AI-ECG Strategy Identifies Patients at Risk for AF

Artificial intelligence (AI)-guided screening for atrial fibrillation (AF) based on ECGs can successfully identify patients who are more likely to have the arrhythmia detected on subsequent monitoring, according to one of the first prospective studies to evaluate such an approach.

Patients tagged by the AI-ECG algorithm as high versus low risk were much more likely to have an AF episode lasting at least 30 seconds identified on subsequent ambulatory monitoring (7.6% vs 1.6%; OR 4.98; 95% CI 2.22-11.75), with similar results when longer-duration AF episodes were used, researchers led by Peter Noseworthy, MD (Mayo Clinic, Rochester, MN), report.

The study, published online this week in the Lancet, builds on the group’s prior work in leveraging millions of ECGs stored at the Mayo Clinic to derive AI-ECG algorithms for various disease states like long QT syndrome and COVID-19. What makes this study unique is its prospective design.

“It’s one thing to say that we can distinguish patients who seem to have a propensity toward atrial fibrillation from those who don’t on the 12-lead ECG. It’s another to say that we can identify high-risk patients and they actually have a higher rate of incident atrial fibrillation diagnosed on prospective cardiac monitoring and follow-up,” Noseworthy told TCTMD.

The study was designed to evaluate the AI-ECG approach as it would be used if ultimately adopted into clinical practice, he said, noting that the participants had previously had ECGs taken for many reasons not necessarily related to AF and were then contacted to take part in screening with remote monitoring.

“We were able to take a population of patients who have existing cardiovascular risk factors, who are probably already at comparatively high risk for atrial fibrillation, and then demonstrate that we can further stratify that population using the ECG,” Noseworthy explained. “So not only is the ECG a marker for these comorbidities that drive AF risk, but also it provides additional prognostic benefit on top of existing clinical risk factors. That means we can identify patients who have risk factors for stroke were they to be diagnosed with atrial fibrillation and target the screening to those who we expect would benefit most.”

We can identify patients who have risk factors for stroke were they to be diagnosed with atrial fibrillation and target the screening to those who we expect would benefit most. Peter Noseworthy

Noseworthy and his Mayo Clinic colleagues reported about 3 years ago that their AI algorithm could identify signs of AF on ECGs taken during normal sinus rhythm, potentially finding a subset of the population that should be targeted for AF screening. But it had remained unclear whether the AI-ECG approach improved risk stratification over clinical factors alone and whether it uncovered AF cases that wouldn’t otherwise have been found through usual clinical practice.

To address those questions in the current study, called Batch Enrollment for an AI-Guided Intervention to Lower Neurologic Events in Patients with Undiagnosed Atrial Fibrillation (BEAGLE), Noseworthy and colleagues prospectively recruited 1,003 patients (mean age 74 years; 38.2% women) who had stroke risk factors but no known AF and who had a 12-lead ECG obtained in routine practice.

The AI algorithm divided participants, who came from 40 US states, into high- and low-risk groups based on the ECGs. All received a continuous ambulatory heart rhythm monitor, which was worn for up to 30 days (mean 22.3 days).

Monitoring showed that patients identified as having a high risk of AF were indeed more likely to have the arrhythmia than those classified as having a low risk. The high-risk risk group also had a significantly greater AF burden (mean 20.32% vs 4.97%; P = 0.016), with no differences between risk groups in terms of the longest AF episode or time to AF diagnosis. About three-quarters of patients who had newly diagnosed AF and available clinical follow-up data started anticoagulation.

In a secondary analysis, the investigators evaluated whether the AI-ECG strategy increased detection of AF over usual care through a median follow-up of 9.9 months. And it did, but only for the high-risk group (10.6% vs 3.6%; P < 0.0001).

“To our knowledge, our study is . . . the first to evaluate the effectiveness of the AI-guided targeted screening program in comparison with usual care, which can inform the design and implementation of atrial fibrillation screening programs at scale,” Noseworthy et al write.

Some Limitations, Clinical Outlook

Commenting for TCTMD, Jagmeet Singh, MD, PhD (Massachusetts General Hospital, Boston), said that as one of the first prospective studies evaluating risk stratification with AI-ECG, “it does add a lot” to the literature on AF screening.

The algorithm “helps risk stratify a relatively high-risk population for stroke and then helps quantify whether that risk stratification was appropriate or not,” Singh said. “It also gives an idea of which patients to really focus on. Because A-fib is such a common arrhythmia that is growing in prevalence, to be able to have scalable risk-stratification strategy you need something that’s relatively easy to do and is done fairly routinely.”

Singh had some caveats when it came to interpreting the study, however. Importantly, there is a question of generalizability, as the AI-ECG algorithm was derived from a Mayo Clinic population that includes a very high proportion of white individuals; in this study, fewer than 4% of participants were nonwhite. There’s also a question about the impact of digital literacy on interpretation of the results, he said, pointing to the fact that the study included 1,003 people out of more than 15,000 initially invited.

“I think AI-based algorithms need to be consumed with some degree of caution because of the data set they come from and the population that they’re eventually being studied in,” Singh said.

The researchers conducted an “elegant and pragmatic” study, but a randomized study with a larger cohort and longer follow-up would help address some of these issues, he said, adding that studies evaluating longer-term monitoring strategies beyond 30 days would be useful to ensure that AF isn’t being missed. “You want to really know if they may end up getting A-fib over the next several months or years,” Singh said.

As a next step before clinical implementation of the AI-ECG approach, and approval from the US Food and Drug Administration, Noseworthy said there is a need for a large, multicenter trial showing that patients diagnosed with AF with screening can be anticoagulated or treated in some other way to reduce their risk for stroke. “Ideally, we would like to get this algorithm in the hands of as many people as possible to execute these kinds of screening programs once we demonstrate that it makes a difference in patients’ lives and their stroke risk,” he said.

Conducting that type of trial will be challenging from a financial perspective, Singh said, but he suggested that a strategy like that studied here, perhaps combined with others, could have a clinical role in the future.

“We should be adopting machine-learning approaches for looking at clinical covariates more elegantly for predicting risk of A-fib. Over and above that, adding the AI-ECG strategy that the Mayo Clinic has may even further add value to the algorithm to really risk stratify patients even better,” Singh said.

The AI-ECG algorithm “actually creates an opportunity for providing remote monitoring for these patients, using not just wearables or patch monitors but even implantables,” he suggested, indicating the importance of “monitoring them for longer periods of time so you can pick up A-fib and in turn also show the impact on clinical outcomes.”