Ensemble-based model to investigate factors influencing road crash fatality for imbalanced data

• A robust framework is developed to analyze imbalanced traffic crash-related datasets. • Three analysis approaches (no bagging, bagging, optimized bagging) under eight machine learning models and eight resampling techniques are used. • Optimized bagging has been compared among three optimization techniques, namely, Grid Search CV, Random Search CV, and Bayesian optimization. • The best performance was shown when data was resampled using Random Under-sampling and trained by Gradient Boosting Machine (GBM) under Bayesian optimized bagging (based on a G-mean score of 65.23 and an F1 score of 60.06). • The feature importance reveals that collision type, vehicle type, road class, vehicle movement, and traffic control are the top five factors influencing crash fatality in Bangladesh. The rapid growth of urbanization and motorization has significantly increased traffic crashes, leading to both loss of life and diminished quality of life for crash survivors and their families. Identifying the factors influencing crash fatality is crucial for reducing such incidents. However, traffic crashes are inherently unpredictable, and crash fatality datasets are often imbalanced. This study provides a comprehensive evaluation of various machine learning (ML) techniques to analyze traffic crash fatality using an imbalanced dataset. It is the first to train eight distinct binary classification models: Classification and Regression Trees (CART), Random Forest (RF), Gradient Boosting Machine (GBM), Extreme Gradient Boost (XGBoost), Logistic Regression (LR), Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Naïve Bayes (NB) under three strategies: in isolation, with bagging, and with optimized bagging techniques (Grid Search CV, Random Search CV, and Bayesian Optimization). To handle data imbalance, eight resampling methods were employed, including SMOTE, Random Under-sampling (RUS), Random Over-sampling (ROS), ADASYN, Tomek Links, Near Miss, SMOTETomek, and SMOTEENN. Results show that GBM, combined with Bayesian optimized bagging and RUS, achieved the best performance with a G-mean score of 65.23 and an F1 score of 60.06. This study offers valuable insights into effective ML techniques, data resampling methods, and advanced optimization strategies for imbalanced crash severity datasets.