跳到正文
原文
MedPage Today·· 2 天前精选AI 评分73

AI生成的出院总结比医生更好?研究提出新思考

AI Writes Better Discharge Summaries Than Doctors, Study Suggests

AI 导读

一项英国研究发现,AI生成的出院总结比医生写的更受临床审查者好评,且错误率更低。然而,研究者强调这些结果仍需进一步验证。

推荐理由

研究显示AI在出院总结上的表现优于医生,但强调AI生成的内容仍需临床验证。

正文

— Physician reviewers gave AI summaries higher marks for accuracy, completeness, clarity, and tone

by Rachael Robertson, Enterprise & Investigative Writer, MedPage Today

November 17, 2026 • 3 min read

  • Physicians were more likely to approve a patient discharge summary generated by a large language model (LLM) than one written by a clinician.
  • More LLM-generated summaries were approved with little to no edits compared with those authored by clinicians.
  • Findings still need prospective validation.

Artificial intelligence (AI) outperformed clinicians in drafting hospital discharge summaries, a U.K.-based retrospective study showed.

A higher proportion of large language model (LLM)-generated discharge summaries were signed off by a clinician reviewer with no or minor changes compared with clinician-authored summaries (92.2% versus 70.0%), meeting prespecified noninferiority criteria, reported Kathleen Goldsmith, MSc, senior data scientist at a partnership between Imperial College London and Imperial College Healthcare NHS Trust in the U.K., and colleagues.

Moreover, clinician reviewers rated LLM-generated summaries higher on accuracy, completeness, tone, and clarity compared to clinician-written summaries, they wrote in JAMA Network Open.

Finally, a smaller proportion of AI-generated discharge summaries contained inaccurate or incomplete information compared with ones written by clinicians (26.2% vs 45.1%), they reported.

"Together, these findings suggest that AI-assisted documentation could strengthen the quality of information transferred at discharge, an important component of safe and effective continuity of care," Goldsmith told MedPage Today.

While this research focused on discharge summaries, Goldsmith said it could be applied to other administrative tasks, but "patient safety and clinical value have to be central to how we adopt technology like this."

In an accompanying editorial, Melanie Molina, MD, of the University of California San Francisco, wrote that this research raises the question not of whether AI can produce acceptable drafts, but whether physicians can verify the information in real-time.

"Lengthy, complex, and fragmented clinical records create a verification paradox: the cases in which automated summarization is most valuable are often the cases in which verification is most difficult," Molina wrote, adding that emergency department care poses particular challenges.

Plus, AI tools have already evolved far beyond the model tested in this study, with newer, more sophisticated models less likely to hallucinate.

Molina argued that while this research shows that "a thoughtfully designed LLM pipeline can produce discharge summaries that clinicians prefer to routine documentation," this does not yet equate to a safer transition of care.

"The enduring clinical task will be ensuring that what it writes is true, complete, and actionable," she wrote.

Goldsmith and team sought to evaluate how AI tools could "free up clinician time for patient-facing care while maintaining clinical safety and accuracy" and to develop a framework for evaluating the use of LLMs in clinical settings.

This retrospective noninferiority study used Imperial Clinical Analytics, Research and Evaluation (iCARE) research data from a network of 5 hospitals in London from Sept. 1, 2022, to Sept. 30, 2023.

Discharge summaries written by clinicians and by GPT-4o mini were blindly reviewed by 12 resident physicians from the Imperial College Healthcare Trust. The primary outcome was clinician sign-off on the discharge summary with little to no edits. Goldsmith noted this endpoint reflects how the technology would be used in real-world settings.

"Having clinician input in both the co-design, use, and evaluation of healthcare AI technologies is crucial to ensure these systems are trusted, and truly deliver the benefits to clinical work," Goldsmith said.

Ultimately, there were 1,294 discharge summaries included, of which 649 were clinician-authored and 645 were LLM-generated.

Patients' median age was 46 and 56.6% were female. The median length of stay was 3 days. On average, each case required the review of 30 clinical documents, or about 4,000 words.

The researchers noted some limitations, including that de-identification may have removed clinically relevant information, that only one LLM was evaluated, and the retrospective study design; prospective assessments at the point of care are still needed.

来源:MedPage Today · medpagetoday.com