Using the Bayesian Network Method to Evaluate the Effectiveness of College Students’ Mental Health Intervention Strategies and Their Impact on Academic Performance

Abstract

Mental health and academic success are increasingly interdependent challenges for university students worldwide. This study developed and validated dynamic Bayesian models to predict academic performance and psychological risk across semesters using probabilistic approaches. We analyzed a cohort of 3,276 undergraduates and externally validated findings against an independent cohort of 5,112 students. Dynamic Bayesian Networks (DBN) and Bayesian Networks (BN) were trained using psychological scores (PHQ-9, GAD-7, PSS-10, CD-RISC) to model psychological risk and academic records to model academic outcomes. Ten-fold temporal cross-validation was conducted internally, and comparative analyses involved Random Forests, XGBoost, Deep Neural Networks, and TabTransformer models. DeLong’s tests compared AUCs and permutation tests assessed Brier scores. Internally, BN achieved 91.0% accuracy, an AUC of 0.84 (95% CI 0.81–0.87), and a Brier score of 0.128, while DBN achieved 94.2% accuracy, an AUC of 0.86 (95% CI 0.84–0.89), and a Brier score of 0.124. In external validation, BN achieved 90.0% accuracy and an AUC of 0.88 (95% CI 0.85–0.90), and DBN achieved 92.0% accuracy and an AUC of 0.91 (95% CI 0.88–0.93). Top predictors included GPA, stress scores, depression scores, and intervention engagement. Posterior predictive p-values exceeded 0.44 across GPA and both outcome domains, indicating adequate calibration. Dynamic Bayesian modeling enables accurate, uncertainty-resilient prediction of both psychological risk and academic outcomes among university students.