Pickup truck crash severity analysis via machine learning: policy insights for developing countries

This study pursues two complementary objectives: first, evaluating machine learning approaches for crash severity prediction to address methodological gaps in pickup truck crash analysis; second, systematically comparing single- versus multi-vehicle crash outcomes to understand distinct risk factors. Using Thailand crash data, the research compares Logistic Regression, Random Forest, XGBoost, and Deep Neural Network models, optimized with K-fold cross-validation and Bayesian Optimization, with SHAP employed for model interpretability. Results demonstrate that model performance varies significantly with injury classification schemes: XGBoost performed best for multiclass injury classification in both crash types, while Random Forest and Deep Neural Networks excelled in binary classification for single- and multi-vehicle crashes, respectively. The methodological analysis reveals the importance of both model selection and classification scheme in achieving optimal predictive performance. When applied to analyze crash factors, the models identified that both crash types are influenced by 4-lane roads, unlit roads, and barriers. Severity in single-vehicle crashes increases with fatigue, 2-lane roads, intra-province highways, and long holidays; in multi-vehicle crashes, severity is influenced by involvement of motorcycles or trucks, head-on collisions, and specific times of day. Factors reducing severity in single-vehicle crashes-such as concrete roads, defective vehicles, and hitting guardrails-do not significantly affect multi-vehicle crashes.