Introduction: Mental Health in Bangladeshi Universities
Mental health is a cornerstone of overall well-being, impacting students’ academic success, social connections, and long-term trajectories. This study from Bangladesh investigates how common depression, anxiety, and stress are among university students and identifies their key risk factors. It also evaluates how well several machine learning (ML) models can predict these mental health outcomes, offering insights for targeted interventions on campus.
Prevalence: How Common Are These Conditions?
Among 1,697 respondents from two public universities, the study found high rates of psychological distress: depression 56.9%, anxiety 69.5%, and stress 32.2%. Severity varied, with a sizable portion experiencing moderate to extremely severe symptoms. These figures align with prior Bangladeshi and regional findings, suggesting pandemic-era pressures, ongoing academic demands, financial concerns, and cultural factors all contribute to elevated distress among students in LMICs.
Associated Factors: Who Is Most at Risk?
The analysis highlighted several demographic, social, and behavioral risk factors. Women showed higher odds of stress, while unfriendly family relationships increased risks across all outcomes (depression, anxiety, stress). Academic factors mattered too: students in arts and commerce faculties reported higher depression rates than science students, and first-year students faced distinct stress patterns. Substance use, including cigarette smoking and illicit drugs, correlated with higher depression and anxiety. Interestingly, socioeconomic status did not show a robust association in this sample, potentially due to a relatively homogeneous student population or campus buffering effects.
Machine Learning: Predicting Mental Health Risk
The study applied seven ML models—KNN, Random Forest (RF), XGBoost, CatBoost, Logistic Regression (LR), Gradient Boosting Machines (GBM), and Support Vector Machines (SVM)—with 80/20 train-test splits and 5-fold cross-validation. SMOTE addressed class imbalance in training. Feature importance analyses (SHAP values and XGBoost gain scores) identified smoking status and family relationship quality as influential predictors, among other factors such as faculty and gender. While several models achieved reasonable accuracy and precision, overall discriminative power (AUC-ROC) remained modest, suggesting that mental health outcomes are shaped by complex, multifactorial factors not fully captured by the available features.
Key ML findings include:
– SVM performed well for depression and anxiety in several metrics, but overall AUCs hovered near 0.5–0.55, indicating limited discriminatory power.
– CatBoost showed strong performance for stress and robust calibration (low log loss), while LR offered solid results for some outcomes due to interpretability.
– Feature importance underscored the relevance of behavioral and psychosocial variables, such as smoking and family dynamics, in predicting mental health status.
Implications for Policy and Campus Practice
These results carry clear implications for university policies in Bangladesh and similar contexts. Universities should implement routine mental health screening, especially for students with identified risk factors like unfriendly family relations, female gender, higher academic years, and substance use. On-campus counseling services must be accessible, confidential, and culturally appropriate, with staff trained to address diverse student needs. Peer-support programs, mental health literacy campaigns, and anti-stigma initiatives can foster help-seeking and resilience. At the policy level, education authorities should prioritize student mental health, allocate resources, and consider national guidelines for university mental health services. Finally, data-driven approaches—such as electronic risk assessments and early warning systems—could complement traditional supports, provided privacy and ethical standards are upheld.
Limitations and Future Directions
The study relies on cross-sectional data and self-reported measures, which limit causal inferences and may introduce response bias. The sample includes only two public universities, excluding private institutions, and used convenience sampling, which may affect external validity. Future research should employ probability-based designs, longitudinal follow-ups, and a broader range of features, potentially including digital behavioral markers, to improve predictive performance and deepen understanding of mental health determinants in Bangladeshi university settings.
Conclusion
Depression, anxiety, and stress are prevalent among Bangladeshi university students, with identifiable risk factors rooted in gender, family dynamics, academic pressures, and substance use. While machine learning offers valuable lenses for risk prediction and feature interpretation, predictive power remains imperfect, highlighting the need for richer data and longitudinal studies. The findings support comprehensive, multi-level strategies—ranging from campus services to national policy—that address mental health with evidence-based, context-specific approaches.
