Exploring e-scooter risk factors based on interpretable machine learning framework

The recent rise in e-scooter usage has reshaped urban mobility but has also led to a significant increase in e-scooter-related injuries, raising critical safety concerns. While existing research has primarily focused on post-crash medical outcomes and general risk comparisons, substantial gaps remain in identifying specific risk factors associated with e-scooter crashes and utilizing interpretable analytical approaches.

This study addresses these gaps by analyzing 2,400 e-scooter crash records from the UK STATS19 database using advanced machine learning models to predict injury severity. SHAP analysis and dependence plots were used to explore key risk factors and their interactions.

The assessment of model performance indicates that LightGBM outperforms other models, while SHAP analysis identified several key factors influencing e-scooter crash severity, including number of vehicles involved, first point of impact, rider's gender, lighting conditions, pedestrian crossings, and road types. Dependence plots revealed that male e-scooter riders are more likely to experience severe crashes with an increasing number of vehicles involved, in low-light conditions, and at higher speed limits. In addition, crashes occurring at higher speed limits on single-carriageways and wet surfaces increase the likelihood of severe injuries.

Based on these findings, the study proposes several recommendations to enhance e-scooter riders' safety.