LLM-enhanced causal graph learning for real-time crash risk prediction

Accurate and reliable risk prediction is crucial for preventing traffic crashes and improving road safety. However, existing approaches often struggle to balance predictive performance with causal interpretability. To bridge this gap, we propose a closed-loop framework that integrates causal discovery, semantic enhancement, and spatio-temporal prediction. First, transfer entropy extracts event-specific causal relationships from dangerous driving scenarios to construct initial causal graphs. These graphs are then refined using a GPT-2-based language model fine-tuned with Low-Rank Adaptation (LoRA), which employs a graph enhancement strategy to reinforce salient causal pathways. Finally, the semantically enhanced graphs are fused with temporal driving features using a spatio-temporal model that combines Graph Attention Networks (GAT) and Long Short-Term Memory (LSTM) networks for real-time risk prediction. Evaluated on car-following scenarios from the highD dataset, our framework achieves an accuracy of 0.956, F1-score of 0.872, and AUC of 0.985, significantly outperforming traditional baselines. Ablation studies further confirm that the GPT-2-based causal enhancement with LoRA fine-tuning is a significant contributor to the model's predictive accuracy. These findings indicate that our framework offers a powerful solution for enhancing both precision and interpretability in collision risk prediction, holding strong potential for deployment in real-time traffic management systems.