Factors influence road accident severity in Bangladesh for heavy and non-heavy vehicles a machine learning-based comparative study

Abstract

Road accidents have become a significant issue in Bangladesh due to the staggering number of incidents yearly. Some major critical factors influence the increase in the severity of accidents which affects the national economy, social life, and public health issues. This study addresses critical gaps in road accident severity research by focusing on rural and suburban areas in Bangladesh, where traffic patterns and infrastructure differ significantly from urban centers. It leverages advanced machine learning techniques, such as Random Forest (RF) and Extreme Gradient Boosting (XGBoost), to enhance predictive accuracy and analyze high-dimensional data. By incorporating vehicle-specific clustering, the study uncovers distinct severity patterns for heavy and non-heavy vehicles, providing targeted safety insights. Additionally, the integration of interpretability methods like feature importance, permutation importance and SHAP (Shapley Additive Explanations) ensures actionable insights into the influence of factors such as vehicle type, weather, and driver behavior, bridging the gap between prediction accuracy and practical, real-world applicability in road safety interventions. Police-reported accident datasets from 2006 to 2015 are used for this purpose. The analysis indicates that for single-vehicle collisions, the type of collision is a critical factor for both heavy and light vehicles. For heavy vehicles, the presence of road dividers significantly influences accidents, while road classification is more impactful for light vehicles. In two-vehicle collisions, factors such as the presence of dividers and movement patterns play important roles, with the availability of fitness certificates particularly affecting collisions between heavy and non-heavy vehicles. Additionally, road class, time of day, and environmental conditions are key contributors to heavy-heavy vehicle collisions, whereas location type and district characteristics are more relevant for light-light vehicle collisions. These insights highlight the importance of context-specific policy measures to improve road safety in emerging economies.