A Data- and Physical- Features Driven Graph Convolution Network for Explainable Prediction of Schizophrenia With Electroencephalography

Abstract

This paper proposes a new multi-branch graph neural network (GCN) architecture for predicting schizophrenia versus healthy subjects via resting-state Electroencephalography (EEG). It leverages on both data and physical driven features derived from power spectral density (PSD) and functional connectivity (FC) indicators of EEG oscillations, which are closely related to neuron oscillations, cognition, and brain operations. The EEG data is first filtered to the delta, theta, alpha, beta and gamma bands for feature extraction using a convolutional neural network (CNN). FC measures and the extracted PSD features at each band are then integrated using a new dynamic GCN (DGCN) structure with residual connections to enhance the classification process. A new contrastive learning-based loss function is proposed for training the resulting multi-branch CNN-DGCN (MCNN-DGCN). Experiments on three EEG datasets of schizophrenia show that the proposed MCNN-DGCN model achieves promising accuracy of 99.55%, 99.40%, and 97.86%, respectively, demonstrating significant advantages over alternative approaches. To study the importance of the physical features identified, ablation studies of features at significant frequency bands and channels are conducted. All datasets suggest that i) parietal and right frontal areas and ii) beta and gamma bands have significant impact on the accuracy of prediction, and hence the disorder, which agree well with the literature, supporting the explainable nature of the proposed approach.