A hybrid machine learning approach for predicting traffic accident collision severity.

Abstract

Accurate prediction of traffic accident severity remains challenging due to feature coupling and class imbalance, which hinder reliable applications in autonomous driving safety systems. This study proposes a Dynamic and Static Cross Entropy Integrated Neural Network (DSCE-INN) to address these issues. Using 857 real-world accident cases from the 2017-2021 China National Automobile Accident In-Depth Investigation System (NAIS), a Weighted Injury Coefficient is developed to enable continuous injury mapping, and K-means clustering reclassifies severity into three levels: property damage only, non-disabling injury, and disabling or fatal injury. Information gain identifies 11 critical features. DSCE-INN employs feature decoupling, transforming the multi-class task into binary sub-models, and introduces a dynamic-static weighted cross-entropy loss to jointly mitigate coupling and imbalance. A soft-hard voting mechanism, combined with L1 regularisation and focal loss, further enhances prediction robustness. Experimental results show accuracies of 0.782, 0.729, and 0.801, significantly outperforming a baseline ANN. Findings demonstrate DSCE-INN's effectiveness and practical value for autonomous driving safety.