Predicting injury severity in truck-involved crashes using machine learning classifiers and cross-validation techniques

This study utilized eight supervised Machine Learning classifiers, Logistic Regression, K-Nearest Neighbors, Gaussian Naïve Bayes, Linear Discriminant Analysis, Support Vector Classifier, and tree-based (Decision Tree, Random Forest, and XGBoost) through hyperparameter tuning to predict injury severity in truck-involved crashes in a developing country. Multi vehicle collisions on straight road segments dominate, with human and environmental factors influencing severity. Feature selection using both filter and embedded methods revealed weak associations of weather and temporal characteristics, emphasizing the need for location specific modelling. Significant features included road lane, collision type, road geometry, crossing lane, and cause of crash. Cross-Validation data splitting techniques proved to improve performance over the standard Single Random Train-Test Split. Hence, two models are highlighted. The optimized XGBoost model succesfully achieved 87% accuracy, despite not using data resampling to address class imbalance, outperforms Random Forest, also in terms of RMSE (0.36) and precision for class 0 (0.87), demonstrating its effectiveness in minimizing false positives. However, Random Forest shows a slightly higher AUC (0.87) compared to XGBoost (0.79), indicating a better overall ability to distinguish between classes. Despite this, XGBoost remains the preferred model for its higher performance in critical metrics, making it more suitable for applications where false alarm prevention is crucial. This study highlights the potential of ML-based prediction tools to enhance toll road safety and guide policy interventions in emerging economies, reducing fatalities while avoiding ineffective safety measures.