Order-based frameworks for predicting and classifying absence duration following an occupational injury

Workplace injuries can lead to prolonged employee absences, generating substantial costs related to lost productivity and wages. Accurate prediction of cases leading to extended leave is essential for efficient resource allocation and return-to-work planning. While machine learning models can be applied to estimate absence duration, their predictions are often not reliable for guiding organizational decision-making. In this study, we propose two novel order-based frameworks for evaluating a regression model’s ability to prioritize injury cases with the longest absences. Using a dataset of workers’ compensation claims, the rank-ordering framework applies a regression model, CatBoost, decision tree, or multiple linear regression, and assesses how well its predicted ranking captures the longest actual absences at varying cutoff points. Then, the order-classification framework assigns binary labels by thresholding the ordered regression outputs, focusing on extended absences. We compare the order-classification performance against three widely used classification models: CatBoost, decision tree, and logistic regression. Results indicate that, despite limited predictive accuracy, the regression models prioritized a substantial share of prolonged absence cases. In addition, the order-classification framework consistently achieves competitive or more effective identification of extended absences compared to traditional classifiers under moderately imbalanced conditions. This study provides the first application of our rank-ordering and order-classification frameworks for prioritizing long-duration workplace injury absences, offering practical insights for occupational health and workforce management.