Development and internal validation of a machine learning model for work-related sick leave in Brazilian workers with mental disorders.

BACKGROUND: Mental disorders are a leading cause of work-related sick leave in Brazil, yet routinely collected surveillance data are seldom used to build prediction tools for occupational health services. We aimed to develop and internally validate machine learning models predicting work-related sick leave among individuals notified with work-related mental disorders. METHODS: We analyzed 21,601 notifications of work-related mental disorder (ICD-10 F99) from Brazil's national surveillance system (SINAN, 2006-2025). Elastic net, random forest, and XGBoost models were tuned by 5-fold cross-validation on a 75% training split and evaluated on a 25% test set. Discrimination, calibration, and clinical utility (decision curve analysis) were assessed; the model with the highest test-set area under the precision-recall curve (AUC-PR) was designated champion and examined for variable importance (SHAP) and fairness across sex, race/ethnicity, education, and region. RESULTS: Workplace absence was granted in 13,896 (64.3%) notifications. Random forest was selected as champion (test AUC-ROC 0.697 [95% CI 0.682-0.710]; AUC-PR 0.802 [0.788-0.816]), similar to XGBoost (AUC-ROC 0.699) and better than elastic net (AUC-ROC 0.680). Decision curve analysis showed positive net benefit over default strategies for threshold probabilities of approximately 30-70%. Notification year, employment status, psychotropic medication use, region, and confirmed work-relatedness contributed most to predictions. Discrimination varied substantially across subgroups, with markedly lower AUC-ROC for Black individuals (0.620) and workers with secondary education (0.664) than overall (0.697). CONCLUSIONS: Routinely collected surveillance data can support moderate-discrimination, potentially useful prediction of work-related sick leave in mental disorders, but subgroup disparities indicate that deployment as a decision-support tool would require targeted recalibration and fairness auditing, particularly for Black workers and those with secondary education. Internal validation alone is not sufficient to support deployment; external, prospective validation is required before any clinical or operational use, consistent with TRIPOD + AI guidance.