Multimodal machine learning for major depressive disorder: Integrating EEG functional connectivity and clinical variables to enhance diagnostic accuracy

The diagnosis of Major Depressive Disorder (MDD) relies heavily on subjective clinical assessments. This study evaluated various machine learning models in differentiating between MDD patients and healthy controls using resting-state electroencephalography (EEG) features and clinical variables as input variables. A total of 123 participants, including 77 MDD patients and 46 sex- and age-matched controls underwent resting-state EEG recording and 11 standardized clinical assessments. From each EEG, we extracted absolute and relative power and functional connectivity metrics, including phase locking value, phase lag index (PLI), and weighted PLI across standard EEG frequency bands. Nineteen ML classifiers were evaluated using leave-one-out cross-validation. The best-performing EEG-only model using absolute power with a medium KNN classifier achieved an area under curve (AUC) of 0.876. The best-performing clinical-only model yielded an AUC of 0.849. A combined model integrating EEG absolute power and clinical variables further improved performance (AUC = 0.896). Integrating EEG features with clinical data significantly enhanced the MDD classification performance. These findings support the potential of using multimodal data fusion and machine learning to develop objective diagnostic tools for the assessment of psychiatric disorders.