Modeling Driver-Caused Road Crash Severity Using SHAP-Enhanced Machine Learning

Road traffic crashes continue to be a significant global issue, causing millions of fatalities and considerable social and economic losses annually. Saudi Arabia consistently reports one of the highest incidences of traffic-related crashes globally. Although numerous prior studies have approached crash prediction mainly as a classification issue centered on the probability of occurrence, few have investigated the fundamental relationships among the factors affecting crash severity, especially those related to driver behavior. This study examined driver-cause road traffic crashes throughout the Kingdom of Saudi Arabia (KSA) utilizing crash data from 2017 to 2022 sourced from the Ministry of Transport and Logistic Services (MOTLS). Crashes severity was categorized into three classifications: fatal, injury and property damage only (PDO). A variety of machine learning models were developed, including Decision Tree, Random Forest, AdaBoost, Gradient Boosting, and XGBoost. Among these, the XGBoost model performance well. Shapley Additive Explanations (SHAP) were examined to ascertain the global contributions of features for improved interpretability. Global SHAP analyses revealed that Accident Type, Accident Cause, Road Surface Condition, Road Type, and Vehicle Type were the most important predictors across all severity levels. Dependence analysis further uncovered strong feature interactions: speeding and freeway road type had the greatest impact on fatal crashes. Fatigue- and sleep-related causes, especially under nighttime conditions, were the contributors to injury crashes, while good road surface conditions and freeway road type were associated with property-damage-only outcomes. The integrated predictive and interpretable modeling framework provides, data-driven insights to inform targeted interventions aimed at reducing road accident severity across Saudi Arabia.