A generalized depression recognition framework based on cross-center and cross-task EEG signals

This study developed a generalized framework for automatic depression recognition using a Dempster–Shafer Theory-based Classification Fusion (DSTCF) method. We collected resting-state electroencephalography (EEG) data from 24 Major Depressive Disorder (MDD) patients and 29 Normal Controls (NC), as well as additional data from 47 MDD and 47 NC in both resting-state and dot-probe task conditions. After extracting eight linear and three nonlinear features, three feature selection methods were applied, and we conducted validations across three datasets and two different tasks. Cross-center validation using the DSTCF method showed that after adopting the Relief method, EEG data in the beta frequency band achieved optimal performance. The best average accuracy achieved was 96.18% on the training datasets and 67.82% on multiple cross-center datasets, representing improvements of over 7% and 8% compared to traditional methods, respectively. Additionally, when evaluating the cross-dataset accuracies of existing popular methods such as mKTAChSel+SVM, PSD+GSTAN, and EEGNet, DSTCF improves accuracy by over 6%. Furthermore, the mean amplitude of peak to peak (Ppmean) and activity features demonstrated high discriminative capability and significant correlation with depressive levels. Based on these two features, abnormal parieto-occipital lobe activation served as a task-independent feature for MDD identification. Additionally, abnormal activation of the frontal lobe in the beta band and the temporal lobe in the theta band can effectively distinguish MDD from NC under resting and task conditions, respectively. These findings not only help in understanding the atypical neural mechanisms of depression but also provide reliable EEG biomarkers for identification and diagnosis.