Comparative Analysis of Machine Learning Based Algorithms for Predicting Injury Severity in Road Accidents

Road crash injury severity prediction is essential for intelligent transportation systems, yet challenged by severe class imbalance, rigid 4-class severity schemes (unhurt/slight/hospitalized/fatal), and optimal methodological selection. This study proposes a structured framework systematically evaluating four machine learning models—CatBoost, HistGradientBoosting, Random Forest, and SVM—across multiclass (native 4-class and ordinal wrapper), binary reduction (non-severe vs. severe), and oversampling techniques using crash data. Multiclass approaches reveal tree ensemble dominance but persistent rare severe class prediction difficulties. Binary class reduction substantially improves severe injury detection performance on this dataset by simplifying decision boundaries, while SMOTE oversampling provides algorithm-specific imbalance mitigation. Random Forest demonstrates the most stable binary performance across evaluation metrics, independent of oversampling strategies. This performance gain comes at the cost of reduced severity granularity compared to the original multiclass formulation. Overall, under imbalance-sensitive evaluation metrics, binary class reduction provides a pragmatic and operationally effective alternative to complex multiclass strategies for severe injury detection.