Predicting the impacts of occupational accidents on the construction sites using machine learning algorithms

Purpose This study aims to develop a machine learning–based framework to predict the retrospectively assessed multidimensional impacts of workplace accidents in the construction industry. Unlike traditional research that primarily focuses on injury severity or fatality outcomes, this study shifts the analytical focus toward post-accident consequences affecting companies and projects, including reputational loss, service delays, cost overruns, workforce restructuring needs, and contractor withdrawal risks. Design/methodology/approach The analysis is based on 203 valid survey responses collected directly from building construction professionals. Nine machine learning algorithms (Ordinal Logistic Regression, K-Nearest Neighbors, Decision Tree, Random Forest, Extra Trees, AdaBoost, XGBoost, LightGBM, and CatBoost) were evaluated using five-fold stratified cross-validation. To address class imbalance, multiple resampling techniques (Raw Data, Synthetic Minority Oversampling Technique, Adaptive Synthetic Sampling, and Synthetic Minority Oversampling Technique–Edited Nearest Neighbors) were tested. Model performance was primarily assessed using the weighted F1 score. Furthermore, SHapley Additive exPlanations (SHAP)-based interpretability analysis was conducted to identify the accident characteristics contributing to higher-impact predictions. Findings The findings indicate that the predictability of accident impacts fluctuates based on the results influenced by class structure and separability. Ensemble-based models frequently exhibit competitive performance, yet no single algorithm has consistently surpassed the others. SHAP-based analysis underscored diverse feature contributions across impact sizes and stressed the importance of interpretable and bias-aware machine learning frameworks in construction safety evaluations. Originality/value This study advances construction safety research by introducing an impact-oriented predictive modeling approach that extends beyond accident occurrence. By integrating systematic model comparison with explainable machine learning techniques, the proposed framework may provide a preliminary data-driven basis for the future development of impact-sensitive decision support systems in construction safety management. The study identified accident characteristics that contribute positively or negatively to model predictions of higher impact levels, while highlighting the significance of data distribution and sample size in interpreting these results.