Overview
Mental health is a cornerstone of overall well-being, yet university students in low- and middle-income countries face rising rates of depression, anxiety, and stress. This cross-sectional study from Bangladesh combines traditional epidemiology with machine learning (ML) to assess prevalence, identify associated factors, and predict mental health risk among students from two public universities.
Key findings: prevalence and severity
Among 1,697 Bangladeshi university students, the study found high levels of mental health symptoms: depression in 56.9%, anxiety in 69.5%, and stress in 32.2%. Severity varied, with substantial portions experiencing moderate to extremely severe symptoms (depression: 45.6% in moderate/extreme ranges; anxiety: 41.2% in moderate to extremely severe ranges; stress: 18.9% in moderate to extreme ranges). These figures align with pandemic-era results and highlight a persistent burden of psychological distress in Bangladeshi campuses.
Associated factors: sociodemographic, health, and behavioral contributors
The study identified several factors linked to mental health outcomes. Female gender was associated with higher odds of stress. Unfriendly family relationships markedly increased the risk of depression and stress, while friendship with family appeared protective. Academic factors mattered: students in later years and those in certain faculties, notably Commerce, showed higher rates of depression and anxiety, suggesting career-related uncertainty and workload pressures as potential drivers. Health behaviors, particularly cigarette smoking and illicit drug use, were significantly tied to depression and anxiety. Alcohol use showed associations with depression and anxiety as well, though not uniformly across all outcomes.
ML-based prediction: methods and outcomes
To explore predictive patterns, the researchers applied seven ML models—KNN, Random Forest (RF), XGBoost, CatBoost, Logistic Regression (LR), Gradient Boosting Machines (GBM), and Support Vector Machine (SVM)—using 80/20 train/test splits with SMOTE handling class imbalance. Five-fold cross-validation and hyperparameter tuning were used to optimize performance. Features included sociodemographic data, health behaviors, and DASS-21 subscale scores (Bangla version).
Model performance varied by outcome. SVM achieved the highest accuracy for depression (about 56.9%), with strong precision, while for anxiety, SVM again performed well (accuracy around 69.5%). CatBoost stood out for predicting stress (accuracy ~67.1%), with favorable calibration indicated by low log loss. Across outcomes, AUC-ROC values hovered near 0.5–0.60, indicating limited discriminative power in this dataset and signaling the need for richer feature sets or longitudinal designs to improve predictive capability.
Feature importance and interpretation
SHAP analyses highlighted key predictors. Cigarette smoking consistently emerged as a strong predictor across outcomes, while family relationship quality and faculty of study also showed significant predictive value. XGBoost emphasized relationship with family as a major factor, underscoring the role of social context in Bangladeshi student mental health. The study’s use of SHAP values enhances transparency, helping practitioners understand which factors most influence predictions in this population.
Implications for policy and practice
The high prevalence of depression, anxiety, and stress among Bangladeshi university students calls for urgent, multi-level interventions. Universities should consider routine mental health screening, expanded on-campus counseling, and targeted programs for students at higher risk (e.g., those in Commerce faculties, first-year vs. advanced-year students, and individuals with strained family relationships). Integrated approaches that address substance use alongside mental health, and engagement with families for supportive environments, may improve outcomes. Policymakers should recognize student mental health as a priority and support data-informed strategies, including digital risk assessments and privacy-protective ML tools, to identify and assist at-risk students early.
Limitations and future directions
Limitations include the cross-sectional design, convenience sampling from two public universities, and reliance on self-reported data. The ML models showed limited discriminatory power, suggesting the need for longitudinal data, more diverse institutions (including private universities), and additional behavioral or digital markers to enhance prediction accuracy.
Conclusion
This study advances our understanding of depression, anxiety, and stress among Bangladeshi university students by integrating epidemiology with ML. The findings illuminate critical risk factors and demonstrate the current potential and limitations of predictive models, guiding future research and informing targeted campus-based mental health strategies in Bangladesh and similar LMIC contexts.
