基于提示引导多跳推理的医学诊断检索增强生成
打开文本图片集
关键词：动态检索增强生成；多跳思维链；医学提示学习；量化重写机制；医学诊断中图分类号：TP391 文献标志码：A 文章编号：1001-3695（2025）10-008-2956-08doi：10.19734/j. issn.1001-3695.2025.04.0089
Retrieval-augmented generation with prompt-guided multi-hop reasoning for medical diagnosis
Qin Le1， Gou Zhinan ，Wang Peiwu³， Zhang Gaofei³，Liu Siyu 1 ， Gao Kai³ （1.SchoolofgementSece&IfoatioEngnering，HebeiUniersitfcomcs&usiness，hjzanO;. DeptfCompeecesiUeitelfee& University of Science & Technology，Shijiazhuang O5oo18，China）
Abstract：Thecomplexityofmedicaldiagnosistasksisparticularlyprominentinthemanifestationsofsymptomsandtheirassociations withdiseases.Duetothecomplexitiesof“diferentsymptomsforthesamedisease”and“samesymptomsfordifferentdiseases”，medical diagnosis tasks impose higherrequirements onthereasoning capabilitiesof models.Traditional RAG technologies，withtheirstaticretrievalandsingle-stepreasoing，truggletocapturemulti-levellogicalrelationsips（uchas symptoms $$ departments $$ diseases $$ differential diagnosis）.To effectively overcome this limitation，this study proposed a novelframeworkforthemedicalfield：PGM-RAG.Byintegratingbasic knowledgeofthemedicalfield，thisframework provided clearreasoning guidancefor the modelthrough thedesignofprompt information foreach reasoning step.Meanwhile，this framework designed aquantitative rewriting mechanism around strictlycontrolthe accuracyof thecontent generatedbylarge language models，therebyenhancingthereliabilityofthereasoningprocessandtheprecisionof diagnosticresults.Experments on two public medical datasets，Huatuo-26MandWebMedQA，showthatthe proposed model outperformsthe existingbest methods by 12.6% and 8.9% in EM and F1 metrics，respectively. Ablation experiments demonstrate that the multi-hop reasoning chain and the quantitative rewriting mechanism significantly improve the model’s performance.
Key words：dynamicretrieval augmented generation；multi hopchain-of-tought；medical promptlearning；quantitativerewriting mechanism；medical diagnosis
0 引言
当前，越来越多的大语言模型（largelanguagemodels，LLMs）在医学领域涌现并取得一定进展，为疾病预防、治疗、个性化医疗等方面提供了强大支持。（剩余24375字）
