Moderate evidence
More
AI-written discharge summaries beat doctors' own in a study of 1,294 London hospital cases
Study title: Clinician Acceptance of Large Language Model–Generated Discharge Summaries
- Evidence type
- Observational study
- Sample
- n = 1,294
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-3039cite or check it
Study title: Clinician Acceptance of Large Language Model–Generated Discharge Summaries
A retrospective study of 5 London hospitals compared discharge summaries written by an AI model with those routinely written by clinicians. Twelve practicing physicians reviewed 1,294 summaries (649 clinician-authored, 645 AI-generated) against the full admission record. The AI summaries, produced by GPT-4o mini with a clinician-codesigned prompt, were 22.2 percentage points more likely to be signed with no or minor edits.
Read the study abstract
Importance Hospital discharge summaries are essential for continuity of care, but fragmented electronic health record (EHR) data can compromise their quality and safety. Large language models (LLMs) are increasingly proposed for clinical documentation tasks despite limited evidence regarding their performance in health care settings compared with the current standard of care. Objective To evaluate clinician acceptability of LLM-generated discharge summaries derived from EHRs compared with routine clinician-writt…
The gain
In a retrospective study of 1,294 discharge summaries from 5 London hospitals, 12 physicians were 22.2 percentage points more likely to sign off GPT-4o mini summaries with no or minor edits than clinician-written ones.
In brief
The AI summaries also scored higher on accuracy, completeness, tone and clarity, with adjusted risk differences for ratings of 4 or above ranging from 19.4 to 29.1 percentage points. Fewer were rated as carrying moderate or greater potential harm: 26.2% versus 45.1% for the clinical summary section, and 26.8% versus 51.5% for plan and requested actions.
The authors call this a promising early use case for generative AI in health care, with close clinician oversight. The evidence comes from retrospective review of records, not from summaries used in live patient care.
Main points
- The study covered admissions at 5 London hospitals between September 1, 2022 and September 30, 2023.
- It included 1,294 summaries (649 clinician-authored, 645 AI-generated), reviewed by 12 practicing physicians.
- Patients had a median age of 46.0 years, 56.6% were assigned female at birth, and median stay was 3.0 days.
- The sign-off difference was 22.2 percentage points (95% CI, 19.0-27.5), clearing the prespecified -5 point noninferiority margin.
- AI summaries scored higher on accuracy, completeness, tone and clarity, by 19.4 to 29.1 percentage points.
- Moderate-or-greater potential harm was rated in 26.2% of AI clinical summaries versus 45.1% of clinician-written ones.
- For plan and requested actions, the figures were 26.8% for AI and 51.5% for clinicians.
The full story
The study used a prespecified noninferiority design to test whether GPT-4o mini summaries matched the current standard of care. Admissions were eligible only if the clinician-authored summary contained identifiable clinical summary, plan and requested actions, and information given sections. Data analysis ran from July 2025 to February 2026.
The primary outcome combined two questions: whether the reviewer would sign the summary, and whether no, minor or major edits were needed. Secondary outcomes used 5-point scales for accuracy, completeness, tone and clarity, and for the potential harm of missing or incorrect information.
The model was GPT-4o mini (July 18, 2024 release) with a prompt designed together with clinicians. Results favoured the AI summaries on every measure reported, including a lower share rated as potentially moderately harmful or worse.
Sources
1 on file- JAMA Network Opendoi:10.1001/jamanetworkopen.2026.44496
Talk it over
The 12 physicians judged 1,294 discharge summaries against the full admission record, not in live patient care. How much should that retrospective setup shape what we take from the 22.2 percentage point gap in sign-off?
AI-written discharge summaries beat doctors' own in a study of 1,294 London hospital caseshttps://truvace.com/item/clinician-acceptance-of-large-language-model-generated-discharge-summaries
The 12 physicians judged 1,294 discharge summaries against the full admission record, not in live patient care. How much should that retrospective setup shape what we take from the 22.2 percentage point gap in sign-off?
Answer in the threadModerate or greater potential harm was rated in 26.2% of AI clinical summaries versus 45.1% of clinician-written ones. How would you weigh that gap against the need for close clinician oversight of GPT-4o mini summaries?
Answer in the threadWhat result from a study of summaries used in live patient care at the 5 London hospitals would change your mind about whether AI-written discharge summaries should replace clinician-written ones?
Answer in the thread
If you have more time
Look up how discharge summaries are checked and signed at a hospital you know. Where would a reviewer fit in if a model like GPT-4o mini drafted them?
Answer in the threadThe patients had a median age of 46.0 years and a median stay of 3.0 days. Which kinds of admissions would you want tested next, and why?
Answer in the thread
ace
What readers say