AI analysis of routine ECGs led to more diagnoses of weak heart pumping

For every 1,000 patients in the EAGLE trial, giving care teams an AI interpretation of routine ECG data produced about five additional diagnoses of low ejection fraction. The result, reported by Mayo Clinic in May 2021, showed how a screening signal can change a real clinical pathway: identify a possible problem, prompt a heart ultrasound and establish whether pumping function is reduced.[2]

Ejection fraction describes the proportion of blood pumped out with each contraction. An echocardiogram uses ultrasound to assess the heart, including its pumping function. An ECG records electrical activity. The AI approach uses patterns in that electrical recording to identify patients more likely to need the imaging examination, making an existing test a route to a different kind of information.[3]

Randomising access to the result

EAGLE included 22,641 adults without previously diagnosed heart failure across 45 clinics and hospitals. The researchers randomized 120 primary-care teams, rather than individual patients, to access the AI results or continue usual care. The primary endpoint was a new diagnosis of ejection fraction at or below 50% within 90 days of an ECG obtained during routine care.[1]

The report used a simple action format: a negative result required no additional action under the study protocol; a positive result suggested considering an echocardiogram. Its design had been refined through clinician interviews and workflow observations over four months. The research team tested versions of the report with practising clinicians, including a final exercise using real patient information.[4]

The pattern is more informative than a single accuracy score. The AI signal concentrated additional imaging among people it had flagged. Most participants were not flagged, so a substantial change inside that smaller group coexisted with a relatively small change in imaging across the whole trial.

The message has to arrive in usable form

A companion qualitative study interviewed 29 providers from teams with high and low uptake. It reveals an implementation detail obscured by the phrase “integrated into the health record”: the report was available in the record, but positive-result alerts arrived by email. Reminders followed when clinicians had neither ordered imaging nor documented a reason to decline.[5]

Some clinicians found repeated emails confusing, particularly when they could not tell which results had been resolved. Alerts were routed to primary care even when another service had ordered the ECG, creating uncertainty about responsibility. Results arriving outside an appointment also created extra calls and coordination. These were reported experiences from the selected interview sample, rather than a count of failures across every clinic.[5]

Those observations make the implementation opportunity concrete: show the result with the relevant patient record, make the next action clear and identify the clinician responsible for closing it. The prediction becomes useful when that person can connect it to what is already happening in the patient’s care.

The commercial device has its own clinical scope

Anumana’s later ECG-AI Low Ejection Fraction 12-Lead Algorithm has a narrower target than EAGLE’s primary endpoint. The FDA clearance dated 28 July 2025 describes software intended to help detect an ejection fraction of 40% or below in adults at risk of heart failure. It analyses ten seconds of a standard 12-lead ECG and supplies a result to clinical software.[6]

The cleared indication requires use alongside clinical judgement. A positive result can prompt further assessment; the software is not a standalone cardiac diagnosis or a monitoring device, and paced rhythms are outside its intended use. Those details prevent the trial’s broad diagnostic endpoint from being mistaken for the specification of a current product.[6]

EAGLE’s demonstrated gain was additional diagnosis. Its lasting contribution is showing the connection from an ordinary electrical recording to a clinician’s decision and a confirmed finding. Whether earlier identification reduces later admissions or deaths needs further outcome evidence.[2]

Report an error or suggest a correction