Improving transparency and predictability of intersection crash severity analysis through Explainable Boosting Machines

Road traffic crashes remain a significant cause of fatalities and injuries worldwide, with intersections being particularly prone to severe collisions. Addressing intersection-related crashes requires an in-depth understanding of contributing factors to develop effective safety interventions. This study focuses on the binary classification of intersection crashes into injury and non-injury categories using interpretable, glass-box Machine Learning (ML) models, including Explainable Boosting Machines (EBM), Logistic Regression (LR), and Decision Trees (DT), to classify and predict injury outcomes while interpreting contributing features. Data augmentation techniques such as SMOTE, and its variants were used to handle the imbalanced data. Utilizing augmented data and Bayesian optimization for hyperparameter adjustment, all glass-box models were trained. Among the empirical modeling techniques, EBM with SMOTE-Tomek treated data provided strong prediction performances and interpretable results, outperforming other glass-box models in terms of Sensitivity (70%), Specificity (61%). F1 score (71%), Balanced Accuracy (66%), and Matthew’s Correlation Coefficient (MCC) score (0.30). Moreover, while predicting the severity of intersection collisions, the EBM approach enables detailed analysis of single and pairwise factor interactions. The most influential variables identified by the model fall into distinct categories: demographic factors (e.g., gender) and infrastructure factors (e.g., junction type, control devices), which can guide policymakers in prioritizing intersection redesign and control improvements. These interpretable insights provide evidence-based directions for reducing crash severity and enhancing intersection safety.