Investigating the factors of three-wheeled autorickshaw (3W-AR) crash severity: comparison of explainable machine learning models.

Three-wheeled autorickshaws (3W-ARs), locally called Bajaj, are a vital mode of public transport in Ethiopia. However, their crash involvement remains an overlooked critical mobility challenge. This study investigates the crash severity influence factors of 3W-ARs using five machine learning models: Random Forest (RF), Decision Tree (DT), Support Vector Machine (SVM), XGBoost, and LightGBM, along with the SHAP interpreter for explainability. Five-year crash record data (2019-2023) from Dire Dawa City were analysed. The results revealed that RF and DT outperformed the other three models with overall accuracies of 92.7% and 92.3%, respectively. Several attributes significantly contribute to the severity of 3W-AR injuries, including driving experience, rider educational level, pedestrian crossing behavior, and vehicle age. Notably, pedestrian-involved collisions particularly those resulting from unsafe pedestrian actions such as sudden road crossing or walking along traffic lanes emerged as one of the most dominant and consistent predictors of fatal and severe injury crashes. By integrating a machine learning model with explainable AI, this study advances data-driven approaches for enhancing 3W-AR safety and crash prevention measures. The findings provide critical insights for policymakers and transportation planners, enabling the development of targeted interventions suited to Ethiopian context.