Spatial–Temporal Prediction of Traffic Accident Risks Using a Fusion of GNN-LSTM

Accurately predicting the spatiotemporal distribution of traffic accident risk is crucial for proactive safety management and intelligent transportation systems. However, the inherent complex spatial dependencies and dynamic temporal patterns in traffic data, coupled with severe class imbalance in accident severity categories, pose significant challenges to conventional prediction models. This study proposes a novel deep learning framework that integrates Graph Neural Networks (GNN) and Long Short-Term Memory (LSTM) networks to address these challenges. The GNN module effectively captures complex spatial dependencies between geographical regions by modeling the traffic network as a graph structure, while the LSTM module learns the temporal dynamics of accident occurrences. To tackle the critical issue of data imbalance caused by scarce severe accident samples, this study employs the Adaptive Synthetic Sampling (ADASYN) technique at the data level to balance sample distributions, and at the model level, utilizes the Focal Loss function to focus the training process on hard-to-classify minority class instances. Furthermore, the model architecture incorporates TransformerConv layers, bidirectional LSTM, and multi-head attention mechanisms to capture more nuanced spatiotemporal features. Experimental evaluation on the large-scale US-Accidents dataset (2016–2021) demonstrates that the proposed GNN-LSTM fusion model significantly outperforms existing methods across multiple key metrics. On the test set, the model achieved a weighted F1-score of 0.7923 ± 0.0123 and a macro-average F1-score of 0.6587 ± 0.0156, representing improvements of 1.12 and 3.53 percentage points, respectively, over the strongest baseline model Graph WaveNet. In addressing class imbalance, the model attained an AUC-PR of 0.7124 ± 0.0098 and a Cohen’s Kappa of 0.6012 ± 0.0134, both significantly superior to conventional approaches. Ablation studies further validated the effectiveness of key components, with the GNN module contributing 13.8% to the performance gain, while ADASYN and Focal Loss contributed 8.7 and 6.5%, respectively. This study provides a robust and efficient solution for traffic risk prediction, offering a scientific foundation for traffic management authorities to develop precise early-warning systems and targeted intervention strategies.