IHILLM-RAG: a safe and private medical large language model based on intelligent hardware interaction and retrieval-augmented generation (RAG)
Zibo Zhou, Yi Yang, Tian‐Ling Ren · 2025
Large language models (LLMs) have shown great potential in medical question answering and health management in the rapidly evolving healthcare landscape these days. However, even with domain-specific pretraining, LLMs may potentially hallucinate and produce factually inaccurate outcomes. In addition, in the field of healthcare, due to the different habits, backgrounds, and physical conditions of patients, it is necessary to strengthen user-oriented medical assistants, but there is relatively little exploration of personalized medical assistants based on LLM. In this study, we propose a medical assistant LLM that combines external medical knowledge base and intelligent monitoring hardware interaction, which is called IHILLM-RAG. On the one hand, we build RAG system through external medical knowledge base to improve the security and accuracy of medical assistant and reduce the phenomenon of hallucination. On the other hand, we design the human physiological signal monitoring chip as a ”human-machine interface” and carry out physical interaction with the human body and data interaction with LLM, so as to enhance the LLM’s learning and memory of the user’s own habits and improve the degree of personalization. We evaluated IHILLM-RAG on the important medical question-answering datasets, our experimental results demonstrate that IHILLM-RAG compares to the state-of-the-art open source models, the accuracy of the three datasets was improved by an average of 62.4%, 50.4% and 48.6%. Through this work, we provide a safer and more personalized method for healthcare which achieves a new paradigm of medical consultation with ”Private AI-doctor” at any time and anywhere.