Pedestrians account for more than one-fifth of global road fatalities each year, with bus stops in metropolitan areas being high-risk locations due to crowding and frequent crossings. Developing effective interventions to enhance pedestrian safety requires a deep understanding of the built environment factors that contribute to pedestrian crashes across different urban settings. Since pedestrian injuries and fatalities are primarily preventable, applying diverse datasets from various countries is essential to identify and compare the key built environment features influencing pedestrian crashes around bus stops. This study evaluates and predicts the key built environment factors affecting pedestrian crashes within buffer distances of 50, 150, and 250 m around bus stops, using datasets from New York (United States), Toronto (Canada), and Greater Melbourne (Australia) between 2012 and 2016. Three modeling approaches, Random Forest (RF), Multi-Layer Perceptron (MLP), and Ordinal Logistic Regression (OLR), were applied with systematic hyperparameter tuning, and the Synthetic Minority Over-sampling Technique (SMOTE) was applied to address class imbalance. Model performance was compared across cities, buffer sizes, and training–validation–testing splits, and external cross-city validation was conducted to evaluate transferability. Results show that RF consistently outperforms MLP and OLR by capturing nonlinear interactions between built environment features more effectively, with the best-performing RF models for New York and Toronto using a 250-meter buffer, demonstrating that larger buffer distances better capture the influence of the built environment on crash occurrences. Furthermore, Common Important Features (CIF) and Common Important Subfeatures (CIS) are extracted to rank and compare the most influential factors affecting pedestrian crashes in each case study. This information may be used to persuade political leaders to develop, implement, and support pedestrian safety measures around bus stops.

Keywords