Modeling and Interpreting Accident Severity in Autonomous Vehicles: A ROSE-LightGBM-SHAP Framework

Although autonomous vehicle (AV) technology is regarded as a transformative force for enhancing road safety, collisions remain inevitable during the transition phase toward large-scale commercial deployment, making it crucial to investigate the causal mechanisms of accident severity for optimizing AV systems and improving transportation infrastructure. To address the dual challenges of extreme class imbalance and the lack of interpretability in existing studies utilizing real-world data, this paper proposes an interpretable accident severity analysis model integrating Random Over-Sampling Examples (ROSE) with the LightGBM algorithm and SHAP (SHapley Additive exPlanations) framework. Utilizing real-world testing data from the California Department of Motor Vehicles (DMV), this study first establishes a rigorous “Split-then-Sample” strategy to effectively mitigate class imbalance while eliminating data leakage risks, and subsequently constructs the ROSE-LightGBM classifier. Experimental results demonstrate the model’s superior robustness in identifying minority class (injury) accidents, significantly outperforming baseline models such as Random Forest and XGBoost with a Recall of 80.00% and an F1-score of 77.42%. SHAP-based multi-dimensional analysis further reveals that severe vehicle damage, high point-of-interest (POI) complexity, and collisions involving vulnerable road users (VRUs) are dominant factors escalating accident severity, whereas high-level sensor configurations equipped with LiDAR and physically separated roads exhibit significant risk-mitigating effects; additionally, the study quantifies the non-linear threshold relationships between environmental features and accident risks. The analytical framework proposed in this paper not only provides an effective paradigm for modeling imbalanced data but also offers actionable, scientifically interpretable insights for the formulation of AV safety strategies.