Multi-scale Residual Graph Attention Network for Major Depressive Disorder Recognition

Major depressive disorder (MDD) has emerged as one of the most prevalent mental disorders. As a typical time series data, electroencephalogram (EEG) is an objective physiological signal. The conventional method for recognizing MDD based on EEG suffers from long-term forgetting. As the length of the time series obtained from a single scale model increases, the time-related hidden states increase, leading to a greater likelihood of flooding previously valid information. Additionally, some existing methods rely on thresholds to weight the brain network connectivity importance, which cannot capture changes in global dynamic interactions and introduces additional biases. To address the above issues, we propose a novel MDD recognition method based on the multi-scale residual graph attention network (MReGAN). On the one hand, this method introduces a multi-scale residual module, which utilizes multi-scale feature representation to obtain complex multi-level changes. It is combined with a dilated causal convolution network to preserve the interaction information of different time periods and solve the problem of long-term forgetting. On the other hand, this method utilizes the multi-scale graph attention mechanism to directly capture the differences between MDD and normal control (NC) in the core topology and significant patterns of the brain functional connectivity network (BFCN), capturing global dynamic interaction patterns. Experimental results on benchmark datasets validate the exceptional performance and computational efficiency of MReGAN. Furthermore, comprehensive analysis shows that the connections between Fp 1 and Fp 2 channels in the Delta frequency band may serve as potential biomarkers.