Explainable Machine Learning for Concurrent Screening of Elevated Work-Related Fatigue Using Routine Working-Hour and...

Background: Work-related fatigue is a preventable occupational health concern in high-intensity manufacturing environments. However, supervisors and occupational health services often lack transparent criteria to support the preventive review of overtime practices. This study developed and internally validated an explainable human-in-the-loop concurrent screening model to identify employees with currently elevated work-related fatigue using routinely collected working-hour records and occupational health examination data. Materials and Methods: De-identified administrative and occupational health data from 314 full-time manufacturing employees were linked with standardized fatigue assessment data. Currently elevated work-related fatigue was defined using the top quartile of the standardized work-related fatigue subscale. To reduce information leakage, the work-related fatigue score used to define the outcome was excluded from the predictor sets. Multivariable logistic regression and XGBoost were compared using nested stratified five-fold cross-validation, with aggregated out-of-fold predictions from the outer folds used for the performance evaluation. Model performance was assessed using discrimination, probabilistic accuracy, calibration, and classification metrics at provisional, study-specific thresholds. Results: XGBoost demonstrated greater discrimination and better overall probabilistic accuracy than logistic regression, with an AUC of 0.77 versus 0.68 and a Brier score of 0.14 versus 0.17, respectively. At the provisional study-specific threshold of τ1 = 0.28, XGBoost achieved a sensitivity of 0.78 and a specificity of 0.70. Using τ1 = 0.28 and τ2 = 0.55, 58.0%, 27.1%, and 15.0% of employees were classified into provisional low-, moderate-, and high-priority tiers, respectively. The observed prevalence of currently elevated work-related fatigue increased across these tiers from 9.3% to 28.2% and 78.7%, respectively. Calibration assessment indicated systematic miscalibration for both XGBoost (intercept α = 0.78, slope β = 1.64) and logistic regression (α = -0.49, β = 0.56). Because τ1 and τ2 were selected and evaluated using aggregated out-of-fold predictions generated from the same single-organization development dataset, their stability and transportability are unknown. Medium-term cumulative overtime, particularly during months -2 to -6, together with routinely measured indicators of vulnerability and recovery, contributed substantially to the model-estimated probability of currently elevated fatigue. Conclusions: Routinely collected working-hour and occupational health examination data may support explainable concurrent screening for currently elevated work-related fatigue and illustrate a potential framework for differentiated preventive reviews. However, the proposed thresholds and review-priority tiers are provisional and study-specific and should not be interpreted as implementation-ready decision rules for occupational health practices. Prospective temporal validation, external validation in independent organizations, and local recalibration are required before threshold-based implementation.