EEG-based epilepsy detection with graph correlation analysis.

Recognizing epilepsy through neurophysiological signals, such as the electroencephalogram (EEG), could provide a reliable method for epilepsy detection. Existing methods primarily extract effective features by capturing the time-frequency relationships of EEG signals but overlook the correlations between EEG signals. Intuitively, certain channel signals exhibit weaker correlations with other channels compared to the normal state. Based on this insight, we propose an EEG-based epilepsy detection method with graph correlation analysis (EEG-GCA), by detecting abnormal channels and segments based on the analysis of inter-channel correlations. Specifically, we employ a graph neural network (GNN) with weight sharing to capture target channel information and aggregate information from neighboring channels. Subsequently, Kullback-Leibler (KL) divergence regularization is used to align the distributions of target channel information and neighbor channel information. Finally, in the testing phase, anomalies in channels and segments are detected by measuring the correlation between the two views. The proposed method is the only one in the field that does not require access to seizure data during the training phase. It introduces a new state-of-the-art method in the field and outperforms all relevant supervised methods. Experimental results have shown that EEG-GCA can indeed accurately estimate epilepsy detection.