AI 中文总结
研究城市医疗保健 OD 预测问题，提出基于知识驱动行为推理的无训练框架，将异构城市信息组织成知识图谱，通过多智能体推理实现预测。实验表明该框架性能优于传统深度学习基线，凸显城市智能在医疗保健出行建模上从数据拟合到知识驱动推理的潜力。
AI 中文摘要
理解城市医疗保健出行对于医疗资源分配、医院容量规划和城市弹性治理至关重要。现有的医疗保健起终点（OD）预测方法主要依赖监督深度学习，需要昂贵的模型训练，且对医院选择行为机制的可解释性有限。本文提出了一种基于知识驱动行为推理的无训练城市医疗保健 OD 预测框架，用基于结构化城市知识的协作式大语言模型推理取代基于梯度的神经网络学习。该框架将包括医疗资源、人口特征、交通可达性、天气状况和历史出行等异构城市信息组织成统一的知识图谱，通过多智能体推理管道实现证据检索、城市情境理解、行为推理、偏好排序和贝叶斯验证。在包含 137 家医院和 6341 个预测任务的深圳真实数据集上的实验表明，该框架实现了 CPC = 0.5449，Top-1 = 66.19%，Top-5 = 68.03%，Top-10 = 74.03%，在所有 Top-K 指标上始终优于传统深度学习基线。与没有结构化推理的相同主干直接大语言模型相比，该框架将 CPC 提高了 87.5%，Top-1 准确率提高了 141%。这些发现凸显了城市智能将医疗保健出行建模从数据拟合转向知识驱动行为推理的潜力。
英文摘要
Understanding urban healthcare mobility is essential for healthcare resource allocation, hospital capacity planning, and resilient urban governance. Existing healthcare origin-destination (OD) prediction methods primarily rely on supervised deep learning, requiring expensive model training while providing limited interpretability into the behavioral mechanisms of hospital choice. This paper presents a training-free urban healthcare OD prediction framework based on knowledge-grounded behavioral reasoning, replacing gradient-based neural network learning with collaborative LLM reasoning over structured urban knowledge. The framework organizes heterogeneous urban information, including healthcare resources, population characteristics, transportation accessibility, weather conditions, and historical mobility, into a unified knowledge graph, enabling evidence retrieval, urban context understanding, behavioral reasoning, preference ranking, and Bayesian verification through a multi-agent reasoning pipeline. Experiments on a real-world Shenzhen dataset containing 137 hospitals and 6,341 prediction tasks demonstrate that the proposed framework achieves CPC = 0.5449, Top-1 = 66.19%, Top-5 = 68.03%, and Top-10 = 74.03%, consistently outperforming conventional deep learning baselines on all Top-K metrics (e.g., Top-5 improves by 22.1% over MLP). Compared with Direct LLM using the same backbone without structured reasoning, the proposed framework improves CPC by 87.5% and Top-1 accuracy by 141%, demonstrating that knowledge-grounded behavioral reasoning, rather than raw LLM capability, is the key to accurate, interpretable, and training-free healthcare OD prediction. These findings highlight the potential of urban intelligence to move healthcare mobility modeling beyond data fitting toward knowledge-driven behavioral reasoning.
Comments16 pages, 9 figures. Submitted to the 4th International Conference on Urban Science and Intelligence
