This study investigates pedestrian–vehicle crashes using explainable machine learning (ML) models and SHapley Additive exPlanations (SHAP) tool to understand the role of pedestrian actions in fatal/severe injury. The dataset consists of 4,343 at-fault pedestrian crashes (2017–2021) in Louisiana state, covering four types of pedestrian actions including crossing at intersection (1,592 crashes), crossing at midblock (1,612 crashes), walking with traffic (820 crashes), and walking against traffic (319 crashes). ML-based models, including gradient boosting machine (GBM), extreme gradient boosting (XGBoost), categorical boosting (CatBoost), adaptive boosting (AdaBoost), and light gradient boosting (LightBoost), were employed to identify top factors. Based on predictive accuracy measures -accuracy, precision, recall, F1, G-Mean, Area Under Receiver Operating Characteristic (AUROC) curve. GBM performed better for “crossing at intersection” and “walking with traffic,” while CatBoost performed better for “crossing at midblock,” and “walking against traffic.” Key findings revealed that dark conditions (with or without streetlight), high-speed settings (40–55 mph), and pedestrian characteristics (>64 years, Caucasian race) were associated with an increased likelihood of fatal/severe injury while crossing at the intersection. For midblock crossing-related crashes, top factors were driver alcohol impairment, highway road, dark-no-streetlight, and posted speed limit of 50–55 mph. Factors such as open country, pedestrian/driver alcohol impairment, 50–55 mph speed limit, two-way road with physical separation, and older pedestrians were found to be associated with increased likelihood of fatal/severe injury while walking with traffic. For walking against traffic, top contributing factors were male drivers, young drivers, presence of passengers, weekend, business/industrial areas, dark-no-streetlight, and pedestrians aged 25–40 years.
Hossain et al. (Sat,) studied this question.