Contribution to the characterization of EEG data for epileptic seizures detection

Abstract

Epileptic seizure Detection is a challenging problem which consists in identifying a seizure among normal brain activity using electroencephalogram (EEG) signals, either by an experienced neurologist or automatically engineered frameworks. In this work, we aim to contribute to the latter to help experts in medical facilities and improve the safety and autonomy of patients. We will strive to understand the effects and contribution of each and all features. We include two types of features: SCSA and nonlinear dynamical features. We will exploit the frequency diversity of EEG and contribute to the optimization of time-embedding hyper-parameters for the dynamical features. Later on, we tackle imbalanced data by introducing 2D-Generative Adversarial Networks for Data Augmentation. Experimental results demonstrate the reliability of the workflow and performance enhancement compared to state-of-the-art accuracy, sensitivity and specificity. The three metrics approach consist scores of 0.99. This is due to two main parts: the introduction, for the first time of the SCSA to characterize epileptic seizures and the careful optimization of the time-embedding hyper-parameters for the nonlinear features.