Evaluating large language models for drafting emergency department encounter summaries.

Area of Science:

• Clinical Informatics
• Artificial Intelligence in Healthcare
• Medical Documentation

Background:

• Large language models (LLMs) show promise for clinical applications like text summarization.
• The increasing deployment of AI scribes necessitates rigorous evaluation of their accuracy in healthcare settings.

Purpose of the Study:

• To evaluate the performance of GPT-4 and GPT-3.5-turbo in generating Emergency Department (ED) encounter summaries.
• To identify the prevalence and types of errors (inaccuracy, hallucination, omission) in LLM-generated ED summaries.

Main Methods:

• Cross-sectional study of 100 randomly sampled adult ED visits (2012-2023).
• Evaluation of GPT-4 and GPT-3.5-turbo generated summaries against three criteria: inaccuracy, hallucination, and omission.
• Analysis of error types and locations within encounter summary sections.

Main Results:

• GPT-4 generated error-free summaries in 33% of cases; GPT-3.5-turbo in 10%.
• GPT-4 summaries had 10% inaccuracies, 42% hallucinations, and 47% omissions.
• Inaccuracies/hallucinations were common in 'Plan' sections; omissions in 'Physical Examination' and 'History of Presenting Complaint' sections.
• Mean potential harm score for errors was low (0.57/7).

Conclusions:

• LLMs can generate clinical encounter summaries, but are prone to hallucinations and omissions.
• Errors in LLM-generated clinical text, while often low-harm, require careful clinician review.
• Understanding error patterns is crucial for safe integration of LLMs in clinical workflows.