AI Readings From Cardiac MRI Matched Survival Risk Better Than Clinician Measurements

Artificial intelligence is edging closer to routine use in cardiac imaging, and a large new study gives that shift real clinical weight. In a cohort of 27,382 patients undergoing evaluation for known or suspected heart disease in the U.K. and the United States, AI-derived ejection fraction measurements were slightly, but significantly, better than clinician measurements at identifying patients at increased risk of death. The difference was not dramatic, but it was consistent after adjustment for age and sex, and it appeared across a large enough population to matter in everyday imaging services.

Ejection fraction remains one of the most heavily used measurements in cardiology because it captures how forcefully the left ventricle is pumping. A low value can influence decisions about heart failure therapy, implantable devices, surgical risk, and the timing of other interventions. When a measurement sits at the center of so many decisions, even a small increase in precision can shift who gets flagged early and who slips through with a falsely reassuring reading.

One sentence captures the practical consequence: AI measured the same scan in seconds.

Mortality Prediction Improved Even Though the Gain Was Modest

Researchers from City St George’s, University of London, and University College London evaluated cardiac MRI scans from a broad clinical population, then followed participants for an average of more than 5 years. During that period, 3,458 deaths occurred. The study then compared AI-generated ejection fraction values with those produced by experienced clinicians interpreting the same imaging data.

The AI system performed slightly better at identifying those at elevated risk of death. In other words, its measurement of heart function carried a somewhat stronger association with subsequent survival than the clinician-derived measurement did. The finding remained statistically significant after accounting for age and sex, which strengthens confidence that the signal was not simply a reflection of demographic differences in the cohort.

George Joy, MBBS, MRCP, the study’s lead author and an NIHR Clinical Lecturer at City St George’s, described the improvement as small but potentially meaningful when scaled across large health systems. That framing fits the data well. Cardiology rarely depends on a single large leap in accuracy; more often, it advances through repeated small gains in triage, interpretation, and workflow that accumulate across thousands of patients.

Speed May Be the Strongest Argument for Clinical Adoption

The performance advantage was paired with a major operational advantage. The AI tool analyzed each cardiac MRI scan in about 2 to 3 seconds per patient, compared with roughly 13 minutes for expert clinician analysis in prior work from the same group. In a service already strained by growing imaging demand, that kind of speed difference changes what a department can realistically process in a day.

Rapid analysis matters because cardiac MRI is not a niche test in modern practice. It is used to characterize ventricular function, evaluate suspected cardiomyopathy, support ischemic assessment, and help clinicians decide whether patients may benefit from medication or devices. A faster readout could reduce bottlenecks between scan acquisition and the moment the result becomes clinically usable, especially when an MRI suite is moving through a high-volume list.

Yet speed alone does not equal better care. A faster calculation only becomes clinically valuable if the output is accurate, interpretable, and embedded into a workflow that still preserves human oversight. That caveat sits at the center of this study’s implications.

The Error Rate Was Low, But Not Negligible

The AI system showed an approximate 1% patient-level error rate. In practical terms, that means misclassification was uncommon, but it was not absent. When errors occurred, ejection fraction was either overestimated or underestimated, and both directions can alter decision-making. An overestimated value may falsely reassure clinicians; an underestimated one can push a patient toward unnecessary concern or downstream testing.

Because ejection fraction is tied to treatment thresholds, even a small error can matter if it lands near a cutoff. For that reason, the study supports a model in which AI performs the first-pass measurement and clinicians verify the result rather than abandoning review altogether. Joy said the goal is routine measuring by AI with physician review when needed, a division of labor that reflects how many medical systems already use automation for triage, quantification, and quality control.

Cardiac MRI interpretation is broader than one number. Symptoms, prior cardiac history, other imaging features, laboratory data, and additional testing still shape the final clinical picture. A single AI output cannot replace the synthesis that experienced clinicians bring to a patient’s overall condition.

The Result Supports First-Line AI, Not Autonomous Decision-Making

The strongest argument for the technology is not that it should replace clinicians, but that it may free them from repetitive manual measurement so they can spend more time on interpretation and complex cases. In a field where imaging volume keeps rising, that redistribution of effort could be valuable. It may also reduce variation between readers and standardize measurements across serial scans, which is particularly important when subtle changes over time affect treatment.

Still, the study does not show that introducing AI into routine practice will reduce deaths or improve symptoms. The endpoint was association between AI-derived measurements and later mortality, not a randomized test of clinical adoption. Real-world implementation will need to answer harder questions: whether the tool performs equally well across different scanners and institutions, whether it behaves the same in different patient groups, and whether an AI-assisted workflow changes treatment timing or outcomes in a measurable way.

Combining Imaging With ECG Data Could Extend Detection Earlier

The investigators are already thinking beyond ejection fraction alone. Their next step is to examine whether AI-based measurements from cardiac MRI can be combined with information from other tests, including the electrocardiogram. Pairing structural imaging with electrical data could help detect abnormalities in the heart muscle before damage becomes more advanced.

That approach makes biological sense. MRI quantifies pump performance and chamber structure, while ECG captures conduction and rhythm abnormalities that may appear earlier in the disease course or alongside evolving myocardial injury. When both are abnormal, the probability of clinically meaningful disease rises. When they diverge, clinicians may be able to investigate more selectively and earlier than they could with either test alone.

The study leaves one central question unresolved: how much better, in actual bedside practice, is an AI-first workflow than the current manual standard once implementation, oversight, and patient outcomes are all taken into account? The answer will depend less on the algorithm’s speed than on whether health systems can integrate it without losing the careful clinical judgment that still determines treatment.


Reporting by Dr. Aishwarya N., MBBS for Medicine World Foundation, with additional data provided by Medical News Today.