Toward precision acupuncture for depression: EEG microstate biomarkers and explainable machine learning for TECAS response.

BACKGROUND: Transcutaneous electrical cranial-auricular acupoint stimulation (TECAS) is an effective and safe neuromodulation therapy for major depressive disorder (MDD). However, interindividual variability in treatment response limits its clinical efficiency. Reliable and interpretable biomarkers for pretreatment response prediction are urgently needed to enable precision acupuncture strategies. OBJECTIVE: To develop an explainable machine-learning framework integrating baseline EEG microstate dynamics and clinical features to predict antidepressant response to TECAS, to translate these predictions into a clinical decision-support tool. METHODS: Fifty first-episode, drug-naïve patients with MDD underwent resting-state EEG assessment prior to 8-week TECAS intervention. Twenty-four EEG microstate features (duration, coverage, occurrence, and transition probabilities of microstates A-D) and six clinical variables were entered into six machine-learning models. Class imbalance was addressed using K-means SMOTE. Model interpretability was achieved using SHAP analyses, and key predictors were validated by conventional statistics and symptom improvement correlations. RESULTS: The random forest model achieved the best performance in discriminating Responders from Non-responders (AUC = 0.912). Explainability analyses consistently identified microstate D metrics, particularly mean duration, coverage, and occurrence, as robust predictors. Responders exhibited significantly shorter duration, lower coverage, and reduced occurrence of microstate D at baseline, and these parameters were negatively correlated with the magnitude of depressive symptom improvement. CONCLUSIONS: This study demonstrates that EEG microstate dynamics, especially microstate D, constitute clinically meaningful and interpretable biomarkers for predicting TECAS treatment response. By integrating neurophysiological signatures with explainable machine learning, we propose a precision acupuncture framework and a practical auxiliary decision-support tool to optimize individualized treatment selection in MDD.