Severity prediction of traffic accidents with recent machine learning paradigms and SHAP

Abstract Traffic-related fatalities are among few leading causes of death worldwide. Crash severity prediction of traffic crashes has become a widespread research subject in recent years. Severity outcomes of crashes are usually influenced by factors such as humans, vehicles, roadway, weather, environment, and their interactions. A better understanding of crash-contributing circumstances and crash severity risk predictors is vital for the timely and effective implementation of appropriate mitigation measures. In the literature, various statistical methods have been widely adopted for crash severity modeling; however, such approaches are based on several underlying unrealistic assumptions, which, if violated, may lead to biased predictions. The current study aims to investigate the applicability of the state-of-the-art gradient boosting algorithm “XGBoost” for crash severity prediction of traffic crashes in the Al-Qassim region, Saudi Arabia. Crash severity classification performance of XGBoost was compared with two traditional machine learning algorithms, including Random Forest (RF) and Logistic Regression (LR), using different statistical measures including accuracy, recall, precision, F-1 scores, receiver operating characteristic (ROC) curves, and area under the curve (AUC). Empirical results revealed that all the models yielded acceptable predictive performance; however, XGBoost achieved the highest accuracy (90.5%), outperforming both the LR and RF models. Models’ comparison using performance metrics also demonstrated superior predictive performance of the XGBoost model. To address the infamous “lack of interpretation” issue frequently raised against the application of machine learning, this study also presents XGBoost-based feature sensitivity and ranking analysis. The results showed that predictors including accident type, weather conditions, number of vehicles involved, lighting conditions, accident cause, and time of the day (TOD) have a strong influence on the crash severity outcome of crashes. SHAP analysis was also performed to assess the consistency of severity risk factors with the XGBoost feature importance technique. The findings of this study are expected to provide crucial insights to safety practitioners and policy-makers for improving highway safety.