An explainable machine learning framework for predicting injury severity in extractive industry accidents

• Explainable ML model predicts severity in extractive industry incidents. • Hybrid framework integrates structured and narrative accident data. • SMOTETomek and Bayesian tuning enhance model classification accuracy. • SHAP ensures transparent links between features and severity outcomes. • Scalable approach supports safety management in high-risk industries. The mining, quarrying, and oil and gas extraction sectors experience disproportionately high rates of severe occupational accidents, frequently resulting in hospitalization and prolonged work loss. This study develops an explainable machine learning framework to predict injury severity outcomes using 2,645 accident investigation reports submitted to the U.S. Occupational Safety and Health Administration between 2015 and 2024. Accident narratives were preprocessed through a multi-step text mining pipeline. A hybrid approach combining TF-IDF vectorization, co-occurrence network analysis, expert screening, and accident taxonomy alignment ensured that both high-frequency descriptors and rare but safety-critical terms were retained. Dimensionality reduction was achieved with Truncated Singular Value Decomposition, and feature selection was conducted using a Random Forest classifier, which integrated structured attributes with enriched textual features. To address class imbalance, SMOTETomek was applied, and hyperparameters were optimized using Bayesian Optimization. Three ensemble models, Random Forest, Gradient Boosting Machine, and eXtreme Gradient Boosting—were compared against a logistic regression baseline. Random Forest achieved the best performance, with an F1-score of 0.94 and an overall area under the curve of 0.95. Beyond predictive accuracy, Interpretability was ensured through SHAP, which linked predictions to identifiable body parts, injury types, and contextual accident factors. The framework provides safety managers with a transparent, data-driven decision-support tool for prioritizing investigations, allocating resources, and identifying recurring high-risk patterns in extractive industries.