首页|期刊导航|数字中医药(英文)|TCM-HyperRAG:面向中医疾病诊断和处方推荐的高阶知识超图

TCM-HyperRAG:面向中医疾病诊断和处方推荐的高阶知识超图OA

TCM-HyperRAG:higher-order knowledge hypergraphs for disease diagnosis and prescription recommendation in traditional Chinese medicine

中文摘要英文摘要

目的 针对知识图谱难以建模中医药知识中的高阶关联,且无法区分不同实体重要性的问题,本研究提出一种加权知识超图增强框架,用于中医智能辅助诊断和处方推荐.方法 本研究构建了基于加权知识超图的方法 TCM-HyperRAG,并基于公开的中医药监督微调数据集(TCM-SFT)进行验证.DeepSeek-V3 逐条从 18 901条疾病诊断记录和 9 290条处方推荐记录中抽取疾病或证候、症状和中药实体,构建加权知识超图,并通过相应的检索策略为骨干大型语言模型(LLM)提供结构化且可追溯的证据.在疾病诊断和处方推荐任务上分别采用 500条未见病例进行测试,以 HuatuoGPT-o1、Llama3.1-8B、Qwen3-8B、GLM4-9B、DeepSeek-V3和 Qwen3-Max为骨干 LLM,对单纯 LLM推理、标准 RAG、4 种 GraphRAG模式和TCM-HyperRAG进行比较.疾病诊断采用平均倒数排名(MRR)和前 k项命中率(H@k k∈{1,3,5},)评价,处方推荐采用前 k项准确率(P@k)、前 k项召回率(R@k)和前 k项F1 分数(F1@k)(k∈{5,10,20})评价.通过重复实验评价模型表现的稳定性,并通过消融实验验证重要性权重的贡献.结果 TCM-Hyper-RAG在两个任务上均取得了具有竞争力的结果.在 疾 病 诊 断 任 务 中,TCM-HyperRAG在 Llama3.1-8B、GLM4-9B、DeepSeek-V3 和 Qwen3-Max取得最高 MRR,平均 MRR为 0.638,高于纯 LLM推理的 0.571和GraphRAG变体的 0.486.在处方推荐任务中,TCM-HyperRAG在全部 6 种大模型上均取得最高 R@20,平均值为 0.544,高于 GraphRAG+Mix的 0.496.在基于 Qwen3-Max的重复实验中,TCM-HyperRAG取得最高的平均 MRR、H@3 和 H@5,分别为 0.769 1、0.862 5和 0.911 8,表明该方法具有稳定的表现.移除实体重要性权重后,疾病诊断平均 MRR 由 0.638降至 0.629,处方推荐平均 F1@10 由 0.438降至 0.425,表明权重感知对整体性能具有积极贡献.结论 TCM-HyperRAG 通过保留多实体关系并优先检索临床重要性较高的证据,在疾病诊断和处方推荐任务中取得了具有竞争力的结果,同时可为基于 LLM的中医智能决策支持提供可追溯的证据.

Objective To address the limitations of knowledge graphs in modeling higher-order associa-tions in traditional Chinese medicine(TCM)knowledge and distinguishing the importance of different entities,we propose a weighted knowledge hypergraph-enhanced framework for in-telligent assistance in TCM diagnosis and prescription recommendation. Methods We developed TCM-HyperRAG,a weighted knowledge hypergraph-based method,and validated it using a publicly available TCM Supervised Fine-Tuning(TCM-SFT)dataset.DeepSeek-V3 extracted disease/syndrome,symptom,and herb entities from 18 901 disease-diagnosis records and 9 290 prescription-recommendation records to construct the weighted hypergraph.The corresponding retrieval strategy then provided the backbone large language models(LLMs)with structured and traceable evidence.The framework was tested on two held-out sets of 500 previously unseen records for disease diagnosis and prescription recom-mendation,respectively.LLM-only inference,standard RAG,four GraphRAG modes,and TCM-HyperRAG were compared using HuatuoGPT-o1,Llama3.1-8B,Qwen3-8B,GLM4-9B,DeepSeek-V3,and Qwen3-Max as backbone LLMs.Disease diagnosis was evaluated using mean reciprocal rank(MRR)and hit rate at k(H@k,k ∈{1,3,5}),whereas prescription rec-ommendation was evaluated using precision at k(P@k),recall at k(R@k),and F1 score at k(F1@k),where k ∈{5,10,20}.Repeated experiments assessed performance stability,and ab-lation experiments examined the contribution of the importance weights. Results TCM-HyperRAG achieved competitive results on both tasks.For disease diagnosis,it obtained the highest MRR with Llama3.1-8B,GLM4-9B,DeepSeek-V3,and Qwen3-Max,and an average MRR of 0.638,higher than 0.571 for LLM-only inference and 0.486 for the GraphRAG variants.For prescription recommendation,TCM-HyperRAG achieved the high-est R@20 with all six backbone LLMs,averaging 0.544 versus 0.496 for GraphRAG+Mix.In the repeated experiments with Qwen3-Max,TCM-HyperRAG obtained the highest mean MRR,H@3,and H@5,reaching 0.769 1,0.862 5,and 0.911 8,respectively,indicating stable perfor-mance.Removing entity importance weights reduced the average disease-diagnosis MRR from 0.638 to 0.629 and the average prescription-recommendation F1@10 from 0.438 to 0.425,indicating a positive overall contribution from importance-aware weighting. Conclusion By preserving multi-entity relations and prioritizing evidence with greater clini-cal importance,TCM-HyperRAG achieves competitive performance in disease diagnosis and prescription recommendation while providing traceable evidence for LLM-based intelligent TCM decision support.

陈伟林;叶绍强;周辉;康棣文;秦莉花;周恺卿

吉首大学计算机科学与工程学院,湖南 吉首 416000,中国Faculty of Computing,Universiti Teknologi Malaysia,Johor Bahru 81310,Malaysia湖南中医药大学信息科学与工程学院,湖南 长沙 410208,中国吉首大学通信与电子工程学院,湖南 吉首 416000,中国||Department of Biomedical Engineering,Faculty of Engineering,Universiti Malaya,Kuala Lumpur 50603,Malaysia湖南中医药大学护理学院,湖南 长沙 410208,中国吉首大学计算机科学与工程学院,湖南 吉首 416000,中国

中医药智能诊断处方推荐检索增强生成超图

Traditional Chinese medicineIntelligent diagnosisPrescription recommendationRetrieval-augmented generationHypergraphs

《数字中医药(英文)》 2026 (3)

362-374,13

National Natural Science Foundation of China(82374437),Natural Science Foundation of Hunan Province(2024JJ7395),and Hunan Province Traditional Chinese Medicine Scientific Research Project(A2024001).

10.1016/j.dcmed.2026.08.004

评论