Optimizing Depression Prediction Models: A Stacked Ensemble Approach with Weighted Feature Importance

Depression is a major mental health issue affecting students, with academic pressure, social support, and personal factors contributing to its development. Early detection of depression is essential to provide timely interventions. This study explores the application of machine learning models, particularly ensemble learning techniques, to predict depression among university students. The study compares the performance of several classifiers, including Random Forest, XGBoost, Extra Trees, Gradient Boosting, and LightGBM. In this study, a stacked model combining the top two classifiers, XGBoost and LightGBM, was developed with the aim of improving predictive performance. Feature importance analysis revealed that academic pressure and a history of suicidal thoughts were the most significant predictors of depression. The stacked model outperformed individual classifiers, demonstrating the power of ensemble methods in improving prediction accuracy. These findings suggest a potential pathway for refining depression prediction models, and incorporating this innovation could offer refined insights into key features to inform future interventions for at-risk students experiencing depression.