Analyzing Factors Influencing Temporary and Permanent Work-Related Accident Disabilities: An Interpretable Machine Learning Approach

This study investigates the use of machine learning (ML) to predict work-related accident injury outcomes in Brazil, specifically distinguishing between temporary and permanent disabilities. A dataset of 4,476 well-documented cases, collected between 2014 and 2022, was constructed from the Brazilian Notifiable Diseases Information System (‘Sistema de Informação de Agravos de Notificação’, SINAN). To identify the 10 most influential predictors, this study integrated the rankings from three distinct algorithms: Kruskal-Wallis, ANOVA, and MRMR. Several classifiers were then evaluated, including Random Forest (RF), Support Vector Machines (SVM), Decision Trees (DT), Naïve Bayes (NB), and K-Nearest Neighbors (KNN). Compared to a baseline model that achieved an accuracy of 66.62% and an F1-score of 53.28%, the RF model recorded the highest overall performance with an accuracy of 86.84% and F1-score of 86.59%. To enhance model interpretability, Shapley Additive Explanations (SHAP) analysis was employed, revealing that affected body part, accident type, and occupational role are critical predictors of disability outcome. These findings highlight the potential of ML, coupled with explainability tools, to facilitate early risk identification and inform targeted occupational health interventions, aligning with regulatory frameworks and established occupational health literature.