ABSTRACT
All over the world, road traffic accidents are one of the biggest problems leading to fatalities, serious injuries and socio-economic consequences. Accurate prediction of severity is crucial in improving the road safety, response and rescue systems, and the consequent fatality rates. In response to this, we propose a robust framework that employs state-of-the-art feature engineering and ML strategies for enhanced accuracy in accident severity prediction. To tackle the imbalanced class, Multi-Distance Synthetic Technique (MDST) is applied to synthesize of data samples, so that underrepresented classes are represented. Furthermore, Implemented and compared three machine learning approach (XGBoost, LightGBM, ensemble model) on balanced dataset. The model demonstrated superior predictive performance (overall accuracy 95%) and improved precision, recall, and F1-scores across all severity levels when compared to other algorithmsThis study shows how data balancing, feature optimization, and ensemble learning can work together to achieve excellence for a real-life problem of accident severity prediction. Overall, we contend that the solution is scalable and objectively adaptable, supplemented with analytics driven insights for the policymakers, transport authorities and emergency services, to develop a data driven approach for enhanced road safety.
