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

Abstract

Road accidents, particularly in urban areas, pose significant challenges due to their complexity and severe consequences. Rapid emergency response is crucial to mitigating their impact, especially for cases requiring urgent medical intervention. This study aims to enhance emergency response strategies by developing a model that prioritizes severe accidents, ensuring faster ambulance dispatch and timely medical assistance. We utilize multi-source accident data from France, incorporating factors related to road conditions, user demographics, vehicle characteristics, and environmental conditions. The dataset is refined using Random Forest Recursive Feature Elimination (RF-RFE) to select key features influencing accident severity. To address class imbalance, we apply the Synthetic Minority Over-sampling Technique with Tomek Links (SMOTE-Tomek) to oversample minority cases and reduce noise. Based on this preprocessed data, we propose a hybrid Convolutional Neural Network with Bidirectional Long Short-Term Memory and Attention Mechanism (CNN-BiLSTM-Attention) model. The convolutional layers capture spatial patterns, while BiLSTM layers learn sequential dependencies. The attention mechanism further refines predictions by emphasizing critical features. This deep learning model significantly outperforms traditional machine learning methods, achieving accuracy score of 94.99%. Finally, a Shapley Additive Explanations (SHAP) analysis ensures interpretability by identifying key factors influencing accident severity. These insights provide actionable guidance for optimizing emergency response strategies and improving road safety measures.