Evaluating large language models on medical evidence summarization.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Natural Language Processing

Background:

• Large language models (LLMs) exhibit strong zero- and few-shot capabilities across tasks.
• High-stakes domains like healthcare are exploring LLM applications.

Purpose of the Study:

• To systematically evaluate LLM performance (GPT-3.5, ChatGPT) in zero-shot medical evidence summarization.
• To assess summary quality using both automatic and human evaluations across six clinical domains.

Main Methods:

• Conducted automatic and human evaluations of LLM-generated medical summaries.
• Defined a taxonomy of error types based on human assessments.
• Analyzed summary quality across different clinical domains and text lengths.

Main Results:

• Automatic metrics showed weak correlation with human-judged summary quality.
• LLMs generated factually inconsistent summaries and exhibited problematic certainty levels.
• Models struggled with identifying salient information and were less accurate with longer texts.

Conclusions:

• LLMs require careful validation for medical evidence summarization due to potential for factual errors and misinformation.
• Human evaluation is crucial for assessing the quality and safety of LLM-generated medical summaries.
• Limitations in handling long contexts and identifying key information need further research.