Early Depression Risk Screening using Leakage-Aware Ensemble Learning with Probability Calibration

Abstract

Depression among university students is an important mental health concern, particularly in technical and engineering education environments characterized by academic workload, competition, and financial stress. Although machine learning approaches are increasingly used in student mental health research, many previous studies rely on leakage-prone or post-outcome variables, limited validation protocols, and discrimination-focused evaluation, which may reduce their suitability for realistic early screening settings. This study presents a leakage-aware and calibration-aware ensemble learning framework for early depression risk screening among technical education students by excluding leakage-prone predictors during model development. Random Forest, Gradient Boosting, and stacking-based ensemble models were evaluated using nested stratified cross-validation, pooled out-of-fold prediction analysis, isotonic probability calibration, threshold sensitivity analysis, subgroup robustness assessment, and SHAP-based interpretability analysis. Experiments conducted on a publicly available dataset filtered to technical degree programs (N = 7,807) showed stable cross-validated performance, with the calibrated stacking ensemble achieving ROC–AUC = 0.8718 ± 0.0081, PR–AUC = 0.8888 ± 0.0078, and recall = 0.8408 ± 0.0100. Statistical testing showed no significant performance difference between stacking and calibrated Gradient Boosting, suggesting that simpler calibrated models may provide comparable screening performance in this dataset. Leakage-ablation analysis showed that inclusion of the post-outcome suicidal-thoughts variable increased pooled out-of-fold ROC–AUC from 0.8701 to 0.9224, highlighting the importance of leakage-aware feature selection for realistic evaluation. A reduced-feature screening model using six early-available predictors also maintained competitive performance (ROC–AUC = 0.8616), supporting the feasibility of lightweight institutional screening. SHAP stability analysis demonstrated consistent feature rankings across validation folds, with academic pressure, financial stress, work/study hours, and sleep duration identified as influential predictors. Overall, the proposed framework provides an interpretable and methodologically transparent approach for depression risk prioritization in educational settings while reducing performance inflation associated with leakage-prone features.