Implementation of an Ensemble Stacking Model for Early Prediction of Depression in University Students

Abstract

The increasing number of mental disorders among university students, particularly depression, highlights the urgency of early and effective interventions to prevent serious consequences on their academic and social performance. This study proposes an Ensemble Stacking-based model that combines Logistic Regression, Random Forest, and Decision Tree to predict depression, using variables such as gender, age, academic performance, and lifestyle factors. The featured model (Case 3) achieved 97.67% accuracy and 97.00% cross-validation, optimized by specific hyperparameter settings. Advanced preprocessing techniques, such as SMOTE for class balancing and PCA for dimensionality reduction, ensured the robustness and generalizability of the model. The methodology included an XGBoost metamodel, leveraging the diversity of the base models to improve accuracy and mitigate overfitting on complex data. These results demonstrate the effectiveness of the proposed approach for early detection of depression in educational contexts, providing practical tools that can facilitate timely and personalized interventions.