Mean amplitude spectrum based epileptic state classification for seizure prediction using convolutional neural networks

An increasing number of algorithms have been proposed for epileptic seizure prediction in recent years. But most of them are based on a partition of the electroencephalograph (EEG) signal of an epileptic patient into preictal, ictal (seizure), and interictal states. In this paper, we propose to further divide the preictal interval into multiple subintervals. Besides discriminating the seizure state from the preictal and interictal states, we also distinguish the preictal subintervals from each other. An epileptic state classification algorithm for epileptic EEG signals is then developed. The amplitude spectrums of EEG signals from 18 channels are firstly calculated and divided into 19 frequency subbands. The mean amplitude spectrum (MAS) on each of the 19 frequency subbands is then computed for each channel to form a MAS map of size 18 $$\times$$ 19. Finally, the MAS map is fed to a convolutional neural network (CNN) for feature extraction and a support vector machine (SVM) is employed for the epileptic state classification. Experiments show that, for a three-subinterval partition of the preictal state, a classification accuracy of 86.25% has been achieved on the CHB-MIT database by the MAS-based epileptic state classification algorithm using CNN and SVM.