Enhancing Emergency Response in Road Accidents: A Severity Prediction Framework Using RF-RFE and Deep Learning Model

Abstract

Road accidents in urban areas pose critical challenges due to their complexity and potential for severe injuries, making rapid severity prediction essential for effective emergency response. Traditional approaches often depend on conventional models with limited feature selection, leading to suboptimal performance and low interpretability. This framework leverages the French Road Accident dataset from 2021–2023, containing multiple features related to accident circumstances, vehicle characteristics, and environmental conditions. Preprocessing involved merging files, removing duplicates and low-variance columns, label encoding, eliminating null values, and selecting the top 20 features using RF-RFE. The dataset was balanced using SMOTE-Tomek and scaled via Standard Scaler. Machine learning models— AdaBoost, XGBoost, LightGBM—and deep learning architectures—GRU, CNN, BiLSTM, CNN-LSTM, CNNBiLSTM, and CNN-BiLSTM with Attention—were implemented. Model evaluation using accuracy, precision, recall, and F1-score showed hybrid deep learning models achieved near-perfect performance, with CNN-BiLSTMAttention attaining the highest accuracy. A Flask-based user interface enables real-time severity prediction, while interpretable AI techniques, including LIME, SHAP, and Alibi, provide insights into feature importance, improving transparency. This approach demonstrates significant enhancement in emergency severity prediction, supporting timely and informed intervention.