Epileptic Signal Classification With Deep EEG Features by Stacked CNNs

Abstract

The scalp electroencephalogram (EEG)-based epileptic seizure/nonseizure detection has been comprehensively studied, and fruitful achievements have been reported in the past. Yet, few investigations have been paid to the preictal stage detection, which is practically more crucial to epileptics in taking precautions before seizure onset. In this article, a novel epileptic preictal state classification and seizure detection algorithm based on deep features learned by stacked convolutional neural networks (SCNNs) is developed. The mean amplitude of sub-band spectrum map (MAS) obtained from the average sub-band spectra of multichannel EEGs is adopted for representation. The probability feature vectors by stacked convolutional neural networks (CNNs) are extracted in the softmax layer of CNNs, where an adaptive and discriminative feature weighting fusion (AWF) is developed for performance enhancement. Following the deep extraction layer, the effective kernel extreme learning machine (KELM) is adopted for feature learning and epileptic classification. Experiments on the benchmark CHB-MIT database and a real recorded epileptic database are conducted for performance demonstration. Comparisons to many state-of-the-art epileptic classification methods are provided to show the superiority of the proposed SCNN+AWF algorithm.