium current and decrease upstroke velocity. or an increase in extracellular potassium [K+]o. a lethargic pump performs this pro