Edge-attention graph convolutional network for depression grading via EEG cross-network interaction analysis

Depression is a serious global social problem. Personalized treatment is an effective way to improve the symptoms of depression. The edge attention graph convolutional network (EAGCN) model of EEG signals effectively extracts features in cross-network interactions for accurate depression grading. This paper proposes a novel fusion method of Time-frequency functional connectivity (TF-FC) and EAGCN to extract cross-network interaction features. This method employs intrinsic time-scale decomposition (ITD) to extract EEG time–frequency features, constructs whole-brain/intra-network/cross-network TF-FC matrices using Pearson correlation coefficients as edge features, integrates traditional EEG features (Hjorth parameters, power spectral density (PSD)) as node features, and inputs them into the EAGCN for grading to establish a fusion model. The results show that with depression severity increasing, the TF-FC connectivity strength decreases. Quantitative analysis shows that the fusion model achieves 93.14% classification accuracy for whole-brain analysis. Especially, the fusion model reaches 97.63% accuracy in VN-VAN and ECN-DMN cross-network interactions. Compared with other methods, our novel fusion method achieves superior performance, ranking among the top positions in the precise classification table. This novel fusion method significantly improves classification accuracy. At the same time, the cross-network interaction features may serve as a novel neurophysiological biomarker for depression severity grading. This study provides a novel theoretical framework and methodological innovation for personalized depression treatment.