Temporal instability of highway pedestrian crash severity: Comparative analysis of machine learning models

Globally, pedestrian-involved crashes on national highways are considered unacceptable due to the clearly defined functional hierarchy of roadways. This study utilizes advanced machine learning (ML) techniques to investigate the severity of pedestrian injuries in traffic crashes, providing a comprehensive analysis of factors influencing crash severity across evening peak and other hours from 2016 to 2023. Statistical analysis reveals a higher fatality rate during evening peak hours (6:00–9:00P.M.) than other time periods. A chi-square test confirms a statistically significant difference in severity proportions between these periods (p < 0.01), supporting the concept of temporal instability—where the influence of contributing factors may vary over time. The study further evaluates the predictive performance of five ML algorithms: Support Vector Machine (SVM), Gradient Boosting (GB), AdaBoost, CatBoost, and Extreme Gradient Boosting (XGBoost). The best-performing model is interpreted using SHapley Additive exPlanations (SHAP), offering transparent insights into the relative importance of key predictors. CatBoost’s results highlight temporal differences in crash frequency and contributing factors, emphasizing the need to account for time-sensitive variations in crash severity. They highlight the importance of incorporating temporal dynamics into risk assessments and demonstrate the value of interpretable ML tools for guiding targeted, time-specific pedestrian safety interventions. This approach provides a robust foundation for developing data-driven policies that better address the complexities of pedestrian crash severity in varying temporal conditions.