TL;DR

Discharge summaries drafted by AI were more often judged ready for clinical use than those written in the normal way, according to a study from Imperial College Healthcare NHS Trust’s iCARE team published in JAMA Network Open and announced on 18 November. Twelve doctors reviewed 1,294 summaries without knowing which were which. They rated 92% of the AI drafts clinically acceptable, compared with 70% of the clinician-written ones.

How the study worked

A discharge summary tells a GP what happened during a patient’s hospital stay: diagnoses, tests, treatment, changes to medication and what follow-up is needed. Writing one means gathering material from every part of a stay, usually while other patients need attention, and the task has grown harder as records have expanded.

The researchers built an application that used large language models to draft each summary from the free-text hospital record for that stay. The evaluation was retrospective, using anonymised data from the trust’s secure research data platform rather than live patients. The doctors who did the reviewing all work at the trust and routinely write discharge documentation.

“Clinically acceptable” meant the summary could be signed off as written or after small fixes. Reviewers also gave the AI versions better marks for tone, clarity, accuracy and completeness, and found fewer omissions and errors with the potential to cause harm.

Clinicians in the design

The trust stresses that resident doctors helped design and test the tool from the start. Dr Benjamin Atkins, one of them, described a “constant tension” on the wards between clinical work and paperwork such as discharge letters, both of which are urgent. Lead researcher Kathleen Goldsmith said clinical AI has to be built “alongside the clinicians who will use them” and evaluated rigorously for safety and reliability.

Looking forward

The study is a retrospective test, not a deployment, and the trust frames AI drafting as a promising use case that still depends on human sign-off. iCARE director Erik Mayer said the same approach could apply to other documentation tasks. In our view, discharge summaries are a sensible target for NHS trusts weighing generative AI: the output is checked by a clinician before it leaves the hospital, and the gain in time is easy to measure. The harder test will be whether the results hold up on a live ward.