ORIGINAL RESEARCH article
Front. Public Health
Sec. Digital Public Health
Volume 13 - 2025 | doi: 10.3389/fpubh.2025.1635381
MEGA-RAG: A Retrieval-Augmented Generation Framework with Multi-Evidence Guided Answer Refinement for Mitigating Hallucinations of LLMs in Public Health
Provisionally accepted
- 1China Academy of Information and Communications Technology, Beijing, China
- 2The University of Sydney, Darlington, Australia
- 3Fudan University, Shanghai, China
Select one of your emails
You have multiple emails registered with Frontiers:
Notify me on publication
Please enter your email address:
If you already have an account, please login
You don't have a Frontiers account ? You can register here
The increasing adoption of LLMs in public health systems has raised significant concerns regarding hallucinations factually inaccurate or misleading information. This study presents the MEGA-RAG framework, a novel approach specifically designed to mitigate hallucinations in LLM based public health applications. MEGA-RAG integrates multi source evidence retrieval, including dense retrieval (via FAISS), keyword based retrieval (via BM25), and knowledge graphs, and employs a cross-encoder reranker to ensure semantic relevance. Additionally, a discrepancy-aware refinement module further enhances factual accuracy. Experimental results demonstrate that MEGA-RAG outperforms four baseline models (PubMedBERT, PubMedGPT, standalone LLM, and LLM with RAG), achieving a reduction in hallucination rates by over 40%, and attaining the highest performance in accuracy (0.7913), precision (0.7541), recall (0.8304), and F1 score (0.7904). These results confirm that MEGA-RAG is highly effective in generating factually reliable and medically accurate responses, thereby improving the credibility of AI-generated health information for applications in health education, clinical communication, and policy development.
Keywords: Public Health, AI in healthcare, LLM Hallucinations, Medical question answering, RAG
Received: 26 May 2025; Accepted: 05 Sep 2025.
Copyright: © 2025 Xu, Yan, Dai and Wu. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
* Correspondence:
Chengxiao Dai, The University of Sydney, Darlington, Australia
Fan Wu, Fudan University, Shanghai, China
Disclaimer: All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.