Can an AI-read ECG detect low ejection fraction? Yes. A Mayo Clinic model scored an AUC of 0.93 for ejection fraction of 35% or less, and the EAGLE randomised trial raised new low-EF diagnoses in primary care by about a third. The unresolved risk is cost: every positive result can trigger an echocardiogram, and possibly years of follow-up scans.
The electrocardiogram is probably the most ordered heart test in the world. It is cheap, fast and done for all sorts of reasons: before surgery, at an annual check, in the emergency department, during a workup for chest pain that turns out to be indigestion. For more than a century, doctors have read it for rhythm, conduction and signs of a heart attack. What they could not read from it was how well the heart pumps.
AI can. Trained on paired ECGs and echocardiograms, neural networks can flag patients whose left ventricle is pumping weakly, a condition that is common, often silent and treatable. This is one of the most promising ideas in clinical AI because it asks for nothing new. No new device, no new appointment, no new sample. Just a second look at a test that was happening anyway.
That is exactly why it deserves careful thought. Opportunistic screening on existing tests is AI’s cheapest win in medicine. It is also a way to start a screening programme that nobody designed, budgeted or agreed to, one positive result at a time.
How accurate is AI ECG for detecting low ejection fraction?
The foundational study came from the Mayo Clinic. Published in Nature Medicine in January 2019 by Zachi Attia and colleagues, it trained a convolutional neural network on paired 12-lead ECG and echocardiogram data from 44,959 patients to identify an ejection fraction of 35% or less. Tested on an independent group of 52,870 patients, the model achieved an area under the curve of 0.93, with sensitivity of 86.3% and specificity of 85.7%.
The paper noted that asymptomatic left ventricular dysfunction affects an estimated 3–6% of the general population. It also reported something that matters a great deal for the cost argument. Among patients whose echocardiogram showed normal function, those with a positive AI screen were four times as likely to develop ventricular dysfunction later. Some of the “false positives” were early true positives.
Accuracy is only the first question. The second is whether putting the result in front of busy clinicians changes anything.
Does AI ECG screening actually find more heart failure?
It does, according to a randomised trial that is still one of the better-designed studies in clinical AI.
The EAGLE trial, published in Nature Medicine in May 2021 by Xiaoxi Yao and colleagues at Mayo Clinic, cluster-randomised 120 primary care teams from 45 clinics and hospitals. Half had access to AI-ECG results; half provided usual care. ECGs came from routine care for 22,641 adults without known heart failure. The primary outcome was a new diagnosis of ejection fraction of 50% or less within 90 days.
The intervention raised new low-EF diagnoses from 1.6% to 2.1% of the whole cohort, an odds ratio of 1.32. Among the 6% of patients the AI flagged as high likelihood, diagnoses rose from 14.5% to 19.5%.
The most interesting number is not in the headline. Overall echocardiogram use was similar in the two arms: 18.2% in control and 19.2% with AI, a difference that was not statistically significant. Among AI-positive patients, echo use rose from 38.1% to 49.6%. The AI did not flood the echo lab. It redirected scans towards the patients most likely to have disease.
A second pragmatic trial pointed the same way. Published in BMC Medicine in June 2025 by a team in Taiwan, it randomised 13,631 inpatients under non-cardiologist care. AI-ECG results raised new low-EF diagnoses within 30 days from 1.1% to 1.5%. Echocardiogram use was essentially unchanged, at about 17% in both groups, but the share of echoes that found low EF rose from 20.2% to 34.2%.
What are the downstream costs of AI ECG screening?
So far, the trials suggest AI-ECG makes echo use smarter rather than bigger. That is encouraging, and it is worth being precise about why it happened.
In both trials, clinicians received a flag and used judgment. Many AI-positive patients did not get an echo, perhaps because they were already known to have other explanations, because they were frail, or because the clinician did not find the flag persuasive. The trials measured AI plus clinician discretion inside organisations that were paying attention. They did not test what happens when health systems run the algorithm on every ECG, automatically, with a protocol that says positive means echo.
Three pressures push towards more testing once AI-ECG leaves the trial setting.
Defensive medicine. Once an abnormal AI result sits in the record, ignoring it carries medico-legal risk. A flag that was optional in a trial becomes, in practice, an order.
The “not yet” positives. Attia’s finding that screen-positive patients with normal echoes are at four times the risk of future dysfunction is scientifically exciting and financially awkward. It invites surveillance echoes for people who are currently healthy. Whether that surveillance improves outcomes has not been established.
Stacked models. Low ejection fraction is only the first of many conditions researchers have trained ECG models to detect. If a health system runs several such models on every ECG, each with a modest positive rate, the share of patients flagged by at least one grows quickly.
A worked example: one health system, one year of ECGs
Hypothetical, with illustrative numbers. A regional health system performs 200,000 ECGs a year in adults without known heart failure. It turns on an AI low-EF model that flags 6% of them, similar to the EAGLE proportion. That is 12,000 positive results.
If clinicians behave as in EAGLE, around half of those patients get an echocardiogram, and roughly one in five is eventually diagnosed with low EF. That means about 6,000 echoes, many of which would have happened anyway, a meaningful number of new diagnoses, and patients started on medicines that are known to help. That is a good trade.
Now change two assumptions. First, the system adopts a protocol under which every positive result gets an echo. Echo volume from this source roughly doubles. Second, it decides that AI-positive patients with normal echoes should have repeat imaging because of their higher future risk. Several thousand people a year join a surveillance list, and the list grows annually because few people leave it.
Then add a third choice. The system licenses five ECG models for different conditions, each flagging 5% of patients independently. The chance a patient is flagged by at least one is about 23%. Nearly a quarter of everyone who has an ECG now has a follow-up question attached.
None of these choices is irrational in isolation. Together they create exactly what we described in wearables data as a screening programme nobody designed: a screening programme with no eligibility rules, no capacity planning and no evaluation.
Isn’t finding more heart failure worth the extra scans?
The strongest counterargument is simple. Heart failure with reduced ejection fraction is serious, common and treatable. Guideline-directed medicines reduce hospitalisation and death. Many patients are diagnosed late, after an emergency admission. If an AI-read ECG can find them earlier using a test that has already been paid for, the cost of some extra echoes is small next to the cost of a heart failure admission. Worrying about over-testing, on this view, is a luxury that patients with undiagnosed disease cannot afford.
This argument is largely right for low ejection fraction specifically. It is a condition with an effective treatment pathway, which is the most basic requirement of any screening programme and one that many “detect it early” ideas fail, as we argued in a test without a treatment is just bad news. The EAGLE and Taiwan trials also show that the feared flood of echoes did not happen when clinicians retained judgment.
Where the counterargument weakens is in extrapolation: from low EF to every condition an ECG model can detect, from clinician discretion to automatic protocols, and from diagnosis rates to outcomes. Neither trial measured hospitalisations or deaths. The case for AI-ECG is strong enough to deploy. It is not yet strong enough to deploy without rules.
What this means for health systems, clinicians and payers
For health system leaders, the key decision is to treat AI-ECG as a screening programme and govern it like one. Decide who is eligible, which conditions are screened, what a positive result triggers, how echo capacity will absorb it and how outcomes will be measured. Start with low EF, where the evidence is best, before adding models.
For clinicians, the trials support using AI-ECG results as a nudge, weighed against what you know about the patient. Documenting why a positive flag did not lead to an echo, when that is the right call, protects both patient and clinician from the drift towards automatic testing.
For cardiology and imaging departments, the realistic risk is not a flood but a steady rise in referrals from sources you do not control. Agree referral criteria in advance with primary care and hospital medicine.
For payers, AI-ECG is an opportunity to pay for a pathway rather than a test: the algorithm, the confirmatory echo and the start of treatment together, with a requirement to report yield. Paying for the flag alone rewards volume.
For patients, a positive AI-ECG result is a reason to talk to a clinician about next steps, not a diagnosis. An echocardiogram is what confirms whether the heart’s pumping function is reduced.
Frequently asked questions
What is an AI ECG?
An AI ECG is a standard electrocardiogram analysed by a machine-learning model that looks for patterns humans cannot reliably see. The best-studied use is detecting low left ventricular ejection fraction, meaning the heart pumps weakly. Models have also been developed for other conditions, though evidence for most of them is less mature.
What is ejection fraction?
Ejection fraction is the percentage of blood the left ventricle pumps out with each beat. A normal value is usually around 50–70%. A reduced ejection fraction can indicate heart failure or heart muscle disease, often before symptoms become obvious. It is normally measured with an echocardiogram, an ultrasound scan of the heart.
How accurate is AI at detecting heart failure from an ECG?
In the Mayo Clinic’s 2019 Nature Medicine study, the AI model identified ejection fraction of 35% or less with an area under the curve of 0.93, sensitivity of 86.3% and specificity of 85.7%. Accuracy varies with the population and threshold, and a positive result needs confirmation by echocardiogram before any diagnosis is made.
Does AI ECG screening increase echocardiogram use?
In the trials so far, not overall. In the EAGLE trial, echocardiogram use was 19.2% with AI versus 18.2% without, a non-significant difference, while more AI-positive patients were scanned. A 2025 inpatient trial in Taiwan found similar echo use in both arms but a higher yield. Automatic follow-up protocols outside trials could change that.
Should I ask my doctor for an AI ECG?
AI-ECG tools are available in some health systems, often applied behind the scenes to ECGs ordered for other reasons. If you have symptoms such as breathlessness, swelling or fatigue, or risk factors for heart disease, the right conversation is with your clinician about whether an ECG or echocardiogram is appropriate for you.
AI-ECG is the rare clinical AI that finds treatable disease, in a test already done, with trial evidence behind it. The risk is not that it fails. It is that it succeeds so quietly that health systems build a screening programme without ever deciding to.
Related on the GHC blog: Primary Care Is the Cheapest Hospital Bed, But Only If You Close One and Paying for Outcomes Rents Behaviour. It Doesn’t Buy It..
Is your health system running AI models on routine ECGs, and does anyone own what happens after a positive result?
