Predictive analysis of construction accident outcomes using explainable artificial intelligence and evidence fusion techniques

Abstract

Effective construction safety management relies on interpretable information to guide practitioner interventions. However, most current predictive models are “black boxes” that provide limited insight into the complex, multifactor causal mechanisms of accidents. Therefore, predictive analysis that provides explainable outcomes is necessary for identifying potential risks early, enabling timely interventions, and improving construction safety. A hybrid analytical framework is proposed that integrates interpretable machine learning using SHapley Additive exPlanations (SHAP) for feature attribution, and the Direct Linear Non-Gaussian Acyclic Model (Direct-LiNGAM) for causal discovery. To integrate these two complementary types of evidence, a modified Dempster–Shafer (D-S) evidence fusion scheme was developed. SHAP analysis revealed that inadequate hazard recognition made a major contribution to the prediction of serious accidents. Specific safety lapses, including inefficient supervision work, were dominant factors driving accident type predictions. The causal network constructed further highlighted the foundational upstream role of contract management in shaping accident outcomes. Through D-S evidence fusion, a too tight schedule and improper accident reporting and handling emerged as the most critical factors influencing accident severity and accident type, respectively. This study develops an integrated framework for construction accident outcome analysis that combines predictive modeling, SHAP-based feature attribution, Direct-LiNGAM-based causal discovery, and modified D-S evidence fusion. Its main methodological contribution is a modified per-variable binary frame of discernment that integrates correlation-based predictive evidence with causal evidence under the multifactor-coupled nature of construction accidents. The framework provides a traceable basis for identifying prediction-relevant risk factors and upstream causal drivers, supporting safety interventions across multiple aspects.