AI heart-scan measurements predict mortality slightly better than clinicians’ in study

An AI tool’s measurements of heart function predicted patients’ risk of death slightly better than measurements by experienced clinicians in a study of 27,382 people. The finding supports a possible first-line role for automated heart-scan analysis, the researchers say, while leaving doctors responsible for checking the results and interpreting what they mean for care.

A treatment measure, judged against survival

The research, led by City St George’s, University of London and University College London, was published in The Lancet Digital Health. It compared AI-generated measurements from cardiac MRI scans with those made by experienced clinicians. The test was not simply whether the two methods produced similar numbers, but how strongly those numbers were linked to subsequent mortality.

The measurement was ejection fraction, a measure of how effectively the heart pumps blood. Doctors use it to help decide whether patients may benefit from specialist heart-failure treatments, implanted heart devices or procedures to improve blood flow. It also helps clinicians assess surgical risk and monitor heart function during chemotherapy. That makes it a number with consequences beyond the scan report.

The retrospective, multicentre study examined patients in the UK and United States who were being investigated for known or suspected heart conditions. Rather than testing AI-directed treatment, the researchers compared measurements against patients’ later survival. This separates the study’s central result—a stronger relationship with mortality—from the question of whether changing clinical practice would improve patient outcomes.

A small predictive edge, a larger speed gap

AI-derived ejection-fraction measurements were slightly but statistically significantly better at identifying patients at increased risk of death. The difference remained significant after accounting for age and sex. Lead author Dr. George Joy described the improvement as small but meaningful, attributing it to the consistency and precision with which AI measured the marker.

The processing-time contrast was much larger. The AI analyzed scans in approximately two to three seconds per patient. Expert clinician analysis typically took around 13 minutes in previous work by the same team. That distinction matters: the clinician timing came from an earlier study, not a fresh side-by-side timing test reported alongside the mortality comparison.

Fast measurements still need clinical judgment

Errors were uncommon, but not absent. The researchers reported an AI error rate of around 1% at the patient level, affecting 0.08% of images. Those figures describe different levels of analysis: the smaller image-level figure does not mean that only 0.08% of patients encountered an error. The team stressed that human oversight remains essential.

Their proposed division of work puts AI first in measurement, with clinicians verifying results and interpreting scans in the wider clinical context. Senior author Dr. Rhodri Davies, who invented the tool, argued that faster, more consistent measurements could reduce costs and streamline decisions. He framed the technology as support for clinicians, not a replacement for them.

The next research step would add information rather than simply accelerate the same task. The team plans to combine AI measurements from heart scans with measurements from other tests, particularly electrocardiograms, or ECGs. Its stated goal is to identify heart-muscle problems earlier, allowing treatment before lasting damage occurs. That remains a research plan, not a demonstrated result of this study.