An Integrated Deep Learning Framework for Crash Injury Severity Prediction With KGAIN‐Based Imputation and WGAN–GP–ENN Balancing

Accurate prediction of crash injury severity is a cornerstone of intelligent transportation systems (ITS), facilitating optimized emergency response and enhanced road safety management. However, the inherent challenges of high‐dimensional missing data and extreme class imbalance in traffic accident records significantly impede the performance of conventional prediction models. To address these issues, this article proposes a novel integrated deep learning framework designed to robustly predict injury severity. Specifically, we develop a K ‐nearest neighbours enhanced generative adversarial imputation network (KGAIN), which synergizes local similarity patterns with global distribution learning to effectively restore complex dependencies in incomplete traffic datasets. Furthermore, to mitigate the bias induced by class imbalance—particularly the rarity of severe injury cases—we introduce a Wasserstein generative adversarial network with gradient penalty and edited nearest neighbours (WGAN–GP–ENN) method. This approach generates high‐quality synthetic minority samples while simultaneously filtering boundary noise, outperforming traditional over‐sampling techniques. Experimental validation using real‐world Chicago traffic collision data demonstrates that the proposed framework significantly enhances predictive performance. The results reveal that gradient boosting classifiers, when integrated with the proposed preprocessing pipeline, achieve superior accuracy (99.80%) and robustness. This study provides a viable technical solution for real‐time accident severity assessment, offering valuable support for decision‐making in future ITS deployments.