Road traffic accident severity prediction based on large-scale data and multi-dimensional factors: an explainable...

Road traffic accidents cause substantial fatalities and economic losses, yet large-scale severity analysis is often constrained by limited access to high-resolution roadway and geometric data. This study constructs a large-scale, multi-source dataset with fine-grained road-alignment characteristics to support data-driven analysis of accident severity patterns. Unlike prior work that uses coarse road descriptors, we automatically extract fine-grained horizontal and vertical geometry from extensive networks, enabling detailed geometric rarely available at a large scale. In total, 26 features were integrated, spanning environmental, roadway, and geometric alignment dimensions. A soft-voting ensemble integrating XGBoost, Random Forest (RF), CatBoost, and LightGBM (LGBM) is employed to support severity prediction, while SHAP (SHapley Additive exPlanations) is used to derive both severity-level importance rankings and sample-level explanations. Rather than focusing solely on predictive accuracy, this study uncovers mechanism-informed insights under rigorous evaluation settings designed to mitigate spatiotemporal dependence. The results show that environmental conditions are more strongly associated with lower-severity accidents, whereas roadway type and geometric alignment features become increasingly important for higher-severity accidents. Furthermore, Accumulated Local Effects (ALE) analyses reveal nonlinear patterns and threshold regions for key continuous variables, providing complementary evidence on how risk varies across operating conditions. These findings provide mechanism-informed insights that can support the development of targeted road safety strategies.