Evaluating Large Language Models on Medical Evidence Summarization.

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Natural Language Processing

Background:

• Large language models (LLMs) demonstrate significant potential in various applications, including high-stakes domains.
• The efficacy of LLMs in medical evidence summarization remains underexplored, necessitating systematic evaluation.

Approach:

• This study systematically assesses the capabilities and limitations of GPT-3.5 and ChatGPT for zero-shot medical evidence summarization.
• Both automatic and human evaluations were employed to assess multiple dimensions of summary quality across six clinical domains.
• A novel terminology of error types for medical evidence summarization was defined based on human evaluation findings.

Key Points:

• Automatic metrics for summary quality assessment show poor correlation with human judgments.
• LLMs may generate factually inconsistent summaries, present information with undue certainty or doubt, and struggle to identify salient details.
• Performance degrades with longer textual contexts, increasing the risk of errors and misinformation.

Conclusions:

• LLMs exhibit limitations in medical evidence summarization, posing risks of factual inconsistency and misinformation.
• Human evaluation is essential for accurate quality assessment and identifying specific error patterns in LLM-generated medical summaries.
• Further research is needed to improve LLM reliability for critical medical information synthesis.