Dynamic medical knowledge graph updating method based on LLM decision control
TL;DR
A Dynamic Medical Knowledge Graph updating framework based on Large Language Model (LLM) decision control is proposed, modeling the knowledge updating process as a sequential decision problem driven by multi-source feedback and stability constraints, thereby achieving a safe and controllable updating mechanism at the mechanism level.
Abstract
Medical knowledge exhibits significant timeliness and evidence-based dependency characteristics, and traditional static Medical Knowledge Graphs are difficult to adapt to the continuous evolution of clinical knowledge. Existing updating methods mostly rely on offline reconstruction or generative models, which suffer from update lag and potential medical safety risks, especially in primary healthcare scenarios with limited data and lack of continuous expert validation. To address this issue, this paper proposes a Dynamic Medical Knowledge Graph updating framework based on Large Language Model (LLM) decision control, modeling the knowledge updating process as a sequential decision problem driven by multi-source feedback and stability constraints. Different from existing methods, this paper innovatively transforms the LLM from a “knowledge generator” into an “update strategy controller,” which, at each time step, performs operations such as Add, Revise, Decay, or Reject on candidate knowledge based on task effectiveness, knowledge consistency, and temporal evolution feedback, thereby achieving a safe and controllable updating mechanism at the mechanism level. In the respiratory disease scenario, small-sample temporal evolution experiments (2,000 records) constructed based on cMedQA v2.0 data and real outpatient cases demonstrate that the proposed method improves Micro-F1 to 78.95% in medical question answering tasks, reduces the error knowledge introduction rate to 3.68%, and achieves a graph utility value of 0.8174, reaching a good balance between performance improvement and structural stability. The results indicate that, within the tested respiratory disease scenario using DeepSeek-R1 as the decision controller, the performance improvement of Dynamic Knowledge Graphs does not depend on knowledge generation capability, but rather on the decision control capability of the updating process. This method provides a new technical pathway for constructing safe and sustainable clinical knowledge systems, especially suitable for resource-constrained primary healthcare environments.
