EEG-Based Prediction of rTMS Treatment Response in Depression: Nonlinear Features and Machine Learning with Minimal Electrode

Repetitive transcranial magnetic stimulation (rTMS) is an established intervention for treatment-resistant depression, but response rates remain highly variable and reliable predictors of outcome are lacking. Resting-state electroencephalography (EEG), combined with machine learning, represents a cost-effective and accessible approach for individualized treatment planning. One hundred patients with major depressive disorder (MDD; 51 responders, 49 non-responders) underwent pre-treatment resting-state EEG prior to rTMS. Recordings were obtained using either 8 or 30 electrodes. Features including band power, coherence, phase-lag index (PLI), and phase-locking value (PLV) were extracted. Machine learning models were constructed using two-stage feature selection (recursive feature elimination with support vector classifier and sequential backward selection) and linear discriminant analysis (LDA). Model performance was assessed with 5-fold cross-validation and 1,000 shuffle iterations. Multi-feature EEG models consistently outperformed single-feature approaches in predicting rTMS treatment response. The best-performing model, based on only 8 electrodes, achieved an accuracy of 78.9% and an AUC of 73.3%, surpassing the 30-electrode configuration. PLI emerged as the strongest individual predictor, but combining multiple EEG features substantially improved classification. LDA provided the most stable performance across limited datasets. Resting-state EEG combined with machine learning can reliably predict rTMS treatment response in MDD. Notably, accurate prediction was achieved with a minimal 8-electrode montage, supporting the clinical feasibility of low-cost EEG assessments. This approach may facilitate personalized treatment selection and improve rTMS outcomes in routine psychiatric care.