Factors associated with casualty injury severity in e-scooter crashes: A machine-learning approach with SHAP explainability

The rapid adoption of electric scooters (e-scooters) as a micro-mobility mode has been accompanied by rising crash-related injuries, yet the factors associated with severe outcomes remain poorly understood. Because the released dataset is organised at the casualty level, the outcome analysed here is the injury severity of an individual casualty involved in an e-scooter crash, classified as slight versus serious/fatal ; the unit of analysis is the casualty record, not the collision. Using 5166 police-reported casualty records from Great Britain between January 2020 and November 2024, we examine 25 crash-scene, rider, casualty, and environmental features. Five supervised classifiers are compared via randomised hyperparameter search with stratified five-fold cross-validation, with all oversampling confined to the training folds; an integer-encoded SMOTE specification is held fixed as a common reference for comparison rather than as a preferred model. The three gradient-boosted tree models have similar aggregate performance: LightGBM attains a weighted F1 of 0.660 (AUC-ROC 0.624), while paired McNemar tests and bootstrap intervals show that LightGBM, XGBoost, and Gradient Boosting each outperform the fitted Logistic Regression baseline; the LightGBM–Random Forest comparison is not robust across the two inferential checks. Cost-sensitive learning raises serious-class recall to 0.536, and a native-categorical CatBoost pipeline raises it to 0.559. Although weighted F1 remains between 0.653 and 0.661 across the four encoding pipelines, minority-class detection and SHAP rankings are encoding-sensitive. SHAP identifies number of vehicles, latitude, casualty sex, time-of-day band, and rider age as the leading predictors in the common reference model, although encoding sensitivity shows that only NumVehicles and DayTime occur in every encoding-specific top five, and the cost-sensitive ranking agrees only moderately with the SMOTE model ( ρ = 0.6662 ). Casualties recorded in single-vehicle incident configurations had higher descriptive serious/fatal injury rates than those recorded in two-vehicle configurations (38.6% versus 26.6%), and a refit restricted to the driver/rider-casualty subset used as a proxy for e-scooter riders — in which NumVehicles remains the top predictor with a stronger single-vehicle effect — indicates the pattern persists within that subset rather than arising solely from mixing riders with pedestrians. A fall or head-impact mechanism is plausible but unmeasured: helmet use, fall occurrence, and head impact are absent from the data, and the subset used to test the pattern is defined by a driver/rider role code that proxies for, rather than verifies, e-scooter rider status. Records involving older riders had higher serious/fatal injury rates (28.8% below age 50 versus 42.7% at or above 50), without establishing age 50 as a policy threshold. DayTime band 0 had the highest observed serious/fatal rate, but the four-band rates are not strictly monotonic and their clock-time meanings are undocumented. Discrimination is moderate and the predicted probabilities are poorly calibrated (Brier 0.219 against a 0.208 prevalence baseline), so these results support cautious, locally evaluated hypotheses rather than causal, operational, or universally transferable prescriptions.