Research on the Construction of an lntelligent Question-Answering System for Alzheimer's Disease Based on Large Language Models and Knowledge Bases
- VernacularTitle:基于大语言模型和知识库的阿尔茨海默病智能问答系统构建研究
- Author:
Wenhu WANG
1
;
Changfa WEI
1
Author Information
- Publication Type:Journal Article
- Keywords: Large language model; Retrieval-augmented generation; Local knowledge base; Alzheimer's disease; Question-answering system
- From: World Science and Technology-Modernization of Traditional Chinese Medicine 2025;27(3):856-866
- CountryChina
- Language:Chinese
- Abstract: Objective Utilizing the Langchain framework in combination with large models and integrating a knowledge base to create a question-answering system,this paper conducts a technical exploration of the application of large models in the field of medical knowledge question and answer.Methods Authenticated knowledge documents,including diagnostic and treatment guidelines for AD endorsed by authoritative organizations such as the Chinese Association of Integrative Medicine and the Chinese Medical Association,along with medical textbooks,were introduced to construct a local knowledge base for AD.This knowledge base enhanced the model's capability in AD-related question-answering.Ultimately,the application of the ChatGLM3-6B model in AD medical question-answering was achieved.Results Using Fact accuracy(FA)and Completeness of response(CR)as evaluation metrics,comparative experiments were conducted between the AD question-answering system and the ChatGLM3-6B and ChatGPT large models.Superior performance is denoted as Win,while equal performance is denoted as Tie.Comparing with the ChatGLM3-6B model,the AD question-answering system achieved a Win rate of 88.09%in FA and a Tie rate of 7.14%,and a Win rate of 85.71%in CR with a Tie rate of 11.90%.Compared to the ChatGPT model,the AD question-answering system attained a Win rate of 54.76%in FA with a Tie rate of 30.95%and a Win rate of 35.71%in CR with a Tie rate of 40.47%.Conclusion The AD question-answering system demonstrated better performance in FA and CR compared to the ChatGLM3-6B and ChatGPT models,confirming the effectiveness of the approach proposed in this study.
