Explainable XGBoost Early-Warning Framework for Academic Stress-Based Student Mental Health Risk Mapping

Abstract

Existing university mental health monitoring often depends on voluntary help-seeking or manual questionnaire interpretation, which may delay early support for students experiencing academic stress. This study proposes an explainable XGBoost-based early-warning framework for non-clinical mapping of student mental health risk from academic stress indicators. The single-site dataset comprised 1,002 anonymized student records from Universitas Muria Kudus. K-Means clustering was used to transform DASS-21 depression, anxiety, and stress scores into low, moderate-, and high-risk categories, while XGBoost predicted the cluster-derived labels using seven single-item academic stress indicators and engineered aggregate and interaction features. On a stratified hold-out testing set of 201 records, the model achieved weighted precision, recall, and F1-score values of 0.8907, 0.8905, and 0.8906, respectively, with class-level F1-scores of 0.9109 for low risk, 0.8900 for moderate risk, and 0.8713 for high risk. Additional ablation, clustering sensitivity, subgroup, threshold, and SHAP stability analyses were conducted to strengthen robustness and interpretability. The findings show that cumulative academic stress and interaction features involving parental expectations, exam anxiety, and learning-method adaptation were consistently influential predictors. The framework is intended to support early institutional prioritization and counseling referral, not clinical diagnosis. Generalization remains limited by the single-institution sample and the use of single-item academic stress indicators; therefore, local retraining and recalibration are required before institutional deployment, including implementation of the Streamlit prototype.