Explainable machine learning for injury severity prediction in two-wheeler and light motor vehicle (LMV) single-vehicle crashes

Road traffic crashes remain a major safety hazard in India, and single-vehicle crashes (SVCs) with two-wheelers and light motor vehicles (LMVs) comprise a significant proportion of severe and fatal injuries. Despite their high risk, SVCs, particularly in India’s northeastern region, remain relatively underexplored in the literature. In addition, most existing studies rely on classic statistical models that assume linearity and offer limited interpretability of the complex crash mechanisms. To fill these gaps, this study used explainable machine learning (ML) methods to identify the influential factors of injury severity levels (fatal vs. non-fatal) resulting from SVCs involving two-wheelers and LMVs in Imphal city, India. Using ten years of police-reported crash data (2011-2020), six ML models were applied, with SHAP-based interpretation used to identify key contributors to injury severity levels in SVCs. Results indicated that different model classes were optimal for two-wheelers and LMVs, reflecting distinct severity mechanisms across vehicle types. The SHAP analysis consistently highlighted temporal factors, young and middle-aged road users, risky driving behaviors, and roadway deficiencies as the most influential determinants of injury severity levels resulting from SVCs. By integrating predictive modeling with interpretability, the study provides actionable insights for road safety policy, supporting targeted enforcement, awareness initiatives for high-risk groups, and prioritization of infrastructure improvements. The proposed framework also offers broader applicability for injury severity analysis of SVCs in other northeastern regions of India.