Identifying factors affecting crash injury severity of pillion riders using interpretable machine learning techniques

In India, motorized two-wheeler (TW) riders account for 44.5% of fatal road crashes. While factors affecting drivers have been studied, research on pillion riders' injury severity remains limited. The study aims to identify factors causing severe injuries to pillion riders by developing an accurate prediction model. The study includes machine learning (ML) models, such as conditional inference tree, random forest (RF), gradient boosting, support vector machine, and a statistical model ordered probit for comparison. The study accounts for the imbalance in injury severity crash data by adopting data balancing techniques. Also, it recommends a combination of ML techniques, variable importance charts, and individual conditional expectation plots for identifying key variables and their effects. The finding suggests that RF trained in up-sampled data performs better than the remaining models. The presence of a central divider on the road reduces fatal injuries to pillion riders. The likelihood of getting severe injury is higher during nighttime crashes, TW-HMV (truck or bus) collisions, and hit-and-run crash cases where the colliding vehicle is unidentified. Older pillion riders are more vulnerable to sustaining fatal injuries in a crash. Crashes involving TWs hitting stationary objects and skidding are more fatal for pillion riders than other collision types.