MULTI-AGENTS SYSTEM ON EPILEPTIC NETWORK

Abel Kinie, Jean-Jacques Montois

LTSI-INSERM/GRAID - IUT de Saint-Malo, Bd déportés BP 195, Saint-Malo Cedex, France

Keywords: Epilepsy, Signal processing, Biomedical signal processing, Multi-agent system, Distributed artificial

intelligence, Behaviour approach.

Abstract: This work is focused on the study and interpretation of epileptic signals, based on the analysis of stereo

electroencephalographic (SEEG) signals with signal processing method and multi-agent approach. The

objective is to use this technique to improve information extraction, representation and interpretation as well

as the implemented control strategies in the different processes. Our approach deals with the information

recorded during the intercerebral exploration and it exploits a dynamic selection of the interest’s

information to optimize the processing without truncating the information. We associated signal processing

algorithms (spectrum analysis, causality measure between signals) approved in the analysis of the epileptic

signal in a multi-agent system.

1 INTRODUCTION

The paroxysmal discharges are initiated by a

network (epileptogen network) distributed in

different cerebral structures linked by dynamic

connections and abnormally facilitated (Chauvel et

al., 1987). The identification of this network, in a

given patient, leads to the definition of an optimal

surgical procedure, that is to say minimizing the size

of the resection of the region of origin of the crisis

(Bartolomei et al., 2005).

The proposed methods for treating the signal

could lead to a better definition of the complex

concepts of the irritative and the epileptogenic zone

(Wendling et al., 1999)

as well as those of the

topography (“where is the source of the signal?”)

and of synchronicity (“these two signals are they

synchronic, therefore reflecting a functional

connectivity?”) The response elements only quantify

a part of the information contained in these signals.

These methods rely generally on a downward

approach and realize with difficulty the dynamic of

interactions between cerebral structures implied in

the epileptic process. The problem of analyzing the

propagation of epileptic activity is difficult to solve

with these approaches because the system must face

varied and unforeseeable situations (epileptic

seizures).

The problem is tackled here by a cooperative

distributive system. (J. Ferber, 1995), (G. Weiss,

1999). the first generation of distributive systems

appears in the middle of the 1970’s with the

development of artificial distributed intelligence

(AID). This generation of systems is characterized

by a distribution of knowledge and processing whilst

conserving a centralized control. The most important

contributions were those of Erman and Roth Hayes

(Erman et al., 1980) with Hearsay II, a system of

word recognition those of Lesser (Lesser and

Corkill, 1983) and actors of Hewitt [8. The second

generation of systems appearing in the 1990’s,

brought the decentralization of control, reutilization

and autonomous cooperative agents (J. Ferber, 1995)

and (G. Weiss, 1999). The cooperative distributed

system relies on a formalism agent who aims to

organize and control the scalar processes and to

insure its coordination by best integrating the

specifics of each process to make the combination of

interests appear between explored cerebral regions.

We hope through analysis by an independent agent,

for better adaptability to the erratic signal changes

and an effective management of complexity (by

local approach). It is therefore the control (Altman et

al., 2002) and follows up of the evolution of

activities from agents which will clarify the

evolution of the activities of explored cerebral

structures.

Paragraph two explains the formal framework of

the planned study. It poses the problem treated. The

methodological approach undertaken is then exposed

336

Kinie A. and Montois J. (2010).

MULTI-AGENTS SYSTEM ON EPILEPTIC NETWORK.

In Proceedings of the 2nd International Conference on Agents and Artiﬁcial Intelligence - Agents, pages 336-340

DOI: 10.5220/0002738603360340

 SciTePress

in detail in the third part which presents the

experimental platform of the multi agents system

(MAS) dedicated to the analysis of SEEG signals.

The fourth paragraph discusses the experimental

results obtained and compares them to traditional

clinical analysis made by a clinician and the last part

proposes a discussion of the potential benefits of this

original approach in the analysis of SEEG signals

and monitoring of epileptic patients.

2 THE MULTI AGENT SYSTEM

IN EPILEPSY

The MAS technical are often used in the artificial

intelligence field, the distributed information

processing systems and the software genius. It is a

discipline which is interested in the collective

behaviours produced by interactions of several

autonomous and flexible entities called agents.

These interactions suppose a co-operation between

these agents. The multi agent systems and signal

processing applying to epilepsy present a close links

according to the entities which compose them.

Indeed, the multi agent systems can allow

modelling and simulation of neurons aggregates or

systems based on autonomous entities and

distributed interactions. However the signals

processing (applying to epilepy) model the

mechanisms which govern the propagation of the

electric activities in the brain (distributed

mechanisms) and also the structure and the

interactions of the nervous cells.

Our paper brings out the common interests to

study the association of these two research fields.

Scalar analysis of the epileptic signal by the

computation of the relations between signals

(Bartolomei et al., 2005), (Wendling et al., 1999)

highlighted the existence of mutual interactions

between the EEGs. In the same way the concept of

epileptogenic network developed by Professor

Patrick CHAUVEL (Chauvel et al., 1987) give raise

to the existence of a co-operation between cerebral

areas in epileptic processes.

In order to take into account the specificity of

each signal and to compute the co-operations

between them, we evaluate 9 descriptors per signal

around a distributed system. These 9 descriptors

evaluated starting from the spectral power density

(PSD) in various frequencies bands from the

epileptic signals which correspond to various

physiological or pathological cerebral cortex states.

These descriptors and the various frequencies

bands used are detailed in paragraph 3.

The MAS from their innovative aspect and their

non-traditional approach (behavioural approach) of

the realization of distributed systems propose an

original method of vectorial processing by

associating existing signal processing scalar

methods. With the MAS technical in the signal

processing study we hope to compute the epileptic

processes mechanism during the paroxysmal

discharge propagation.

3 AGENTIFICATION OF THE

PROBLEM

The experimental platform used in this work is

MadKit (Multi-agent development Kit) (J. Ferber,

1995). The implementation of MAS requires using

iterative algorithms to define the behaviours of the

various agents which composed the system. For

conveniences reasons we used a ready to

employment platform (MadKit). In this model, an

organization is regarded as a structural relationship

between collections of agents. Thus, an organization

can be described only on the basis of its structure.

MADKIT is conceived by Jacques FERBER and al

(J. Ferber, 1995). It implements its three central

concepts Agent, Groups and Role.

Our approach is built starting from several

groups of agents whose properties and missions

must allow:

 To classify the signals having the same SEEG

activities and/or contained similar spectral

components (groups).

 to classify the signals whose activities change

in the same temporal interval

(segmentation)(groups).

 To represent the seizure as a coloured image

allowing locating in space and in time the

SEEG signals having the pathological or the

similar activities (groups and roles...).

 To associate all these partial results to bring

out a global behaviour of the analyzed seizure

(groups and roles).

To formalize the problem we consider a vectorial

signal S (t) made up of a recording SEEG signals on

N channels and an interval [0, T]. S (t) = [S1 (t)......

SN (t) S

(t) ∈ . k = 1...... N, t = {0, 1/Fe... T-1/Fe},

N is the number of explored cerebral structures and

Fe is the sampling rate.

Our architecture is made of a reactive agent’s

community which is made up of various agents

MULTI-AGENTS SYSTEM ON EPILEPTIC NETWORK

337

groups whose are able to analyze, in parallel, the

signals coming from N recording SEEG channels (N

≈100). In this approach, to each SEEG channel

(epileptic signal) we associated a "Descriptor" agent.

Its role is to analyze the spectral components of the

segment of signal recorded on this channel and to

seek possible links with his/her "colleagues" on the

[t, t+Δt] interval. All the "Descriptors" agents

communicate the results of their analysis to a

"Classifying” agent which is in charged to compute

the components of each SEEG segment then to

classify it according to its a priori knowledge.

Indeed no exhaustive trainings classes were built and

transmitted to the "Classifying" agent for each

epileptic activity. When T described the [0, T]

interval (T being the total recording duration), a

"Visualization" agent produces a coloured space

time and spectral tablecloth representation, with

each agent at the t moment. Parallel to each activity

class identified by the "Descriptors" agent on the [t,

t+Δt] interval we associated a "Causality" agent

whose role is to evaluate the degree of relationship

between brain’s entities bound by affinities.

3.1 Descriptors Agent

The agent associated with each SEEG channel uses

the energy of the signal calculated in various

frequencies bands to characterize the activity during

the computation time. To take into account the

specificity of SEEG signal, we chose the 9 bands

classically used in the field (δ1, δ2, θ1, θ2, α1, α2,

β1, β2, γ ).

Energy is calculated starting from the

power spectral density (psd to see (1) and (2))

estimated starting from the periodogram method A

characteristic vector (9 parameters representing the

agent state at the t moment) is produced for each

SEEG channel for the current segment.

)( fSdsp

nbPtsFFTf

∑

(1)

() ()

ifn

nbPtsFFT

Sf Sn e

−

∑

(2)

Where S (f) is the Fast Fourier Transform the s(t)

signal.

3.2 Classifying Agent

During each cycle, the "Classifying" agent receives

the characteristic vector produced by each

"Descriptors" agent and associates a scalar code to

the segment considered. . From this agent state (its

scalar code) and its “neighbours” state (information

on the proximity between SEEG sensors), a decision

is taken according to the identification of all agents

components (strong probability of being in the same

state as its “neighbours” belonging to the same

brain‘s structure).

3.3 Visualization Agent

It associates during the computation time a colour to

each "Descriptors" agent and produces a space time

and spectral tablecloth representation of the spectral

components of each SEEG signal. The colour

attribution is carried thanks to each agent state

(spectral components). “Cold "and" hot "colours

then respectively code the "low frequencies

activities and the "high frequencies" activities and

the black colour codes the not classified activities.

Figure.1: Association between activity and colour.

3.4 Causality Agent

The agent associated with each class and activity

identified on the [t, t+Δt ] interval uses the nonlinear

coefficient of regression h

calculated on a pair of

X(t) and Y(t) observed signals on the limited

temporal support [ t, t+Δt ]. It makes it possible to

characterize the functional couplings between

cerebral structures and to consider a time delay of

propagation possible between these two

observations.

The Load Agent (Scheduler) is charged to plan

the execution order and to activate in a synchronized

way all the operated agents. The principle is rather

simple to each agent or agents groups (G) equipped

with a role (R) we associated an activator (A

) on

the level of Scheduler. Scheduler is useful of these

activators to plan the tasks while launching to each

activators (A

) cycle the ones after the others

according to the desired execution order. A cycle

corresponds to a full rotation of the various

activators which one need to activate.

The global architecture of the establishment is

summarize in 3 phases (A Phase, B Phase and C

Phase)

A Phase

The Server agent (data base) feeds the Descriptors

agents and "Rate/rhythm" in signal samples. The

Descriptors, the "Rate/rhythm» and the Causality

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

338

(the signal processing algorithms) are charged to

carry out the traditional computation of low level.

The "Rate/rhythm" deals with the segmentation

(detections of the moments of rupture in the

analyzed signal).

B Phase

Observant agents (Watcher) (Erman et al., 1980) are

charged to recover the data (results provided by the

agents of low level. They also provide a more

elaborate processing on these data and contribute to

the emergence of a global organization. The

Classifying agent allows the junction of the other

agents with the Descriptors agent, it also orders the

results generated by the latter. The Visualization

agent provides a representation of the seizure where

information is made available. The observers

"Rupture" and the Causality respectively represent

for the agents "Rate/rhythm" and "the Causality

what "To classify it" represents for the Descriptors

agent. Load Agent (interface) allows choosing the

patient seizure to be analyzed, to create and to

parameterize the execution architecture. It also

manages the scheduling of the spots to be carried out

by the agents.

C Phase

This phase which is not presented here will have to

join together information resulting from phases A

and B to order the seizure propagation. It will have

to consider the evolution of the interactions between

agents (affinities) to raise the evolution of the

couplings between brain’s structures during the

seizures.

4 RESULTS

The SEEG signals are recorded by a BMSI-

NICOLET system which allows a simultaneous

acquisition of SEEG signals on 128 channels, at a

sampling rate of 256Hz. The studs of the electrodes

are numbered from 1 to 15 from the internal end to

the external end. Each electrode is located by a letter

(A, B, C....), affected of a "premium" (A', B', C '...)

if it is established on the left dimensions of the brain.

We obtained a complete image of the analyzed

epileptic seizure where each activity is well

identified according to the frequencies activities of

its localization SEEG sensors and its appearance

moment (time).

The x-axis corresponds to the temporal space

(time in second), the y-axis space (studs or signals),

the various electrodes are separated by white

features and each electrode studs is presented at the

extreme right-hand side of the diagram. The

following paragraphs propose the results analysis.

In a patient presenting temporal lobe epilepsy,

according to the clinician; the brain’s structures

involved in the initiation of its seizure are the

internal temporal pole (TP), the former hippocampus

(B), the posterior hippocampus (C) and the

entorhinal cortex (TB).

Figure 2: Space –time and -spectral representation of P1.

Figure 2 shows the localization of various

activities (ictal and normal) but also allow

identifying for each SEEG sensor, the activity type.

It also highlights about several initiating brain’s

structures. We observed quite localized ictal

activities and this clearly highlight again about the

implication of a great number of SEEG channels in

the propagation of the paroxysmal discharge during

the computation time.

5 CONCLUSIONS

On patients study the results highlight a zone

initiating the seizures including the internal temporal

brain’s structures (former and posterior

hippocampus, entorhinal cortex) and confirm the key

role of these structures in the partial temporal lobe

epilepsy.

Our work formalization problems and an analysis

made of each entity’s behaviour towards co-

operation, association and competition which bring

out these relevant images, can lead us to a better

understand some epileptic observed phenomena.

Difficulty to interpret some others phenomena come

from the artefacts strongly present in the signals and

a good management of these measurements noises

can improve the results.

The technological platform thus created is

situated at the borders of several research domains

(signal processing, computer engineering and

artificial distributed intelligence) and opens the

perspectives notably for monitoring epileptic

MULTI-AGENTS SYSTEM ON EPILEPTIC NETWORK

339

patients. The original procedure proposed and the

first results obtained give us hope to go further in the

exploitation of MAS techniques to help with the

diagnosis of epilepsy by means of the description of

the propagation of the discharge during epileptic

seizures and the vectorial processing of the signal in

its entirety. In the long term we envisage to place

more interactivity between the system and its user so

as to better link the clinical symptoms to the

evolution of certain physiological parameters.

REFERENCES

P. Chauvel, P. Buser., J.M. Badier, C.Liegiois-Chauvel.,

P. Marquis et Bancaud La zone épileptogène chez

l’homme : représentation des événements inter

critiques par cartes spatio-temporelles. Revue

Neurologique, n° 143, pp. 443-450, 1987.

F. Bartolomei, P. Chauvel, F. Wendling, “Dynamique des

réseaux neuraux dans les épilepsies partielles

humaines”, Revue Neurologique, Vol. 161, 2005, p.

767-780.

F. Wendling, M. Shamshollahi, et al, “Time-frequency

matching of warped depth-EEG seizure observations”,

IEEE Trans. Biomed. Eng, Vol. 46(5), 1999, p. 601-

605.

J. Ferber. Les Systèmes Multi Agents. Vers une intelligence

collective. InterEditions, 1995.

G. Weiss, “Multiagent System. A Modern Approach to

Distributed Artificial Intelligence”, The MIT Press,

Cambrige, Massachussetts (Ed) 1999.

L. D. Erman F. Hayes-Roth, V. R Lesser et al, "The

Hearsay II speech understanding system; Integrating

knowledge to resolve uncertainty",ACM Computing

Survey 12, 1980, p.213-253.

V. R. Lesser, D. Corkill: The Distributed Vehicle

Monitoring Testbed: A Tool for Investigating

Distributed Problem Solving Networks. AI Magazine

4(3), 1983, p.15-33.

Hewitt, C. E., “Viewing control structures as patterns of

passing messages”, Artificial Intelligence 8, 323-364,

1977.

E. Altman, T. Basar, R. Srikant, “Nash equilibria for

combined flow control and routing in networks :

Asymptotic behaviour for a large number of users”,

IEEE Transactions on Automatic Control, Special

Issue on Control Issues in Telecommunication

Networks,Vol.6, #47, 2002 p.917-930.

ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence

340