Attention Fusion and Abnormal Brain Topology Neural Network for Mild Depression Recognition

Many studies attempt to explore the underlying mechanisms of depression and distinguish between depression patients and normal controls (NC) using electroencephalography (EEG) techniques. With the advancement of deep learning methods, an increasing number of studies aim to design Computer-Aided Diagnosis (CAD) systems for mild depression (MD) to achieve early identification. However, few studies construct models based on abnormal brain topological structures specific to MD patients. In this study, we investigate the abnormal brain topological structures of individuals with MD based on EEG data obtained during an emotional face paradigm. Functional connectivity analysis reveals a higher proportion of inter-hemispheric connections in the MD group compared to intra-hemispheric connections. Additionally, intra-hemispheric connections are primarily observed within the frontal and parietal lobes of both groups. Hierarchical clustering analysis results indicate impairments in the frontal and parietal lobes in the MD group compared to the NC group. Based on these findings, we propose a novel feature called 'cross-brain feature' and introduce a multi-cross-brain attention fusion mechanism to integrate information between brain regions. We train and test our models using 5-fold cross-validation. The results demonstrate that the classification model based on abnormal brain topological structures achieves the highest performance among the three state-of-the-art algorithms, with an accuracy of 80.1%, an area under the ROC curve (AUC) of 80%, and a sensitivity (SEN) of 86.3%. These findings suggest that combining abnormal brain topological structures derived from functional connectivity matrices with deep learning techniques can provide an effective objective approach to the early detection of depression.