Abstract

Rapid urbanization and increasing traffic density have increased the crash risk of vulnerable road users (VRUs), particularly pedestrians and cyclists. Meanwhile, the conventional statistical models have difficulties in capturing the nonlinear and complex nature of crash data, limiting their safety analysis effectiveness. This study utilized advanced Gradient Boosting machine learning and integrated it with SHapley Additive exPlanations (SHAP) using five years of crash data from Michigan, USA, employing a two-tiered modeling design consisting of a 4-class joint structure and binary subset frameworks. Rigorously evaluated using 10-fold stratified cross-validation to predict crash severity for VRUs, the CatBoost model had better predictive performance (AUC = 0.917) than LightGBM, Random Forest and the traditional Logistic Regression models. The analysis further indicated that prior crash actions, particularly risky crossing behaviors, are the most significant determinants of injury severity for both user groups. However, the pedestrian crash severity is strongly associated with lighting conditions and speed limits, while cyclist crash severity is more heavily influenced by intersection involvement and roadway geometry. Moreover, SHAP interaction analysis showed that the speed effect on severity significantly increases when it interacts with hazardous actions or poor visibility. The findings provide a critical insight into the implementation of effective measures and infrastructure improvements to increase the safety of VRUs.

IPC Classification

Keywords