Abstract
Major depressive disorder (MDD) is highly prevalent, yet antidepressant selection remains largely trial and error because of limited predictive biomarkers. Resting-state electroencephalography (EEG) combined with graph theory-based functional networks offers a noninvasive approach to probing brain dysconnectivity. This study aimed to delineate source-level functional network differences between antidepressant responders and non-responders, and to evaluate whether machine learning models using nodal network features can discriminate treatment outcome. Resting-state EEG was recorded at baseline from 47 patients with MDD and 41 healthy controls (HCs). After 12 weeks, patients were classified as responders (n = 20) or non-responders (n = 27) based on a ≥ 50% reduction in Hamilton Depression Rating Scale scores. Source-level networks were assessed using path length (PL), strength, clustering coefficient (CC), and eigenvector centrality (EC). Machine learning with nested cross-validation classified responders and non-responders using nodal features. At the global level, non-responders showed lower strength and CC and higher PL in the high beta band than responders. Responders showed higher theta-band CC than HCs. A support vector machine achieved 82.98% accuracy (95% CI, 72.34%-93.62%; p < 0.001), 85.19% sensitivity (95% CI, 70.37%-96.30%; p < 0.001), and 80.00% specificity (95% CI, 60.00%-95.00%; p < 0.001). In responders, theta-band EC in the precuneus and pars orbitalis correlated significantly with psychological assessment scores. Source-level functional network analysis combined with machine learning may identify pretreatment EEG features associated with subsequent antidepressant response.