Explainable AI-driven hybrid modeling for crash severity analysis near bus stops in urban areas

Crashes near bus stops in urban areas present critical safety risks for pedestrians and vehicle occupants, especially in high-traffic areas with frequent transit activity. This study investigates crash severity patterns near Texas bus stops from 2017 to 2023 using a hybrid explainable AI framework. XGBoost and Random Forest models were applied, as part of feature selection process, to identify key contributing factors, which were subsequently used in Cluster Correspondence Analysis (CCA) to determine latent crash typologies. Six distinct crash clusters were identified, including intersection-related collisions, midblock crashes involving parked vehicles, impairment-related incidents, and fatal pedestrian or bicyclist crashes at crosswalks and bike lanes. Contributing factors such as inadequate traffic control, poor lighting, vehicle type, and impaired driving were prominent. Vulnerable road users, including older adults, cyclists, and pedestrians, were disproportionately affected in areas with limited visibility and substandard infrastructure. To enhance interpretability, SHAP values were computed within each cluster by providing localized insights into variable importance across crash scenarios. The findings inform targeted countermeasures, such as improved lighting, intersection redesign, and bus stop reconfiguration, to mitigate crash severity near bus stops and promote safer multimodal travel environments.