JMIR Publications

ABSTRACT

Background:

Medical postgraduates carry a high burden of comorbid depression and anxiety, yet few accessible, low-cost screening tools exist. Social capital could serve as an efficient stratification signal; however, whether it can be operationalized as a brief digital screening module for university mental-health platforms remains untested.

Objective:

To develop and validate a parsimonious, digitally deliverable social-capital–based machine-learning model for screening depression–anxiety comorbidity among medical postgraduates and to quantify the incremental value of social capital over other low-cost variables.

Methods:

In a cross-sectional survey of 2,587 medical postgraduates from five Chinese medical schools, we collected demographic, lifestyle, communication-behavior, and six-dimension social-capital data. Comorbidity was defined as CES-D ≥ 16 and SAS ≥ 50. Participants were stratified into training (60%, n=1,551), validation (20%, n=518), and test (20%, n=518) sets. Five models—logistic regression, random forest, XGBoost, CatBoost, and TabPFN—were compared; screening thresholds (sensitivity ≥ 0.85) and isotonic calibration were fixed on the validation set before a single test-set evaluation. Ablation, leave-one-school-out (LOSO) validation, SHAP/odds-ratio, learning-curve, and subgroup analyses were performed.

Results:

Comorbidity prevalence was 29.7% (768/2,587). Social-capital dimensions alone (AUROC 0.749) outperformed 13 demographic and lifestyle variables (0.641; ΔAUROC +0.108), whereas adding all non-social-capital variables yielded only marginal gain (+0.017). A parsimonious 8-variable model achieved screening-grade discrimination with no significant loss versus the full model (bootstrap ΔAUROC 95% CIs spanned zero). All five models converged near AUROC 0.74–0.77 (e.g., TabPFN 0.765, 95% CI 0.720–0.805; sensitivity 0.896, specificity 0.398, NPV 0.901). Leave-one-school-out internal–external cross-validation indicated cross-school stability (0.751 ± 0.020), and the learning curve plateaued at ~40% of training samples. At the pre-specified operating point, decision-curve analysis showed a positive net benefit relative to both treat-all and treat-none strategies across threshold probabilities of approximately 0.15–0.55, supporting the model's utility as a first-stage digital screener that can spare high-risk students from full-scale assessment.

Conclusions:

social-capital variables provided the largest incremental improvement in discrimination in this dataset. A parsimonious 8-variable model achieved an AUROC of approximately 0.74–0.77 with sensitivity ≥ 0.85 and NPV ≈ 0.90, and performance remained stable in leave-one-school-out internal–external cross-validation; neither more complex models nor additional non-social-capital features conferred meaningful incremental value. As a brief, digitally administered module, the model offers a low-burden first-stage screening option for university mental-health platforms. Independent external validation in other settings and prospective evaluation of the proposed two-stage screening workflow are warranted. Clinical Trial: none.