Abstract: The real electronic medical record data covering diagnosis related groups (DRG) coding are too scarce to support language models in learning text features, and the existing disease coding models are difficult to interpret the results for complex text. Therefore, this paper designs a medical question answering system model GLM-2B-DRAGON (generative language model-deep bidirectional language-knowledge graph pretraining) that integrates medical knowledge graph and large language model. Firstly, ChatGLM-6B model is employed to extract and update medical entities and entity relationships, and a knowledge graph DRG-Net covering medical knowledge such as DRG coding is obtained. Secondly, the cross-modal encoder is used to jointly encode the QA pairs and the knowledge graph to realize the complementary text-graph bidirectional information flow to capture the characteristics of medical text. Finally, the interpretability of the answer results is verified through the visual analysis of the path weights of the knowledge graph. The experimental results show that the proposed system model is superior to the existing knowledge graph enhanced language models on the public dataset CommenSenseQA and the self-built medical dataset MedicalQA.

Key words: disease diagnosis related grouping, medical Q&, A, medical knowledge graph, large language model