Explaining Road Crash Severity with Machine Learning: A Predictive and Interpretable Framework

Accurate and interpretable crash severity prediction is necessary for enhancing road safety and data-driven transportation policy. This study proposes a machine learning (ML)-based framework for binary crash severity classification using 9,000 real-world crash records from Great Britain, comprising both categorical and numerical attributes related to roadway, vehicle, environmental, and driver factors. To address class imbalance, the original severity labels were transformed into a binary outcome and balanced using the Synthetic Minority Over-sampling Technique (SMOTE) in combination with class-weighting. Three classifiers Extra Trees Classifier (ETC), Random Forest Classifier (RFC), and Decision Tree Classifier (DTC) were developed and comparatively evaluated. The results show that ETC had the best performance, with an F1-score of 0.93 and recall of 0.93 for the severe crash class. RFC produced comparable performance, with an F1-score of 0.93 and a higher recall of 0.97, indicating strong sensitivity to severe crashes. Although DTC yielded lower predictive performance, it offered greater structural simplicity and interpretability. To enhance transparency, SHapley Additive exPlanations (SHAP) analysis was employed to identify the most influential predictors of crash severity. Receiver Operating Characteristic (ROC) curves, Precision-Recall (PR) curves, and confusion matrices further supported model evaluation. The findings demonstrate the effectiveness of ensemble learning combined with explainable artificial intelligence (XAI) for reliable crash severity prediction.