Multi-Bands Joint Graph Convolution EEG Functional Connectivity Network for Predicting Mental Disorders

As a typical representative of mental disorders, anxiety and depression disorders have occupied a large number of people after the outbreak of the new crown epidemic. However, anxiety and depressive disorders are detected by screening with clinical scales, potentially undermining the efficacy of detecting mental disorders, as it is heavily contingent upon the personal condition of the individual participant and the expertise of diagnosing physician. Therefore, to predict the feasibility of individual-level mental disorders, we developed a multi-bands joint graph convolution network (MBJ-GCN) framework based on electroencephalogram (EEG) brain functional connectivity (FC) networks. The functional connectivity networks are fused with orthogonal power envelopes correlation across five different frequency bands. Furthermore, the five frequency bands of brain FC networks were then integrated using locally weighted clustering coefficients (LWCC), which were then combined into feature vectors to represent the vertexes. After that, we design parallel GCN layers with different inputs via random embeddings, which can recognize intrinsic mental disorder maps from the embeddings in GCNs. Finally, we concatenate the outputs of all GCN layers in the fully connected layer for prediction. The experimental results have shown that our proposed method has low mean absolute error (MAE) and high Pearson correlation coefficient (PCC) performance compared with other prediction methods. It is possible to represent both individual characteristics and clinical information of potential patients using the proposed method. In conclusion, we extend traditional scale interviews to FC-based individual predictions, thereby taking a step towards the development of applicable techniques for quantitative real-world monitoring of mental disorders.