A Hybrid K‐Means Clustering and Ensemble Model for Predicting Road Traffic Accident Severity

Road traffic accidents remain a critical global challenge, causing millions of deaths and economic losses. Understanding critical factors and accurately predicting accident severity are essential to mitigating these impacts. However, heterogeneity, severe class imbalance, and high dimensionality of accident data limit both traditional and stand‐alone machine learning models. This study proposes a novel hybrid of K‐Means clustering and ensemble classifiers to enhance severity prediction and reveal significant factors. K‐Means clustering extracts latent spatiotemporal patterns and creates a new feature in the training set, while ensemble classifier models: XGBoost, Random Forest, LightGBM, and Decision Tree, are applied to perform the classification task. We used the optimal K equal to 5 clusters, determined through the refined elbow method with maximum perpendicular distance to the line connecting the first and last points. Class imbalance is addressed effectively using the SMOTE‐Tomek and variational autoencoder resampling techniques. The proposed hybrid KXGB, tested on 13,064 real‐world accident records in Addis Ababa, achieved an accuracy of 98.61%, a precision of 98.75%, a recall of 98.74%, an F1‐score of 98.68%, and an AUC of 99.9%. It clearly outperformed standalone models and recent hybrid approaches. The study identified time, junction, location, day of week, collision type, driver age, and vehicle type as the most influential factors of accident severity using SHAP explainable analysis. The proposed framework offers superior predictive accuracy combined with actionable insights for targeted road safety interventions.