DKD-MARL: a data-knowledge dual-driven multi-agent reinforcement learning framework for traffic crash severity...

Accurate prediction of traffic crash severity is critical for post-crash emergency response and proactive safety interventions. However, existing methods either rely on structured data-driven statistical learning and overlook domain semantic knowledge, or use single large language models (LLMs) suffering from unstable prediction. To address these limitations, this paper proposes DKD-MARL, a data-knowledge dual-driven multi-agent reinforcement learning framework for traffic crash severity prediction. The framework first constructs a dual-channel representation module that generates structured feature vectors and knowledge-grounded textual descriptions. A multi-agent inference system is then built, comprising one global statistical agent and four domain-specific LLM agents focusing on human, vehicle, environment and time factors. To adaptively integrate the predictions from these agents, a deep Q-network (DQN)-based fusion module is developed, which formulates the fusion task as a state-dependent decision problem. A reward shaping mechanism incorporating class imbalance, confidence support, and ordinal misclassification costs enables the DQN to learn sample-specific fusion policies. Extensive experiments on the Victoria Road Crash dataset demonstrate that DKD-MARL achieves superior performance across all evaluation metrics, with an accuracy of 0.763 and a macro F1-score of 0.703, outperforming machine learning baselines, zero-shot LLMs, prompting strategies, and static multi-agent fusion methods. Ablation studies confirm the complementary contributions of each agent and reward component. Few-shot and extreme-imbalance experiments further validate the framework's robustness under limited data and long-tailed distributions. Interpretability analyses reveal how the DQN dynamically adjusts agent contributions according to crash scenarios. This work offers a promising solution for reliable and accurate crash severity prediction.