Epileptic Signal Classification with Deep Transfer Learning Feature on Mean Amplitude Spectrum

Epilepsy, as a sudden and life-threatening nervous system disease, seriously affected around 6% population in the world. Epileptic classification has attracted wide attention in the past and a number of methods have been developed. But currently studies are mainly on three epileptic states classification (preictal, ictal, interictal) or seizure/non-seizure detection. Among them, the one hour before seizure onset was generally considered as preictal, where the division is actually not fine enough for some practical applications. In this paper, the epileptic signal classification with a more granular time-scale of the preictal stage is studied and a novel deep Electroencephalogram (EEG) feature extraction with the convolutional neural network (CNN) based transfer learning is developed. The subband mean amplitude spectrum map (MAS) of multichannel EEGs is computed for signal representation and three popular deep CNNs are exploited for feature transfer learning, respectively. Experiments on the benchmark CHI-MIT epilepsy EEG database show that the proposed algorithm achieves a highest overall accuracy of 92.77% when the one hour preictal stage is divided into small segments with a fine resolution of 20-minutes scale.