You are on page 1of 13

Annals of Biomedical Engineering, Vol. 29, pp. 47–59, 2001 0090-6964/2001/29共1兲/47/13/$15.

00
Printed in the USA. All rights reserved. Copyright © 2001 Biomedical Engineering Society

An Experimentalist’s Approach to Accurate Localization


of Phase Singularities during Reentry
ANAND N. IYER and RICHARD A. GRAY
Department of Biomedical Engineering, University of Alabama at Birmingham, 1670 University Blvd.,
Volker Hall B 140, Birmingham, AL

(Received 30 December 1999; accepted 27 October 2000)

Abstract—A phase variable that uniquely represents the time the ‘‘state’’ of a cell at any instant of time.16 Phase maps
course of the action potential has been used to study the 关 ␪ (x,y,t) 兴 are obtained by calculating phase at each
mechanisms of cardiac fibrillation. A spatial phase singularity
共PS兲 occurs during reentrant wave propagation and represents
instant of time for each recorded site on the heart. A
the organizing center of the rotating wave. Here, we present an unique spatial pattern exists near the tip of rotating
error analysis to investigate how well PSs can be localized. waves. The phase at the exact center of rotation is un-
Computer simulations of rotating spiral waves scaled appropri- defined, but neighboring sites exhibit a continuous pro-
ately for cardiac tissue were studied with various levels of gression of phase equal to 2␲ around the center. In the
noise added. The accuracy in identifying and localizing singu-
larities depended on three factors: 共i兲 the point chosen as the
phase maps, this center corresponds to a phase singular-
origin in state space used to calculate the phase variable; 共ii兲 ity 共PS兲. PSs are a necessary 共but not sufficient兲 condi-
signal to noise ratio; and 共iii兲 discretization 共number of levels tion for functional reentry in the heart and thus play a
used to represent data兲. We found that for both simulation as significant role in sustaining cardiac arrhythmias.16,25 A
well as experimental data, there existed a wide range for the better understanding of phase and PSs may provide valu-
choice of origin for which PSs could be identified. Discretiza-
tion coupled with noise affected this range adversely. However,
able insight into the mechanisms of cardiac fibrillation.
there always existed a range for choice of the origin that was However, in order to analyze the phase variable appro-
20% or more of the action potential amplitude within which the priately during cardiac arrhythmias, it is necessary to
accuracy of localizing PSs was better than 2 mm. Thus, a provide a formal analysis of the phase mapping tech-
precise determination of origin was not necessary for accurately nique. In addition, it is essential to quantify the errors
identifying PSs. © 2001 Biomedical Engineering Society.
关DOI: 10.1114/1.1335538兴
associated with the identification of PSs including the
effect of signal noise.
Keywords—Spiral wave, Computer simulations, Two state Here, we test the ‘‘robustness’’ of the accuracy in
variable model, BRDR, State space, Optical mapping. localizing PSs using automated algorithms to calculate
phase and identify PSs. How the choice of origin for the
INTRODUCTION phase transformation, the effect of noise, and the level of
discretization affect the ability to localize PSs are stud-
Sudden cardiac death is the major cause of fatalities ied. The algorithms are first applied to computer simu-
in the industrialized world.20 It is believed that ventricu- lations of stable rotating spiral waves using two state
lar fibrillation in the form of reentrant spiral waves is the variable models with parameters chosen such that the
underlying cause of these deaths.15 Typically, spiral results are appropriately scaled for cardiac tissue. The
waves in the whole heart are unstable, causing and sus- algorithms are then tested on stationary and nonstation-
taining fibrillation.16,17 In the past, the phenomenon of ary spiral waves generated using ionic models, as well as
fibrillation has been considered the result of primarily data from rabbit hearts in tachyarrhythmias, recorded us-
random events. Only recently, investigators have been ing current video-imaging technology.
able to show evidence that organization exists during
fibrillation.12,16,26 Still, the mechanisms underlying fibril- THEORETICAL FORMULATION
lation remain poorly understood. One approach to under- Models of Cardiac Tissue
stand the mechanisms of cardiac fibrillation is the use of
a cardiac ‘‘phase’’ variable, ␪, which uniquely defines Complex ionic models6,10,19 and simpler two-state
variable models1,4,11,22 共TSVMs兲 of the cardiac cell have
Address correspondence to Anand N. lyer, 1670 University Blvd.,
been studied over the years by mathematicians and elec-
Volker Hall B 140, Birmingham, AL 35294-0019. Electronic mail: trophysiologists to provide a better understanding of the
ani@crml.uab.edu electrical activity in the heart. Each of these models
47
48 A. N. IYER and R. A. GRAY

TABLE 1. Values of parameters used for the BAR model


computer simulations.

Parameter Value

a0 ⫺0.84
a1 0.15
a2 0
a3 0
b0 1
b1 ⫺6.75
b2 13.5
b3 ⫺6.75
␶e 1
␶g 33

reproduces some 共but not all兲 features of cardiac cells.


TSVM dynamics are entirely described using two state
variables and thus allow a formal analysis of phase
共which is computed using two variables兲 using the
‘‘true’’ state space. Using TSVMs is also important in
order to form a link between calculating phase from the
true and ‘‘reconstructed’’ state spaces 共described later兲.
TSVMs consist of two differential equations that de-
scribe the dynamics of a fast variable 共e兲, analogous to
the transmembrane potential (V m ) and a slow variable FIGURE 1. Representation of one action potential cycle. „a…
共g兲, representing recovery. We use a generic representa- State space representation of the action potential „solid line…
tion of TSVMs. The equations for a single cell are as using two variables, e and g. The dotted lines are the
nullclines for the FitzHugh model „a TSVM…. The intersection
follows: of the nullclines is called the fixed point. In this case, the
fixed point is stable and is equal to the resting potential.
de/dt⫽ 关 F 共 e 兲 ⫺g 兴 / ␶ e , Phase, ␪, is measured with respect to an origin „ e * , g * … cho-
sen in the state space. Inset shows the fast variable, analo-
共1兲 gous to transmembrane potential, which is used to represent
dg/dt⫽ 关 H 共 e 兲 ⫺g 兴 / ␶ g , the action potential cycle. „b… The true state space „nullclines
denoted by dotted lines… for the BAR model, which has three
fixed points „denoted by diamonds…, only one of which is
where F(e)⫽a 0 ⫹a 1 e⫹a 2 e 2 ⫹a 3 e 3 and H(e)⫽b 0 stable. Inset shows an action potential obtained using this
model.
⫹b 1 e⫹b 2 e 2 ⫹b 3 e 3 , ␶ e and ␶ g are the time constants for
the fast and slow variables, respectively.
In the present study, we chose a TSVM by Bär and
True State Space and Phase
Eisworth 共BAR model兲4 and modified the equations to fit
Eq. 共1兲. The parameter values were scaled appropriately Transmembrane potential or e alone is not a unique
共see Table 1兲 for cardiac tissue. This was done by choos- measure of the state of the heart. For example, during
ing the diffusion coefficient 共0.6 cm2/s兲 to match that of one action potential, the transmembrane potential has a
cardiac tissue and ensuring that the rotation period of the value of ⫺20 mV twice, during both depolarization and
spiral wave 共150 ms兲 was within the range observed in repolarization. However, using two variables to represent
the heart. The time constant of the fast variable was set the time course in the true state space diagram 共see Fig.
to 1 ms so that the conduction speed was appropriate and 1兲 can eliminate this duality; each point on the trajectory
the full range of the action potential amplitude was has a unique value in state space. In order to utilize a
scaled to range from ⫺80 to 20 mV. single variable that can uniquely define the phase of each
Detailed mathematical models of the ionic currents in action potential, we use a cardiac phase variable, ␪,
cardiac tissue have been developed. In order to ensure which is calculated from two variables. A detailed de-
that the method of calculating phase and localizing PSs scription of the calculation of phase is given later.
can be applied to cardiac data, we will analyze reentrant In the past, phase has most often been defined as the
waves generated with an ionic model. In this paper, we time elapsed since the last activation or event.25 During
use the Drouhard–Roberge modification9 of the Beeler– cardiac fibrillation, however, all sites do not undergo
Reuter model6 共BRDR model兲 to generate stationary as ‘‘full-blown’’ action potentials, which makes proper
well as nonstationary spiral waves. picking of activation times very difficult. In this work,
ACCURATE LOCALIZATION OF PHASE SINGULARITIES 49

we use a definition for phase that does not depend on some trajectories and outside others. If any trajectory
activation time, but instead uses two variables to calcu- does not traverse around the origin, then the phase for
late phase.16 In the case of TSVMs, the two true vari- that site does not change by 2␲ during one rotation.
ables, e and g, can be used to calculate phase for any site Therefore, for stable reentry, the ideal point for comput-
at any instant of time. To calculate phase, we choose a ing phase should be the one such that all trajectories
point in state space as the origin, denoted by (e * ,g * ), traverse around it exactly once during each rotation, i.e.,
and make a transformation to a polar coordinate system * ,g ideal
(e ideal * ). We approximated this ideal state by choos-
in true state space. Phase is calculated as follows: ing a state within the smallest trajectory 共i.e., smallest
amplitude oscillations兲 during stable reentry. For a stable
␪ 共 t 兲 ⫽arctan 2 关 g 共 t 兲 ⫺g * , e 共 t 兲 ⫺e * 兴 . 共2兲 rotating spiral in a TSVM, the state for each site is
defined by a closed loop trajectory in the true state space
The arctan 2 function returns a value of phase between diagram during one complete rotation and Fig. 5共a兲
⫺␲ and ␲ depending upon the value and signs 共⫹/⫺兲 of shows trajectories plotted on state space from every site
the two input variables. of the array during stable reentry for the BAR model.
In order to study the TSVM in two dimensions, we * ,g ideal
The ideal origin (e ideal * ) is within the smallest tra-
incorporate a spatial diffusion term to account for the jectory, denoted by an arrow.
coupling between cells.21 In the heart, only the fast vari-
able diffuses. Thus, the diffusion term, Dⵜ 2 e 共where D Calculation of Phase using One Variable
is the diffusion coefficient in cm2/s and ⵜ 2 is the La-
For numerical simulations of TSVMs, we can calcu-
placian operator兲, is added to the equation for the fast
late phase directly from the two state variables, e and g.
variable in Eq. 共1兲. This gives us the following equation
However, for experimental data, typically only one vari-
for the two-dimensional case:
able is recorded. Accordingly, the calculation of phase
and the identification of PSs have been carried out using
⳵ e/ ⳵ t⫽ 关 F 共 e 兲 ⫺g 兴 / ␶ e ⫹Dⵜ 2 e, one variable.16 This is accomplished using a ‘‘recon-
共3兲 structed’’ state space, that is topologically equivalent to
⳵ g/ ⳵ t⫽ 关 H 共 e 兲 ⫺g 兴 / ␶ g . the true state space.23 The reconstructed variables can be
obtained for the original time series e(t) as e(t),
For a perfectly stable spiral wave 共i.e., rotational sym- ⳵ e(t)/ ⳵ (t), ⳵ 2 e(t)/ ⳵ t 2 ... or e(t), e(t⫹ ␶ ), e(t⫹2 ␶ ),...
metry兲 the time derivatives of both e and g will be zero Here, we use a time delay-embedding technique to cal-
at the exact center of rotation. By setting the left-hand culate phase from a two-dimensional reconstructed state
side of Eq. 共3兲 to zero, one can solve for the values of e space. Phase can be calculated as follows:
* ,g ideal
and g at the center (e ideal * ); these values will be
different than the fixed point for a single cell due to the ␪ 共 t 兲 ⫽arctan 2 关 e 共 t⫹ ␶ 兲 ⫺e * ,e 共 t 兲 ⫺e * 兴 . 共4兲
diffusion term. We denote these values as ‘‘ideal’’ be-
cause this is the state we wish to use as the origin for the It can easily be seen that the ideal origin in the recon-
transformation to polar coordinates and calculating * ,e ideal
structed state space will then be (e ideal * ).
phase.
Each site on the heart 共physical space兲 has a value of
Time Delay-Embedding
e 共activation兲 and g 共recovery兲 at each instant of time.
Over time, the values of e and g for each site go through The reconstructed state space is obtained by plotting
a periodic cycle. A single trajectory in state space is e(t) against a recreated variable, in this case, e(t⫹ ␶ ).
obtained by plotting the ‘‘e’’ and ‘‘g’’ values for each The choice of time-delay 共␶兲 should be such that the
site/pixel through time. Figure 5共a兲 was obtained by plot- phase trajectories in state space uniquely define the time
ting several such trajectories 共from several different sites/ course of the action potential. One method is to pick ␶
pixels兲. Any point in this state space can be used as the such that the autocorrelation between e(t) and e(t⫹ ␶ ) is
origin (e * ,g * ) to calculate the instantaneous phase ac- zero, ensuring that the variables are linearly independent.
cording to Eq. 共2兲. But, we cannot pick just any point if In our analysis for the BAR model and for the experi-
we want an accurate phase map. Why? For stable reen- mental data, we used this method to choose ␶. However,
try, the phase for each site should progress through ex- Gray and Jalife have shown14 that if the value of ␶ is
actly 2␲ during each complete rotation, in order to define greater than the upstroke of the action potential, the
phase uniquely. The action potential amplitude of sites phase loops cross over in state space. For the BRDR
near the center of rotation in a spiral wave is smaller model, using the ‘‘zero autocorrelation method’’ resulted
than those in the periphery. If we pick a random point in in a value of ␶ ( ␶ ⫽14 ms) that was larger than the up-
the state space as the origin, this origin may lie within stroke which resulted in phase loops crossing over as
50 A. N. IYER and R. A. GRAY

shown in Fig. 2共a兲. When this occurs, phase cannot be


defined uniquely. Figure 2共b兲 shows the phase loop for
the same data for ␶ ⫽2. Thus, we chose a value of ␶ that
was equal to the upstroke 共␶ ⫽2 ms for the BRDR
model兲.

Spatial Phase Singularities


A phase map shows the instantaneous value of phase
at each site of the spiral wave. Phase maps 关 ␪ (x,y,t) 兴
can be obtained by using Eqs. 共2兲 or 共4兲 to calculate
phase at every site 共x,y兲 of the spiral wave for each
instant of time. A snapshot of the spiral wave generated
using the BAR model and its corresponding phase map
are shown in Fig. 3; phase was computed using the ideal
origin from reconstructed state space.
A spatial PS occurs at a site where all phase values
converge. Specifically, a PS occurs when the line integral
of the change of phase around a site equals ⫾2␲ where
the sign indicates chirality. The mathematical definition
of a PS is

冖 “ ␪"dr⫽⫾2 ␲ . 共5兲

This definition of the PS is used in an automated algo-


rithm that identifies the locations of PSs from a phase
map. A schematic representation of the algorithm for
FIGURE 2. Choice of ␶. „a… For a value of ␶ Ä14 ms, three
identifying a PS is given in Fig. 4. The PS-finding algo- phase loops, drawn for the BRDR model, crossover in recon-
rithm evaluates the difference in phase over a discrete structed state space. When this occurs, phase does not
path. If the phase difference is equal to ⫾2␲, the path uniquely define every portion of the action potential cycle.
v „ t … denotes time course of transmembrane potential and
encloses a PS. The path length can be varied, however,
v „ t ¿ ␶ … denotes the recreated variable. „b… For a value of ␶
smaller path lengths are desirable for precision in local- Ä2 ms, which is equal to the upstroke time, phase trajecto-
izing PSs. Here, we used a path length of eight pixels, ries for the same data are full loops. In this case, phase is
which will identify 4 pixels that satisfy the condition for unique for every portion of the action potential cycle.
a PS 关Eq. 共5兲兴 if one exists, and the PS lies at the
BRDR model, the time constants of the d and f gates
common junction of these four pixels 共‘‘䊉’’ in Fig. 4兲.
were reduced compared to their original values to obtain
stable 共50% reduction兲 and unstable 共67% reduction兲 spi-
METHODS ral waves as done previously.8 Spiral waves were gener-
Model Simulations ated by imposing a linear gradient of two variables 共e
and g for the BAR model, and V m and h for the BRDR
The algorithms and software were implemented using model兲 on the x and y axes, respectively,18 which al-
PV-Wave version 6.10 共Visual Numerics兲 software. A lowed the system to evolve into a spiral wave. In all
100⫻100-element array was used for the TSVM simu- cases, the spiral waves were allowed to go through a few
lations. A 200⫻200-element array was used to simulate rotations before performing analyses, to ensure stability.
spiral waves using the BRDR model, but the data was Henceforth, the term e * will represent the origin cho-
spatially averaged to result in a 100⫻100 array. A spa- sen using the BAR model. The comparable origins for
tial discretization 共dx兲 of 0.25 mm and a temporal dis- the BRDR model and experimental data will be denoted
cretization 共dt兲 of 0.01 ms were used for all simulations by v * and f * , respectively. These terms can be used
and the data were sampled every 1 ms. The diffusion interchangeably in Eq. 共4兲.
coefficient 共D兲 for all models was 0.6 cm2/s. No-flux
boundary conditions were imposed. For the BAR model Methods of Analysis
parameters were chosen to ensure stable rotation 共values
are given in Table 1兲; the rotation period was 152 ms Method of Calculating the Actual Locations of PSs. The
similar to that observed in cardiac tissue.2,17,22 For the method of calculating the actual locations of the PSs
ACCURATE LOCALIZATION OF PHASE SINGULARITIES 51

FIGURE 3. Snapshot from a spiral wave simulation and the corresponding phase map. „a… Snapshot from a spiral wave obtained
using the BAR model. Color bar indicates values in millivolts. „b… Snapshot from a phase map for the same instant of time as
in „a…. Phase is calculated from reconstructed state space. Color bar indicates values of phase in radians.

described later uses only one variable, e(t) 共or trans- actual location of the PS was calculated and plotted
membrane potential兲; the second variable was obtained graphically as a function of the chosen e * . However,
using the time delay technique discussed before. We beyond a certain range of e * , PSs were not identified by
wanted to use the same method for all calculations so the algorithm.
that we could compare results quantitatively. Phase for
every pixel was calculated using Eq. 共4兲 关replacing e * The Effective Range. We consider the PS-finding algo-
* in Eq. 共4兲兴 with e pixel
with e pixel * equal to the mean am- rithm to have failed if it did not identify any PSs for
plitude of that pixel over approximately 500 ms. Next, more than two of the eight instants in one rotation. For
phase maps were generated and the PS-finding algorithm all models there existed a range of e * beyond which the
was used to determine the positions of the PS for eight PS-finding algorithm failed. We call this range the ‘‘ef-
intervals of one complete rotation. These are the actual fective range.’’
positions of the PSs during the rotation, denoted by
(x 1 ,y 1 ),(x 2 ,y 2 ),...,(x 8 ,y 8 ). For the BAR model, this Addition of Noise. Levels of noise in experimental data
method of identifying PSs was identical to that using depend on several factors. Cinelli et al.7 have found that
e * ⫽e ideal
* 共see method described under ‘‘Calculation of
Phase using One Variable’’兲. For the nonstationary spiral
wave, another step was added to the earlier method in
order to locate a single pixel as the organizing center of
the spiral wave for each rotation. One pixel was identi-
fied as the average location 关 (x 1 ⫹x 2 ⫹....⫹x 8 )/8,(y 1
⫹y 2 ⫹....⫹y 8 )/8兴 of these eight positions and the mean
amplitude of this pixel for one rotation was obtained.
This mean value was used as the reference origin
*
(e reference) to calculate phase maps for the entire record-
ing area and then the ‘‘actual’’ positions of the PSs were
located.

Error Due to Choice of Origin. In order to ascertain the FIGURE 4. Schematic representation of the algorithm that
identifies a PS. „a… Using a path length of eight pixels, the
sensitivity of accurately localizing PSs on the choice of line integral or change in phase around a site „along the
origin, the choice of the origin (e * ) was varied in incre- arrows… is computed, and four pixels „stars… are identified by
ments of 5 mV. This is equivalent to introducing an error the PS-finding algorithm. The PS exists at the intersection of
these four pixels. The diagram denotes the algorithm applied
in the choice of origin which is equal to the difference to the center pixel. This path length of eight pixels is applied
between the chosen e * and e ideal* 共or e reference
* for the to each pixel in the data array to search for a PS. „b… The
nonstationary case兲. Phase maps were recalculated and change in phase between two adjacent points with values
‘‘a’’ and ‘‘b’’ is calculated according to the flowchart. This
PSs were computed for the same eight intervals. The algorithm ensures that the value of the result is within the
average distance of the location of these PSs from the range À␲ to ␲.
52 A. N. IYER and R. A. GRAY

for acquisition of voltage-sensitive dye fluorescence sig- and the effective range was not affected by SNR. For the
nals using video imaging techniques, the ‘‘distribution of discretized data, however, the addition of noise caused a
noise across the pixels of a frame was Gaussian, indicat- decrease in the effective range. For increasing noise lev-
ing that pixel variability was randomly distributed.’’ In els, the error in localizing PSs increased more abruptly
order to study the effect of noise on localizing PSs in with error in choice of origin than the float data. Using
model simulations, a normally distributed random data discretized data, for 20% added noise (SNR⫽5), the
set, n(x,y,t), was used to simulate experimental noise. effective range reduced by 35 mV as compared to no
The mean value was zero and the standard deviation was noise. For SNR⫽10, the effective range was ⫺55 to
set to different levels to simulate various signal to noise ⫺20 mV.
ratios 共SNRs兲. This noise was added to the fast variable
data as follows: BRDR Model (Stable and Unstable Spirals). The results
of the error analysis for the BRDR model were similar
e n 共 x,y,t 兲 ⫽e 共 x,y,t 兲 ⫹n 共 x,y,t 兲 /SNR. 共6兲 for both the stationary and nonstationary cases as seen in
Figs. 7共b兲 and 7共e兲. The trajectories in reconstructed state
We define SNR as SNR⫽1/␣ , where ␣ is fraction of space for both cases are shown in Figs. 7共a兲 and 7共c兲.
action potential amplitude. The resulting data, e n (x,y,t), The trajectory of the PS for the nonstationary case is
was then filtered using a spatio-temporal boxcar filter of plotted in Fig. 7共d兲. Using float data type, noise had little
width five. Localization of PSs was done for SNRs of 5, effect on the ability to localize PSs and there was a wide
10, 20, and ⬁ 共no noise兲. effective range over which PSs could be identified 共over
75% of action potential amplitude for the stationary and
Discretization. The precision of experimental data is lim- 85% for the nonstationary case, with an SNR of 10兲.
ited by the number of levels used for the analog-to- However, discretization had a profound effect on the
digital 共A/D兲 conversion. To study if discretization itself ability to localize PSs 关Fig. 7共f兲兴 as seen in the case of
had an effect on the accuracy of localizing PSs, all of the BAR model simulations. The effective range for dis-
our model simulation data, except for the BRDR station- cretized data, obtained from the nonstationary spiral
ary case, were discretized to 4096 levels 共12 bits兲. wave using BRDR model, was 40 mV for SNR⫽10.
Analyses were performed on the original and discretized
data. Experimental Data. Experimental data was obtained from
four rabbit hearts mapped using video imaging technol-
ogy 共for details see Ref. 3兲. Data was sampled at 480 Hz
RESULTS
using a charge coupled device 共CCD兲 camera with a
Error Analysis spatial resolution of 0.31⫾0.03 mm. We chose only
those episodes of tachyarrhythmias that were terminated
Phase Calculated from True State Space Versus successfully by application of an electric shock to ensure
Reconstructed State Space. The effect of the choice of proper normalization. The data were passed through a
origin in localizing PSs was studied for the BAR model five-point median temporal filter followed by a spa-
by computing phase using true state space 关i.e., two tiotemporal conical filter (3⫻3⫻3 mask). The SNR for
variables e and g; i.e., Eq. 共2兲 with e * ⫽e ideal* and g * our experimental data was approximately 9.0 for the raw
⫽g ideal兴 and reconstructed state space 关i.e., one variable
* data. Figures 8共a兲 and 8共b兲 shows a snapshot from the
e; Eq. 共4兲兴. The trajectories for reconstructed state space experimental data and the corresponding phase map at
for the BAR model are shown in Fig. 5共b兲. For this the same instant of time. The effect of varying the origin
* was ⫺35.5 mV. The results are shown in
model, e ideal for the four sets of experimental data are summed to-
Fig. 5共c兲. It can be seen that there is very little difference gether in Fig. 8共c兲. The effective range for all four ani-
between the two plots. The effective range is also the mals was very similar 共approximately 40% of the action
same for both cases. This suggests that using a recon- *
potential amplitude兲, while the f reference varied from
structed variable is an effective method to obtain phase ⫺56.3 to ⫺42.9 mV.
data when only one variable is available.
Comparison with Other Algorithms
BAR Model (Stable Spiral). The results for the BAR As the method of calculating phase and identifying
model for float data type 共before discretization兲 and in- PSs is relatively new, it is necessary to compare the
teger data type 共after discretization兲 are shown in Figs. results with previous methods of locating the tip of re-
6共a兲 and 6共b兲. For float data, there was very little effect entrant waves. In order to do this, we calculated isoch-
of increase in noise level 共decreased SNR兲 on the ability rones for the stationary cases using the BAR and BRDR
to localize PSs. An increase in error in picking the origin models, and isopotential maps and for the BRDR non-
led to a gradual increase in the error in localizing PSs stationary case. The isochrones were computed from
ACCURATE LOCALIZATION OF PHASE SINGULARITIES 53

FIGURE 5. One variable versus two variables for the BAR model. „a… Trajectories from every element in the array for stable,
stationary spiral waves. Trajectories are plotted onto true state space for the BAR model. Note that there is a central area
„indicated by arrow…, around which all trajectories traverse. „b… Trajectories in reconstructed state space. Trajectories from every
element in the array are plotted onto recreated state space for a stable spiral wave simulation using the BAR model. Note that
there exists a central area, around which all trajectories traverse. „c… Results from the BAR model simulations, comparing the
use of only one variable to identify PSs against the ‘‘gold’’ standard method of using two variables. Note that the results for
both cases are very similar in PS error as well as effective range.

maps of activation times 共defined as the time of crossing and their exact location corresponds to where the local
50% of the height of the action potential兲.3 The PSs curvature of the isopotential changes sign. Thus it ap-
identified for each model are plotted onto the resulting pears that our method of calculating PSs is consistent
graphs 共Fig. 9兲. From Fig. 9共a兲, one can see that for the with other methods of identifying the tip of the spiral
BAR model the PS is located at the exact center of wave.
rotation as indicated by the isochrones. For the stationary
spiral wave with the BRDR model, there is a central DISCUSSION
‘‘core’’ area 关Fig. 9共b兲兴 that does not activate during the
rotation and the PS was exactly in the center of this Recently developed video imaging technology has
inactivated area. For the BRDR nonstationary case, a made it possible to record the transmembrane potential
series of isopotential maps are shown instead of iso- from thousands of sites over the whole heart with high
chrone maps due to the limitation of isochrone maps in spatial and temporal resolution.5 This provides new av-
showing subsequent positions of the wave front. Isopo- enues of research to explore. The method of using a
tential maps for eight equal intervals of one rotation and phase variable to study cardiac activity is a direct off-
the corresponding PSs are plotted on the maps in Fig. spring of the advancements in video imaging and a rela-
9共c兲. The PSs are very close to these isopotential lines tively new concept. This ‘‘phase mapping’’ technique is
54 A. N. IYER and R. A. GRAY

FIGURE 6. BAR model results. Results


from the BAR model simulations of stable
spiral waves, showing the effects of
choice of origin, discretization and noise
on the accuracy in localizing phase sin-
gularities. ‘‘n’’ indicates noise level in
percentage of action potential amplitude.
SNRÄ100Õn . „a… BAR model with float
data type. „b… BAR model for discretized
data. Comparing panels „a… and „b… shows
the effect of discretization.

based on certain fundamental properties of ‘‘excitable’’ while Ikeda et al.17 have calculated the area of the core
cells and nonlinear wave propagation.24 Here, we show to be 12 mm2 in canine atria.
that the localization of PSs is the same whether phase is
computed from the true state space or a reconstructed
one. This result ensures that analyzing PSs computed
from the dynamical patterns of transmembrane potential Method of Calculating the Cardiac Phase Variable
共only兲 is sound. In addition, we show that although noise
level, discretization, and choice of origin affect the abil- Previously, the core of a rotating spiral wave has been
ity to localize PSs, they can be localized with an accu- calculated by identifying the region of low amplitude
racy of 2 mm over a wide range of these parameters variation of potential during one complete rotation.17,22
even for unstable spiral waves. The method of choosing an ideal origin for the BAR
Perhaps the most applicable result for experimentalists model used in this paper is similar to previous work by
is that, for data recorded with a video imaging system Winfree24 who plotted trajectories in true state space to
with SNR ⬃10, there exists a range of over 20 mV, track the tip of rotating waves in TSVMs. However, this
which is over 20% of action potential amplitude, for method 共using true state space兲 cannot be extrapolated to
picking the origin such that the error in localizing PSs is experimental data where only one variable is recorded.
less than 2 mm. For example, if one were to pick an In this work we calculate a second reconstructed variable
origin of ⫺50 mV 共for example兲 for all four data sets, using the time-delay embedding method. An advantage
PSs could still be identified with an accuracy of 2 mm. of our method 共also shared by Winfree’s technique, but
The magnitude of this accuracy is within the range of not by Pertsov’s22 method兲 is that PSs can be identified
values reported for the central core area during reentry in within a very short time interval 共as little as two video
cardiac tissue. Pertsov et al.22 calculated the core size to frames兲. This enables a description of rotating waves on
be 17 mm2 for canine and sheep ventricular myocardium, a time scale much shorter than the period of rotation.
ACCURATE LOCALIZATION OF PHASE SINGULARITIES 55

FIGURE 7. BRDR model results. „a… Trajectories plotted onto reconstructed state space for the stationary BRDR model. „b… Error
analysis for the stationary spiral wave showing the effect of choice of origin and noise on the ability to localize phase
singularities using float data type. „c… Trajectories plotted onto reconstructed state space for the nonstationary BRDR model. „d…
The movements of the PS of the nonstationary spiral wave over 450 ms for the BRDR model. „e… Results of the error analysis
for the nonstationary spiral wave using float data type and „f… integer data type „discretized data…. Note that for no noise „ n
Ä0…, discretization has little effect on the PS error. However, discretization coupled with noise tends to cause shortening of the
effective range as well as increase in the PS error.

Identifying PSs for Nonstationary Rotating Waves should be noted that this step is not absolutely necessary
In the section ‘‘Method of Calculating Actual Loca- to identify instantaneous PSs for the nonstationary case.
tions of PSs,’’ we have added a second step for the The second step was added to be consistent with the
nonstationary case for defining an organizing center. It stationary case. For the nonstationary case, f * 共or e * 兲
56 A. N. IYER and R. A. GRAY

FIGURE 8. Experimental data results. „a… Snapshot from optically recorded data from rabbit heart in stable tachyarrhythmia.
Data is scaled from À80 to 20 mV. Thin white lines show the movement of the PS over one rotation. „b… Snapshot from the
corresponding phase map for the same instant of time as in „a…. The phase singularity exists at the region where all colors
converge. Note that it is also easier to visually identify the general location of the tip of the rotating wave from the phase data
rather than transmembrane potential. The movement of the PS over one complete rotation †same as in „a…‡ is shown in thin
black lines. „c… The results from four sets of experimental data are pooled together. There is a range of origin greater than 20%
of action potential amplitude, for which the error in identifying PSs is less than 2 mm.

can be obtained by calculating the mean amplitude of all increased effective range and little effect due to noise
pixels over time.16 The optimal time interval 共window兲 to compared to those obtained with integer data. It is inter-
compute f * 共or e * 兲 deserves to be studied in greater esting to note that discretization by itself had little effect
detail. Here, we averaged the amplitudes of all pixels for with no added noise. However, discretization coupled
two full rotations to obtain f * . with noise produced a shortening of the effective range
and an increase in the error in localizing PSs. But where
Effects of Discretization does the discretization error come from? In order to find
A comparison of the graphs in Figs. 6 and 7 for float the source of this error, we performed the following
and integer data types show that, in the presence of analysis on the same set of float data from computer
noise, discretization had a large effect on the ability to simulations with added noise. First, the float data was
localize PSs. The results using the float data type showed used to calculate phase in floats, which was then used to
ACCURATE LOCALIZATION OF PHASE SINGULARITIES 57

FIGURE 9. Comparison with alternative method. „a… Isochrones for one complete rotation of the spiral wave using the BAR
model. Arrow shows the position of the PSs calculated. „b… Isochrones for one complete rotation of the stationary spiral wave
using the BRDR model. „c… Isopotential maps „50% of maximum APA… at eight equal intervals of one rotation of the non-
stationary spiral wave using the BRDR model. PSs calculated using the PS-finding algorithm are denoted by ‘‘*.’’

localize PSs 关see Figs. 6共a兲, 7共b兲, and 7共e兲兴. Second, the sults suggest that investigators carrying out simulations
float data was used to calculate phase in floats, but the may want to save transmembrane potential maps as float
phase map was discretized and PSs were identified 共not data 共or calculate phase ‘‘on-the-fly’’兲 in order to pre-
shown兲. In the third case, the float data was discretized cisely localize PSs.
before calculating phase 关Figs. 6共b兲 and 7共d兲兴. The re-
sults from the first two cases were very similar and Limitations
showed almost no effect of discretization whereas there
was a marked effect on the third case. This leads us to Many different problems can occur during the identi-
believe that the precision of the data used to calculate the fication of PSs. First, a PS can be identified when none
phase variable plays an important role in the accurate existed. Second, a PS could exist and not be identified.
localization of PSs. Although discretization cannot be Third, a PS that existed could be identified, but at a
avoided during data acquisition, newer, more sensitive nearby location. The first two cases usually occur near
共i.e., lower noise levels and more A/D bits兲 CCD the extremities of the effective range and their frequency
cameras13 will increase the effective range and reduce increased with noise and discretization. However, these
the effect of noise in localizing PSs accurately. Our re- cases were infrequent enough that the affect of these
58 A. N. IYER and R. A. GRAY
5
factors could be considered minimal in our analysis. For Baxter, W. T., J. M. Davidenko, L. M. Leow, J. P. Wuskell,
example, for the nonstationary BRDR case with SNR of and J. Jalife. Technical features of a CCD video camera
system to record cardiac fluorescence data. Ann. Biomed.
10 using discretized data, multiple PSs were identified 5
Eng. 25:713–725, 1997.
times, and no PSs were identified twice 共out of 72 cal- 6
Beeler, G. W., and H. Reuter. Reconstruction of the action
culations兲 within the effective range. The goal of our potential of ventricular myocardial fibers. J. Physiol. (Lon-
study was to study only the third case in detail 共we don) 268:177–210, 1977.
7
studied the affect of three parameters兲. Further work is Cinelli, A. R., S. R. Neff, and J. S. Kauer. Salamander
required to study the first two issues as well as extending olfactory bulb neuronal activity observed by video rate,
voltage-sensitive dye imaging. I. Characterization of the re-
our results to the situation where multiple PSs exist. cording system. J. Neurophysiol. 73:2017–2032, 1995.
8
Courtemanche, M. Complex spiral wave dynamics in a spa-
tially distributed ionic model of cardiac electrical activity.
SUMMARY Chaos 6:579–599, 1996.
9
Drouhard, J., and F. A. Roberge. Revised formulation of the
The objective of this work was to quantify the accu- Hodgkin-Huxley Representation of the sodium current in car-
racy of localizing PSs in heart tissue and the factors that diac cells. Comput. Biomed. Res. 20:333–350, 1987.
10
affect this accuracy. The results obtained from the dis- Ebihara, L., and E. A. Johnson. Fast sodium current in car-
cretized BRDR with a SNR of 10 are very similar to diac muscle: A quantitative description. Biophys. J. 32:779–
790, 1980.
those obtained from our experimental data and suggest 11
FitzHugh, R. Pulses and physiological states in theoretical
that a considerable amount of noise and discretization is models of nerve membrane. Biophys. J. 1:445–466, 1961.
tolerable in the localization of PSs. During fibrillation, 12
Garfinkel, A., P. S. Chen, D. O. Walter, H. S. Karagueuzian,
multiple rotating waves are often observed on the heart. B. Kogan, S. J. Evans, M. Karpoukhin, C. Hwang, T.
Our results suggest that it may be possible to identify Uchida, M. Gotoh, O. Nwasokwa, P. Sager, and J. N. Wiess.
Quasiperiodicity and chaos in cardiac fibrillation. J. Clin.
two PSs which are as close as 2 mm. Here, we felt it was
Invest. 99:305–314, 1997.
necessary to analyze the case in which we know that one 13
Gray, R. A., and Banville, I. Video imaging of fibrillation
and only one PS exists, by which we can ascertain the and defibrillation. In: Quantitative Cardiac Electrophysiology,
capabilities and limitations of the method of calculating edited by C. Cabo and D. S. Rosenbaum. New York: Marcel
phase and identifying PSs. Here we show that it is pos- Dekker 共in press兲.
14
sible to accurately localize PSs in TSVMs, ionic models, Gray, R. A., and Jalife, J. Video imaging of cardiac fibrilla-
tion. In: Optical Mapping of Cardiac Excitation and Arrhyth-
and experimental data and the accuracy is similar for
mias, edited by D. S. Rosenbaum and J. Jalife, New York:
both stationary and nonstationary spiral waves. Futura 共in press兲.
15
Gray, R. A., J. Jalife, A. V. Panfilov, W. T. Baxter, C. Cabo,
J. M. Davidenko, and A. M. Pertsov. Mechanisms of cardiac
ACKNOWLEDGMENTS fibrillation. Science 270:1222–1223, 1995.
16
Gray, R. A., M. P. Arkady, and J. Jalife. Spatial and temporal
The authors would like to thank Frederick Evans and organization during cardiac fibrillation. Nature (London)
Isabelle Banville for the experimental data and Kyle Jus- 392:75–78, 1998.
17
tice for the graphical user interface in the data analysis Ikeda, T., T. Uchida, D. Hough, J. J. Lee, M. C. Fishbein, W.
J. Mandel, P. S. Chen, and H. S. Karagueuzian. Mechanism
software. This work was supported in part by grants of spontaneous termination of functional reentry in isolated
from the Whitaker Foundation 共RAG兲 and the American canine right atrium. Circ. Res. 94:1962–1973, 1996.
Heart Association 共RAG兲. 18
Kuramoto, Y. Chemical Oscillations, Waves and Turbulence.
Berlin: Springer, 1984.
19
Luo, C. H., and Y. Rudy. A dynamic model of the cardiac
ventricular action potential. Circ. Res. 74:1097–1113, 1994.
20
Myerburg, R. J., K. M. Kessler, A. Interian, Jr., P. Fernan-
REFERENCES dez, S. Kimura, P. L. Kozlovskis, T. Furukawa, A. L. Bas-
sett, and A. Castellanos. Clinical and experimental patho-
1 physiology of sudden cardiac death. Cardiac
Aliev, R. R., and A. V. Panfilov. A simple two-variable
model of cardiac excitation. Chaos, Solitons Fractals 7:293– Electrophysiology, From Cell to Bedside. Philadelphia: W. B.
301, 1996. Saunders, 1990.
2 21
Athill, C. A., T. Ikeda, Y. Kim, T. Wu, M. C. Fishbein, H. S. Nagumo, J. S., S. Arimoto, and S. Yoshizawa. An active
Karagueuzian, and P. Chen. Transmembrane potential prop- pulse transmission line simulating nerve axon. Proc. IRE
erties at the core of functional reentrant wave fronts in iso- 50:2061, 1962.
22
lated canine right atria. Circ. Res. 98:1556–1567, 1998. Pertsov, A. M., J. M. Davidenko, R. Salomonsz, W. T. Bax-
3
Banville, I., R. A. Gray, R. E. Ideker, and W. M. Smith. ter, and J. Jalife. Spiral waves of excitation underline reen-
Shock-induced figure-of-eight reentry in the isolated rabbit trant activity in isolated cardiac muscle. Circ. Res. 72:631–
heart. Circ. Res. 85:742–752, 1999. 650, 1992.
4 23
Bar, M., and M. Eiswirth. Turbulence due to spiral wave Takens, F. Detecting strange attractors in turbulence. In Dy-
breakup in a continuously excitable medium. Am. Physiol. namical Systems and Turbulence, edited by D. A. Rand and
Soc. 48:1635–1637, 1993. L. S. Young. Lect. Notes Math. 898:366–381, 1981.
ACCURATE LOCALIZATION OF PHASE SINGULARITIES 59
24 26
Winfree, A. T. When Time Breaks Down. Princeton: Witkowski, F. X., K. M. Kavanagh, P. A. Penkoske, R.
Princeton University Press, 1987. Plonsey, M. L. Spano, W. L. Ditto, and D. T. Kaplan. Evi-
25
Winfree, A. T. Varieties of spiral wave behavior: An experi- dence for determinism in ventricular fibrillation. Phys. Rev.
mentalist’s approach to the theory of excitable media. Chaos Lett. 75:1230–1233, 1995.
1:303–334, 1991.

You might also like