Journal de recherche sur le cerveau

Wavelet Time Scattering Based Classification of Interictal and Preictal EEG Signals

Abstract

Afreda A. Susu, H.A. Agboola, C. Solebo, F.E.A. Lesi and D.S. Aribike

If it were possible to reliably identify the preictal brain state from dynamical changes in EEG data of epilepsy patients, then the age long problem of actualizing a fully automated closed-loop seizure – warning or seizure-prevention system that is clinically deployable would have been resolved. Accordingly, through feature engineering, a great deal of effort has been invested over the discovery of EEG features or measures that are always indicative of the preictal brain state. However, this has proven to be difficult, time consuming and apparently unsuccessful. Therefore, lately, attention has shifted to feature learning-methods that automatically learn and extracts useful discriminatory features from raw data. This paper studies the efficacy of wavelet time scattering learned EEG features for interictal and preictal EEG classification. Wavelet time scattering network developed in Matlab and two different EEG datasets: CHB-MIT scalp EEG and AES intracranial EEG datasets were used for the study. The learned interictal and preictal EEG features were used to train and evaluate a simple binary support vector machine classifier. Three different classification accuracy results namely ordinary cross validation, true cross validation and test classification accuracy results were reported for the analysis. Mean classification accuracy values of 93.15%, 97.57% and 91.33% were obtained respectively for the scalp EEG while mean classification accuracy values of 98.33%, 100% and 96.73% were obtained respectively for the intracranial EEG. A general comparison showed that the combination of wavelet time scattering learned EEG features and a simple binary support vector machine classifier performed equally or even better than deep convolutional neural networks in EEG classification tasks. Finally, wavelet time scattering has proven to be a very good EEG feature learner and may greatly improve the sensitivity and specificity of seizure prediction algorithms.