Severity Prediction Scheme for Temporal Dependency Learning in Traffic Accidents

Abstract

This study presents a novel hybrid deep learning framework, Distil-BERT-CNN, designed to optimize spatial pattern extraction and temporal dependency learning for predicting traffic accident severity. However, this hybridization is achieved by integrating the lightweight, attention-based Distil- BERT model for efficient sequence processing with Convolutional Neural Networks (CNN) for robust spatial feature extraction, the proposed approach addresses the limitations of traditional statistical and machine learning models, which often struggle with class imbalance and complex, nonlinear relationships in real-world traffic data. Extensive experiments on multi-source datasets, including historical accident records, real-time traffic flow, and weather data, demonstrate that Distil-BERT-CNN significantly improves precision and recall, particularly for minority accident severity classes, compared to existing methods such as Random Forest, Gradient Boosting, LSTM, and cascade deep learning models. The enhanced predictive performance of this framework has the potential to support traffic management authorities in proactive accident prevention, resource optimization, and rapid emergency response, ultimately contributing to safer and more efficient transportation systems.