Could AI improve discharge summaries?

The challenge

When a patient leaves hospital, the right information matters

A discharge summary is one of the most important links between hospital and community care. It tells GPs and patients what happened during a hospital stay, including diagnoses, test results, treatments and follow-up plans.

But producing these summaries can be time-consuming. Information is often spread across different clinical notes, teams and systems, and important details can sometimes be missed or recorded inaccurately.

For patients, the consequences can be serious. A missing diagnosis, unclear treatment plan or overlooked follow-up can create gaps in care at the point when someone is moving from hospital back into the community.

Putting AI to the test

Could Generative AI help clinicians write better discharge summaries?

iCARE researchers wanted to find out whether a Large Language Model (LLM) could help doctors draft high-quality discharge summaries - without taking clinical judgement out of the process.

The team developed an AI application that uses GPT-4o-mini to generate a draft summary from information in the electronic patient record. The draft was designed to be reviewed and edited by a clinician before it could be signed off.

The study used 1,294 real hospital stays, with half receiving an AI-generated summary and the other half using the original clinician-written summary.

The summaries were then independently assessed by 12 doctors who routinely write discharge summaries. Without knowing which summaries had been generated by AI, they assessed whether each was suitable for sign-off and rated its accuracy, completeness, clarity, tone and potential for harm.

A striking result

Doctors were more likely to approve AI-assisted summaries

92% of AI-generated summaries were approved with only minor changes, compared with 70% of clinician-written summaries.

That means doctors approved 22 more summaries in every 100 with only minor changes when they were AI-generated.

The AI-generated summaries were also rated more highly for accuracy, completeness, clarity and tone, and were less likely to contain potentially harmful missing or inaccurate information.

Why it matters

More than a better discharge summary

The findings suggest that Generative AI could help clinicians with time-consuming documentation while improving the quality of information passed between hospital, GPs and patients.

Clearer discharge summaries could also make it easier for patients to read and understand important information about their care, helping them feel more informed about what happened in hospital and what they need to do next.

But the significance of this work goes further.

By testing AI using real-world healthcare data and with clinicians firmly in control, the study provides evidence for how Generative AI can be evaluated safely before being introduced more widely.

This could help shape how AI is responsibly adopted across the NHS.

Keeping people in control

AI can help - but clinicians remain responsible

The study also showed why human oversight matters. AI-generated summaries sometimes contained incorrect information, so they cannot simply be accepted without review.

The aim is not to replace clinical judgement. Instead, AI can provide a first draft, giving clinicians more time to focus on checking the information, making changes where needed and ensuring the final summary is safe and appropriate for the patient.

What's next

From promising results to real-world use

The next step is to understand how AI-assisted discharge summaries work in practice.

The team will explore whether the summaries are useful and easy for GPs to work with, while continuing to develop ways to monitor and audit AI-generated content.

The goal is not simply to develop an AI tool, but to understand how new technologies can be introduced into healthcare safely, responsibly and with evidence behind them.

People involved in this research