Answers are grounded in this paper and cite their sources.
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-written summaries. Design, Setting, and Participants In this retrospective comparative effectiveness study using a prespecified noninferiority framework, acceptability of LLM-generated discharge summaries was compared with clinician-authored discharge summaries. Admission records from 5 London hospitals for September 1, 2022, to September 30, 2023, were evaluated against the full admission record by practicing physicians. Admissions were eligible only if the clinician-authored discharge summary contained identifiable clinical summary, plan and requested actions, and information given sections. Data analysis was conducted from July 2025 to February 2026. Intervention LLM-generated summaries were produced using GPT-4o mini (July 18, 2024 release; OpenAI) and a clinician-codesigned prompt. Main Outcomes and Measures The primary outcome was clinician signoff with no or minor edits, derived from 2 sequential questions: (1) whether the reviewer would sign the summary and (2) whether no, minor, or major edits were required. A 5–percentage-point noninferiority margin was prespecified. Secondary outcomes assessed accuracy, completeness, tone, and clarity (using a 5-point Likert scale: 1, poor; 2, fair; 3, good; 4, very good; and 5, excellent) and potential harmfulness of missing or incorrect information (using a 5-point Likert scale: 1, no harm; 2, minor harm; 3, moderate harm; 4, high harm; and 5, extremely high harm). Results The study sample included 1294 (649 clinician-authored and 645 LLM-generated) unique case discharge summaries that were evaluated against the full admission record by 12 practicing physicians. Case patients had a median age of 46.0 years (IQR, 26.2-69.8 years), 733 (56.6%) were assigned female at birth, and the median length of stay was 3.0 days (IQR, 1.0-7.0 days). The absolute difference in model-estimated probability of signoff with no or minor edits (LLM generated minus clinician authored) was 22.2 percentage points (95% CI, 19.0-27.5 percentage points); the lower bound exceeded the prespecified noninferiority margin of −5 percentage points, demonstrating noninferiority. LLM-generated summaries received higher ratings for accuracy, completeness, tone, and clarity (adjusted risk differences for scores ≥4 ranged from 19.4 [95% CI, 14.8-24.1] to 29.1 [95% CI, 25.1-33.6] percentage points; all 95% CIs excluded 0). LLM-generated summaries also had a lower proportion of cases rated 3 or greater for potential harm on the clinical summary and plan and requested actions sections (140 of 534 [26.2%] and 124 of 462 [26.8%]) compared with clinician-authored summaries (271 of 601 [45.1%] and 278 of 540 [51.5%]). Conclusions and Relevance In this study, LLM-generated discharge summaries met the current standard of care, representing a promising early use case for generative artificial intelligence in health care with close clinician oversight.
No takes yet. Share an insight, caveat, or question.
Goldsmith et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: