Williams Christopher Y K, Bains Jaskaran, Tang Tianyu, Patel Kishan, Lucas Alexa N, Chen Fiona, Miao Brenda Y, Butte Atul J, Kornblith Aaron E
Bakar Computational Health Sciences Institute, University of California, San Francisco, California, United States of America.
Department of Emergency Medicine, University of California, San Francisco, California, United States of America.
PLOS Digit Health. 2025 Jun 17;4(6):e0000899. doi: 10.1371/journal.pdig.0000899. eCollection 2025 Jun.
Large language models (LLMs) possess a range of capabilities which may be applied to the clinical domain, including text summarization. As ambient artificial intelligence scribes and other LLM-based tools begin to be deployed within healthcare settings, rigorous evaluations of the accuracy of these technologies are urgently needed. In this cross-sectional study of 100 randomly sampled adult Emergency Department (ED) visits from 2012 to 2023 at the University of California, San Francisco ED, we sought to investigate the performance of GPT-4 and GPT-3.5-turbo in generating ED encounter summaries and evaluate the prevalence and type of errors for each section of the encounter summary across three evaluation criteria: 1) Inaccuracy of LLM-summarized information; 2) Hallucination of information; 3) Omission of relevant clinical information. In total, 33% of summaries generated by GPT-4 and 10% of those generated by GPT-3.5-turbo were entirely error-free across all evaluated domains. Summaries generated by GPT-4 were mostly accurate, with inaccuracies found in only 10% of cases, however, 42% of the summaries exhibited hallucinations and 47% omitted clinically relevant information. Inaccuracies and hallucinations were most commonly found in the Plan sections of LLM-generated summaries, while clinical omissions were concentrated in text describing patients' Physical Examination findings or History of Presenting Complaint. The potential harmfulness score across errors was low, with a mean score of 0.57 (SD 1.11) out of 7 and only three errors scoring 4 ('Potential for permanent harm') or greater. In summary, we found that LLMs could generate accurate encounter summaries but were liable to hallucination and omission of clinically relevant information. Individual errors on average had a low potential for harm. A comprehensive understanding of the location and type of errors found in LLM-generated clinical text is important to facilitate clinician review of such content and prevent patient harm.
大语言模型（LLMs）具备一系列可应用于临床领域的能力，包括文本摘要。随着环境人工智能抄写员和其他基于大语言模型的工具开始在医疗环境中部署，迫切需要对这些技术的准确性进行严格评估。在这项横断面研究中，我们从2012年至2023年在加利福尼亚大学旧金山分校急诊科随机抽取了100例成年患者的急诊就诊病例，旨在研究GPT - 4和GPT - 3.5 - turbo生成急诊就诊摘要的性能，并根据三个评估标准评估就诊摘要各部分错误的发生率和类型：1）大语言模型总结信息的不准确；2）信息幻觉；3）相关临床信息的遗漏。总体而言，GPT - 4生成的摘要中有33%在所有评估领域完全无错误，GPT - 3.5 - turbo生成的摘要中有10%完全无错误。GPT - 4生成的摘要大多准确，只有10%的病例存在不准确情况，然而，42%的摘要出现了幻觉，47%遗漏了临床相关信息。不准确和幻觉最常见于大语言模型生成的摘要的“计划”部分，而临床信息遗漏集中在描述患者体格检查结果或现病史的文本中。错误的潜在危害评分较低，7分制的平均得分为0.57（标准差1.11），只有三个错误的评分达到了4分（“有永久伤害的可能性”）或更高。总之，我们发现大语言模型可以生成准确的就诊摘要，但容易出现幻觉和遗漏临床相关信息。平均而言，个别错误造成伤害的可能性较低。全面了解大语言模型生成的临床文本中错误的位置和类型，对于促进临床医生对此类内容的审查以及防止患者受到伤害非常重要。
