An explainable machine learning framework for analyzing and predicting mental health problems among university students in Bangladesh.

Mental health problems (MHPs) among university students are an increasing public health concern globally, including in Bangladesh. While machine learning (ML) methods can capture complex patterns in mental health data, their application to non-probability, cross-sectional student samples requires cautious interpretation. This study applies an exploratory explainable ML framework to examine patterns of depression, anxiety, and stress within a high-symptom sample of Bangladeshi university students. A cross-sectional online survey was conducted among 424 students during a period of significant social and political unrest in July 2024. Within the study sample, high proportions of moderate-to-extremely severe symptoms were observed for depression (66.0%), anxiety (73.4%), and stress (52.4%), with substantial symptom overlap (48.1% meeting criteria for all three). Eight ML models were evaluated using repeated train-test splits and cross-validation. Support vector machine (SVM) models showed relatively stable performance for depression and stress prediction, while XGBoost demonstrated comparatively stronger discriminative ability for anxiety within the sample. Explainable ML analyses using SHAP and LIME indicated that features related to sleep disturbance, cognitive complaints, and self-reported behavioral difficulties contributed prominently to model predictions across outcomes. Academic discipline and internet addiction severity were more influential in anxiety-related predictions, whereas gender and university type showed greater relevance in depression-related models. The primary contribution of this study is methodological and exploratory, not predictive or clinical. These feature contributions reflect model behavior and overlapping symptom dimensions rather than independent or causal risk factors. Overall, this study demonstrates the feasibility of using explainable ML approaches for exploratory pattern recognition in student mental health data under exceptional contextual conditions, providing a transparent framework, benchmark performance metrics, and a methodological template for future research. The findings are exploratory and sample-specific and should not be interpreted as population-level prevalence estimates or deployable prediction tools.