AI readings of heart MRIs slightly beat doctors at flagging death risk

A computer program that reads heart MRI scans was slightly better than experienced clinicians at picking out patients who went on to have a higher risk of dying, according to a large study led by researchers in the United Kingdom.¹ The work looked back at scans from 27,382 patients in the UK and the United States and was published in the journal The Lancet Digital Health.¹ The researchers also stressed that doctors still need to check the program’s results.¹

What happened

The study was led by specialists from City St George’s, University of London and University College London.¹ It was a retrospective, multicentre study. That means the team used scans and health records that already existed, gathered at more than one hospital, instead of signing people up and following them forward from the start.¹

All the patients had a heart MRI because they had known or suspected heart disease.¹ An MRI uses magnets and radio waves to make detailed moving pictures of the heart. The patients were then followed for an average of more than five years, and 3,458 of them died during that time.¹

The team focused on one number that heart doctors rely on every day: the ejection fraction. It shows how effectively the heart pumps blood out with each beat.¹ It is used to choose treatment for heart disease, to judge the risk of surgery and to keep watch on the heart during chemotherapy.¹ The researchers compared the ejection fraction worked out by an AI tool with the measurements made by specialists reading the same MRI scans.¹

The evidence so far

According to the authors, the AI’s measurements were slightly but statistically significantly better at identifying patients at higher risk of death.¹ “Statistically significant” means the difference is unlikely to be down to chance alone. It does not mean the difference is large, and the authors themselves describe it as slight.¹

The result held after the researchers accounted for patients’ age and sex, two things that strongly affect the risk of dying on their own.¹

Speed was the other big difference. The program analysed a scan in about two to three seconds per patient.¹ In earlier work by the same research group, expert assessment of the images took about 13 minutes.¹

The program was not perfect. The researchers reported AI errors in about 1% of cases at the patient level, affecting 0.08% of individual images.¹ For that reason they said clinicians should verify the results and interpret scans in light of each patient’s full medical picture.¹

What it does not show yet

This kind of study can show which measurement lines up better with who later died. It cannot show that using the AI in real hospitals would lead to different treatment decisions, or that it would help patients live longer. No patients were treated differently because of the AI in this study.

The report available to Biohack does not give the size of the gap between the AI and the clinicians in plain numbers, the name of the AI tool, or how the patients were split between the UK and the US.¹ It also does not say how the AI performed in specific groups, such as people with particular heart conditions. Those details would matter for judging where the tool might be most and least useful.

Because the study looked back at past scans, it also reflects how those hospitals scanned and treated patients at the time. Results may differ in hospitals with other scanners, other patient groups or other ways of working.

Who is behind it

The research was carried out by academic teams at City St George’s, University of London and University College London, using data from patients in the UK and the US.¹ The coverage Biohack reviewed was reported by UA.NEWS, citing the medical news site News-Medical.¹ That coverage does not say who funded the work or whether any authors have ties to companies that make imaging software.¹ Biohack did not have access to the published paper itself, where such disclosures would normally appear.

How it fits the field

Several research groups are trying to get AI to read heart MRI scans, largely because expert readers are in short supply. A Carnegie Mellon University and Cleveland Clinic team has described cardiac MRI as widely regarded as the gold standard for checking the heart’s structure, function and tissue health.² The same team noted that each scan can contain hundreds to thousands of images, that interpreting one exam can take trained specialists 40 minutes or more, and that the technology is expensive and concentrated in major medical centres.³ That 40-minute figure refers to reading a whole exam, while the 13 minutes cited by the UK group referred to expert assessment of images in their own earlier work, so the two numbers describe different tasks.¹³

In May 2026 the Carnegie Mellon and Cleveland Clinic group announced a system called CMR-CLIP, trained on more than 13,000 patient studies from Cleveland Clinic by matching moving MRI images with doctors’ written reports.² The group said it reached accuracy as high as 99% for certain heart conditions and still performed strongly on separate datasets from France and Cleveland Clinic Florida.²

In March 2026 a Penn Medicine-led team reported a model in Nature Biomedical Engineering trained on more than 300,000 MRI video clips from roughly 20,000 patients.⁴ It estimated ejection fraction with what the team called expert-level accuracy and diagnosed 39 heart conditions; in a screen of more than 40,000 scans it flagged 112 previously undiagnosed cases of hypertrophic cardiomyopathy, a thickening of the heart muscle.⁴ That team said it plans prospective studies, meaning tests that follow patients forward in time.⁴

What sets the new UK-led study apart is its focus. Instead of asking whether the AI matches an expert’s diagnosis, it asked whether the AI’s measurement of heart pumping lines up with who actually died over the following years, in a large group followed for more than five years on average.¹

What to watch next

The authors see AI as a possible first-pass tool for reading heart MRI scans, with experts reviewing the results when needed.¹ Their next step is to combine AI readings of heart scans with other tests, including electrocardiograms (ECGs, the common test that records the heart’s electrical activity), in the hope of spotting problems in the heart muscle earlier.¹

Key questions for future work include whether the AI’s small edge holds up in studies that follow patients forward in time, whether it changes the decisions doctors make, and whether those changes help patients. How the occasional AI errors are caught in everyday practice will also matter, since the researchers themselves say human oversight remains necessary.¹