<![CDATA[Depression among college students is a growing public health concern, with existing screening methods often limited in sensitivity, scalability, and interpretability. This study developed and validated an explainable machine learning framework for early depression risk identification and tiered intervention planning in universities. We propose a Dynamic Weighted Ensemble Model (DWEM) that integrates five tree-based algorithms, with weights optimized via Bayesian search and cost-sensitive learning. Informed by the diathesis–stress framework, features were engineered and interpreted using SHAP to provide global and local explanations. The model was evaluated using stratified five-fold cross-validation with careful control of data leakage. The DWEM achieved an accuracy of 94.96% and an AUC of 98.95%, with balanced sensitivity and specificity, outperforming both single-model benchmarks and traditional questionnaire-based screening. SHAP analysis stably identified academic performance, stress-burnout, sleep problems, and protective factors as key risk determinants. Based on these outputs, a probability-based three-tier intervention framework was designed to translate risk stratification into actionable clinical support. This study demonstrates that an optimized ensemble approach, combined with theory-driven features and robust explainability, can provide a reliable, transparent, and practical tool for scalable mental health screening, supporting a shift toward proactive, data-driven prevention and efficient resource allocation in campus settings.]]>
Copyrights © 2026
