Traffic Accident Injury Severity Forecasting with a Graph Neural Network Approach

Accurate prediction of injury severity in road crashes is essential for proactive traffic safety management and effective emergency response. Traditional predictive models often fail to capture the intricate spatial-temporal dependencies and contextual factors that influence crash outcomes. This study introduces a novel approach using Graph Neural Networks (GNNs) to model and predict injury severity in road traffic incidents. By treating crash records as interconnected nodes within a graph structure—linked by temporal, spatial, and environmental similarities—the GNN captures contextual relationships that linear models often ignore. The proposed framework integrates multiple features, including crash location, weather, road type, and vehicle characteristics, and applies graph convolution to learn meaningful embeddings that inform severity predictions. Experimental evaluations on real-world traffic crash datasets demonstrate that the GNN-based model outperforms traditional machine learning approaches in classification accuracy and generalization, particularly in complex urban environments. This system offers a powerful tool for transport authorities aiming to reduce the public health impact of road accidents.

Keywords : Graph Neural Network Framework, road crashes