Hybrid Explainable Deep Ensemble Framework with self-attention mechanism for early prediction of depression severity using PHQ-9 clinical and behavioral data.

Background: Depression is a major global mental health disorder that requires accurate, reliable, and interpretable prediction models to support early identification and timely intervention. This study proposes a Hybrid Explainable Deep Ensemble Framework for multiclass prediction of depression severity using PHQ-9 clinical responses and behavioral characteristics. Methods: A retrospective cross-sectional machine learning study was conducted using a dataset comprising 682 anonymized participants. Predictor variables included demographic characteristics, PHQ-9 questionnaire items, sleep quality, academic pressure, and financial pressure. A hybrid ensemble integrating TabNet, multilayer perceptron (MLP), and XGBoost was developed with a self-attention mechanism for adaptive feature weighting and a weighted soft-voting strategy for final prediction. Model performance was evaluated using five-fold stratified cross-validation and an independent test set. Accuracy, precision, recall, macro F1-score, ROC-AUC, Matthews correlation coefficient, confusion matrix analysis, and SHAP-based explainability were employed to assess predictive performance and model interpretability. Results: The proposed framework achieved an overall accuracy of 91.3%, precision of 0.907, recall of 0.898, macro F1-score of 0.902, and a macro-ROC-AUC of 0.940, outperforming the baseline machine learning models. Class-wise evaluation demonstrated the strongest performance for severe depression (F1-score = 0.934), whereas the mild depression category exhibited comparatively lower predictive performance (F1-score = 0.861) because of overlap with adjacent severity levels. Confusion matrix analysis indicated that misclassification primarily occurred between neighboring depression categories while clinically implausible predictions were rare. SHAP analysis identified sleep disturbance (PHQ-4), fatigue (PHQ-3), academic pressure, depressed mood (PHQ-2), and financial pressure as the most influential predictors. The ablation study further demonstrated that incorporating the self-attention mechanism and ensemble learning strategy improved both predictive performance and model robustness. Conclusion: The proposed Hybrid Explainable Deep Ensemble Framework combines high predictive accuracy with transparent decision-making for multiclass depression severity assessment. By integrating deep learning, ensemble learning, self-attention, and explainable artificial intelligence, the framework provides accurate and clinically interpretable predictions and represents a promising approach for future clinical decision-support systems and precision mental health screening.