Enhancing Large Language Models for Improved Accuracy and Safety in Medical Question Answering: Comparative Study

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Decision Support

Background:

• Large language models (LLMs) show promise for healthcare but face accuracy and safety limitations.
• Current LLMs struggle with outdated knowledge and can produce unreliable information in clinical settings.
• Enhancing LLM reliability is critical for their adoption in healthcare.

Purpose of the Study:

• To evaluate the Med-RISE tool's effectiveness in improving LLM accuracy and safety for medical question answering.
• To compare Med-RISE's performance against baseline large language models (LLMs) across diverse medical domains.
• To assess the impact of Med-RISE on reducing hallucinations in LLM-generated medical answers.

Main Methods:

• Developed Med-RISE, a retrieval-augmented generation framework with query rewriting, real-time information retrieval, summarization, and a fact/safety filter.
• Integrated Med-RISE with four LLMs (GPT-3.5, GPT-4, Vicuna-13B, ChatGLM-6B).
• Assessed performance on four medical question datasets (MedQA, PubMedQA, MedMCQA, EYE500), measuring accuracy and hallucination rates (factuality and faithfulness).

Main Results:

• Med-RISE integration increased LLM accuracy by 9.8%–16.3% (mean 13%) across datasets.
• Accuracy improvements were 16.3% for GPT-3.5, 12.9% for GPT-4, 13% for Vicuna-13B, and 9.8% for ChatGLM-6B.
• Med-RISE reduced hallucinations by 11.8%–18% (mean 15.1%), with factuality hallucinations decreasing by 13.5% and faithfulness by 5.8%.

Conclusions:

• The Med-RISE framework substantially enhances LLM accuracy and reduces hallucinations in medical question answering.
• Real-time retrieval and filtering capabilities improve LLM reliability and interpretability in medicine.
• Med-RISE presents a promising solution for clinical practice and decision support systems.