EEG microstate dynamics and source-space functional connectivity features for machine learning classification of major depressive disorder

Diagnosis of major depressive disorder (MDD) primarily heavily relies on clinical interviews and subjective symptom reporting, highlighting the urgent need for objective biomarkers. This study aimed to characterize MDD-related alterations in rapid electrophysiological network dynamics by integrating electroencephalogram (EEG) microstate analysis and source-space functional connectivity (FC) analysis, and to evaluate the classification performance of these features and their combination. Resting-state EEG data were collected from 115 patients with MDD and 43 healthy controls (HCs). EEG microstate and source-space FC features were extracted and compared between groups. Group differences were assessed with correction for multiple comparisons. A support vector machine (SVM) classifier was applied to evaluate the classification performance of microstate features, FC features, and their combination. A sensitivity analysis adjusting for age, gender, and body mass index was also performed. Compared to HCs, MDD patients showed a significant increased transition probability from microstate D to B after false discovery rate (FDR) correction ( P FDR = 0.048). In the source-space FC analysis, several edge-wise differences were observed at the uncorrected threshold of P < 0.05; however, none of these FC differences survived FDR correction. In the SVM classification analysis, the weighted fusion model achieved the numerically highest area under the curve (AUC) of 0.759 (95% confidence interval (CI): 0.660–0.847), compared with the FC-only model (AUC = 0.747, 95% CI: 0.642–0.841) and the microstate-only model (AUC = 0.724, 95% CI: 0.623–0.818). Patients with MDD showed altered EEG microstate transition dynamics, particularly an increased transition probability from microstate D to B. Several edge-wise FC differences between patients with MDD and HCs were observed, however, no edge-level FC differences survived correction for multiple comparisons. The integration of microstate and FC features may provide potentially useful EEG-derived information for machine learning classification, but its generalizability and clinical utility require validation in larger independent cohorts. The study was registered on http://www.chictr.org.cn/ and the registration number was ChiCTR2200057365 (registration date: March 9, 2022).