Model-Agnostic Interpretability for Student Depression Screening: Integrating XAI and Multi-Criteria Model Selection

Depression in university students has emerged as a critical public health issue, necessitating scalable and comprehensible methodologies for early identification. This research introduces an Explainable Artificial Intelligence framework to forecast depression severity among Bangladeshi undergraduate students aged 18 to 28 years, utilizing sociodemographic, lifestyle, behavioral, and psychometric characteristics obtained from validated instruments such as the Beck Depression Inventory, Second edition, Patient Health Questionnaire 9, Center for Epidemiologic Studies-Depression scale, and the University of California Los Angeles Loneliness Scale. We train tree-based machine learning models like Decision Tree, Random Forest, XGBoost, and LightGBM on a 27-dimensional feature space and then test them using a wide range of metrics that measure predictive performance, calibration, stability, fairness, and interpretability. The results show that Decision Tree and LightGBM have the highest classification accuracy (96.0%) and the best ability to tell the difference between classes. XGBoost, on the other hand, gives the best calibrated probability estimates. SHapley Additive exPlanation-based explainability analysis shows that composite psychometric scores are the most important predictors of severe depression. This is backed up by large effect sizes and statistically significant differences between groups. The explanations are very stable across subsamples, and the fairness evaluation shows that there is very little bias between gender groups. These results show that combining accurate machine learning models with strong interpretability and fairness assessment can be a dependable and clear way to screen for depression in academic settings with limited resources.