A hierarchical Bayesian approach for predictive analytics of depression severity in medical students

Abstract

Medical students experience disproportionately high rates of depression due to intense academic pressures and clinical demands. Without timely, targeted intervention, they face increased risks of academic underperformance and adverse outcomes. Existing predictive models often adopt a one-size-fits-all approach to predict depression for the entire student population. This approach may perform poorly for sparsely represented subgroups, such as medical students. To address this limitation, we propose a hierarchical Bayesian predictive model that estimates medical students' depression severity, even when medical students comprise only a small fraction of the overall dataset. Our hierarchical Bayesian modeling framework generalizes across subgroups via partial pooling, offering a novel analytical contribution to healthcare modeling in which subgroup imbalance is prevalent. Based on data from nearly 168,000 students, our model reduces the mean absolute error of predictions by at least 31% compared to baseline models, including XGBoost and deep neural networks. Statistical analysis using Wilcoxon Rank-Sum Tests with Bonferroni correction across more than 650 previously unseen medical students confirms that our model's performance is significantly superior to established baselines. Beyond improved predictive accuracy, our model identifies key depression-related stressors, including financial hardship, international student status, smoking frequency, and eating disorders. Accurate predictions and identified stressors help clinicians and academic administrators to recognize at-risk medical students and deliver timely, targeted interventions.