Explainable Prediction of Prolonged Work Disability After Occupational Injury and Disease: Temporal Validation and Sex-Specific Performance.

PURPOSE: To compare explainable models for predicting prolonged work absence after occupational injury or disease and assess temporal calibration and sex-specific performance.

METHODS

We analysed 182,051 closed work-absence episodes recorded in Spain during 2022-2025. Prolonged absence was defined as ≥101 days. Penalised logistic regression, Random Forest, XGBoost and LightGBM were fitted on 2022-2023, selected in 2024 and refitted on 2022-2024. The previously examined 2025 cohort was reevaluated without using it for fitting or selection. Discrimination, calibration, Brier score and operating characteristics were assessed, with 2,000 paired episode bootstrap resamples. Additional analyses considered sex, alternative duration thresholds and nonlinear age and tenure; SHAP described predictor contributions.

RESULTS

The 2025 cohort included 45,773 episodes and 3,756 outcomes (8.2%). LightGBM had the highest ROC-AUC (0.829; 95% CI 0.823-0.836), logistic regression the highest PR-AUC (0.389; 0.373-0.406), and Random Forest the lowest Brier score (0.063). XGBoost had the highest F1-score (0.342). Sex contrasts did not establish equivalence. Injury description, body region, accident mechanism, age and tenure were the leading XGBoost SHAP contributors. Selection of closed episodes limited interpretation, particularly at 365 days.

CONCLUSION

No machine-learning model was uniformly superior to penalised logistic regression. Calibration, operational trade-offs, external validation and prospective impact evaluation remain important before implementation; availability of predictors at notification must also be verified.