Machine Learning-Based Road Accident Fatality Prediction Under Imbalanced Data Conditions: An Evaluation of Resampling and Classification Techniques

Road traffic accidents remain a leading cause of mortality worldwide and a major public safety concern, particularly in developing regions where data imbalance affects predictive modelling performance. This study proposes a machine learning framework for fatal accident prediction under imbalanced class conditions using multiple resampling strategies and classification algorithms. Multiple classification algorithms, including Logistic Regression, Random Forest, XGBoost, and LightGBM, were evaluated with both data-level and algorithm-level imbalance-handling techniques that include Synthetic Minority Oversampling Technique (SMOTE), ADASYN, Random Oversampling and Random Undersampling. Performance was measured using imbalance-aware metrics such as accuracy, precision, recall, F1-score, and Area Under the Receiver Operating Characteristic Curve (ROC-AUC). In addition, statistical significance was tested using the Friedman and Wilcoxon signed-rank tests. To enhance interpretability, SHAP analysis was used to explain model decisions at both global and local levels. Experimental results indicate that ensemble boosting models significantly outperformed conventional methods, with LightGBM achieving the best overall performance with 99.35% accuracy, 99.25% F1-score, and ROC-AUC of 0.999736 without resampling. XGBoost also demonstrated strong robustness to class imbalance. Although resampling techniques slightly improved minority class learning, the improvements were not statistically significant. SHAP analysis reveals that vehicle make, crash location and engine type were key determinants of fatal outcomes. The findings demonstrate that advanced boosting algorithms possess strong predictive capability for imbalanced crash severity prediction and provide insight into developing robust predictive systems for traffic safety management for policymakers and intelligent transportation systems. It also demonstrates that integrating imbalance-aware learning with explainable AI enhances both predictive performance and interpretability in road safety analytics.