Attention-based Spatio-Spectral Network for seizure detection

Epilepsy is a neurological disorder of the brain caused by abnormal neuronal discharges that exist worldwide. Seizure classification based on multichannel electroencephalography (EEG) is an essential pathway for seizure diagnosis and treatment. In previous seizure detection tasks, many methods focus on directly analyzing data from different channels, often ignoring the relationship between signal channels and failing to utilize the brain network structure fully. In this paper, we design an epileptic seizure detection network based on an attention mechanism and graph convolutional networks to efficiently aggregate spatio-temporal features from channels as well as spectro-temporal features from time-frequency images. Specifically, we first develop a channel attention weighting algorithm that adaptively learns inter-channel weights. Then a time-frequency graph feature extraction aggregation module based on graph convolutional network with frequency attention mechanism is used to aggregate time-frequency image features from different channels, and finally the two features are combined and fed into a classifier for seizure detection. Experimental results on a publicly available dataset show that the proposed method outperforms existing methods, indicating that it is a viable solution for seizure detection.