Prioritising construction accident risk factors using an integrated Random Forest, XGBoost, and ANN framework with sensitivity analysis

Abstract

Construction safety remains a critical concern due to accidents influenced by multiple interacting factors. While previous studies have identified key risk drivers, robust, data-driven approaches for prioritising them remain limited. This study proposes an integrated machine learning and sensitivity-based framework to systematically prioritise construction accident risk factors. A dataset of key safety-related variables was analysed using Random Forest Regression, Extreme Gradient Boosting, and Artificial Neural Networks. Model performance was evaluated using standard metrics, and sensitivity analysis quantified the relative importance of each risk factor. Within the available dataset, the Random Forest model achieved the highest observed test correlation coefficient of 0.813; however, statistical testing did not indicate significant differences between model prediction errors. Sensitivity analysis revealed that physical strain sub-factors, particularly mechanical and electrical strain, gravity-related tasks, and motion-related exertion, collectively account for 33.74% of the total influence on predicted accident risk. Cognitive demand factors, specifically perceived time pressure and workflow interruptions, represent the most influential secondary drivers. The top five sub-factors alone account for nearly half of the cumulative risk impact, providing a preliminary, data-driven basis for targeted safety interventions within the studied context, subject to further validation with larger, more geographically diverse datasets. The proposed framework offers a promising decision-support tool for construction stakeholders, enabling data-driven prioritisation of safety risks and supporting more informed resource allocation for risk mitigation, subject to further validation in real industrial environments.