Predicting Depression Risk in University Students Using Transparent Machine Learning (Preprint)

UNSTRUCTURED Depressive symptoms among university students in India are increasing, yet routine screening using standardized clinical instruments remains limited in large educational systems. This work developed an interpretable machine-learning frame- work to estimate depressive symptom burden using a theory- driven proxy index derived from secondary psychosocial and contextual data. A cross-sectional secondary analysis was con- ducted using an anonymized student-level dataset (N = 27,901) and aggregated regional mental-health indicators across 231 administrative entities. Psychosocial variables representing sleep disruption, academic pressure, financial stress, social support, and related domains were combined into a proxy depressive- symptom burden index with acceptable internal consistency (Cronbach’s α = 0.78). Four classifiers (logistic regression, ran- dom forest, support vector machine with radial basis kernel, and gradient boosting) were evaluated using cross-validated and held- out test performance, calibration, and decision-curve analysis. Suicidal ideation was included only in sensitivity analyses to avoid target leakage. Gradient boosting demonstrated the strongest discrimination (test AUC=0.88, 95% CI 0.87–0.89) with good calibration and positive net benefit across threshold probabilities of 0.20–0.70. In the leakage-controlled model, academic pressure, sleep disruption, financial stress, and reduced social support were the most influential predictors, while regional anxiety prevalence and substance-use burden provided additional contextual signals. Risk stratification classified 16% of students as low, 46% as moderate, and 38% as high burden. Interpretable, theory-aligned machine learning can estimate depressive symptom burden from secondary data and may support resource prioritization in settings where standardized screening is not routinely available; such models should function as human-in-the-loop decision- support tools rather than diagnostic systems, with prospective validation and fairness evaluation required before deployment.