(© Aliaksandr Marko - stock.adobe.com)
Measuring a Heart Takes Doctors 13 Minutes. AI Needs Seconds.
In A Nutshell
- AI measured heart scans in three to four seconds, versus a median of 13 minutes for doctors in earlier research.
- Across 27,382 patients, AI’s numbers tracked who later died slightly better than doctors’ did.
- Doctors did better with a thickened-heart-muscle disease, and AI erred more on extremely low readings.
- Researchers see a first-pass role with doctor oversight, but whether it improves care remains untested.
Measuring a heart on an MRI scan by hand took doctors a median of 13 minutes in earlier research. An artificial intelligence program did the same job in three to four seconds, and its numbers also tracked which patients later died slightly better than the doctors’ own did, according to a study of more than 27,000 heart patients in the United Kingdom and the United States.
Ejection fraction, one of the key measurements, is the percentage of blood pushed out of the heart’s main pumping chamber with each beat. Doctors use it to guide decisions about implanted heart devices, heart failure treatments, procedures that restore blood flow, chemotherapy, and surgical risk. Clinicians trace the chamber’s outline on an MRI image by hand, a process that can vary from one doctor to the next. Earlier research found software measures more consistently, so researchers tested whether its numbers also do a better job of flagging patients at higher risk of dying.
Published in The Lancet Digital Health by researchers from University College London and other institutions, the results amounted to a cautious yes.
AI Heart Scan Numbers Tracked Survival Slightly Better
Records covered 27,382 patients who had a heart MRI. Most, 26,106, were referred from nine hospitals to Barts Heart Centre in London between late 2014 and 2019. Another 1,276 were scanned at the University of Pittsburgh Medical Center between 2010 and 2016. Typical patient age was 56, and 62% were men. Each patient’s first scan counted, including images blurred by breathing or irregular heartbeats, as in a busy real-world clinic.
An AI program re-measured each scan after training on other patients’ images, never seeing data from these two centers. Doctor measurements came from original clinical reports by accredited senior clinicians with more than five years of experience.
Deaths were tracked through national records in the UK and hospital records in Pittsburgh. Follow-up typically lasted about five and a half years, and 3,458 patients, or 13%, died. The AI’s numbers did a modestly better job of separating patients who survived from those who died. Age and sex are strong predictors on their own, and once they were factored in, the gap shrank to a sliver.
On average, the AI recorded an ejection fraction about 4 percentage points lower than doctors did. Starting with the doctor’s number and adding the AI’s improved the match with who actually survived substantially. Starting with the AI’s number and adding the doctor’s helped only slightly. That pattern fits with the AI capturing most of what doctors’ measurements revealed about survival, plus some extra information, though that remains an interpretation of the statistics. The edge also held after accounting for which clinician read each scan.
AI Erred Far More Often on Extremely Low Readings
Reviewers checked the AI’s tracings for mistakes likely to throw off a measurement. Among 600 randomly chosen patients, 1% had one. Among scans where AI and doctor numbers disagreed widely, about 2% did. Among scans with extreme readings, 23% did, and every mistake involved an extremely low reading.
Doctors had an odd habit of their own. Recorded ejection fractions clustered at exactly 60%, which the researchers say may reflect rounding to a medically meaningful cutoff. The AI showed no such habit.
Doctors Outperformed AI With Thickened Heart Muscle
Researchers also split patients by diagnosis. Among nearly 5,000 patients with blocked heart arteries, the AI’s number tracked survival better. Among the roughly 3,500 who had blood-flow testing, the edge held after accounting for doctors’ findings on blood-flow problems and damaged heart tissue, though it faded once age and sex were added too. Among about 2,200 patients with weakened heart muscle unrelated to blocked arteries, the two sources came out nearly even.
Doctors came out ahead among 1,611 patients with hypertrophic cardiomyopathy, a disease that thickens the heart muscle. Their numbers followed survival more closely there, and the AI added nothing meaningful. Every measurement, from either source, predicted survival poorly in that group. The authors point to one possible reason: the AI’s tracing method counts small muscle ridges as part of the blood-filled chamber, which could miss features of the disease. More broadly, doctors may have been swayed by patient history and other test results the AI never saw.
Across all patients, combining AI and doctor measurements matched survival better than either alone, hinting that each catches something the other misses.
AI Heart Scan Speed Still Needs Doctor Oversight
Even small gains, the authors argue, could matter at the scale of a whole population of patients. In their words, “AI offers much higher throughput thereby reducing clinician workload and removing a substantial pressure on clinical services.”
One limit matters most. This was a retrospective study, looking back at existing scans and records. Nobody changed treatment based on the AI’s numbers, so the work shows better prediction, not that AI-guided care helps patients live longer.
AI that reads heart scans in seconds deserves a trial run as a first-pass reader, with a clinician checking the work and forward-looking studies tracking whether patients benefit.
Disclaimer: This article summarizes published research and is not medical advice. Consult a qualified health professional about heart health or treatment.
Paper Notes
Limitations
This was a retrospective study of patients already referred for cardiac MRI, so findings need confirmation in prospective research before AI replaces clinician measurement in practice. The study measured how well each source’s numbers tracked mortality, not whether using AI changes treatment or survival. It had no data on cause of death, only whether and when patients died. Disease categories came from clinical reports rather than independent review. At the London site, a language-model tool sorted diagnoses with a small error rate (under 2%), while Pittsburgh diagnoses were coded manually by clinicians without duplication or blinding. Deaths were tracked through a national registry in the UK but through local hospital records in the US, so some US deaths outside that health system may have gone uncounted. Performance by ethnicity was not assessed because the information was not routinely recorded. The authors assessed model fit and discrimination but not calibration, did not evaluate clinical utility, and applied no hierarchical control for multiple comparisons. Differences in how clinicians traced the heart could also have contributed to the AI’s advantage, though an analysis adjusting for clinician technique found the AI’s edge remained.
Funding and Disclosures
Funding came from a National Institute for Health and Care Research Clinical Lectureship and a British Heart Foundation Accelerator Award. The funders had no role in study design, data collection, analysis, interpretation, or writing of the report. Several authors have financial ties to MyCardium AI, which commercializes the AI software (1CMR) used in the study through an exclusive license from University College London. One author is the company’s Chief Executive Officer and another is a Director. Seven authors own company shares, and eight are paid consultants. The AI image analysis itself was performed by one of the shareholding authors. Other authors disclosed research fellowships, consulting or speaker fees, or grants from health research bodies and pharmaceutical or medical device companies, and one reported a pending patent on an automated cardiac MRI planning method, as detailed in the paper’s declaration of interests.
Publication Details
Title: “Performance of artificial intelligence versus clinician left-ventricular segmentation of cardiac MRI for mortality prediction: a retrospective, multicentre, prognostic study.” Authors: George Joy, Karan Punjabi, George Searle, Zaid Abdulelah, Tuguldur Gantur, Iain Pierce, George D Thornton, Benjamin Meredith, Hunain Shiwani, Miles Curtis, Shane Cashin, João B Augusto, Inês Miranda, Marianna Fontana, Peter Kellman, James P Howard, Matthew Shun-Shin, Thomas A Treibel, Charlotte Manisty, Michele Orini, Erik B Schelbert, Alun D Hughes, James C Moon, and Rhodri H Davies. Corresponding author: Dr. Rhodri H Davies, Institute of Cardiovascular Science, University College London. Journal: The Lancet Digital Health, published online Nov. 2, 2026. DOI: https://doi.org/10.1016/j.landig.2026.101042
