AI-Enabled Ensemble Learning Framework for Predicting Road Accident Severity

Predicting Road accident severity remains a critical public safety challenge, particularly in densely populated cities like Dhaka, Bangladesh, where diverse risk factors, such as road infrastructure, weather, and traffic conditions, amplify crash consequences. This study proposes a scalable ensemble machine learning framework that leverages 49,566 multi-source accident records collected from 2007 to 2024. Data sources include the Accident Research Institute (ARI), Bangladesh Road Transport Authority (BRTA), and Dhaka Metropolitan Police (DMP). The preprocessing pipeline incorporates structured encoding, statistical imputation, and class balancing using the Synthetic Minority Over-sampling Technique (SMOTE), along with detailed hyperparameter tuning for reproducibility. Baseline models logistic regression, decision trees, and support vector machines (SVMs), are benchmarked to contextualize performance improvements. The final ensemble model integrates Random Forest, Gradient Boosting, and XGBoost using soft voting and stacking. Our stacked ensemble achieved an accuracy of 82.55%, significantly outperforming individual learners and classical baselines. Explainability was addressed using SHAP to identify influential features. The proposed framework demonstrates strong generalizability, achieving consistent results across both internal and external datasets, and is suitable for real-time deployment in Intelligent Traffic Management Systems (ITMS).