Clinical Large Language Model Evaluation by Expert Review (CLEVER): Framework Development and Validation

Area of Science:

• Artificial Intelligence in Medicine
• Natural Language Processing
• Clinical Informatics

Background:

• Evaluating large language models (LLMs) is challenging due to data contamination and the gap between benchmark tasks and clinical practice.
• Existing LLM evaluation methods like public benchmarks and LLM-as-a-judge are limited by data issues and self-preference bias.
• There is a critical need for robust evaluation frameworks that reflect real-world clinical utility.

Purpose of the Study:

• To introduce CLEVER (Clinical Large Language Model Evaluation-Expert Review), a novel methodology for evaluating LLMs in healthcare.
• To conduct a blind, randomized, preference-based evaluation of LLMs using practicing medical doctors.
• To compare the performance of a general-purpose LLM against healthcare-specific LLMs on clinical tasks.

Main Methods:

• The CLEVER methodology was employed to compare GPT-4o with two healthcare-specific LLMs (8B and 70B parameters).
• Evaluations were performed on three distinct clinical tasks: text summarization, information extraction, and question answering.
• Practicing medical doctors provided preference-based evaluations focusing on factuality, clinical relevance, and conciseness.

Main Results:

• Medical doctors preferred the smaller, healthcare-specific LLM over GPT-4o in 45% to 92% of cases across key clinical dimensions.
• The healthcare-specific LLM demonstrated superior performance in factuality, clinical relevance, and conciseness compared to GPT-4o.
• Performance was comparable in open-ended medical question answering, indicating specialized LLMs excel in context-dependent tasks.

Conclusions:

• Healthcare-specific LLMs can outperform larger, general-purpose LLMs in tasks requiring clinical context understanding.
• The CLEVER methodology provides a valid and reliable approach for evaluating clinical LLMs, confirmed by interannotator agreement and correlation analyses.
• This study underscores the importance of specialized LLMs and expert review for advancing AI in medicine.