Data Set and Benchmark (MedGPTEval) to Evaluate Responses From Large Language Models in Medicine: Evaluation...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Natural Language Processing

Background:

• Large language models (LLMs) show promise in clinical applications but risk generating unreliable responses (hallucinations).
• Hallucinations in medical LLMs pose significant patient safety risks.
• Systematic evaluation of medical LLMs is crucial for identifying and mitigating these risks.

Purpose of the Study:

• To develop and validate a comprehensive evaluation system, MedGPTEval, for assessing large language models in the medical domain.
• To benchmark the performance of leading LLMs, including ChatGPT, ERNIE Bot, and Dr PJ, using the developed system.

Main Methods:

• Designed evaluation criteria through literature review and expert consensus (Delphi method).
• Created Chinese medical datasets, including dialogues and case reports, for LLM interaction.
• Conducted blind evaluations of LLM-generated responses by licensed medical experts using established criteria.

Main Results:

• The developed MedGPTEval system includes 16 indicators across medical, social, contextual, and robustness capabilities.
• Dr PJ demonstrated superior performance over ChatGPT and ERNIE Bot in multi-turn medical dialogues and case report scenarios.
• Dr PJ showed better robustness in terms of semantic consistency and error rates, though ChatGPT slightly outperformed it in medical professional capabilities.

Conclusions:

• MedGPTEval offers a robust framework for evaluating medical LLMs, including open-source datasets and benchmarks.
• Experimental results indicate Dr PJ's superior performance in social and professional medical contexts compared to ChatGPT and ERNIE Bot.
• The MedGPTEval system is readily adoptable by researchers to enhance open-source medical LLM datasets and facilitate further development.