Predicting comorbid anxiety in adolescents with major depressive disorder: an EEG-based machine learning approach with SHAP interpretability.

Comorbid anxiety in adolescents with major depressive disorder (adMDD) is linked to higher suicide risk and poorer prognosis, necessitating precise screening tools. This study developed an interpretable machine learning (ML) framework using electroencephalography (EEG) biomarkers to predict anxiety symptoms in 200 adMDD (aged 12-18, drug-free for ≥ 2 weeks). Participants, stratified by Hamilton Anxiety Rating Scale (HAMA ≥ 14 for comorbid anxiety, n = 172; HAMA < 14, n = 28), underwent EEG recordings. Thirty-two EEG features (time-domain, frequency-domain, nonlinear, Hjorth, and entropy-based) were extracted, and seven machine learning models were trained and validated using 3-fold nested cross-validation. Results indicated the Light Gradient Boosting Machine (LightGBM) have superior predictive performance, with AUC value of 0.72 ± 0.05, accuracy value of 0.83 ± 0.02, precision value of 0.91 ± 0.02, recall value of 0.89 ± 0.03, F1 score of 0.90 ± 0.14, and AUPRC of 0.92 ± 0.03. In addition, SHAP analysis highlighting Normalized first difference, the ratio of Alpha to Beta, and Theta-band power as key predictors. This framework, enhanced by SHAP's global and local interpretability, identifies individualized EEG patterns for anxiety risk, offering a transparent, data-driven tool for personalized diagnostics in adMDD. These findings align with translational psychiatry goals to advance biomarker-based diagnostics in youth mental health.