Estimated reading time: 6 minutes · Last updated:
Large language model drafts of AI discharge summaries were judged clinically acceptable 92 per cent of the time in a blinded review. Summaries written by clinicians in usual practice reached 70 per cent. The study was led by iCARE (Imperial Clinical Analytics, Research and Evaluation) and published in JAMA Network Open on 17 November 2026. In it, 12 doctors at Imperial College Healthcare NHS Trust assessed 1,294 summaries without knowing which ones software had produced. The AI drafts also scored higher on accuracy, completeness, clarity and tone, and reviewers flagged fewer potentially harmful errors in them. The result is strong but has limits. The evaluation looked back at anonymised past records, and every draft was judged by one test: whether a doctor would sign it.
This work is an important step towards making AI a meaningful and trusted part of clinical practice.
Kathleen Goldsmith, lead researcher and senior data scientist at iCARE
Key takeaways
- Acceptability gap: In the iCARE evaluation, 92 per cent of AI-generated discharge summaries were clinically acceptable, compared with 70 per cent of clinician-written ones.
- Blinded review: Twelve doctors at Imperial College Healthcare NHS Trust independently reviewed 1,294 summaries without knowing their origin.
- Fewer harm flags: Reviewers found fewer potentially harmful omissions and inaccuracies in the AI drafts. They also rated those drafts higher on accuracy, completeness, clarity and tone.
- Retrospective design: The tool drafted summaries from anonymised past admissions held in the Imperial Secure Data Environment. It was not tested on patients being discharged at the time.
Table of contents
- Key takeaways
- What the 92 Per Cent Acceptability Figure Measures
- Why Discharge Summaries Strain Doctors on the Ward
- How iCARE Built and Tested the Tool With Resident Doctors
- What the Study Leaves Unanswered
- Will AI-drafted discharge summaries reach routine use?
- What to be careful about
- Frequently asked questions
What the 92 Per Cent Acceptability Figure Measures
Clinical acceptability had a practical definition in this study. A summary passed if a doctor could sign it off with no changes, or with only small ones. On that test, 92 per cent of AI drafts passed against 70 per cent of summaries produced in normal practice. That is a gap of about 22 percentage points. Put the other way, roughly 8 per cent of AI drafts needed more than light editing, compared with about 30 per cent of the clinician-written set.
The reviewers were 12 doctors at Imperial College Healthcare NHS Trust who write discharge paperwork as part of their usual jobs. Each worked independently and blind. None of them knew whether a person or the software had produced any of the 1,294 summaries. That design matters because it removes any bias for or against machine-written text.
The AI drafts also did better than a simple pass rate shows. Reviewers rated them higher on accuracy, completeness, clarity and tone. They also flagged fewer potentially harmful omissions and inaccuracies in them. That last measure bears most directly on patient safety once someone has left hospital.
Why Discharge Summaries Strain Doctors on the Ward
A discharge summary is the handover note a hospital sends when a patient leaves. It tells the GP and other professionals what was diagnosed and which tests and treatments were given. It also records how medicines changed and what follow-up is planned. Patients and families can use the same document to understand what happened during the stay and what comes next.
Writing one means pulling together material scattered across an entire admission. That includes notes from several clinicians, test results and input from specialist teams. Doctors usually do this while other patients still need them. The job has grown harder as patient records have become longer and more detailed.
Dr Benjamin Atkins, a resident doctor at the trust, was one of the study's participants. He described a pull between work that moves a patient's treatment forward and paperwork that gets them home. Each of those jobs is urgent, frequently for several patients at the same time, and neither can be put off. In his view, that clash is exactly why automating administrative tasks would matter so much on hospital wards.
How iCARE Built and Tested the Tool With Resident Doctors
iCARE is a data and analytics initiative set up by Erik Mayer and Ben Glampson. Its team built an application that uses large language models, AI systems that generate written text, to draft a summary. The drafts are built from the unstructured electronic health record data attached to each hospital admission.
The work ran on the Imperial Secure Data Environment (SDE). This is a platform where approved researchers, clinicians and analysts can work with anonymised health records. The evaluation was retrospective: the software produced drafts for past admissions rather than for patients leaving hospital at the time. Feeding it many documents per stay mirrored the scattered sources doctors work from.
Resident doctors who write these summaries routinely helped design and assess the application, among them Atkins and Dr Fidan Damirova. Kathleen Goldsmith, lead researcher and senior data scientist at iCARE, said such applications must be developed "alongside the clinicians who will use them". She added that they must then be tested rigorously for safety, reliability and clinical value.
What the Study Leaves Unanswered
The results come from historical records and from reviewers at one trust. They show how drafts compare on paper, not how the tool behaves on a busy ward. The study also reports no measured time savings. Erik Mayer, director of iCARE, said the approach could apply to other tasks and would ultimately give health workers more time for patient-centred care. That is a statement about future use, not a finding.
Human sign-off stays at the centre of the method. Acceptability was defined by whether a doctor would put their name to a draft. That keeps professional accountability with the clinician rather than the software. Damirova described the potential as turning fragmented clinical information into clear discharge communication that could support safer transitions of care. Her wording is conditional, and the evidence supports keeping it that way.
The obvious next test would be a prospective one. In that kind of study, the tool would draft summaries for current patients and doctors would edit them in real time. No such study has been announced.
| Measure | AI-generated | Clinician-written |
|---|---|---|
| Met clinical acceptability threshold | 92% | 70% |
| Ratings for accuracy, completeness, clarity, tone | Higher | Lower |
| Potentially harmful omissions and inaccuracies | Fewer identified | More identified |
Will AI-drafted discharge summaries reach routine use?
The case for
- A 22-point acceptability margin in a blinded review gives adopters a concrete benchmark rather than a demonstration.
- Fewer potentially harmful omissions points to better information reaching GPs at handover.
- Resident doctors helped design the tool, which makes the workflow fit more credible.
- Mayer says the same method could extend to other documentation tasks.
The case against
- The evaluation was retrospective and has not shown performance on live wards.
- All 12 reviewers worked at one trust, which may limit how far the results generalise.
- Time savings, the main promise for doctors, were not measured.
- Every draft still needs a clinician's review and signature, so workload gains depend on how fast that check is.
What to be careful about
- The drafts were tested on past anonymised records, and live admissions may expose errors the retrospective set did not.
- Results from one trust's doctors and records may not carry over to hospitals with different systems or note-writing habits.
- If clinicians come to trust drafts that usually pass, the remaining 8 per cent needing real rework could slip through with lighter checking.
The bottom line
The iCARE results give generative AI in hospitals something it often lacks: a blinded comparison against real clinical work. Drafts from the system passed the sign-off test 92 per cent of the time, against 70 per cent for usual practice, and drew fewer harm flags. That is a meaningful margin on a task doctors describe as urgent and repetitive. It is also a retrospective result from one trust, with no time savings measured and no live deployment described. The fair reading is that discharge documentation looks like a credible place for these tools. That holds as long as clinicians stay involved in the design, the review and the final signature. The evidence that matters next is how the tool performs on the ward itself.
What to watch
- Watch whether Imperial College Healthcare NHS Trust moves the tool from retrospective testing into a live ward pilot; no date has been set.
- Watch for iCARE applying the same co-design and blinded-evaluation method to documentation beyond discharge summaries, as Erik Mayer suggested; no timetable has been given.
- Watch for any follow-up study that measures how much doctor time the drafts actually save; none has been announced.
Frequently asked questions
Who funded the iCARE discharge summary research?
iCARE is funded by the NIHR Imperial Biomedical Research Centre (BRC). It runs as a partnership between Imperial College London and Imperial College Healthcare NHS Trust, and was set up by Erik Mayer and Ben Glampson.
What is the JAMA Network Open paper on AI discharge summaries called?
The paper is titled Clinician Acceptance of Large Language Model–Generated Discharge Summaries. JAMA Network Open published it on 17 November 2026.
Could the iCARE approach be used for other hospital paperwork?
Erik Mayer, director of iCARE, said the method is applicable to other use cases beyond discharge summaries. He said it could ultimately give health workers more time for patient-centred care. No specific next use case has been named.
Related reading