A heterogeneous stacking ensemble with explainable AI for mental health risk prediction in university students

To address the scarcity of scalable and interpretable mental health screening tools for university students, we developed MHSP-Stack (Mental Health Status Prediction via Stacking) and applied multi-level explainability analysis to identify key risk factors. Using the LIME (Local Interpretable Model-agnostic Explanations) framework complemented by a SHAP-based sensitivity analysis, we found that PHQ-9 depression score (28.6% relative importance) and GAD-7 anxiety score (21.9%) were the strongest predictors of mental health status, followed by exercise frequency (9.0%) and online stress (8.2%)—findings that challenge the assumption that academic pressure is the dominant risk factor. The Spearman rank correlation between LIME and SHAP global rankings was \(ρ _s = 0.86\) , confirming that these findings are robust to the choice of XAI technique. MHSP-Stack is a two-layer heterogeneous ensemble that combines six machine learning algorithms—logistic regression, support vector machine, random forest, gradient boosting, histogram gradient boosting, and LightGBM—in a base layer, with a logistic regression meta-learner integrating their outputs via stratified five-fold cross-validation to prevent data leakage. Trained on a dataset of 1,800 university students described by 15 validated mental health indicators, MHSP-Stack was benchmarked against 14 state-of-the-art algorithms and achieved the highest F1-score (0.8688) and area under the ROC curve (0.9241). We additionally benchmarked four alternative meta-learners, confirming that the linear meta-learner offers the best trade-off between performance, interpretability, and overfitting resistance, and evaluated robustness to distribution shifts showing F1 remains above 0.81 under severe measurement noise. MHSP-Stack offers an accurate, transparent, and deployable decision-support tool for university mental health early-warning systems.