Epileptic Seizure Detection and Anticipation using Deep Learning with Ordered Encoding of Spectrogram Features

Abstract

Electroencephalogram (EEG) signals of the brain play a vital role in the detection of epileptic seizures. This paper proposes a new spectrogram-based deep learning method for the detection and anticipation of epileptic seizures. Unlike the existing methods, the proposed method formulates the feature descriptor such that it retains the neighborhood order of spectrograms both in time and frequency, while significantly reducing the dimensionality of the feature descriptor. The spectrogram in each of the 18 EEG channels is constructed by dividing each EEG signal into 3 time-blocks and 19 frequency-blocks. The mean magnitude value of each of these blocks is computed and thereby compactly representing the input EEG signal by a 3D tensor of size 18×19×3. This tensor descriptor is given as an input to the proposed convolution neural network for learning high-level features. Evaluations are performed on a publicly available EEG dataset of 23 patients and the results from the proposed method are compared with 9 other existing methods. Further, a five-class classification is performed using the proposed method for the anticipation of seizures. The proposed method is found to outperform the existing state-of-the-art methods both in detection and anticipation of epileptic seizures.