A Computational Intelligence GNN–LSTM Framework for Spatiotemporal Prediction of Traffic Accident Severity in Smart Cities Using SHAP XAI

Traffic accidents pose a critical public safety concern worldwide, leading to substantial fatalities, injuries, and economic losses, particularly in developing countries. The complex interplay of spatial, temporal, environmental, and driver-related factors makes the accurate prediction of accident severity a challenging task. This study introduces a novel computational intelligence hybrid deep learning framework that integrates Graph Neural Networks (GNN), Long Short-Term Memory (LSTM) networks, and Multi-layer Perceptrons (MLP) to model the spatiotemporal dynamics of traffic accidents. Using a comprehensive U.S. road accident dataset enriched with geospatial, temporal, environmental, and vehicle features, the proposed architecture captures both the spatial dependencies within the road network and the temporal evolution of risk patterns. Comparative experiments demonstrate that traditional machine learning models perform well when optimized with selected features; their predictive capabilities remain limited. In contrast, the proposed GNN–LSTM hybrid model achieves outstanding predictive results, with 99.97% accuracy and 99.99% precision, recall, and F1-score significantly surpassing existing state-of-the-art methods. Additionally, SHAP-based Explainable AI (XAI) techniques are integrated to provide interpretable insights into model predictions, enhancing transparency and supporting practical deployment. These results highlight the effectiveness of incorporating spatial structure and temporal trends in accident severity prediction. The proposed framework generalizability is further tested using K-fold cross validation techniques and further testing on an independent dataset. The proposed model exemplifies the potential of computational intelligence to address complex, real-world challenges in the advancement of AI-powered smart cities.