Predicting safety accident costs in construction projects using ensemble data-driven models

The construction industry suffers from frequent and expensive safety accidents, significantly affecting construction project performance. Numerous data-driven classification models have been developed to categorize construction accident outcomes. While critical influencing factors provide insights for safety prevention, existing models have given less attention to the cost of accidents-an important indicator influencing management decisions. This study aims to develop accident cost prediction models that examine crucial precursors of safety accidents, offering guidance for construction safety prevention from a financial perspective. This study collected 1,606 accident reports from the Chinese construction industry between 2005 and 2022 to address this gap. Three ensemble data-driven methods, namely random forest, extreme gradient boosting regressor (XGBoost), and natural gradient boosting regressor (NGBoost) were employed to develop accident cost prediction models. Based on the performance comparison, the random forest regression model for accident cost was determined to be the best prediction model. To extract the critical attributes affecting safety accident costs, this study utilized shapely additive explanations (SHAP) value to analyze the sensitivity and influence of input variables of data-driven models. The findings showed that collapse has the greatest impact on accident costs, as indicated by the highest mean SHAP value, followed by falling from height. Furthermore, factors such as year, safety supervision, drawing, and construction plan are noteworthy in affecting accident cost prediction. Safety department, protection, and work conditions hold a slightly higher degree of influence compared to contracting arrangement, safety culture, safety supervision, training and examination, and mechanical equipment on the model output. This study provides a dimension that might be overlooked in the investigation of safety accidents in the construction industry and the insights provided by findings will contribute to the development of targeted safety accident prevention strategies.