Interpretable machine learning for imbalanced pedestrian injury severity prediction in urban Jordan

Abstract This study presents a machine learning framework for predicting pedestrian accident severity using Amman, Jordan's first complete 10-year traffic dataset (2014–2023). Addressing the critical class imbalance where minor injuries predominate (85%), causing standard models to poorly detect severe cases (< 25% recall), we implement cost-sensitive algorithms and specialized undersampling techniques, such as XGBoost with Balancing the Loss Function (XGBLF) and Random Data Partitioning with Voting Rule (RDPVR), which enhanced learning from underrepresented Major/Fatal cases while maintaining data authenticity. Through mixed-type correlation analysis and statistical testing, vehicle speed, road illumination, vehicle type, driver age, and road conditions emerged as the most significant predictive factors. RDPVR achieved a 63% true positive rate for Major/Fatal injuries, a 2.78-fold improvement over standard classifiers, and XGBLF achieved 95%, but this achievement was on account of the accuracy of the minor cases. Comprehensive interpretability analysis (SHAP, LIME, and Permutation Importance) revealed that heavy vehicles, poor lighting, and high-speed driving strongly predict Major/Fatal outcomes. Notably, the analysis demonstrates Jordan's improved safety trajectory, with 2020–2023 showing reduced severe accidents compared to 2014–2016, indicating measurable policy impact. This study delivers the first interpretable, context-sensitive AI framework for Amman/Jordan pedestrian safety, translating technical insights into actionable recommendations for targeted interventions, urban planning, and data-driven enforcement strategies to reduce pedestrian injury severity in high-risk zones.