AI Analyzes Cardiac MRI Scans in 3 Seconds, Slightly Outperforming Doctors at Predicting Patient Death Risk

Artificial intelligence can measure a key indicator of heart function from MRI scans in roughly 2 to 3 seconds and produce results that are slightly but significantly better than those of experienced clinicians at predicting which patients face an elevated risk of death, according to a large retrospective study published in The Lancet Digital Health.

The research, led by investigators from City St George's, University of London, and University College London, examined cardiac MRI scans from 27,382 consecutive patients in the United Kingdom and the United States who were being evaluated for known or suspected heart disease. Patients were followed for a median of 5.6 years, during which 3,458 deaths occurred. The study compared AI-generated measurements of left-ventricular ejection fraction with measurements recorded by clinicians in routine clinical reports.

Ejection fraction describes the percentage of blood the left ventricle pumps out with each beat. It is one of the most widely used metrics in cardiology, serving as a gatekeeper for decisions about heart-failure therapies, implantable devices, revascularization procedures, surgical risk assessment, and monitoring of patients receiving chemotherapy. Small differences in how the measurement is obtained or reported can influence treatment pathways.

Researchers found that AI-derived ejection fraction showed better statistical model fit and marginally better discrimination for all-cause mortality than clinician-derived values. The advantage remained after adjustment for age and sex. When AI measurements were added to clinician measurements, model performance improved substantially; adding clinician measurements to the AI values produced only a slight further improvement. Similar patterns appeared in a subgroup of patients with obstructive coronary artery disease.

The AI tool completed its analysis in approximately 2 to 3 seconds per patient. Earlier work by the same research group had found that expert clinician analysis of a cardiac MRI typically requires about 13 minutes. The software used in the study is a deep-learning system previously developed to segment the left ventricle across different scanner vendors and clinical conditions. It was applied without post-hoc manual corrections.

"Ejection fraction is one of the most important measurements we make in cardiology because it helps determine how patients are treated and how closely they should be monitored," said Dr. George Joy, an NIHR Clinical Lecturer in the Cardiovascular Clinical Academic Group at City St George's and lead author of the study. "Our findings suggest AI can measure this important marker with remarkable consistency and precision, resulting in a small but meaningful improvement in identifying patients at higher risk."

Senior author Dr. Rhodri Davies, an associate professor at University College London and inventor of the AI tool, emphasized the practical implications at scale. "Even modest improvements in accuracy can have a significant impact when applied across thousands of patients," Davies said. "AI is not replacing clinicians, but it can provide faster, more consistent measurements at low-cost and support more streamlined clinical decision-making that ultimately benefits patients."

The researchers reported that AI errors were uncommon. An error rate of about 1% was observed at the patient level, affecting 0.08% of images. They stressed that clinicians should continue to verify results and interpret scans in the full clinical context. The paper's interpretation section states that the findings support the potential utility of AI as a first-line method for measuring ejection fraction, provided clinician oversight is maintained, while noting that the approach should be tested in prospective studies.

The analysis was retrospective and drew on existing clinical reports rather than a standardized re-reading of every scan by multiple observers under controlled conditions. Clinician measurements in routine practice can vary because of differences in technique, contouring conventions (such as whether trabeculations and papillary muscles are included in the blood pool), and digit preferences around common thresholds. The AI system produced systematically lower ejection-fraction values on average than the original reports, a difference the authors attribute in part to more consistent handling of these anatomical details.

Cardiac MRI remains the reference standard for measuring ventricular volumes and function, but its wider use is constrained by the time and specialized expertise required for image analysis. In the United Kingdom alone, roughly 120,000 heart MRI scans are performed each year. Health systems in both the UK and the United States face growing demand for cardiac imaging amid workforce pressures. Automated tools that reduce the time spent on quantitative measurements could free clinician capacity for image interpretation, patient consultation, and decision-making, the researchers suggested.

The study was funded by an NIHR Clinical Lectureship, the NIHR University College London Hospitals Biomedical Research Centre, and a British Heart Foundation Accelerator Award. It builds on earlier work by members of the same group demonstrating that AI can analyze cardiac MRI with high precision while patients are still in the scanner. Those earlier efforts focused on speed and measurement reproducibility; the current paper examines the downstream relationship between those measurements and long-term survival.

The authors noted that the prognostic advantage, while statistically significant, was modest. Discrimination metrics improved only slightly, and the clinical importance of the difference would depend on how the information is used in practice—particularly near treatment thresholds. They also observed that AI measurements improved model performance more than clinician measurements did when both were available, suggesting that the automated values capture additional prognostic information in this dataset.

Limitations include the retrospective design, reliance on historical clinical reports rather than blinded re-analysis, and the absence of a prospective evaluation of whether using AI as the first-line measurement actually changes patient management or outcomes. The cohort was drawn from two high-volume centers, one in London and one in Pittsburgh, and may not fully represent practice in all settings. Complex congenital heart disease was excluded. The paper calls for prospective studies to determine whether the observed statistical advantage translates into measurable clinical benefit.

The team has indicated plans to combine AI-derived measurements from cardiac MRI with data from other tests, particularly electrocardiograms, in an effort to identify abnormalities earlier. For now, the published findings add to a growing body of evidence that automated analysis of cardiac MRI can match or exceed manual methods on certain quantitative tasks while operating at a fraction of the time. Whether that capability is adopted as a routine first step, with clinicians retaining responsibility for verification and clinical correlation, will depend on further validation, regulatory pathways, and local workflow design.

As imaging volumes continue to rise and specialist capacity remains constrained, tools that deliver consistent, rapid measurements of core metrics such as ejection fraction are likely to attract continued scrutiny. The current study supplies one of the larger real-world comparisons to date linking AI versus clinician cardiac MRI measurements to subsequent mortality, while underscoring that speed and modest gains in prognostic performance do not eliminate the need for human oversight.