Fact-Checking Large Language Model Responses to a Health Care Prompt: Comparative Study

Area of Science:

• Artificial Intelligence in Healthcare
• Natural Language Processing
• Clinical Decision Support

Background:

• Large language models (LLMs) offer potential applications in healthcare, including patient education and diagnosis.
• Current evaluations of LLMs in healthcare settings are limited.
• This study addresses the need for rigorous assessment of LLM capabilities in clinical contexts.

Purpose of the Study:

• To evaluate the accuracy and efficiency of automated fact-checking using two distinct LLMs.
• To demonstrate how an LLM can assist patients in refining prompts for improved clinical safety.
• To compare LLM performance against human expert evaluations.

Main Methods:

• A comparative study involving two LLMs (GPT-4o, OpenBioLLM-70B) and three human experts.
• A clinical scenario focused on retinoid safety for acne treatment, involving prompt refinement and fact-checking.
• Evaluation of LLM fact-checking accuracy and time efficiency on 20 diverse clinical statements.
• Outcome measures included accuracy percentage, time to fact-check, and prompt redrafting success.

Main Results:

• GPT-4o and OpenBioLLM-70B achieved 86% agreement with experts in the clinical scenario, though with some omissions regarding critical safety information (isotretinoin, folic acid).
• For 20 clinical statements, GPT-4o demonstrated 100% accuracy compared to human experts, while OpenBioLLM-70B achieved 95% accuracy.
• LLMs significantly reduced fact-checking time compared to human experts (seconds/minutes vs. 18 minutes on average).

Conclusions:

• GPT-4o can effectively enhance patient prompts for better health information capture and safety.
• Both evaluated LLMs provide efficient fact-checking with accuracy levels approaching those of human experts.
• Human expert review remains essential to verify accuracy and ensure comprehensive patient safety, especially for critical information.