A Spatiotemporal Channel Attention Residual Network With Extended Series Mean Amplitude Spectrum for Epilepsy Detection

Abstract

Accurately detecting electroencephalogram (EEG) signals in a specific period before the epileptic seizure can effectively predict epilepsy and reduce the harm caused by epilepsy to the patient. However, the current research rarely pays attention to the influence of spatio-temporal features on the detection of EEG signals before seizures. To accurately predict epilepsy before seizures, this paper proposes a spatio-temporal channel attention residual network (STCARN) with extended series mean amplitude spectrum (MAS) of EEG signals. Specifically, the extended series MAS feature representation is first developed to rationally combine the temporal relevance of multiple MAS and the spatial relevance of EEG channels, fully representing the related activities of the brain. Furthermore, STCARN is proposed to extract the spatio-temporal information of extended series MASs by rationally fusing residual convolutional structure, channel attention mechanism, and recurrent network structure, which significantly improves the performance of epilepsy detection. STCARN is evaluated in the three-classification and five-classification tasks for epilepsy detection. The results indicate that STCARN achieves 99.98% and 99.51% accuracy in the two tasks, respectively.