.jpg)
.jpg)
Generative artificial intelligence can write hospital discharge summaries that senior clinicians judge to be safer and more reliable than those produced by doctors working against the clock on busy wards. That is the conclusion of new research from Imperial College London and its NHS partners, which found that 92 per cent of the AI-drafted summaries reached a clinically acceptable standard, against 70 per cent of those written by hospital staff.
The work was led by iCARE, a joint data initiative between Imperial College London and Imperial College Healthcare NHS Trust, and has appeared in a peer-reviewed journal. For a field in which enthusiasm for AI has frequently run ahead of proof, the result offers something firmer than a demonstration. On the figures, roughly one in five more summaries needed little or no rework when a machine had written the first draft.
The researchers used genuine patient records instead of invented scenarios. The material was anonymised and drawn from hospitals in north-west London, then held in the Imperial Secure Data Environment, a protected NHS research setting. A large language model was set up to read the untidy mixture of free-text notes, results and medication changes that accumulates during an admission, and to turn it into a coherent summary.
Assessment was blinded. Twelve resident doctors, each of whom writes discharge summaries as a matter of routine, reviewed 1,294 documents without being told who or what had produced them. The AI drafts were rated higher for accuracy, completeness, clarity and tone. They were also less prone to leaving out information that could have harmed a patient.
The reviewers did not find the software flawless. Some AI drafts contained mistakes, a reminder that the tool is intended to prepare a draft for a doctor to check and sign, never to replace the doctor's judgement. That arrangement keeps responsibility with a named clinician, which regulators and hospital boards will expect.
The subject was chosen with care. A discharge summary is the document that carries a patient's story from the hospital to the community. It tells a GP what was found, what was treated, which medicines were started or stopped and what should happen next. It also informs patients and families about what to expect after leaving. When the summary is thin, late or wrong, the consequences surface weeks afterwards in primary care, or in a readmission.
Yet the document is usually written at the margins of the working day. Resident doctors such as Dr Benjamin Atkins and Dr Fidan Damirova, who took part in the study, spend much of their time on urgent clinical tasks, with paperwork building up behind them. Summaries are frequently completed at the end of a long shift, when concentration is at its lowest. Automating the first draft promises to return some of that time to patients, and to lift the quality of a record on which many other clinicians depend.
Importantly, the technology was shaped by the people who will use it. Resident doctors helped design the system and refined its instructions over several rounds, so that its output matched how a summary is meant to read in practice. Kathleen Goldsmith, who led the research, has said that involving medical staff throughout development and testing is a necessary condition for clinical technology that can be trusted and held to account. Tools imposed from outside tend to meet resistance on the wards, and the study suggests that early clinical ownership may be as valuable to adoption as raw performance.
The implications reach beyond one trust. NHS organisations are under steady pressure to modernise while demonstrating that new tools are safe, lawful and effective. A model of evaluation conducted inside the health service, on its own data and judged by its own practising doctors, gives other trusts a method they can copy.
Erik Mayer, director of iCARE, believes that method has wider use. Testing AI against real working routines, with blinded review by clinicians, could be applied to other administrative burdens such as clinic letters and referrals. Each application would need separate evidence, but the template now exists.
The next phase will look at what happens once summaries leave the hospital. Researchers intend to find out whether GPs consider the AI-assisted documents helpful and easy to use, and to build better ways of monitoring and auditing what the software produces. Those questions will show whether the gains recorded under study conditions survive the demands of everyday practice.
The research was supported by the National Institute for Health and Care Research (NIHR) Imperial Biomedical Research Centre, which links the work to the NHS's established research infrastructure.