Extraction of reproducible seizure patterns based on EEG scalp correlations

Abstract

The objective of this work is to identify similarities in the spatio-temporal dynamics of epileptic seizures, record with scalp EEG. A comprehensive method is proposed and applied in EEG of the patients who suffer from temporal lobe epilepsy. The method is based on the computation of the time-varying degree of non linear correlation between scalp electrodes at seizure onset and during seizure spread, determined by a nonlinear regression analysis. The quantification and coding of these similarity relations allow the comparison between two epileptic networks. Results show that reproducible patterns may be extracted from different seizures of the same patient and confirm the existence of different subtypes of temporal lobe epilepsy.

Introduction

Epilepsy is characterized by recurring seizures in which sudden abnormal synchronization between large groups of neurons occurs [1]. Temporal lobe epilepsy (TLE), the most common form of partial epilepsy in adults, is refractory to medical treatment in 20–30% of the cases. Surgery, which consists in removing the epileptogenic zone (EZ), can be proposed in about half of patients suffering from drug-resistant epilepsy [2]. Localization of the epileptogenic zone EZ, characterized by structures involved at seizure onset is thus a major issue before considering surgical operation [3], [4]. Many techniques are used to produce a reliable diagnosis, from analysis of clinical signs (semiology) to anatomical and functional imaging (MRI) [5], [6], [7]. Among these techniques, electrophysiological methods are the most adapted to better understand the spatio-temporal dynamics of the brain electrical activity during the interictal to ictal transition [8]. Electrophysiological methods can be divided into two groups depending on the location of electrodes: electroencephalography (EEG, scalp electrodes), stereoelectroencephalography (SEEG, intracerebral electrodes) and electrocortigraphy (ECoG, cortical grids).

Signal processing methods have been successfully used on SEEG recording to complete visual inspection, in many studies [8], [9]. It was showed that seizure evolution was closely linked to EZ organization and allowed to objectively (i) quantify seizure patterns and (ii) extract reproducible phenomena from different seizures recorded in a same patient [10]. In SEEG exploration, electrodes are positioned closer to sources, compared to scalp EEG. Consequently, spatio-temporal evolution of seizures involving both mesial and lateral parts of the temporal lobe is more easily analysed using such invasively recorded electrophysiological signals.Conversely, only few studies with similar intent were conducted on scalp EEG [11]. This may be explained by at least two factors. First, scalp EEG is a projection of the global cerebral activity on electrodes positioned over the head. Consequently, recorded signals correspond to complex mixing between ictal activity rising from relatively circumscribed regions (partial seizures) and the activity of structures that are not involved in seizures. Second, many artifacts are present in the surface EEG signals, especially during seizures where uncontrolled muscular and eyes movements may occur. From the information processing viewpoint, the main difficulty is to extract invariant information from non-stationary noisy signals. In this context, classical pattern recognition techniques that have been applied to scalp EEG did not prove to be sufficiently robust. We chose to address the problem of characterizing the reproducibility of seizures from the quantification of time-varying statistical relationships between signals recorded from distant scalp electrodes and from the comparison of these relationships from one seizure to the other.

Measurement of correlations between electrophysiological signals has been used in neurophysiological applications to reveal functional connectivity between cerebral structures. Regarding partial TLE, several methods have been successfully applied to SEEG signals [12], [13], [8], [14] in order to better understand interdependencies between brain structures during the transition from interictal to ictal state.

Although the transfer function between brain activity and observed scalp signals is complex, we assume that abnormal synchronization processes between brain structures that occur during epileptiform activity is also reflected in correlations measured between scalp electrodes.

Several parametric and non-parametric methods were described to estimate synchronizations at seizure onset and during seizure: directed transfer function in electrocorticograph ECoG [15], the non-linear regression method introducedby Pijn (1990) phase synchronization [14]. In the proposed method, measure were preferred as they take into account for possible non-linear relationship between analyzed signals [16]. Our method also contrasts with that described in [11] where authors developed algorithms to group similar seizures using scalp EEG. Their method was based on the computation of the edit distance between EEG signals quantified and coded as sequences of symbol vectors, as proposed in [10]. In this paper, the matching method is applied on statistical relationships computed from pairs of signals. The paper is organized as follows: section two describes the procedure used for data collection and provides details about the method proposed to extract similarities (in terms of time-varying correlation in scalp signals) among different seizures. Results obtained in 43 patients are presented in Section 3 and discussed in Section 4.