You are on page 1of 12

Abstract.

In the eld of epilepsy, the analysis of


stereoelectroencephalographic (SEEG, intra-cerebral re-
cording) signals with signal processing methods can help
to better identify the epileptogenic zone, the area of the
brain responsible for triggering seizures, and to better
understand its organization. In order to evaluate these
methods and to physiologically interpret the results they
provide, we developed a model able to produce EEG
signals from ``organized'' networks of neural popula-
tions. Starting from a neurophysiologically relevant
model initially proposed by Lopes Da Silva et al. [Lopes
da Silva FH, Hoek A, Smith H, Zetterberg LH (1974)
Kybernetic 15: 2737] and recently re-designed by Jansen
et al. [Jansen BH, Zouridakis G, Brandt ME (1993) Biol
Cybern 68: 275283] the present study demonstrates that
this model can be extended to generate spontaneous
EEG signals from multiple coupled neural populations.
Model parameters related to excitation, inhibition and
coupling are then altered to produce epileptiform EEG
signals. Results show that the qualitative behavior of the
model is realistic; simulated signals resemble those
recorded from dierent brain structures for both inte-
rictal and ictal activities. Possible exploitation of simu-
lations in signal processing is illustrated through one
example; statistical couplings between both simulated
signals and real SEEG signals are estimated using
nonlinear regression. Results are compared and show
that, through the model, real SEEG signals can be
interpreted with the aid of signal processing methods.
1 Introduction
About 0.51% of the population suers from epilepsy,
which is, next to strokes, the most common neurological
disease (Martin 1993). Epilepsy is the result of abnormal
synchronous discharges from large ensembles of neurons
in brain structures. In partial epilepsies, the most
common forms, a region of the brain becomes hyperex-
citable and synchronized and forms a robust circuit of
neural structures generating seizure. This abnormal
region is called the ``epileptogenic zone.'' During
epileptic seizures, stereotyped and uncontrolled altera-
tions occur in the behavior of the patient. These clinical
symptoms vary greatly depending on the brain site
aected by the original paroxystic discharge and its
propagation through the brain. In an large number of
suerers, partial epilepsies are drug-resistant. In these
cases, methods of investigation used are aimed at
anatomically dening the epileptogenic area (Chauvel
et al. 1987) and may possibly lead to surgical treatment
(Engel 1987; Chauvel 1989). These methods may be
divided into three groups, depending on the type of
information they provide: anatomical (imaging tech-
niques), clinical (analysis of ictal signs and symptoms)
and physiological (surface or depth recording of
the spontaneous electrical activity of the brain, dur-
ing or between seizures). Stereoelectroencephalography
(SEEG, intracranial electrodes) belongs to the latter
group. SEEG remains a gold standard method in
presurgical evaluation of epileptic patients, providing
markers of brain activities in the form of time series
signals directly recorded from various cerebral regions.
The analysis of SEEG signals is performed either
visually or by means of signal processing techniques.
These techniques have the advantage not only of
providing quantied information on signals (for exam-
ple, on their morphological or spectral properties) but
also of computing meaningful quantities that are not
accessible by visual inspection (for example, signals
interdependencies). When carefully used (Wendling
et al. 1997; Zaveri et al. 1999), these techniques can
provide crucial information in the study of seizures.
However, if the goal is to progress in the understanding
of the organization of the epileptogenic zone, then signal
processing techniques must help us to physiologically
interpret SEEG signals. Indeed, questions about this
organization remain open: how does the transition from
interictal to ictal activity operate? Can it be detected?
How are involved structures connected to one another?
Can abnormal couplings be identied? How do these
Correspondence to: F. Wendling
(e-mail: fabrice.wendling@univ-rennes1.fr,
Fax: +33-299-286917)
Biol. Cybern. 83, 367378 (2000)
Relevance of nonlinear lumped-parameter models in the analysis
of depth-EEG epileptic signals
F. Wendling
1
, J. J. Bellanger
1
, F. Bartolomei
2
, P. Chauvel
2
1
Laboratoire Traitement du Signal et de L'Image, INSERM Universite de Rennes 1, Campus de Beaulieu, 35042 Rennes Cedex, France
2
Laboratoire de Neurophysiologie et Neuropsychologie, INSERM Universite de la Me diterrane e, 13385 Marseille Cedex 5, France
Received: 3 January 2000 / Accepted: 24 March 2000
couplings evolve as the epileptic activity propagates
(from medial to lateral structures in temporal lobe
epilepsy, for instance)? Do some structures play a
specic role in this propagation?
A means to address some of these questions attempts
to relate quantities provided by signal processing algo-
rithms to physiological mechanisms implied in the gen-
eration of seizure signals. This relationship can be
established through physiological modeling of SEEG
signals. However, designing such a model is not a simple
task. First, the level of modeling (from microscopic to
macroscopic) is a crucial parameter for explanation of
the model. It must be deep enough to provide infor-
mation on the underlying mechanisms that are not di-
rectly accessible. However, for parameters to be
identied and results of modeling to be validated, the
level must also be related to the nature of available real
observations (for example, micro-electrodes or macro-
electrodes recording isolated cells or assemblies of cells).
Second, to be used in the interpretation of SEEG seizure
signals, the model must include parameters directly re-
lated to important mechanisms in epilepsy, such as ex-
citation, inhibition or coupling.
Various models have been proposed to study brain
activities along with their relation to the resulting EEG.
These models may be divided into two groups: those
dedicated to the study of single neurons (Traub 1979)
and those dedicated to the study of biological neural
networks.
In the present work, real observations are SEEG
signals (recorded from intracerebral macroelectrodes;
see Sect. 3) that reect the global post-synaptic activity
of large populations of neurons.
Two dierent modelling approaches were used to
study the activity in neural populations along with its
relation to the observed EEG. The rst is based on
networks built with a large number of elementary cells to
represent both the spatial and temporal properties of the
activity in a given system (Miles et al. 1988). The second
is a lumped-parameter approach in which neural popu-
lations, able to represent the generation of spontaneous
EEG, are modeled as nonlinear oscillators.
This second approach is followed in the present
study, which starts from a model initially developed by
Lopes Da Silva et al. (1974, 1976) for alpha rhythm
generation, formally studied by Zetterberg et al. (1978)
and recently re-designed by Jansen and Rit (1995) to
represent the generation of evoked potentials in the vi-
sual cortex. The objectives of this paper are (1) to extend
the capabilities of the previous model such that it can
generate vectorial EEG signals from multiple coupled
neural populations, (2) to alter model parameters to
make it produce signals that qualitatively resemble real
SEEG signals recorded on multiple channels during in-
terictal or ictal periods and (3) show that simulations
may help to interpret results obtained from signal pro-
cessing methods on real signals. The paper is organized
as follows. In Sect. 2, the original model of EEG signals
and the extension to multiple populations are presented
along with a method to alter model parameters (taking
into account general hypotheses related to seizure gen-
esis) to produce signals reecting epileptiform activity.
In Sect. 3, signals simulated from various scenarios (e.g.
increased excitation, increased coupling strength, unidi-
rectional/bidirectional couplings) are qualitatively com-
pared to real SEEG signals. Results, discussed in Sect. 4,
show that, in the scenarios studied, the behavior of the
model seems particularly relevant; although a precise
waveform t was not the primary intent, simulated sig-
nals may strongly resemble those recorded during sei-
zures. An example of the potential use of the model to
study and interpret the behavior of signal processing
algorithms is detailed.
2 Modeling of epileptiform EEG signals
from multiple coupled neural populations
Realistic macroscopic models of EEG generation have
been proposed since the early 1970s. Substantial progress
was made by Freeman (1978, 1987) and co-workers
(Eeckman and Freeman 1991) in their study and
modeling of the olfactory system. From a long series of
experiments including both works on experimental data
and on models, they established that the central olfactory
system, composed of three parts (i.e. the olfactory bulb,
anterior nucleus and prepyriform cortex), can be mod-
eled as a system of interconnected ensembles of excit-
atory and inhibitory neurons. The dynamics of each
ensemble is represented by a second-order ordinary
dierential equation having a static nonlinearity. Free-
man (1978, 1987) underlines that the simplicity and
generality of the elements of his model make it useable in
the study of other neural systems. Similar ideas were
developed by Lopes Da Silva et al. (1974) and led to the
design of a lumped-parameter model able to generate
alpha rhythm. These lumped-parameter models are able
to represent particular activities of interacting popula-
tions of neurons. Further works dealt with the perfor-
mances of such models in terms of stability (Zetterberg
et al. 1978) or with their adaptation to specic neural
structures such as the hippocampal CA1 region (Leung
1982). Starting from Lopes Da Silva's lumped-parameter
model, Jansen et al. (1993) designed a computer model
that produces spontaneous EEG and evoked potentials.
Recently (Jansen and Rit 1995), they extended the model
such that it can generate EEG and visual evoked
potentials from two coupled populations of neurons. In
the present study, we use Jansen's model with altered
parameters to produce epileptiform signals. However, in
the context of SEEG exploration, the number of
recorded brain structures is generally greater than two.
Thus, the number of coupled populations must be
increased to represent multiple distant sites and their
interactions. This extension of the model is presented
below (Sect. 2.2).
2.1 Model of one neural population
The EEG primarily reects summated synaptic poten-
tials in the activated pyramidal cells. In the single
368
population model, a cluster of neurons is considered that
contains two subsets: subset 1 is composed of the main
cells (i.e. pyramidal cells) which receives feedback (either
excitatory or inhibitory) from subset 2, composed of
local interneurons (i.e. other nonpyramidal cells, stellate
or basket cells) which receives excitatory input only. The
inuence from neighboring or distant populations is
modeled by an excitatory input p(t) that globally
represents the average density of aerent action poten-
tials. Subset 1 is characterized by (1) two dynamic linear
transfer functions that change pre-synaptic information
(i.e. the average pulse density of action potentials) into
post-synaptic information (i.e. an excitatory or inhibi-
tory post-synaptic membrane potential, EPSP or IPSP,
respectively) and by (2) a static nonlinear function that
relates the average membrane potential of a given subset
into an average pulse density of potentials red by the
neurons. Subset 2 is modeled in a similar manner, except
that only one linear transfer function is needed since
input to interneurons is excitatory. As detailed in Jansen
and Rit (1995), the impulse response of the linear
transfer function is given by h
e
(t) = u(t) Aate
)at
in the
excitatory case, and by h
i
(t) = u(t) Bbte
)bt
in the
inhibitory case, where u(t) is the Heaviside function.
Quantities A/a
2
and B/b
2
are the static gains of lters h
e
and h
i
respectively. Lumped parameters a and b are
linked to both the membrane average time constant and
the average distributed delays in the dendritic tree.
Given time constants 1/a and 1/b, parameters A and B
can be used to adjust the sensitivity of excitatory and
inhibitory synapses, respectively.
The static nonlinear function is represented by the
sigmoid function S(v) = 2e
0
=(1 e
r(v
0
v)
), where 2e
0
is
the maximum ring rate, v
0
is the post-synaptic potential
corresponding to a ring rate of e
0
and r is the steepness
of the sigmoid.
Finally, interactions between main cells and inter-
neurons are summarized in the model by four con-
nectivity constants, C1C4, which account for the
average number of synaptic contacts. The above de-
scription is summarized in Fig. 1a. Starting from the
fact that each linear transfer function h
e
(t) and h
i
(t)
introduces a pair of rst-order ordinary dierential
equation of the form
_ z(t) = z
1
(t) ;
_ z
1
(t) = Ggx(t) 2gz
1
(t) g
2
z(t) ;
Fig. 1ac. Structure of the multiple coupled
populations model. a Single population
model. b Elementary model for a given
population i within c the multiple coupled
populations model. Appropriate setting of
parameters K
mn
allows systems to be built
inside which neural populations can be
unidirectionally and/or bidirectionally
coupled. Each population produces an
EEG signal
369
where G = A or G = B and g = a or g = b, depend-
ing on the excitatory or inhibitory case, and where x(t)
and z(t) are the input and output signals, respectively, of
the linear transfer functions, the following set of six
dierential equations that govern the model can be
easily established:
_ y
0
(t) = y
3
(t) ;
_ y
3
(t) = AaS(y
1
y
2
) 2ay
3
(t) a
2
y
0
(t) ;
_ y
1
(t) = y
4
(t) ;
_ y
4
(t) = Aap(t) C
2
S[C
1
y
0
(t)[ 2ay
4
(t) a
2
y
1
(t) ;
_ y
2
(t) = y
5
(t) ;
_ y
5
(t) = BbC
4
S(C
3
y
0
(t) 2by
5
(t) b
2
y
2
(t) :

This set of equations is solved by classical numerical


integration methods (Runge-Kutta, for example). In
linear mode, alpha-like rhythms are produced by the
model. However, as mentioned by Jansen (1996),
experiments conducted by the authors showed that
``the model is capable of much more complex behavior
when it is allowed to operate in a nonlinear mode''. Our
experiments are in agreement with these previous
studies; as shown in Sect. 3, altering some of the model
parameters caused the model to produce signals that
strongly resemble those recorded with intracranial
electrodes, ictally or interictally.
2.2 Model of multiple coupled neural populations
Jansen and co-workers (1993) extended the above model
by coupling two populations in order to explore the
hypothesis that certain visual evoked potentials are due to
an interaction between cortical columns. Similarly, in our
working hypothesis of the epileptogenic network, neural
populations belonging to multiple (N > 2) brain struc-
tures may be implied in the generation and the propaga-
tion of epileptic activities. The purpose of this section is
(1) to describe how to increase the number of coupled
populations and (2) to establish the set of ordinary
dierential equations that describe the model for a given
number of coupled populations. Pyramidal cells are
excitatory neurons that project their axons to other areas
of the brain (Martin 1993). The model can account for
this organization, by using the average pulse density of
action potentials from the main cells of one population as
an excitatory input to another population of neurons.
However, as neural populations may belong to distinct
and distant cerebral areas, new parameters must be
introduced that take into account the dierent ways of
connecting populations and the delays associated with
these connections. Again constant K
ij
is used to dene the
strength of coupling between population i and population
j while a lter with an impulse response h
d
(t), similar to
h
e
(t), can be used to model the delay of connections from
population i: h
d
(t) = u(t) Aa
d
te
a
d
t
(Fig. 1).
This transfer function, chosen here to be independent
of i, introduces a new second-order equation that splits
into two rst-order ordinary dierential equations.
Hence, the model is described by a set of eight ordinary
dierential equations per population (the two equations
describing the output of a given population along with
the six previous equations describing the intra-popula-
tion behavior).
_ y
n
0
(t) = y
n
3
(t) ;
_ y
n
3
(t) = AaS[y
n
1
(t) y
n
2
(t)[ 2ay
n
3
(t) a
2
y
n
0
(t) ;
_ y
n
1
(t) = y
n
4
(t) ;
_ y
n
4
(t) = Aa p
n
(t) C
2
S(C
1
y
n
0
)

i=1;...;N;i,=n
K
i
y
i
6
(t)
( A
2ay
n
4
(t) a
2
y
n
1
(t) ;
_ y
n
2
(t) = y
n
5
(t) ;
_ y
n
5
(t) = Bb C
4
S[C
3
y
n
0
(t)[

2by
n
5
(t) b
2
y
n
2
(t) ;
_ y
n
6
(t) = y
n
7
(t) ;
_ y
n
7
(t) = Aa
d
S(y
n
1
(t) y
n
2
(t)) 2a
d
y
n
7
(t) a
2
d
y
n
6
(t) :

Here, superscript n denotes the population under


consideration; population n receives aerent informa-
tion from populations i; i = 1; . . . ; N; i ,= n, as well as
from the neighborhood [pulse density p
n
(t)]. Quantity
y
n
0
(t) is the output of the interneuron EPSP transfer
function h
e
; y
n
1
(t) and y
n
2
(t) are the outputs of the main
cells' EPSP and IPSP transfer functions h
e
and h
i
,
respectively; y
n
6
(t) is the output of the EPSP transfer
function h
d
. For simplicity, superscript index n was not
added to local parameters A, B, a, b, a
d
, C1C4.
However, these parameters depend on population n and
may vary from one population to another. One can also
notice here that, for n = 2, the obtained set of
2 8 = 16 dierential equations is equivalent to that
established by Jansen et al. (1993). The ability to couple
multiple neural populations, each one having its own
parameter adjustment, is of interest in the context of
epileptic signals analysis, since particular hypotheses,
especially those related to the local balance of excitation
and inhibition and to the strength/direction of cou-
plings, may be explored using the model.
2.3 Generation of epileptiform signals from dened
underlying organizations
All parameters in the model are set on a physiological
basis. The meaning of each parameter along with its
standard value is summarized in Table 1. Readers may
refer to Jansen and Rit (1995) for detailed information
on the way these values were obtained. In the present
study, some of these parameters are altered to produce
signals that resemble those recorded by intracerebral
electrodes during epileptic seizures. There exists an
innite number of ways to alter and combine model
parameters. We chose to restrain this study by trying to
take into account some hypotheses about general
macroscopic mechanisms of epilepsy.
As far as local mechanisms are concerned, the balance
between excitation and inhibition is at the center of
370
numerous studies demonstrating that interictal or ictal
spike discharges (on the EEG) in focal or partial seizures
originate in an area of the cortex that is excessively ex-
citable (reviews in Dichter 1997). This increased excit-
ability may be due to increased excitation, a loss of
inhibition or both (Prince 1996). Numerous studies have
also identied possible mechanisms leading to hyperex-
citability, the most important of these probably being
the existence or development of recurrent excitatory
circuits. These circuits may play a role in the production
of epileptic activities (Jereys 1990; review in Mac Na-
mara 1994; Wilson et al. 1998). In comparison with
changes in excitatory and inhibitory neurotransmitter
functions, less is known about the alterations in voltage-
gated currents and ions channels in epileptic disorders.
However, these alterations, although they are not ex-
plicitly taken into account in the model, are also re-
garded as important elements in epileptic activity. For
example, changes in Na
+
voltage-gated function or ex-
pression have been recently described in animals models
(Bartolomei et al. 1997; Vreugdenhil et al. 1998). In
addition, the depression in extracellular Ca
2+
concen-
tration might contribute to the seizure initiation and
the synchronization of epileptic activity (Louvel and
Heinemann 1983). Other ionic or osmolar extra-
cellular changes can facilitate epileptiform activity
(Schwartzkroin et al. 1998).
In the model, the intra-population behavior is pri-
marily inuenced by parameters A and B; they deter-
mine the amplitude of the post-synaptic potential,
excitatory and inhibitory, respectively. Our experiments
show that the ratio A/B, which can be related to the
balance between excitatory and inhibitory processes,
directly inuences the type of signal dynamics generated
by the model. This ratio can be used to control the
``degree of excitability'' within a given population, al-
lowing ``active'' and ``passive'' neural sites to be dened
in the model of multiple coupled populations. An active
site will be dened as a site for which the excitation/
inhibition ratio is increased, producing sustained dis-
charges of spikes, whereas in a passive site the excita-
tion/inhibition ratio is normal or slightly increased in
order to produce either normal background activity or
sporadic spikes.
As far as interactions between neural groups are
concerned, some studies show that connections which
have become strengthened (Chauvel et al. 1987; Gloor
1990; Fish et al. 1993) may be implied in the production
of epileptic activity. The way circuits are built, also,
seem to be important. Buchanan and Bilkey (1997) for
example, underlined that some structures which recip-
rocally connect with others may play a role in epile-
ptogenesis. In the model, the connectivity constant K
ij
can be used to adjust the strength of coupling between
population i and population j. This constant may also be
used to interconnect populations in dierent ways. In-
deed, as shown in Fig. 2, in the simple case of three sites,
the model allows more complex systems to be built with
underlying organizations that use unidirectional con-
nections and/or bidirectional connections between neu-
ral populations. Using both the excitatory/inhibitory
ratio and the coupling mode, the causality relations
between signals from such systems can be determined
a priori and can be qualitatively controlled. For example,
unidirectional coupling (of given strength) from a rst
active site to a second passive site produces two signals
which reect a situation where the activity of the second
site strongly depends on that of the rst one.
3 Results
In the sequel, real data are referred to as SEEG signals.
These signals were recorded using SEEG in patients
(candidates for epilepsy surgery) suering from tempo-
ral lobe epilepsy. Recordings were performed using
intracerebral multiple lead electrodes that spatially
sample the temporal region of the brain. Along a given
electrode, 1015 leads (length: 2.5 mm, diameter:
0.8 mm, 1.5 mm apart) record signals from medial
(amygdala, hippocampus, parahippocampal gyrus) and
from lateral neocortical (midtemporal gyrus, superior
Table 1. Physiological inter-
pretation and standard values
of model parameters (adapted
from Jansen and Rit 1995)
Parameter Interpretation Standard value
A Average excitatory synaptic gain 3.25 mV
B Average inhibitory synaptic gain 22 mV
a, b Membrane average time constant and
dendritic tree average time delays
a = 100 s
)1
b = 50 s
)1
C1, C2 Average number of synaptic contacts in
the excitatory feedback loop
C1 = C
C2 = 0.8 C
(with C = 135)
C3, C4 Average number of synaptic contacts in the
inhibitory feedback loop
C3 = 0.25 C
C4 = 0.25 C
(with C = 135)
m
0
, e
0
, r Parameters of the nonlinear sigmoid function
(transforming the average membrane
potential to an average density of action potentials)
m
0
= 6 mV
e
0
= 2.5 s
)1
r = 0.56 mV
)1
a
d
Average time delay on eerent connection from
a population
a
d
= 33 s
)1
K
ij
Connectivity constant associated with the connection
between two populations i and j
No standard value.
Used to adjust the coupling
strength between populations
371
temporal gyrus) structures. The position of each elec-
trode is strategically dened from the analysis of surface
EEG signals, clinical data and imaging in order to
record structures that play a potential role during the
seizures.
3.1 Inuence of the excitation/inhibition ratio
A Gaussian white noise was used as the model input,
p
n
(t). The mean and variance (corresponding to a rate of
30150 pulses/s) were adjusted so that the model
produced a signal similar to the spontaneous EEG
recorded from neocortical structure electrodes during
interictal periods when other parameters of the model
are set to standard values (Table 1). This signal (Fig. 3a)
reects a normal activity which resembles that reected
by real SEEG signals (Fig. 3f). Starting from this point,
all parameters were kept standard except for A (gain
constant in the excitatory loop) which was progressively
increased. This increase (Fig. 3be) makes the spikes
that appear sporadically (from A = 3.4), then frequently
(A = 3.6) and nally rhythmically (A = 3.8). The
average frequency of this rhythmic discharge of spikes
increases as A is increased (from A = 3.8 to A = 8.5).
Visual inspection of real signals shows that sporadic
spikes (usually appearing before seizures start) as well as
rhythmic discharges of spikes (usually appearing after
seizures start) can be encountered in SEEG recordings
(Fig. 3gi). For this type of activity, one also notices
that real waves are quite accurately reproduced by the
model. From A = 8.5, the model switches from spiking
activity to sinusoidal activity (alpha-like) whose enve-
lope also depends on the value of A. Again, this type of
activity may be reected by SEEG signals during
seizures (see the rst 5 s of simulated signal in Fig. 4c,
described below). The excitation/inhibition ratio may
also be adjusted to position the model on the border
between two types of activity. In that case, spontaneous
transitions occur (Fig. 4). The upper part of Fig. 4
presents a real SEEG signal recorded in the amygdala at
the beginning of a temporal lobe epilepsy seizure. The
SEEG signal shows high amplitude spikes followed by
faster activity (quasi-sinusoidal around t = 24 s) which
slows down to resume spiking activity with a lower
average frequency. Figure 4b provides a detailed view of
this period of time. A spontaneous transition between
quasi-sinusoidal activity and spikes is shown in Fig. 4c.
One notices that the simulated signal behaves qualita-
tively as the real one; the sinusoidal activity of lower
amplitude and higher frequency changes to a lower-
frequency, higher-amplitude spiking activity.
3.2 Inuence of the coupling strength and direction
A key advantage of the multiple populations model is its
ability to produce signals reecting causality relations
between underlying populations. These relations can be
easily controlled via the strength (parameters K
ij
) and
thedirectionof couplings (unidirectional or bidirectional).
Figure 5 illustrates the eect of increasing K
ij
on a
simple example in which three populations are consid-
ered. The rst one was made hyperexcitable by increas-
ing its excitation/inhibition ratio. Parameters of the two
other populations remained at standard values. For null
couplings between these three populations, the resulting
signals correspond to normal background activity with
sporadic spikes appearing randomly from the rst
population and only normal background activity from
the two others. When parameters K
12
and K
23
(repre-
senting the unidirectional coupling from population 1 to
population 2 and the unidirectional coupling from
population 2 to population 3) are increased
(K
12
= 200, K
13
= 200), the propagation of a spike
from population 1 to population 2 and population 3 is
simulated (Fig. 5a). As expected and as shown by time
delays, spikes in populations 2 and 3 are due to that in
population 1. Now, introducing a reciprocal coupling
Fig. 2. Building systems from dened underlying organizations. This
example gives the possible ways to interconnect three neural
populations from unidirectional couplings (single-sided arrow) and/
or bidirectional couplings (double-sided arrows)
372
from population 3 to population 1 (K
31
= 200) causes
the model to produce a dierent activity. The result is
shown in Fig. 5b and c. The introduction of a reciprocal
coupling in the model is sucient to produce sustained
discharges of spikes (involving the three populations)
that closely resemble those observed in SEEG signals
during the spread of a seizure to neocortical structures in
temporal lobe epilepsies (Bartolomei et al. 1999). The
same eect is obtained when a retro-coupling from
population 2 to population 1 is introduced. Interestingly,
reciprocal coupling is not the only way to produce
rhythmic synchronized activity. When passive popula-
tions (i.e. generating a normal type activity) are
connected to active populations (i.e. generating rhyth-
mic epileptiform activities), the activities tend to syn-
chronize for suciently high coupling values. Figure 6cf
gives an example of that situation in a two-coupled
populations model. In population 1, the excitation/
inhibition ratio was increased to make it periodically
generate spikes. In population 2, all parameters were
kept standard. Increasing parameter K
12
makes the
activity of population 2 dependent on that of population
1. Once K
12
has reached a certain value, rhythmic
synchronized activities are generated by the system. The
Fig. 3ai. Simulated signals
(top) and real stereoelectroen-
cephalography (SEEG) signals
(bottom). Starting from a
``normal EEG'' (a), the
progressive increase of the
excitation/inhibition ratio makes
the model generate sporadic
spikes (b) which appear more
frequently (c and d) before
transforming in a rhythmic
discharge of spikes (e). These
simulated signals resemble real
SEEG signals recorded before
(f,g) and during (h,i) a seizure
recorded from a neocortical
structure (mid temporal gyrus).
See text for values of modied
parameter A
373
Fig. 4ac. Transitions between
activities reected in real and
simulated signals. a Transitions
of activities in a real SEEG
signal recorded from mesial
structures (amygdala) at the
beginning of a temporal lobe
seizure. b Zooming on the
transition between sinusoidal
and spike activities. c
Spontaneous transition
occurring in the simulated signal
when the excitation/inhibition
ratio is adjusted to position the
model on the frontier between
two types of activities
Fig. 5ac. Inuence of the
coupling mode. Population 1
was made hyperexcitable by
increasing its excitation/
inhibition ratio. a For
unidirectional coupling (of
certain strength) between
population 1 and population 2
and between population 2 and
population 3, sporadic spikes
originating from population 1
propagates to population 2, then
from population 2 to population
3. b the introduction of a
reciprocal connection is a
sucient condition to make the
model generate sustained
discharges of spikes that
resemble some reected in real
SEEG signals during the
propagation of temporal lobe
seizures (c)
374
following example illustrates how to exploit these
signals.
3.3 Using the model to interpret signal processing
results: an example
A key advantage of the model is that it is based on
physiologically relevant parameters. Thus, it can be used
to study, compare or validate signal processing algo-
rithms aimed at providing statistics related to model
parameters, in the hope of designing algorithms able to
help in the physiological interpretation of real data. To
illustrate this idea, a simple example will be briey
detailed. Figure 6a shows two SEEG signals s
1
(t) and
s
2
(t), recorded from hippocampal and amygdaloid
formations, respectively, at the beginning of a temporal
lobe epileptic seizure. Spikes start to appear in the
hippocampus around t = 22 s and give rise to a
sustained rhythmic activity not only in the hippocampus
but also in the amygdala. Visual inspection of signals
shows that hippocampal spikes seem to precede those of
the amygdala. A large variety of signal processing
methods have been proposed to study multichannel
EEG signals. Roughly, methods (either linear or non-
linear) can be divided into two groups: those dedicated
to the analysis of the morphological or spectral
properties of signals (detection, characterization,
Fig. 6ag. Example of use of the
model in the analysis of results
from signal processing
algorithms. a Real SEEG signals
recorded from the hippocampus
(bottom) and amygdala (top) at
the beginning of a seizure. b
Nonlinear regression used in the
measurement of the statistical
coupling between real signals
(solid line) and surrogates (dash
line). c,d,e,f Simulated signals
from an ``active site'' (rhythmic
discharge of spikes, bottom) and
a ``passive site'' (standard
parameter values, top) for
unidirectional coupling of
increasing strength. g
Measurement of the statistical
coupling between simulated
signals (solid line) and surrogates
(dash line). The behavior of the
coupling estimator (nonlinear
correlation coecient) is similar
in real and simulated signals.
Results tend to demonstrate that
the hippocampus plays a
predominant role in the
analyzed seizure
375
classication), and those dedicated to the analysis of the
statistical coupling between signals (cross correlation,
coherence, comparison of state space trajectories). This
second category is of particular interest in this example,
since it is centered on structure interactions during
seizures. We chose to analyze both simulated and real
signals using a nonlinear method. This can be justied
by at least two facts. First, some studies (e.g. Le Van
Quyen et al. 1998) show that subtle relations between
signals are not accessible by linear methods. Second, the
model used here (coupled nonlinear oscillators) includes
nonlinear elements. Among the possible nonlinear
methods, nonlinear regression is perhaps the simplest.
It was rst used by Pijn and Lopes Da Silva (1993) in the
eld of EEG signals. They introduced a nonlinear
correlation coecient (referred to as h
2
), a measure of
the degree of coupling between two signals.
Figure 6b shows the evolution of the nonlinear cor-
relation coecient (h
2
) when measured on real data and
surrogate data with Fourier methods obtained from real
signals as time increases. This statistical coupling glob-
ally increases with time except during a transition period
(from t = 35 s to t = 50 s), when nonstationary activ-
ity appears in s
2
(t) (mixings of spikes and noisier activ-
ity). h
2
reaches a high value (0.92 at t = 51 s) when
spikes appear to be quite synchronized. h
2
measured on
surrogate signals is lower than that measured on real
signals, showing that part of the relation between signals
is nonlinear. This was also conrmed by a coherence
analysis (not shown). Readers may refer to Pijn and
Lopes Da Silva (1993) for details about nonlinear re-
gression used in EEG analysis and may refer to Popi-
vanov and Mineva (1999) for use and interpretation of
surrogate data techniques. Figure 6cg summarize the
results of a simulation done with the model for two
unidirectionally coupled populations. The rst one was
changed for its excitation/inhibition ratio to obtain an
activity of sustained spikes. The second was maintained
at standard values. The coupling strength K
12
was pro-
gressively increased (0600 with a step of 10). For each
value of K
12
, h
2
was measured on simulated signals and
on surrogate signals obtained from simulated signals.
From K
12
= 0 to K
12
= 150, h
2
follows the evolution of
K
12
quite well. Over this interval, subtle changes take
place in the signal from population 2 (small amplitude
waves that progressively increase in number and in
amplitude; see Fig. 6d). Then, from K
12
= 150 to
K
12
= 220, a non-stationary activity appears in the
second signal (mixing of spikes and previous activity;
Fig. 6e). Over this interval, the h
2
(Fig. 6g) decreases.
This phenomenon is linked to the behavior of popula-
tion 2: the operating region of the associated model
changes with its average input level, modied by the
increase of K
12
. Due to the nonlinear characteristics of
the model, these changes of operating regions, which
induce important modications in the signal generated
by population 2, are not progressive as a function of K
12
.
This is reected by the measurement of h
2
. Finally, for
K
12
> 220, spikes generated in population 2 progres-
sively synchronize, i.e. temporally align, with those of
population 1 (Fig. 6f). The general behavior of h
2
looks
similar on both real and simulated signals. Interpreted
through the model in which epileptiform activity in
population 2 (driven) depends on that in population 1
(driver), this result tends to show that the hippocampus
plays a leading role in this seizure while the amygdala is
under inuence. Additional measurements of time delays
and evaluation of coupling directions could help in the
study of such hypotheses (see Sect. 4). Of interest are
also h
2
values measured on surrogate data. These values
are lower than those computed on simulated data (as for
real signals). This observation conrms that part of the
relationship between the two nonlinear coupled oscilla-
tors is nonlinear, as expected.
4 Discussion
We have presented the extension of an EEG generation
model initially proposed by Lopes Da Silva (1974) and
recently re-designed by Jansen and Rit (1995) to study
the generation of evoked potentials in the visual cortex.
This approach, based on a neurophysiologically relevant
modeling, diers from other approaches in which
external models, either linear or nonlinear, are used to
simulate EEG activities (Kaipio and Karjalainen 1997).
In its extended version, the model produces vectorial
signals reecting epileptiform activities (sporadic spikes,
rhythmic spikes, quasi-sinusoidal discharges) that
closely resemble activities recorded from dierent cere-
bral structures with intracranial electrodes (i.e. SEEG
exploration) during seizures. In the model, neural pop-
ulations are represented by nonlinear oscillators that
may be coupled via excitatory connections. To us, this
level of modeling (based on a lumped parameter ap-
proach) is particularly suited to the nature of SEEG
signals. Numerous studies show that epileptic activities
result from a collective behavior of neural assemblies
which synchronize to produce bursts of spikes at the
cellular level and spikes (sporadic or sustained bursts) on
the EEG. Obtained results are in accordance with results
obtained in various studies performed during the last
decade. Lopes da Silva et al. (1994) present evidence
that local neural networks behave as nonlinear dynamic
systems and show, through paired-pulse experiments,
that ``the excitation/inhibition ratio increases in the
course of the establishment of a kindled epileptogenic
focus.'' Zetterberg et al. (1978) studied a single popula-
tion model whose structure is very close to the single
population model presented here. They noticed that the
model was able to generate epileptiform activities
(spikes) that are ``borderlike cases between normal
background activity and seizure activity.'' Our results
are also in accordance with those obtained by Freeman
(1987) based on his model of the olfactory system (based
on the same concepts as those developed here) that is
able to produce chaotic signals that t (more or less)
epileptiform patterns extracted from EEG seizure sig-
nals. Similar ideas may also be found in work developed
by Babloyantz and Destexhe (1986) that points to non-
376
linear properties (deterministic nature of brain activity,
existence of chaotic attractors, change of dimensionality
between normal or epileptic state) of the EEG using
mathematical tools initially dedicated to the study of
nonlinear dynamical systems.
However, if our study conrms the results of Zetter-
berg (1978), it also goes slightly further, showing that (1)
instability is caused by introducing an imbalance be-
tween excitation and inhibition, as evidenced in the single
population model and (2) spikes may propagate from
one population to another for suciently high values of
couplings and transform in sustained bursts of spikes for
certain underlying organizations, as shown in the multi-
ple populations model. Our study also diers from that
of Freeman (1970) in several points. First, our main
objective is to progress in the understanding of the or-
ganization of the epileptogenic zone. In this context,
modeling is a way to evaluate and to interpret infor-
mation from signal processing algorithms. Second, our
approach is more general. A simplied model of the
neural population is used and several models can be
coupled in numerous ways to represent more complex
systems and to simulate signals from these systems, the
hope being to design signal processing algorithms able to
nd an underlying organization. In this context, we are
currently working on the estimation of coupling direc-
tions in simulated signals, using methods that provide
asymmetric quantities. Results are also compared with
those obtained from the measurement of time delays. In
the last simulation, our intent was not to alter parameters
in order to accurately reproduce real activities but to
show that the model can be used to study the behavior of
a signal processing algorithm. The simple chosen exam-
ple shows that one must be cautious with the interpre-
tation of measured quantities. The performed simulation
is particularly interesting in illustrating this idea: the
observed decrease in h
2
, which could have been inter-
preted as a decrease of the coupling, corresponds, in fact,
to an increase of this coupling generating a more complex
nonstationary situation. It also shows that some quan-
tities are able to characterize the evolution of some model
parameters. A detailed analysis of such quantities on
both real and simulated signals could help in the inter-
pretation of underlying physiological mechanisms im-
plied in the generation of signals. Several studies
developing similar ideas (use of physiological models in
the design of signal processing algorithms) have already
been presented in the eld of sleep electroencephalogram
analysis (Da Rosa et al. 1991) or alpha-state detection
(Kemp and Blom 1981). More recently, Roessgen et al.
(1998) improved the performance of seizure detection in
the EEG of a newborn by using an approach based on a
physiologically relevant model whose structure is very
close to a linear version of the model described here.
In its present form, the model cannot represent some
epileptic activities. Precisely, low voltage rapid dis-
charges (LVRD) are not taken into account. Even if the
origin of LVRD remains controversial, we think that the
model can be improved in that direction. Interesting
concepts about lumped descriptions and cerebral
rhythms may be found in Wright's recent study (1999),
which reports a new extension allowing his model to
generate fast activities in the order of 40 Hz.
5 Conclusion and future work
The present study was performed in the context of the
analysis of SEEG seizure signals using signal processing
techniques. Our goal is to progress (1) in the identica-
tion of the zone responsible of the triggering of seizures
and (2) in the understanding of its organization. To
achieve this goal, we think that quantities provided by
signal processing algorithms must be (1) well understood
in a context of nonstationary SEEG signals exhibiting
nonlinear properties and (2) related, if possible, to the
underlying organization of neural populations belonging
to brain structures that are involved in the generation of
these signals. To us, modeling may be a way to establish
such a relation, i.e. to bridge the gap between signal
processing techniques and physiological interpretation
of real data. A step in this direction has been accom-
plished with the model presented here, which seems
particularly relevant in the study of SEEG signals
recorded during interictal or ictal periods. In this model,
genericity and modularity allow multiple neural popu-
lations to be coupled (in dierent ways) in order to build
more complex systems from which spontaneous EEG is
simulated. For some parameter adjustments, we ob-
tained simulated signals that resemble those observed in
the brain, and we believe that such results would be
dicult to obtain from purely external models (even
nonlinear), i.e. not based on neurophysiological infor-
mation. In its current version, the model only generates
stationary activities. In the near future, higher-level
mechanisms, such as short-term potentiation, will be
incorporated to change parameters as time advances,
oering the possibility to study controlled transitions
between activities.
Acknowledgement. The authors wish to thank Daniel Drake for
useful comments on the manuscript.
References
Babloyantz A, Destexhe A (1986) Low-dimensional chaos in an
instance of epilepsy. Proc Natl Acad Sci USA 10: 35133517
Bartolomei F, Gastaldi M, Massacrier A, Planells R, Nicolas S,
Cau P (1997) Changes in the mRNAs encoding subtypes I, II
and III sodium channel alpha subunits following kainate-in-
duced seizures in rat brain. J Neurocytol 26: 667678
Bartolomei F, Wendling F, Vignal JP, Kochen S, Bellanger JJ,
Badier JM, Le Bouquin-Jeannes R, Chauvel P (1999) Seizures
of temporal lobe epilepsy: identication of subtypes by coher-
ence analysis using stereo-electro-encephalography. Clin Neu-
rophysiol 110: 17411754
Buchanan JA, Bilkey DK (1997) Transfer of epileptogenesis be-
tween perirhinal cortex and amygdala induced by electrical
kindling. Brain Res 771: 7179
Chauvel P (1989) Indications et me thodes du traitement chirurgical
des e pilepsies, Epilepsies 1: 258275
Chauvel P, Buser P, Badier JM, Liegeois-Chauvel C, Marquis P,
Bancaud J (1987) La ``zone e pileptoge ne'' chez l'homme:
377
repre sentation des e ve nements intercritiques par cartes spatio-
temporelles. Rev Neurol (Paris) 143: 443450
Da Rosa AC, Kemp B, Paiva T, Lopes da Silva FH, Kamphuisen
HA (1991) A model-based detector of vertex waves and K
complexes in sleep electroencephalogram. Electroencephalogr
Clin Neurophysiol 78: 7179
Dichter MA (1997) Basic mechanisms of epilepsy: targets for
therapeutic intervention. Epilepsia 38: S2S6
Eeckman FH, Freeman WJ (1991) Asymmetric sigmoid non-lin-
earity in the rat olfactory system. Brain Res 557: 1321
Engel J Jr (1987) Surgical treatment of the epilepsies. Raven Press,
New York
Fish DR, Gloor P, Quesney FL, Olivier A (1993) Clinical responses
to electrical brain stimulation of the temporal and frontal lobes
in patients with epilepsy. Pathophysiological implications.
Brain 116: 397414
Freeman WJ (1978) Models of the dynamics of neural populations
Electroencephalogr Clin Neurophysiol [Suppl] 34: 918
Freeman WJ (1987) Simulation of chaotic EEG patterns with a
dynamic model of the olfactory system. Biol Cybern 56:
139150
Gloor P (1990) Experimental phenomena of temporal epilepsy.
Facts and hypotheses. Brain 113: 16731694
Jansen BH (1996) Nonlinear dynamics and quantitative EEG
analysis. Electroencephalogr Clin Neurophysiol [Suppl] 45: 39
56
Jansen BH, Rit VG (1995) Electroencephalogram and visual
evoked potential generation in a mathematical model of cou-
pled cortical columns. Biol Cybern 73: 357366
Jansen BH, Zouridakis G, Brandt ME (1993) A neurophysiologi-
cally-based mathematical model of ash visual evoked poten-
tials. Biol Cybern 68: 275283
Jeerys JG (1990) Basic mechanisms of focal epilepsies. Exp
Physiol 75: 127162
Kaipio JP, Karjalainen PA (1997) Simulation of nonstationary
EEG. Biol Cybern 76: 349356
Kemp B, Blom HA (1981) Optimal detection of the alpha state in a
model of the human electroencephalogram. Electroencephalogr
Clin Neurophysiol 52: 222225
Le Van Quyen M, Adam C, Baulac M, Martinerie J, Varela FJ
(1998) Nonlinear interdependencies of EEG signals in human
intracranially recorded temporal lobe seizures. Brain Res 792:
2440
Leung LS (1982) Nonlinear feedback model of neuronal popula-
tions in hippocampal CAl region. J Neurophysiol 47: 845868
Lopes da Silva FH, Hoek A, Smith H, Zetterberg LH (1974) Model
of brain rythmic activity. Kybernetic 15: 2737
Lopes da Silva FH, van Rotterdam A, Barts P, van Heusden E,
Burr W (1976) Models of neuronal populations: the basic
mechanisms of rhythmicity. Prog Brain Res 45: 281308
Lopes da Silva FH, Pijn JP, Wadman WJ (1994) Dynamics of local
neuronal networks: control parameters and state bifurcations
in epileptogenesis. Prog Brain Res 102: 359370
Louvel J, Heinemann U (1983) Changes in Ca++, K+ and
neuronal activity during oenanthotoxine induced epilepsy in cat
sensori motor cortex. Electro encephalogr Clin Neurophysiol
56: 457463
Mac Namara JO (1994) Cellular and molecular basis of epilepsy.
J Neurosci 14: 34133425
Martin JH (1993) The collective electrical behavior of cortical
neurons: the electroencephalogram and the mechanisms of
epilepsy. In: Kandel ER, Schwartz JH, Jessel TM (eds) Prin-
ciples of neural science, 3rd edn. Prentice Hall, Englewood
Clis, New Jersey pp 777791
Miles R, Traub RD, Wong RK (1988) Spread of synchronous
ring in longitudinal slices from the CA3 region of the hippo-
campus. J Neurophysiol 60: 14811496
Pijn JP, Lopes Da Silva FH (1993) Propagation of electrical ac-
tivity: nonlinear associations and time delays between EEG
signals. In: Zschocke S, Speckmann EJ (eds) Basic mechanisms
of the EEG, Birkauser, Boston, pp 4161
Popivanov D, Mineva A (1999) Testing procedures for non-sta-
tionarity and non-linearity in physiological signals. Math Bio-
sci 157: 303320
Prince D (1996) Basic mechanisms of focal epileptogenesis
Avanzini G, Fariello R, Heinemann U, Mutani R (eds) Epile-
togenic and excitotoxic mechanisms. Libbey, London,
pp 1727
Roessgen M, Zoubir AM, Boashash B (1998) Seizure detection of
newborn EEG using a model-based approach. IEEE Trans
Biomed Eng 45: 673685
Schwartzkroin PA, Baraban SC, Hochman DW (1998) Osmolarity,
ionic ux, and changes in brain excitability. Epilepsy Res 32:
275285
Traub RD (1979) Neocortical pyramidal cells: a model with den-
dritic calcium conductance reproduces repetitive ring and
epileptic behavior. Brain Res 173: 243257
Vreugdenhil M, Faas GC, Wadman WJ (1998) Sodium currents in
isolated rat CA1 neurons after kindling epileptogenesis. Neu-
roscience 86: 99107
Wendling F, Badier JM, Chauvel P, Coatrieux JL (1997) A
method to quantify invariant information in depth-recorded
epileptic seizures. Electroencephalogr Clin Neurophysiol 102:
472485
Wilson CL, Khan SU, Engel J Jr, Isokawa M, Babb TL, Behnke
EJ (1998) Paired pulse suppression and facilitation in human
epileptogenic hippocampal formation. Epilepsy Res 31: 211
230
Wright JJ (1999) Simulation of EEG: dynamic changes in synaptic
ecacy, cerebral rhythms, and dissipative and generative ac-
tivity in cortex. Biol Cybern 81: 131147
Zaveri HP, Williams WJ, Sackellares JC, Beydoun A, Duckrow
RB, Spencer SS (1999) Measuring the coherence of intracranial
electroencephalograms. Clin Neurophysiol 110: 17171725
Zetterberg LH, Kristiansson L, Mossberg K (1978) Performance of
a model for a local neuron population. Biol Cybern 31: 1526
378

You might also like