Development and validation of a probability-strength stacking ensemble model for depression risk prediction in college students: A SHAP-based interpretability analysis

Abstract

Depressive symptoms are common among college students, yet scalable screening tools that adequately capture nonlinear risk patterns remain limited. This study developed and validated a Probability-Strength Stacking (PSS) ensemble framework for depression risk prediction using a publicly available, anonymized college student mental health dataset from Kaggle (n = 27,901). Thirteen sociodemographic, psychological, behavioral, and economic predictors were analyzed. Data were split 80:20 into training and held-out test sets. XGBoost, LightGBM, Random Forest, SVM, CatBoost, hard-label stacking, and PSS Stacking-v2 were compared. PSS Stacking-v2 used continuous probability outputs from four base learners as meta-features for a logistic regression meta-learner, preserving inter-model confidence information discarded by conventional hard-label stacking. Performance was assessed by AUC, accuracy, sensitivity, specificity, F1, and Brier score, with 5-fold cross-validation, 1000-iteration bootstrap resampling, SHAP analysis, and decision curve analysis. On the held-out test set (n = 5581), PSS Stacking-v2 achieved AUC=0.9193 (95% CI: 0.9125–0.9266), accuracy=0.8461, sensitivity=0.8843, specificity=0.7920, F1=0.8706, and Brier score=0.1112, comparable to standalone CatBoost (AUC=0.9197, Brier=0.1098) and substantially outperforming hard-label stacking (AUC=0.8605; Brier=0.1251). SHAP identified suicidal ideation, academic pressure, and financial stress as the strongest predictors. The model showed positive net benefit across decision thresholds from 0.04 to 0.96, indicating good potential for practical risk stratification and early identification of depression among college students.