Machine Learning-Based Analysis of Fatal Construction Accidents Using SHAP: Insights for Safety-Assistive Vehicle Applications

Abstract

The construction industry is among the most hazardous sectors, with frequent serious injuries and fatalities. This study investigates the key factors contributing to fatal accidents and explores how safety-assistive vehicles—currently limited to basic alarm and control functions—can be advanced into comprehensive safety management tools. Utilizing Korea's national accident database (CSI) from 2019 to 2023, we analyzed 15,807 cases, including 807 fatal incidents (5.1%). Predictive models employing CatBoost and AdaBoost yielded strong performance (AUC: CatBoost 0.912; AdaBoost 0.908). SHAP analysis identified top predictors of fatality: falls, worker negligence, hazardous objects, small-scale sites (<20 workers), and high-value projects (>$76.9M). Our results indicate that integrating predictive analytics may enable safety-assistive vehicles to go beyond alarms, facilitating real-time detection of accident risks, hazardous zones, and unsafe behaviors. This proactive capability can enhance safety management at construction sites. The study demonstrates the practical utility of machine learning for identifying high-risk conditions and guiding the development of smarter safety-assistive systems. Future research will focus on applying computer vision and detection technologies to further improve real-time accuracy.

Keywords: Construction safety, accident analysis, machine learning, CatBoost, AdaBoost, SHAP

DOI: 10.54941/ahfe1007050