Predicting injury severity in vehicle-to-pedestrian collisions: evidence from Madhya Pradesh state, India

The present research examines characteristics of vehicle-to-pedestrian collisions and assesses the factors affecting such crashes in Madhya Pradesh using accident record data from Madhya Pradesh Road Development Corporation and Accident Response System, encompassing 6104 accident entries across 16 variables. To address data heterogeneity, Latent Class Clustering was employed, and a Random Forest Algorithm was used for predictive modelling to accurately classify injury severity based on various predictors. The latent clustering identified three distinct classes: Class 1 (27.7%) associated with fewer severe injuries, attributed to effective traffic controls; Class 2 (8.7%), characterized by reduced injury rates, linked to well-organized traffic systems; and Class 3 (63.5%), which showed a higher incidence of severe injuries, primarily due to narrow, poorly designed, or infrastructure-deficient roads and excessive speeding. The Random Forest Classifier effectively used predictors such as age, gender, violation type, traffic control, and vehicle type, achieving 97.2% precision and 89.2% recall, and an overall accuracy of 88.8%. The Decision Tree model further highlighted the relative importance of each variable in predicting injury severity. The model’s performance was benchmarked against baseline classifiers including logistic regression, Support Vector Machine, and Decision Tree, confirming the superior accuracy and robustness of the Random Forest approach. This study emphasizes the need for proactive road safety measures and provides a robust analytical framework for injury severity classification, supporting the development of targeted interventions to improve pedestrian safety in India.