Neural networks involving the medial temporal structures intemporal lobe epilepsy

Fabrice Bartolomeia,*, Fabrice Wendlingb, Jean-Jacques Bellangerb,Jean ReÂgisa, Patrick Chauvela

aUnite d'eÂpileptologie et Laboratoire de Neurophysiologie et Neuropsychologie, INSERM EMI 9926, Universite de la MeÂditerraneÂe, 13385

Marseille Cedex 5, FrancebLaboratoire Traitement du Signal et de L'Image, INSERM EPI 9934, Universite de Rennes 1, Campus de Beaulieu, 35042 Rennes Cedex, France

Accepted 27 April 2001

Abstract

Objectives: In a previous study using the averaged coherence technique to study interactions between medial/limbic and lateral/neocor-

tical regions, we observed that epileptogenic networks in temporal lobe epilepsy seizures (TLES) could be divided into 4 subtypes, i.e. medial

(M), medial±lateral (ML), lateral±medial (LM), and lateral (L). In the ML and LM subtypes, medial structures and the anterior temporal

neocortex are co-activated at the onset of seizures. However, using this approach, we were unable to determine the direction of coupling and

may have overlooked non-linear variations in interdependency. The purpose of the present study using non-linear regression for analysis of

stereoelectroencephalographic (SEEG) signal pairs was to measure the degree and direction of coupling between medial and neocortical

areas during TLES in patients with the M, ML, and LM subtypes.

Methods: Eighteen patients with drug-resistant TLEs who underwent SEEG recording were studied. We used a non-linear correlation

method as a measure of the degree and the direction of coupling on SEEG signal pairs. Patients with pure lateral TLEs were not studied. We

analyzed the functional coupling between 3 regions of the temporal lobe: the anterior temporal neocortex, the amygdala, and the anterior

hippocampus. A physiological model of EEG generation was used to validate the non-linear quanti®cation method and assess its applicability

to real SEEG signals.

Results: Results are ®rst based on a physiological model of EEG data in which both degree and direction of coupling are explicitly

represented, thus allowing construction of the neural systems inside which causality relationships are controlled and generation of multi-

channel EEG signals from these systems. These signals provide an objective way of studying the performance of non-linear regression

analysis on real signals.

In medial networks (10 patients), the ictal discharge is limited to the medial limbic structures and may propagate secondarily to the cortex.

Quanti®ed results demonstrated no signi®cant coupling between medial and lateral structures at the beginning of the seizures. Conversely,

almost constant unidirectional or bidirectional coupling was observed between hippocampus and amygdala.

In medial±lateral (5 patients) and lateral±medial (3 patients) networks, the initial ictal discharge includes both limbic and neocortical

regions. A rapid `tonic' discharge is observed over the temporal neocortex at the onset of seizure. Quantitative analysis showed an initial

increase in the non-linear correlation coef®cient between neocortex and medial structures. Quanti®cation of the coupling direction demon-

strated in¯uence of medial over lateral structures (medial±lateral) or of the lateral neocortex over medial structures (lateral±medial).

Conclusions: These results con®rm the existence of several generic and organized networks involving the medial structures during TLE

seizures. q 2001 Elsevier Science Ireland Ltd. All rights reserved.

Keywords: Temporal lobe epilepsy; Non-linear regression; Neural networks; Stereoelectroencephalography; Classi®cation

1. Introduction

Temporal lobe epilepsy (TLE) is a form of partial

epilepsy involving seizure originating in one or several

anatomic divisions of the temporal lobe and propagating

through interconnected neuronal networks within or beyond

its boundaries. TLE is classi®ed into two main subtypes

(Commission on Classi®cation and Terminology of the

International League Against Epilepsy, 1989; Williamson

et al., 1998). The ®rst and most common subtype is called

medial TLE (MTLE) because of primary involvement of

medial (limbic) temporal structures. The second subtype is

Clinical Neurophysiology 112 (2001) 1746±1760

1388-2457/01/$ - see front matter q 2001 Elsevier Science Ireland Ltd. All rights reserved.

PII: S1388-2457(01)00591-0

www.elsevier.com/locate/clinph

CLINPH 2000792

* Corresponding author. Service de Neurophysiologie Clinique et UniteÂ

d'Epileptologie, CHU Timone-264 Rue St Pierre, 13005 Marseille, France.

E-mail address: fbartol@aol.com (F. Bartolomei).

called lateral (or neocortical) TLE (LTLE) because of

primary involvement of the temporal neocortex (Ebner,

1994; Burgerman et al., 1995;Walczak, 1995). MTLE has

been extensively studied over the last decade resulting in

description of the so-called `MTLE syndrome' (French et

al., 1993; Wieser et al., 1993; Williamson et al., 1993). The

cardinal histopathological and neuroimaging marker of the

MTLE syndrome is hippocampal sclerosis that can be read-

ily detected by MRI.

Drug treatment is often inef®cient in TLE patients, but

surgical excision of the epileptogenic region and/or seizure

propagation pathways can stop or signi®cantly reduce the

frequency of seizures. However, the rate of surgical failure

in this population is far from negligible (Engel et al., 1993;

French et al., 1993; Williamson et al., 1993), suggesting that

current criteria of patient selection are not entirely satisfac-

tory. In this regard, several studies have demonstrated

hippocampal atrophy in patients with LTLE (Cascino et

al., 1993) and patients with temporal or extratemporal

lesions (Levesque et al., 1991). This observation suggests

that this MRI ®nding may not be a speci®c diagnostic tool

for MTLE.

Analysis of the epileptogenic zone in TLE remains the

important step for understanding seizure-triggering

mechanisms, identifying implicated brain structures, and

improving surgical planning. Several authors have high-

lighted the richness and variability of epileptogenic region

organization in TLE patients (Bancaud et al., 1965;

Bancaud, 1981; Wieser, 1983; Munari et al., 1994; Walc-

zak, 1995). Particular emphasis has been placed on the

complexity of interactions within limbic or paralimbic

structures such as the temporal pole (Munari et al., 1994;

Chabardes et al., 1999) or the insular cortex (Isnard et al.,

2000) and between limbic and neocortical structures (Barto-

lomei et al., 1999). Synchronization of neuronal networks

has been shown to be essential for generation of epileptic

events at a cellular level (Lopes da Silva and Pijn, 1995).

Synchronization of several spatially separate neural struc-

tures in the brain could actually form the substratum of the

epileptogenic zone at a macroscopic level (Badier and

Chauvel, 1995; Wendling et al., 1997, 2001).

This speculation has led to the hypothesis that the epilep-

togenic zone results from a network formed by neuronal

populations in different cerebral structures. However,

there is little direct evidence for the existence of these

neuronal networks in the literature. Proof would require

demonstration of abnormal functional links between several

neural structures at the onset of seizure. Occurrence of these

links could be con®rmed by analyzing interdependency of

electrical activities in these neural assemblies. A number of

techniques have been described for this purpose based on

either linear (e.g. the coherence function) or non-linear

calculation. Traditionally, non-linear calculation has

involved either mutual information (Mars and Lopes da

Silva, 1983) or non-linear regression (Pijn et al., 1991,

1992; Pijn and Lopes da Silva, 1993). A developing new

approach to non-linear calculation has been based on the

study of non-linear dynamical systems and chaos (Le Van

Quyen et al., 1998).

In a previous study using the linear averaged coherence

technique to study interactions between medial/limbic and

lateral regions, we demonstrated that epileptogenic

networks in TLE patients should be divided into 4 subtypes,

i.e. medial (M), medial±lateral (ML), lateral±medial (LM),

and lateral (L) (Bartolomei et al., 1999). In the ML and LM

subtypes, medial structures and the anterior temporal

neocortex are co-activated at the onset of seizures.

However, using this approach, we were unable to determine

the direction of coupling and may have overlooked non-

linear variations in interdependency. Recent studies have

suggested the existence of such non-linear events during

TLE seizures (Casdagli et al., 1997; Van der Heyden et

al., 1999).

The purpose of the present study using non-linear regres-

sion for analysis of stereoelectroencephalography (SEEG)

signal pairs was to measure the degree and direction of

coupling between medial and neocortical areas during

temporal lobe epilepsy seizures (TLES) involving the

medial limbic structures in patients with the M, ML, and

LM subtypes. A physiological model of EEG generation

was used to validate the non-linear quanti®cation method

and assess its applicability to real SEEG signals.

2. Methods

2.1. Patients and SEEG study

Eighteen patients were selected from a group undergoing

evaluation for surgical treatment of intractable TLE

between 1994 and 1999. All of them had a comprehensive

evaluation, including a detailed history and neurological

examination, neuropsychological testing, routine MRI

study, interictal/ictal SPECT, and video-EEG recording of

seizures. At the end of this ®rst non-invasive phase, patients

were separated into two groups: (1) when all the localization

criteria converged to a pure MTLE, patients were operated

on (anterior temporal lobectomy or amygdalo-hippocam-

pectomy); and (2) when discrepancies between the exis-

tence of hippocampal atrophy or other lesion and

electroclinical or/and SPECT data occurred, SEEG was

indicated for de®ning the contribution of remote temporal

or extratemporal areas to the epileptogenic zone. In the

latter group, it could happen that SEEG validated the exis-

tence of pure MTLE. SEEG recordings were performed as

described elsewhere (Chauvel et al., 1996). Multiple leads

(10±15) depth electrodes were placed intracranially accord-

ing to Talairach's stereotactic method (Talairach et al.,

1974) based on available information (determined from

non-invasive data) on the epileptogenic zone (de®ned as

the cerebral regions simultaneously discharging at seizure

onset). The strategy of electrode placement was individua-

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±1760 1747

lized depending on clinical, neurophysiological, and anato-

mical characteristics of each patient. However, all the

patients had electrodes that spatially sampled some

medial/limbic regions (amygdala and hippocampus) and

lateral/neocortical regions of the middle temporal gyrus

(T2). In some cases (see Table 1), electrodes were also

placed in other structures such as the temporal pole, the

ento/perirhinal cortex, the parahippocampal gyrus, the

superior temporal gyrus (T1), and the insula, as well as

various frontal and parieto-occipital regions to determine

the role of these different areas in the epileptogenic zone

and in the propagation of the discharges. Electrodes explor-

ing the temporal pole and the entorhinal region were more

recently introduced in our explorations. This explains why

in this series only 3 patients had these electrodes (patients 7,

8 and 15). Patients with bilateral TLE or multifocal epilepsy

(e.g. with independent temporal and frontal epileptogenic

zones) were excluded. In addition, patients with LTLE as

de®ned in our previous study (Bartolomei et al., 1999) were

not selected for this study.

2.2. Non-linear regression analysis

Non-linear regression analysis is aimed at quantifying,

using a parameter that takes its values within [0,1], how a

signal X from a given channel can be predicted from a signal

Y from another channel. The estimation of this parameter,

referred to as the non-linear correlation coef®cient h2*xy , is

performed on a temporal window of ®xed duration and slid-

ing on both signals in order to follow the temporal evolution

of the statistical relationship (linear or non-linear) between

both signals. Besides the estimation of h2*xy , a second quan-

tity is evaluated to obtain information on the causal property

of the relationship. This quantity, referred to as the direction

index Dxy, takes into account both an estimated time delay

between signals X and Y and the asymmetrical nature of the

non-linear correlation coef®cient. In order to study the

statistical properties of estimated quantities h2*xy and Dxy

and in order to interpret their behavior on real SEEG signals,

a physiological model of EEG generation is introduced

(Section 2.2.1). A summarized description of the non-linear

correlation coef®cient h2*xy and the direction index Dxy is then

given in Section 2.2.2.

2.2.1. Simulation of EEG signals from multiple coupled

neural populations

In the present work, an EEG model is used to physiolo-

gically interpret measures performed on SEEG signals. A

summarized description of this model is given in the follow-

ing. Readers may refer to our previous studies (Wendling et

al., 2000) for a detailed description.

As depicted in Fig. 1, the model follows a lumped-para-

meter approach in which neural populations are modeled as

non-linear dynamical systems based on neurophysiological

concepts (excitation, inhibition, synaptic ef®cacy, connec-

tivity constants). This approach has already led to realistic

macroscopic models of EEG generation, such as those

reported by Lopes da Silva et al. (1974, 1976) for alpha

rhythm representation.

In the context of SEEG exploration, the number of

recorded brain structures is always greater than one. Thus,

starting from the aforementioned works, we recently

extended the capabilities of this class of model such that it

can now generate vectorial EEG signals from multiple

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±17601748

Table 1

Patients' characteristics and chronology of brain structure involvement during seizuresa

No. Side of TLE MRI or histopathology A Hros Hcau TP EC T2 T1 F TPO P

1 R HS 1 (S,LVD) 1 (S,LVD) 2 (S) 3 (SW) 4 (SW) ± ± ±

2 L N 2 (S) 1 (LVD) 3 (S) 3 (SW) 4 (SW) ± ±

3 L HS 1 (S) 1 (LVD) 2 (S) 3 (SW) 4 (SW) ± ± 4 (SW)

4 L HS 2 (S) 1 (S) 2 (S) 3 (SW) 4 (SW)

5 R HS 1 (S) ± 3 (S) 3 (SW) ± 3 (S)

6 L Hippocampal lesion 1 (S) 1 (S) 2 (LVD) 2 (LVD) 2 (LVD) ± ± ±

7 L HS 1 (S) 1 (S) 1 (S) 1 (S) 1 (LVD) 2 (SW) ± ± ± ±

8 R N 1 (S) 1 (S) 2 (S) 1 (LVD) 2 (SW) 3 (SW) 3 (SW) 4 (SW)

9 R HS 1 (S) 1 (S) 2 (S) 2 (SW) 3 (SW) 4 (SW)

10 R HS 1 (S) 1 (S,LVD) 2 (S) 3 (SW) 4 (SW) 4 (SW) 4 (SW)

11 R HS 2 (LVD) 1 (LVD) 2 (LVD) 2 (LVD) 1 (LVD) 3 (SW) ± 3 (SW)

12 L HS 1 (S) 1 (S) 3 (S) 2 (LVD) 4 (S) ± ± 4 (SW)

13 L HS 2 (LVD) 2 (LVD) 1 (LVD) 2 (LVD) 3 (S) 3 (S) ± ±

14 L HS 1 (LVD) 4 (SW) 2 (LVD) 1 (LVD) 5 (SW) 4 (SW) ± ±

15 L Ganglioglioma PHG 1 (LVD) ± 1 (LVD) 1 (LVD) 1 (LVD) 1 (LVD) 2 (S) ± 3 (SW) ±

16 L N 1 (LVD) 1 (LVD) 1 (LVD) 2 (SW) 4 (SW) ± ± ±

17 L Cortical dysplasia MTG 1 (LVD) 1 (LVD) 1 (LVD) 2 (SW) 3 (SW) ±

18 L Cortical dysplasia MTG 1 (S) 1 (S) 2 (S) 1 (S) 4 (SW) ± 3 (S) 3 (S)

a F, female; M, male; (), aspect of the discharge; S, spikes; LVD, low-voltage fast discharge; slow W, slow waves; SW, spike-waves; le, lesion identi®ed by

MRI; A, amygdala; Ha, anterior part of the hippocampus; Hp, posterior part of the hippocampus; T2, middle temporal gyrus; T1, superior temporal gyrus; F,

prefrontal region; PHG, parahippocampal gyrus; EC, entorhinal cortex; HS, hippocampal sclerosis; N, normal.

distant coupled neural populations. Taking into account

current hypotheses about epileptic mechanisms, we also

altered model parameters to make the model produce

signals that qualitatively resemble real SEEG signals

recorded on more than one channel during interictal or

ictal periods. Consequently, the model offers a framework

in which it is possible (i) to build networks with a priori

known couplings between populations and (ii) to generate

realistic epileptiform signals from these networks with

appropriate alteration of model parameters (local balance

of excitation and inhibition, degree and direction of

couplings between populations). In this study, the model

is used to relate coupling parameters (degree/direction)

between neural populations to relations between EEG

signals they produce, these relations being evaluated using

non-linear regression.

2.2.2. Non-linear regression analysis and characterization

of couplings (degree/direction)

Given two SEEG signals X�t� and Y�t� recorded from two

neural groups Gx and Gy, respectively, two questions are to

be addressed.

1. How are we to quantify the synchronization (in a general

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±1760 1749

Fig. 1. In the model, a set of interconnected neural populations is considered. (a) Each population contains two subsets of neurons that mutually interact via

excitatory or inhibitory feedback: the main cells (i.e. pyramidal cells) and the local interneurons (i.e. other non-pyramidal cells, stellate or basket cells). The

in¯uence from neighboring neurons is modeled by an excitatory input that globally represents the average density of afferent action potentials. Each subset is

characterized by: (i) two dynamic linear transfer functions, he and hi that change the average pre-synaptic pulse density of action potentials into post-synaptic

membrane potential (either excitatory, EPSP, or inhibitory, IPSP); and (ii) a static non-linear function (sigmoid) that relates the average post-synaptic potential

of a given subset to an average pulse density of potentials ®red by the neurons. Moreover, interactions between main cells and interneurons are summarized in

the model by 4 connectivity constants C1±C4, which account for the average number of synaptic contacts. (b) Finally, since pyramidal cells are excitatory

neurons that project their axons to other areas of the brain, the model accounts 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. A connection from a given population i to a population j is

characterized by parameters Kij and hd, which represent the degree of coupling and the average time delay associated with this connection, respectively.

Kij is dimensionless. When Kij � 0, populations i and j are uncoupled and produced activities are independent. When Kij . 0, an excitatory connection from

population i to population j is simulated. Activity produced by population j is partially driven by that produced by population i. Details about the model and its

extension to multiple coupled neural populations may be found in Wendling et al. (2000).

sense) between the activity from Gx and that from Gy?

2. How are we to evaluate, from computation, which of the

following situations is the most plausible:

± Gx leads Gy,

± Gy leads Gx,

± Gx and Gy in¯uence each other,

± activities in Gx and Gy are independent?

To answer questions 1 and 2, the simplest and most classical

method consists in evaluating a quantity based on the linear

correlation coef®cient between random samples X�t� and

Y�t� present in the analysis window sliding on both signals.

This quantity is generally normalized such that it is equal to

0 when there exists no linear dependence between X�t� and

Y�t� and it is equal to 1 when Y�t� is an af®ne function of

X�t�.When close to 1.0, this quantity provides an answer to

question 1 because one can consider, in that case, that activ-

ities in Gx and Gy are strongly coupled or dependent. The

interpretation of a low value is more intricate. Indeed, it

only indicates a weak statistical linear relationship between

X�t� and Y�t�. However, it does not allow the non-linear

relationship hypothesis to be rejected and thus for us to

conclude that activities in Gx and Gy are independent.

Non-linear regression is a way to address the above dif®-

culty. In this method, a non-linear function is used in the

regression of Y�t� over X�t�. It allows two quantities, namely

the non-linear correlation coef®cient and the time delay,

denoted in the following by h2*xy and txy, respectively, to

be calculated from signals X�t� and Y�t�. These quantities

were ®rst introduced in the ®eld of EEG for the study of

associations between signals by Pijn and Lopes da Silva

(1993). A summarized description of these two quantities

is given in the following. Readers may refer to Pijn and

Lopes da Silva (1993) for a detailed description.

The non-linear correlation coef®cient h2*xy is obtained by

considering the amplitude y of signal Y�t 1 t� as a perturbed

function of the amplitude x of signal X�t� (i.e. the condi-

tional mean of Y�t 1 t� given X�t� � x). The variance of the

perturbation formally corresponds to the conditional

variance of Y�t 1 t� given X�t� (i.e. the reduction of

variance of Y�t 1 t� obtained by predicting y values from

x values). This conditional variance can be estimated from

the construction of a piecewise linear regression curve

Y�t 1 t� � h�X�t��. Here, it is also normalized by the

variance of Y�t 1 t� and the normalized value is subtracted

from 1 to obtain the following quantity:

h2xy�t� � 1 2

VAR�Y�t 1 t�=X�t��VAR�Y�t 1 t��

where

VAR�Y�t 1 t�=X�t�� � arg minh

�E�Y�t 1 t�2 h�X�t���2�

The above procedure can be reiterated for different values of

t between X and Y to obtain the time shift txy for which

h2xy�t� is maximum:

h2*xy � max�h2

xy�t�� � h2xy�txy�

tmin , t , tmax

h2*xy is called the `non-linear correlation coef®cient'. Its

values are contained between 0 (Y is independent of X)

and 1 (X and Y are linearly or non-linearly dependent) and

provide an answer to question 1. Question 2 can also be

answered using non-linear regression. Unlike a linear corre-

lation coef®cient, h2*xy is not symmetric and generally h2*

xy is

different from h2*YX . As described by Arnhold et al. (1999)

and although the relationship is less evident, this asymmetry

can provide insight into causality relations: for given X and

Y signals of the same order of complexity and for a given

prediction error obtained from a non-linear asymmetric

interdependence estimator, if Y is partially driven by X,

then the difference Dh2 � h2*xy 2 h2*

YX is positive. In this

same case, the difference between time delays when

measured from X to Y�txy� and from Y to X�txy�,Dt � txy 2 tYX , is also positive.

Since both quantities (asymmetry and time delay) contain

causality information, we proposed to conjointly use them

in a new quantity (referred to as the `direction index' Dxy)

aimed at characterizing the direction of coupling between

neural populations that produce signals:

Dxy � 12�sgn�Dh2�1 sgn�Dt��

where sgn denotes the sign function.

In theory, if there exists a relation between signals X and

Y and if Dxy � 11 (21), then Y (X) is in¯uenced by X (Y). In

practice, signals X�t� and Y�t� are non-stationary and a slid-

ing window is used to locally analyze the relationship

between signal X�t� and signal Y�t�. Over this window,

Dh2and Dt are computed at each sliding step and Dxy is

estimated from the averaged sum of their respective signs

Dxy, over a small number of steps. Consequently, Dxy values

are contained between 21 and 11.

As described in Section 3.1 on signals simulated with the

model, a Dxy value close to 11 (21) conjointly with a

signi®cantly high value of h2*xy indicates a unidirectional

coupling from population Gx to population Gy (from popu-

lation Gy to population Gx), whereas a value close to 0

indicates a bidirectional coupling between the two popula-

tions. Finally, it must also be mentioned that we showed that

Dxy provides a better estimate of the direction of coupling

than the time delay or the asymmetry property taken sepa-

rately (Wendling et al., 2001).

As far as the recording technique is concerned, our

system allows SEEG signals to be recorded on 128 channels

at a sampling frequency of 200 Hz with respect to a

common reference signal. In our signal processing analysis,

we do not use monopolar but bipolar signals that are built by

subtracting signals recorded on two adjacent sites located in

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±17601750

a given cerebral structure. This operation allows the refer-

ence signal common to all recorded signals to be removed.

3. Results

3.1. Results obtained on simulated signals

3.1.1. Unidirectionally coupled populations

The model con®guration used to simulate signals from

two populations unidirectionally coupled is presented in

Fig. 2a. The E=I ratio (that was shown to control the degree

of excitability in the model) in population 1 is increased

such that the activity from this population becomes epilepti-

form (rhythmic discharge of spikes). Population 2 is main-

tained at standard values; the corresponding EEG re¯ects a

normal background activity. Starting from zero, the

coupling value from population 1 to population 2, denoted

by K12, is progressively increased in steps of 25 units. For

small values of K12, sporadic spikes start to appear in the

signal generated by population 2. As K12 increases, these

spikes become more and more frequent and tend to synchro-

nize with those originating from population 1. Finally, for

the highest values of K12, both populations generate rhyth-

mic synchronized spikes. For each value of K12, the non-

linear correlation coef®cients h2xy and h2

yx and the direction

index Dxy are computed on long-duration simulated signals

(300 s) using a 5 s duration window (i.e. 1000 samples at a

sampling rate of 200 Hz) sliding over signals by steps of

0.5 s (i.e. 100 samples).

Experimental averaged values and standard deviations

(SDs) of these quantities are represented in Fig. 2b,c. Both

averaged h2xy and h2

yx increase with K12 from a null value to a

value of 0.85. One can notice that h2xy (solid line) is slightly

greater than h2yx (dashed line). Fig. 2c gives the evolution of

the average direction index Dxy. Some important points

about this quantity can be emphasized. First, it signi®cantly

deviates from 0 (2SD at least) even for relatively small

values of coupling between population 1 and population 2.

Second, from K12 � 175, h2xy is greater than 0.5 and Dxy is

greater than 0.8, clearly indicating a unidirectional coupling

of relatively high degree from population 1 to population 2.

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±1760 1751

Fig. 2. Simulation of EEG signals from two neural populations (a) unidirectionally coupled and (b) bidirectionally coupled. Computation of non-linear

regression coef®cient h2xy (solid line) and h2

yx (dashed line) and direction index Dxy for increasing degree of coupling in both con®gurations. See text for detailed

interpretation (Section 3.1).

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±17601752

Fig. 3. Non-linear regression analysis in a patient (patient 1) with a medial

seizure (MTLEs). (a) Representation of the onset of seizures and the corre-

sponding epoch on SEEG traces. (b) Three epochs of 10 s are analyzed at

the onset of seizures. A, amygdala; H, anterior hippocampus; NC, neocor-

tex of the middle temporal gyrus. h2 refers to the non-linear correlation

coef®cients h2*xy and h2*

YX , represented by two traces. Dxy, direction index. See

text for detailed interpretation.

Fig. 4. Non-linear regression analysis in a patient (patient 11) with a

medial±lateral seizure (MLTLEs). (a) Representation of the onset of

seizures and the corresponding epoch on SEEG traces. (b) Three epochs

of 10 s are analyzed at the onset of seizures. A, amygdala; H, anterior

hippocampus; NC, neocortex of the middle temporal gyrus. h2 refers to

the non-linear correlation coef®cients h2*xy and h2*

YX , represented by two

traces. Dxy, direction index. See text for detailed interpretation.

In other words, these two values characterize the fact that

the activity generated by population 2 is causally deter-

mined (or `driven') by that generated by population 1.

3.1.2. Bidirectionally coupled populations

The model con®guration used to simulate signals from

bidirectionally coupled populations is represented in Fig.

2d. Now, E=I ratios are slightly increased (relative to the

standard value) to make both populations more excitable;

epileptiform events (spikes) sporadically appear in the

activity produced by both populations. Moreover, for the

two populations to in¯uence each other identically, coupling

parameters are set to equal values �K12 � K21 � K�.Starting from null couplings �K � 0� between the two

populations for which spikes are not correlated, we progres-

sively increased the value of K, as in the previous case.

Visual inspection of simulated signals shows that both activ-

ities synchronize faster when couplings are bidirectional.

For the highest values of K, both populations generate

sustained discharges of spikes whose average frequency is

also higher than in the unidirectional case. Fig. 2e gives

averaged values of h2xy (solid line) and h2

yx (dashed line) as

K increases. Now, values range from 0.0 to 0.96, which is

higher than for unidirectional coupling. Fig. 2f gives the

curve representing the evolution of the averaged direction

index Dxy as a function of K. The combined features of

asymmetry and time delays are re¯ected in this curve: it

does not diverge from 0 and the SD is higher than for

unidirectional coupling. The behavior of quantities

described above can be explained by the model con®gura-

tion used to generate signals: since both populations

mutually and symmetrically in¯uence each other, neither

of the two activities is determined by the other. The method

clearly re¯ects this situation through the behavior of h2xy and

Dxy measured on signals and characterizing the degree and

direction of coupling, respectively.

3.2. Interactions between the middle temporal gyrus, the

amygdala, and the anterior hippocampus in patients with

TLE

Main patient characteristics and results of SEEG record-

ings are summarized in Table 1. Visual analysis of ictal

patterns demonstrated that interindividual variability was

high but intraindividual reproducibility was good. The inter-

actions between medial (hippocampus and amygdala) and

lateral neocortical (middle temporal gyrus) structures during

ictal periods at the onset of seizure were quanti®ed by non-

linear regression analysis as described earlier using bipolar

SEEG signals (determination of non-linear correlation coef-

®cients h2xy and h2

yx followed by analysis of direction index

Dxy for values with signi®cantly high correlation coef®-

cients).

Tables 2±4 summarize correlation coef®cients and direc-

tion indexes obtained during onset of seizure. Maximal

values statistically different from the interictal period (.2

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±1760 1753

Fig. 5. Non-linear regression analysis in a patient (patient 17) with a lateral±

medial seizure (LMTLEs). (a) Representation of the onset of seizures and

the corresponding epoch on SEEG traces. (b) Three epochs of 10 s are

analyzed at the onset of seizures. A, amygdala; H, anterior hippocampus;

NC, neocortex of the middle temporal gyrus. h2 refers to the non-linear

correlation coef®cients h2*xy and h2*

YX , represented by two traces. Dxy, direc-

tion index. See text for detailed interpretation.

SD) are given. The coupling pattern was divided into 3

epochs of 10 s. Full results for a typical patient in each

group are shown (Figs. 3±5). Seizure networks can be clas-

si®ed under 3 `generic' headings as follows.

3.2.1. Medial networks

Patients 1±10 presented seizures characterized by initial

discharge limited to medial limbic structures. MRI and/or

histopathological examination after surgery revealed hippo-

campal sclerosis in most of these patients (Table 1). All but

one patient presented ictal discharges in both the hippocam-

pus and amygdala. The exception was patient 2 who

presented exclusive involvement of the hippocampus at

the onset of seizure. Table 2 summarizes the results of

non-linear regression analysis in the 10 patients with

seizures characterized by medial networks. A characteristic

®nding in this group was the absence of signi®cant interac-

tion between medial and neocortical structures at the onset

(epoch 1). In contrast, h2xy values between amygdala and

hippocampus were high. Based on the coupling direction

index Dxy, the `leader' structure was the amygdala in 3

cases and the hippocampus in 3 cases. Coupling was bidir-

ectional in 3 cases.

Subsequent functional coupling between amygdala and

hippocampus followed several patterns. The most frequent

was a strong decrease concurrent with the appearance of

low-voltage fast discharges (patients 1±5 and 9). Coupling

between medial and lateral structures (usually the amygdala

and neocortex) sometimes appeared during epoch 2 or 3. In

these cases, medial structures rather than the neocortex

always played the leader role although subsequent inversion

of this relationship during epoch 3 was observed in 3 cases

(patients 6, 8, and 9), suggesting that the neocortex can be

an active secondary relay. Fig. 3 shows the course of non-

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±17601754

Table 2

Non-linear correlation analysis of coupling between amygdala (A), hippocampus (H), and neocortex (NC) in patients with medial networksa

Cases Epoch 1 Epoch 2 Epoch 3

A±H A±NC H±NC A±H A±NC H±NC A±H A±NC H±NC

1 ± ± ± ± ± ± ±

h2xy (max) 0.71 0.89

Dxy A! H (0.7) A! H (0.6)

2 ±* ± ± ± ± ± ±

h2xy (max) 0.35 0.40

Dxy A Ã H N! A (0.7)

3 ± ± ± ± ±

h2xy (max) 0.40 0.30 0.53 0.39

Dxy A! H (0.8) A! NC (0.5) H$ A A! NC (0.7)

4 ± ± ± ± ± ± ±

h2xy (max) 0.40 0.48

Dxy H! A (0.7) A! H (0.6)

5 ± ± ± ± ±

h2xy (max) 0.45 0.25 0.50 0.30

Dxy H$ A NC! H (0.5) H$ A NC! H (0.6)

6 ± ±

h2xy (max) 0.49 0.28 0.57 0.64 0.74 0.50 0.52

Dxy H$ A A! H (0.5) A! NC (0.4) H! NC (0.6) A$ H NC! A (0.7) H!NC (0.8)

7 ± ± ± ± ±

h2xy (max) 0.45 0.60 0.45 0.70

Dxy A$ H A$ H A$ H A! NC (0.7)

8 ± ± ± ± ±

h2xy (max) 0.60 0.56 0.32 0.40

Dxy A! H (0.7) H! A (0.6) A! H (0.6) A! NC (0.5)

9 ± ± ± ± ±

h2xy (max) 0.30 0.27 0.49 0.37

Dxy H! A (0.6) A! NC (0.8) H! A (0.7) NC! A (0.7)

10 ± ± ± ± ±

h2xy (max) 0.59 0.32 0.55 0.30

Dxy H! A (0.8) H! A (0.5) A$ H NC! H (0.6)

a Three epochs of 10 s are shown in which the maximal values of the non-linear regression coef®cient h2xy are given when they are statistically different from

interictal period (.2 SD). For unidirectional coupling, the maximal value of the index Dxy is indicated.

linear regression analysis during a medial seizure in patient

1. A clear-cut increase in interdependency between the

amygdala and hippocampus was observed at the same

time as appearance of spiking activities in these two struc-

tures. The direction index Dxy increased to a value of 10.8,

suggesting that the discharge was driven by the amygdala.

When the fast low-voltage discharges emerged, the h2xy

value decreased (epoch 2). Late occurrence of a second

phase of synchronization (epoch 3) with high h2xy (maximum

0.7) and Dxy values indicated that the amygdala played the

leader role. The end of the seizure was marked by a rebound

in the h2xy value. No signi®cant interdependency was

observed between medial and neocortical structures.

3.2.2. Medial±lateral networks

Medial±lateral networks were observed in patients 11±15.

MRI and/or histopathological examination of surgical speci-

mens revealed hippocampal sclerosis in 4 cases (patients

11±14) and a dysplastic lesion (ganglioglioma) involving

the parahippocampal gyrus in one (patient 15) (Table 1).

Seizure started either by a short period of spiking activity

over the amygdala and hippocampus (cases 11±14) or by a

slow wave complex in the parahippocampal region (case 15)

followed by fast low-voltage discharges affecting both the

medial structures and the neocortex. Table 3 shows the

result of h2xy analysis in these patients. During epoch 1,

high h2xy values were observed between medial and lateral

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±1760 1755

Table 3

Non-linear correlation analysis of coupling between amygdala (A), hippocampus (H), and neocortex (NC) in patients with medial±lateral networksa

Cases Epoch 1 Epoch 2 Epoch 3

A±H A±NC H±NC H±A A±NC H±NC H±A A±NC H±NC

11 ± ± ± ± ±

h2xy (max) 0.35 0.55 0.45 0.33

Dxy A! H (0.8) A! NC (0.7) A! NC (0.7) A! H (0.8)

12 ± ±

h2xy (max) 0.80 0.60 0.70 0.41 0.75 0.40 0.40

Dxy A! H (0.9) A! NC (0.8) H! NC (0.7) N$ A H! A (0.8) N! A (0.6) N!H (0.5)

13 ± ± ±

h2xy (max) 0.36 0.33 0.61 0.61 0.45 0.35

Dxy A$ H A! NC (0.7) H! NC (0.8) H!NC (0.9) A! H (0.8) A! NC (0.7)

14 ± ± ± ±

h2xy (max) 0.61 0.40 0.60 0.39 0.90

Dxy A! NC (0.6) H! A (0.7) A$ H (0.8) NC! A (0.7) H$NC

15 ± ± ±

h2xy (max) 0.43 0.53 0.41 0.30 0.50 0.50

Dxy H! A (0.6) A! NC (0.7) H! NC (0.8) H! A (0.7) NC! A (0.8) H!NC (0.8)

a Three epochs of 10 s are shown in which the maximal values of the non-linear regression coef®cient h2xy are given when they are statistically different from

the interictal period (.2 SD). For unidirectional coupling, the maximal value of the index Dxy is indicated.

Table 4

Non-linear correlation analysis of coupling between amygdala (A), hippocampus (H), and neocortex (NC) in patients with lateral±medial networksa

Cases Epoch 1 Epoch 2 Epoch 3

A±H A±NC H±NC A±H A±NC H±NC A±H A±NC H±NC

16 ± ± ±

h2xy (max) 0.53 0.40 0.42 0.40 0.53 0.50

Dxy H! A (0.8) NC! A (0.6) NC! H (0.7) A! NC (0.6) H! A (0.6) H$NC

17

h2xy (max) 0.45 0.69 0.55 0.40 0.5 0.60 0.60 0.67 0.30

Dxy H! A (0.6) NC! A (0.6) NC! H (0.4) H! A (0.5) NC! A (0.7) NC! H (0.6) H! A (0.4) A! N (0.6) H!NC (0.4)

18 ± ± ± ±

h2xy (max) 0.54 0.40 0.4 0.41 0.65

Dxy A! H (0.6) NC! A (0.4) A! H (0.5) NC! A (0.4) NC$ A (0.4)

a Three epochs of 10 s are shown in which the maximal values of the non-linear regression coef®cient h2xy are given when they are statistically different from

interictal period (.2 SD). For unidirectional coupling the maximal value of the index Dxy is indicated.

structures as well as between medial structures. It is note-

worthy that coupling with the neocortex was already present

even when rhythmic modi®cations appeared over medial

regions. The fundamental difference between this group

and the preceding one was medial±lateral coupling.

Coupling was found to be either bidirectional or unidirec-

tional to the amygdala. The neocortex always appeared to be

driven by medial structures (amygdala or amygdala and

hippocampus) at the onset of seizure.

Fig. 4 depicts the typical pattern of non-linear regression

analysis in a patient with MLTLE. The ®rst epoch was

characterized by an increase in h2xy values between medial

structures (amygdala and hippocampus). The direction

index Dxy indicated that the amygdala was the leader struc-

ture in the initial phase of coupling. This second epoch was

marked by a decrease in coupling between the amygdala and

hippocampus coinciding with the appearance of fast low-

voltage discharge.

Signi®cant coupling was also demonstrated in the ®rst

epoch between the medial structures and neocortex (h2xy

value reaching 0.6 (amygdala±neocortex) and 0.8. (hippo-

campus±neocortex)). The direction index demonstrated that

the medial structure played the leader role in this coupling.

Initial coupling was followed by a decrease (again concur-

rent with low-voltage discharges) followed by coupling

between these regions.

3.2.3. Lateral±medial networks

The lateral±medial network subtype was observed in

patients 16±18. Results of non-linear analysis are shown

in Table 4. These networks were associated with lesions

involving the anterior temporal neocortex in two cases

(Table 1). The seizure pattern appeared to consist of a

rapid low-voltage discharge involving both medial and

lateral structures. The main characteristic of this group

was the existence of initial coupling between the neocortex

and medial structures (amygdala and/or hippocampus). The

main difference from the MLTLE group was the direction of

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±17601756

Fig. 6. Seizure and medial network according to the interactions between several temporal lobe structures in patient 7. The maximal h2xy values during the ®rst

10 s of seizure (epoch 1) are given for signi®cant interactions. The direction of the coupling indicated by the arrows is given by the direction index Dxy. A,

amygdala; Ha, anterior hippocampus; ant T2, anterior part of the middle temporal gyrus; midT2, mid-part of the middle temporal gyrus; TP, temporopolar

region; T1, ®rst temporal gyrus; Ins, insula; TP l, lateral part of the temporal pole; TPm, medial part of the temporal pole.

coupling since the neocortex was the initial leader and

medial structures appeared to be driven. Major non-linear

interdependency was also observed between amygdala and

hippocampus in these patients.

Fig. 5 shows the course of couplings associated with

LMTLE seizure in patient 17. During epoch 1, h2xy coef®-

cients between the neocortex and amygdala reached values

up to 0.5 while Dxy rose to 10.9. This ®nding indicates that

the leader role was played by the neocortex. During epoch 2,

h2xy rose signi®cantly in comparison to interictal values

(maximum 0.8). At the same time, Dxy values remained

around 11.0. Interpretation of simulation results suggests

that neural populations in lateral structures tended to couple

with ± and drive ± neural populations from medial struc-

tures. During epoch 2, major coupling was observed

between the amygdala and hippocampus and between the

hippocampus and neocortex.

3.2.4. An approach to the complexity of involved neural

networks

The results presented in Section 3.1 are based on non-

linear regression evaluation of the interactions between the

amygdala, anterior hippocampus, and neocortex of the

middle temporal gyrus. Patients 7, 8 and 15 had additional

electrodes exploring other medial limbic regions, i.e. the

temporal pole and the entorhinal/perirhinal cortex. Patients

7 and 8 presented medial seizures and non-linear regression

analysis of the coupling revealed that the networks involved

all explored limbic neural structures including the medial

part of the temporal pole. The leader structure in these

networks was the amygdala±hippocampus complex (Figs.

6 and 7). In patient 8, strong bidirectional coupling occurred

between the amygdala, hippocampus, and entorhinal cortex,

which drove the temporal pole. These two cases show that

medial networks can extend beyond the classical amygdala±

hippocampus complex.

Fig. 8 shows the network involved in patient 15. The

onset of seizure was characterized by a slow wave complex

in various structures, including the medial structures, para-

hippocampal gyrus (lesional site), and temporal neocortex,

followed by a fast low-voltage discharge. Analysis of h2xy

values at the onset of the seizure showed that these struc-

tures are functionally linked in a complex medial±lateral

network. In this network, the lesional site played the leader

role.

4. Discussion

This study was focused speci®cally on interactions

between the temporal neocortex and limbic structures and

the relative contribution of these structures to the generation

of TLE seizures. On the basis of Pijn and Lopes da Silva

(1993), our ®ndings showed that non-linear regression

allows not only reliable quanti®cation of relationships

between signals in epileptic partial seizures. The results

described here are based on a physiological model of EEG

data. This model differs from the more `external' models

(autoregressive, for example) classically used to evaluate

EEG signal processing (Lopes da Silva et al., 1975; Kocarev

and Parlitz, 1996). Our model allows us to explicitly repre-

sent both degree and direction of coupling, thus allowing

construction of the neural systems inside controlling caus-

ality relationships and generation of multichannel EEG

signals from these systems. These signals provide an objec-

tive way of studying the performance of non-linear regres-

sion analysis and allow comparison of h2xy and Dxy

coef®cients computed on real signals.

In a previous study we de®ned 4 subtypes of TLE, namely

medial, medial±lateral, lateral±medial and lateral (Bartolo-

mei et al., 1999). The results of this study con®rm and

extend our knowledge about these spatio-temporal patterns

in several respects. One of the most important results is that

the epileptogenic zone giving rise to initial ictal discharge

can be more accurately de®ned as a network of neural struc-

tures rather than as a focus of abnormal activity. This de®-

nition implies the existence of `abnormally facilitated

connections' between neuronal groups that display stereo-

typed behavior during seizures.

In medial networks, the ictal discharge is limited to the

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±1760 1757

Fig. 7. Medial network in patient 8 including the entorhinal cortex and the

medial part of the temporal pole. The maximal h2xy values during the ®rst

10 s of seizure (epoch 1) are given for signi®cant interactions. The direction

of the coupling indicated by the arrows is given by the direction index Dxy.

A, amygdala; Ha, hippocampus anterior part; T2, anterior part of the second

temporal gyrus; T1, ®rst temporal gyrus; Ins, insula; TP l, lateral part of the

temporal pole; TPm, medial part of the temporal pole; EC, entorhinal

cortex.

medial limbic structures and may propagate secondarily to

the cortex. The initial discharge may be phasic (clonic) or

tonic (rapid) but propagation is always clonic (Bartolomei et

al., 1999). Quanti®ed results in the present study demon-

strated no signi®cant coupling between medial and lateral

structures at the beginning of the seizures. Conversely,

almost constant coupling was observed between hippocam-

pus and amygdala, thus demonstrating that these two struc-

tures were involved in generation of seizures. Our data also

indicated that this coupling could be unidirectional or bidir-

ectional. Previous studies using linear methods (Lieb et al.,

1987; Duckrow and Spencer, 1992; Gotman and Levtova,

1996) described major interactions between medial struc-

tures during TLE seizures. Findings of a recent study (Le

Van Quyen et al., 1998) involving a new method of non-

linear evaluation of SEEG signal interdependence (mutual

non-linear prediction) showed that non-linear relationships

take place during seizures between several regions of the

medial structures during TLE seizures. However, this study

was limited to the interactions within the limbic medial

structures of the temporal lobe.

In MTLE, initial coupling between medial structures

seems to be a prerequisite condition to initiate ictal activity.

This initial phase is generally followed by a second phase in

which a strong decrease in coupling is observed in parallel

with development of fast low-voltage discharges. These

®ndings have already been reported in linear (Lieb et al.,

1987; Duckrow and Spencer, 1992) as well as other non-

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±17601758

Fig. 8. Seizure and complex medial±lateral network (patient 15) including the medial limbic structures (A, amygdala; Ha, anterior hippocampus; PHG,

parahippocampal gyrus posterior part; EC, entorhinal cortex) and the neocortical part of the left temporal lobe (T3, inferior temporal gyrus; T2, middle

temporal gyrus). The maximal h2xy values during the ®rst 10 s of seizure (epoch 1) are given for signi®cant interactions. The direction of the coupling indicated

by the arrows is given by the direction index Dxy.

linear studies (Le Van Quyen et al., 1998). A possible

mechanism could be a loss of synchronization in the activity

of the two structures. This phenomenon could result from

functional disconnections between structures. In this regard,

experimental evidence has shown that generation of fast

rapid discharges in neuron clusters depletes extracellular

calcium concentrations to levels no longer consistent with

synaptic transmission (Jefferys, 1990; Wheal, 1990).

Compensatory mechanisms as well as slowing of the

discharge rate could re-establish the connections.

In medial±lateral and lateral±medial networks, the ictal

discharge includes both limbic and neocortical regions. A

rapid `tonic' discharge is observed over the temporal

neocortex at the onset of seizure (Bartolomei et al., 1999).

Quantitative analysis showed an initial increase in the non-

linear correlation coef®cient and direction index. This ®nd-

ing would indicate an in¯uence of medial over lateral struc-

tures (medial±lateral) or of the lateral neocortex over medial

structures (lateral±medial). The amygdala probably plays a

central role in these networks since cortical±medial inter-

dependence always involves this structure. This result is in

agreement with previous data (Gotman and Levtova, 1996)

showing that the amygdala was the most common

discharge-initiating structure in patients presenting neocor-

tical involvement during TLE seizures. This result can be

interpreted in terms of connectivity between these struc-

tures. Connections between the amygdala and temporal

cortex can be described as strong, direct, and bidirectional

while connections with the hippocampus are indirect,

passing through the entorhinal cortex (Amaral and Insausti,

1990). The existence of these dense direct connections

could explain why the amygdala is more often involved

than the hippocampus in these cortico-limbic networks.

Our methodological approach seemed to be well suited to

de®ning the involved networks and identifying leader struc-

tures in these networks. In the present study, investigation of

limbic structures was limited mainly to interactions between

the amygdala and hippocampus. However, involvement of

other functionally interconnected structures in these

networks is likely. In this regard, our preliminary results

from the patients with electrodes in the temporal pole and/

or peri/entorhinal region indicated that these regions might

be involved in the epileptogenic network in addition to the

amygdalo-hippocampal complex. Further study in a larger

population of patients will be needed to assess involvement

of other limbic or paralimbic regions in medial networks or

in medial±lateral networks.

Our ®ndings are in agreement with previous reports

including those of Bancaud et al. who were the ®rst to

propose the existence of several TLE subtypes (Bancaud

et al., 1965; Bancaud, 1981) and with other reports describ-

ing simultaneous involvement of the temporal neocortex

and medial structures during TLE seizures (Wieser, 1983;

Spencer et al., 1990; Munari et al., 1994). It is noteworthy

that neuroimaging and/or pathological features in patients

presenting medial and medial±lateral networks are consis-

tent with primary disease in temporal limbic structures,

while those in patients presenting lateral±medial networks

are consistent with primary pathology of the neocortex. This

suggests that hippocampal sclerosis would not be the discri-

minating factor since it is usually found in the ®rst two

subtypes. Thus, there is convincing evidence that the clas-

sical distinction between medial and lateral TLE is insuf®-

cient and oversimpli®ed.

A more appropriate classi®cation and better understand-

ing of the network organization of the epileptogenic zone

would assist the surgeon in achieving `optimal' delineation

of the volume of tissue to be excised. It is generally

acknowledged that MTLE has a better surgical prognosis

even though there is little supportive evidence in the litera-

ture. Standardized surgical procedures limited to medial

temporal structures (amygdalo-hippocampectomy) or

extended to the anterior temporal lobe (anterior temporal

lobectomy) are performed in most cases. In about 35% of

cases, however, patients are not seizure-free after such

procedures (Engel et al., 1993). Failure may be due to the

complexity of interactions between neocortical and limbic

temporal lobe structures within the epileptogenic zone,

particularly when a standardized surgical procedure is

performed.

References

Amaral D, Insausti R. Hippocampal formation. In: Paxinos G, editor. The

human nervous system, San Diego, CA: Academic Press, 1990. pp.

711±757.

Arnhold J, Grassberger P, Lenhertz K, Elger CE. A robust method for

detecting interdependences: application to intracranially recorded

EEG. Physica D 1999;134:419±430.

Badier J, Chauvel P. Spatio-temporal characteristics of paroxysmal inter-

ictal events in human temporal lobe epilepsy. J Physiol (Paris)

1995;89:255±264.

Bancaud J. Epileptic attacks of temporal lobe origin in man. Jap J EEG

EMG 1981;Suppl:61±71.

Bancaud J, Talairach J, Bonis A, Schaub C, Szikla G, Morel P, Bordas-

Ferrer M. La steÂreÂoeÂlectroenceÂphalographie dans l'eÂpilepsie: informa-

tions neurophysiopathologiques apporteÂes par l'investigation fonction-

nelle stereotaxique Paris: Masson, 1965.

Bartolomei F, Wendling F, Vignal J, Kochen S, Bellanger J, Badier J, Le

Bouquin-Jeannes R, Chauvel P. Seizures of temporal lobe epilepsy:

identi®cation of subtypes by coherence analysis using stereo-electro-

encephalography. Clin Neurophysiol 1999;110:1741±1754.

Burgerman R, Sperling M, French J, Saykin A, O'Connor M. Comparison

of medial versus neocortical onset temporal lobe seizures: neurodiag-

nostic ®ndings and surgical outcome. Epilepsia 1995;36:662±670.

Cascino G, Jack C, Sharbrough F, Kelly P, Marsh W. MRI assessments of

hippocampal pathology in extra-temporal lesional epilepsy. Neurology

1993;43(Suppl 2):A383.

Casdagli M, Iasemidis L, Savit R, Gilmore R, Roper S, Sackellares J. Non-

linearity in invasive EEG recordings from patients with temporal lobe

epilepsy. Electroenceph clin Neurophysiol 1997;102:98±105.

Chabardes S, Kahane P, Hoffman D, Munari C, Benabid AL. Role of the

temporo-polar region in the genesis of temporal lobe seizures. Epilepsia

1999;40(Suppl 7):78.

Chauvel P, Vignal J, Biraben A, Badier J, Scarabin J. Stereoelectroence-

phalography. In: Pawlik G, Stefan H, editors. Multimethodological

assessment of the epileptic forms, New York, NY: Springer, 1996.

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±1760 1759

Commission on Classi®cation and Terminology of the International League

Against Epilepsy. Commission on Classi®cation and Terminology of

the International League Against Epilepsy: proposal for revised classi-

®cation of epilepsies and epileptic syndromes. Epilepsia 1989;30:389±

399.

Duckrow R, Spencer S. Regional coherence and the transfer of ictal activity

during seizure onset in the medial temporal lobe. Electroenceph clin

Neurophysiol 1992;82:415±422.

Ebner A. Lateral (neocortical) temporal lobe epilepsy. In: Wolf P, editor.

Epileptic seizures and syndromes, London: John Libbey and Co, 1994.

pp. 375±382.

Engel J, VanNess P, Rasmussen T, Ojemann L. Outcome with respect to

epileptic seizures. In: Engel J, editor. Surgical treatment of the epilep-

sies, 2nd ed. New York, NY: Raven Press, 1993. pp. 609±622.

French J, Williamson P, Thadani V, Darcey T, Mattson R, Spencer S,

Spencer D. Characteristics of medial temporal lobe epilepsy: I. Results

of history and physical examination. Ann Neurol 1993;34:774±780.

Gotman J, Levtova V. Amygdala±hippocampus relationships in temporal

lobe seizures: a phase coherence study. Epilepsy Res 1996;25:51±57.

Isnard J, Guenot M, Ostrowsky K, Sindou M, MauguieÁre F. The role of the

insular cortex in temporal lobe epilepsy. Ann Neurol 2000;48:614±623.

Jefferys J. Basic mechanisms of focal epilepsies. Exp Physiol 1990;75:127±

162.

Kocarev L, Parlitz U. Generalized synchronization, predictability and

equivalence of unidirectionally coupled dynamical systems. Phys Rev

Lett 1996;76:1816±1819.

Le Van Quyen M, Adam C, Baulac M, Martinerie J, Varela F. Nonlinear

interdependencies of EEG signals in human intracranially recorded

temporal lobe seizures. Brain Res 1998;792:24±40.

Levesque M, Nakasato N, Vinters H, Babb T. Surgical treatment of limbic

epilepsy associated with extratemporal lesions: the problem of dual

pathology. J Neurosurg 1991;75:364±370.

Lieb J, Hoque K, Skomer C, Song X. Interhemispheric propagation of

human medial temporal lobe seizures: a coherence/phase analysis. Elec-

troenceph clin Neurophysiol 1987;67:101±119.

Lopes da Silva F, Pijn J. Epilepsy: network models of generation. In: Arbib

M, editor. The handbook of brain theory and neural networks,

Cambridge, MA/London: Bradford/The MIT Press, 1995. pp. 367±369.

Lopes da Silva F, Hoek AHS, Zetterberg L. Model of brain rhythmic

activity. Kybernetic 1974;15:27±37.

Lopes da Silva F, Dijk A, Smits H. Detection of non-stationarities in EEGs

using the autoregressive model ± an application to EEGs of epileptics.

In: Dolce H, Kunkel, editors. CEAN ± computerized EEG analysis,

Stuttgart: Fischer, 1975. pp. 180±199.

Lopes da Silva F, van Rotterdam A, Barts P, van Heusden E, Burr W.

Models of neuronal populations: the basic mechanisms of rhythmicity.

Prog Brain Res 1976;45:281±308.

Mars N, Lopes da Silva F. Propagation of seizure activity in kindled dogs.

Electroenceph clin Neurophysiol 1983;56:194±209.

Munari C, Tassi L, Kahane P, Francione S, DiLeo M, Quarato P. Analysis

of clinical symptomatology during stereo-EEG recorded mesiotemporal

lobe seizures. In: Wolf P, editor. Epileptic seizures and syndromes,

London: John Libbey and Co, 1994.

Pijn J, Lopes da Silva F. Propagation of electrical activity: nonlinear asso-

ciations and time delays between EEG signals. In: Zschocke, Speck-

mann, editors. Basic Mechanisms of the EEG, Boston, MA: Birkauser,

1993.

Pijn J, Van Neerven J, Noest A, Lopes da Silva F. Chaos or noise in EEG

signals; dependance on state and brain site. Electroenceph clin Neuro-

physiol 1991;79:371±381.

Pijn J, Velis DN, Lopes da Silva F. Measurement of interhemispheric time

differences in generalised spike-and-wave. Electroenceph clin Neuro-

physiol 1992;83:169±171.

Spencer S, Spencer D, Williamson P, Mattson R. Combined depth and

subdural electrode investigation in uncontrolled epilepsy. Neurology

1990;40:74±79.

Talairach J, Bancaud J, Szickla G, Bonis A, Geier S. Approche nouvelle de

la chirurgie de l'eÂpilepsie: methodologie steÂreÂotaxique et reÂsultats theÂr-

apeutiques. Neurochirurgie 1974;20(Suppl 1):1±240.

Van der Heyden M, Velis D, Hoekstra B, Pijn J, Van Emde Boas W, Van

Veelen C, Van Rijen P, Lopes da Silva F, DeGoede J. Non-linear

analysis of intracranial human EEG in temporal lobe epilepsy. Clin

Neurophysiol 1999;110:1726±1740.

Walczak T. Neocortical temporal lobe epilepsy. Epilepsia 1995;36:633±

635.

Wendling F, Badier J, Chauvel P, Coatrieux J. A method to quantify invar-

iant information in depth-recorded epileptic seizures. Electroenceph

clin Neurophysiol 1997;102:472±485.

Wendling F, Bellanger J, Bartolomei F, Chauvel P. Relevance of nonlinear

lumped-parameter models in the analysis of depth-EEG epileptic

signals. Biol Cybern 2000;83:367±378.

Wendling F, Bartolomei F, Bellanger J, Chauvel P. Interpretation of inter-

dependencies in epileptic signals using a macroscopic physiological

model of EEG. Clin Neurophysiol 2001;112:1201±1218.

Wheal H. Membrane electrophysiology of epileptiform activity in the

hippocampus. Acta Neurochirurgica 1990;50(Suppl):6±13.

Wieser H. Electroclinical features of the psychomotor seizures. London:

Butterworths, 1983.

Wieser H, Engel JJ, Williamson P, Babb T, Gloor P. Surgically remediable

temporal lobes syndromes. In: Engel Jr J, editor. Surgical treatment of

the epilepsies, 2nd ed. New York, NY: Raven Press, 1993. pp. 49±63.

Williamson P, French J, Thadani V, Kim J, Novelly R, Spencer S, Spencer

D, Mattson R. Characteristics of medial temporal lobe epilepsy: II.

Interictal and ictal scalp electroencephalography, neuropsychological

testing, neuroimaging, surgical results and pathology. Ann Neurol

1993;34:781±787.

Williamson P, Engel P, Munari C. Anatomic classi®cation of localization-

related epilepsies. In: Engel J, Pedley T, editors. Epilepsy: a compre-

hensive textbook, New York, NY: Lippincott-Raven, 1998.

F. Bartolomei et al. / Clinical Neurophysiology 112 (2001) 1746±17601760