FDA Clears AI-ECG Screening Tools for CV Care: What’s Next Is Up for Grabs

Even a decade ago, the notion of software as medical device (SAMD) sounded like science fiction. But in 2023, the US Food and Drug Administration cleared the first two algorithms based on artificial intelligence (AI) designed to mine electrocardiogram data for use as cardiovascular screening tools—one for detecting low ejection fraction and the other for identifying hypertrophic cardiomyopathy.

In doing so, the agency pushed open the doors to this newer regulatory category for the field of cardiology.

Back in 2013, the International Medical Device Regulators Forum (IMDRF), of which FDA is a member, defined SAMD as "software intended to be used for one or more medical purposes that perform these purposes without being part of a hardware medical device." It wasn’t until late 2018, though, that the FDA began entertaining the notion of approving SAMDs, and it took until 2021 for the agency to publish its action plan for clearing AI and machine learning-based SAMDs. While the agency doesn’t have a comprehensive list of all SAMDs to which it has given its stamp to date, it does have a published database of almost 700 AI-enabled devices it has cleared since 1995, many of which are classified as SAMDs.

Cardiology has seen a stream of AI tools developed in recent years. With the advent of the SAMD designation and the commercialization of the technology, however, AI will soon be in the hands of clinicians. How eager physicians are to adopt AI-based tools and, just as pivotal, what they’ll cost have yet to be seen.

“We are on the cusp of dramatic transformations in medicine; this is just the leading edge of it,” Paul Friedman, MD (Mayo Clinic, Rochester, MN), co-inventor of the Anumana ECG-AI Low EF algorithm, told TCTMD. The product, one of the two AI-based SAMDs approved this year, can diagnose reduced ejection fraction from an ECG. He’s optimistic that uptake will be swift. “Clinicians will have powerful additional diagnostic tools [on top of] currently available tests, which means they are already available in the workflows and we know how to order them, they are readily available, [and] will offer additional diagnostic information that can help us identify who needs advanced care and who doesn't,” he explained.

We are on the cusp of dramatic transformations in medicine; this is just the leading edge of it. Paul Friedman

The Anumana algorithm, which got the FDA’s signoff in October through the 510(k) premarket pathway, is indicated for the detection of patients who might be at risk for heart failure. Previously, the randomized EAGLE study demonstrated the algorithm’s success in pinpointing patients with low EF by 32% compared with usual care.

Before that, in August, the FDA approved the Viz HCM module (Viz.ai), the first ECG algorithm to be cleared in the SAMD category, through its 513(f)(2) de novo pathway. This tool also mines ECG data, but to identify patients with hypertrophic cardiomyopathy (HCM).

Friedman said he foresees wide use of AI-based algorithms in a variety of practice settings. “Academic centers will always be interested in new technology and new ways to help patients,” he pointed out, while community hospitals and clinics may see them as a way to benefit more from their existing infrastructure without the need to buy new equipment.

“This could potentially address healthcare equity,” Friedman predicted. “If you are a clinic in an under-resourced environment in whatever part of the world, an echo machine, a CT scanner, and a MRI scanner are very expensive, plus the personnel who are properly trained on their use.” ECGs, on the other hand, are more commonplace and their data could be leveraged to identify which patients need more specialized care or referrals for advanced imaging elsewhere.

Cautious Excitement

The heart failure physicians who spoke with TCTMD for this story seem, for the most part, excited about the new technology, especially as a means to help diagnose patients who would benefit from earlier intervention. However, there was a consensus that at this point it’s too early for much tangible change to happen.

Mitchell Psotka, MD, PhD (Inova Heart and Vascular Institute, Falls Church, VA), who has studied digital solutions in HF care, told TCTMD that AI has the potential to change medicine but it’s premature to say exactly how or to what extent these specific algorithms will sway practice. The way AI is sometimes “marketed in the lay press, and even in the medical literature, is at times really hyperbolic in terms of what it can and cannot do,” he said. “A realistic assessment of the capabilities and limitations of this type of software is appropriate to keep in mind when evaluating its potential benefit for patients and for the clinicians caring for them.”

Further, Psotka said, what the FDA has cleared to date isn’t artificial intelligence per se, but rather “just algorithmic screening of large data sets,” an approach that still requires clinicians to provide substantial input. While he potential for benefit can’t be denied, he added, use of these algorithms could also “create a lot more echocardiogram ordering and usage” if inappropriately calibrated.

The value proposition will have to be that health systems will save money by taking better care of their patients, preventing rehospitalizations, and perhaps this could be part of a move to value-based care. Mitchell Psotka

Because of this, Psotka stressed the importance of general FDA oversight, in which the agency has so far been investing. “It’s appropriate that these tools are treated as tests that need to be evaluated and regulated so that they cannot just be produced and broadcast and sold at the discretion of a manufacturer without any sort of oversight,” he said. “But I do think that they fit into the larger digital landscape in terms of how digital tools are being used to help with patient care and to help with overall efficiency of health system function.”

David Ouyang, MD (Smidt Heart Institute, Cedars-Sinai Medical Center, Los Angeles, CA), too, said he is excited about the potential for AI-based technologies to change the field of cardiology, but cautioned against reading too much into FDA approval of these algorithms. “The bar for the FDA is relatively low, and that's why I often recommend further validation or confirmation before using it in clinical practice,” he told TCTMD, adding that while Anumana seems to have done this prospectively, Viz.ai has yet to do so.

Indeed, Amanda Vest, MBBS (Tufts Medical Center, Boston, MA), is optimistic that, even though it’s “yet to be attained,” AI holds a lot of promise in medicine.

“The real value of AI analytic techniques is in taking very large quantities of data—particularly nonnumerical data such as images, movie clips, geographic locations—and integrating and analyzing them in a manner that traditional statistical techniques would not have accommodated,” she said. While the two cardiovascular-related SAMDs approved in 2023 show the potential for AI to wring more information out of tests that are already being performed, there are likely myriad new ways for tests to be applied prospectively. “Opportunities such as bringing electrocardiogram readings or echocardiogram images into research questions that utilize AI is very interesting,” said Vest, “and I think holds significant promise for the future.”

For her, the approval of these two algorithms offers “a glimpse of where we may be going in the future and how the electronic medical records may be harnessed in a meaningful way to triage and connect patients who will benefit from it.” Further validation, though, is needed in additionally health settings “to understand how image quality or patient characteristics could affect the performance of the AI algorithm,” Vest added.

A Competitive Landscape

New technology has never come cheap. While clinicians and patients alike might be curious at how AI-ECG algorithms might change practice, it’s not clear who will pay for the use of SAMDs and how they are being acquired.

Anumana is currently marketing their algorithm as a subscription service to institutions, but it is “flexibly priced based on the value it delivers to the health system through reimbursement and the improvement in healthcare resource utilization, ensuring the health system and patients are the primary beneficiaries,” according to a spokesperson.

Viz.ai is also marketing its HCM algorithm as an annual subscription with deployment financially supported by Bristol Myers Squibb.

Whether those costs will be passed along to patients remains to be seen. Psotka doesn’t envision insurance companies paying for individual runs of an algorithm, but rather foresees hospitals investing in SAMDs as institutional tools.

“The value proposition will have to be that health systems will save money by taking better care of their patients, preventing rehospitalizations, and perhaps this could be part of a move to value-based care,” he said. For that to happen, though, hospitals will be looking for choices and cost-efficacy analyses to back up those options, Psotka commented. “This is going to be a competitive landscape where these companies will have to have evidence that their product is going to be beneficial for the consumers, and the ‘consumers’ I do think are going to be health systems.”

The bar for the FDA is relatively low, and that's why I often recommend further validation or confirmation before using it in clinical practice. David Ouyang

Vest agreed. “The cost is going to fall upon the health systems to utilize such algorithms in a way that achieves the population health goals that that institution is pursuing,” she said. “For example, finding more cost-efficient ways to screen for heart failure that isn't already clinically overtly evident is a very important intervention that many health systems are trying to roll out on a population health level.” Additionally, identifying HFrEF earlier could lead to patients being prescribed on guideline-directed medical therapies sooner and thus yield longer-term cost savings, she added.

But as SAMDs become more integrated into patient workflow, health systems may want to recoup some costs, particularly if algorithms are able to do some of the work of other, more lucrative tests. As Friedman observed, these algorithms are “essentially converting an ECG to do the partial work of an echocardiogram or CT MRI. . . . So I do anticipate there will be likely additional billing generated by this.”

While the initial adoption of an AI-ECG dashboard may be a pricy investment for institutions, once that is in place, that infrastructure will make it “very, very inexpensive to add algorithms,” Friedman said. “We'll see how that unfolds, [but] over time, that will be what happens; there will be increasing value very rapidly.”

In time, once these algorithms have shown proven value, insurance should “definitely pay” for them, Ouyang stressed.

Additional Barriers

Plenty of other barriers stand in the way of widespread clinical adoption of AI-based algorithms, experts agreed.

For starters, Vest said, a strong IT infrastructure and the personnel needed to oversee that will be key. “We need to know more about the computing power and memory space required for such algorithms, and they certainly need to be deployed in a way that supports continued learning, ongoing refinement, and detection of any erroneous functions within the software system,” she said. “In many hospitals and cardiology divisions, there may not be a computer analyst who has expertise in AI techniques or is able to advise on the appropriateness of the data inputs that are going into the software.”

As with any algorithm, the inevitability of “garbage in, garbage out” means that “we need to make sure that the technology is being deployed in a similarly refined way as it was by the academic teams that developed them,” Vest added.

“There will probably be growing pains to figure out how to actually deploy it wisely in health systems,” Ouyang predicted, adding that ongoing performance monitoring will be needed. “Our experience has been some of these algorithms perform better in a retrospective setting than actually when they are being used.”

Other digital health technologies have shown variable performance or applicability across health systems, based on the type of electronic health record software being used, Psotka noted, calling implementation the “hidden elephant” in the room for the uptake of AI-based algorithms. “There are multiple different electronic health records that exist and there is not always interoperability between them,” he said. “And even, say, among health systems that have purchased the same type of electronic health record, the versions of that electronic health record may be different at different health systems and may have different capabilities of interacting with a software algorithm such as this.”

We need to make sure that the technology is being deployed in a similarly refined way as it was by the academic teams that developed them. Amanda Vest

Other implementation barriers relate to the workflow for dealing with algorithm outputs, hiring staff to manage them, and understanding their downstream impact, all of which still need to be figured out, said Psotka. “Depending on the volume, it can be a large undertaking for a health system.”

Just the Beginning

So far, Friedman estimates that the Anumana algorithm has been used “over 500,000 times by clinicians, [and] mostly not by cardiologists, interestingly.” To TCTMD, a company spokesperson said it is “having active discussions with multiple health systems in the US” and that it received “a great deal of interest in the product from clinicians” at the American Heart Association (AHA) Scientific Sessions meeting in November.

Viz.ai also hosted launch events at AHA and the European Society of Cardiology Congress in 2023, and claims to have spurred interest through those opportunities. “We have a large and growing footprint of 1,500 leading healthcare systems and hospitals already using Viz.ai for neuro and cardiovascular disease detection and care coordination,” the company’s chief clinical officer, Jayme Strauss, RN, MSN, MBA, SCRN, told TCTMD in an email.

While Anumana has plans to market the AI-ECG low EF algorithm in other regions, starting with Europe, Viz.ai told TCTMD they are not currently planning to do so for the HCM module.

“As with any tool, you need to figure out where it works really well, where you need to be more mindful of it,” Friedman said. “It is incredibly powerful, but like any test, it is just that: it is a test. And we have to put it in perspective in our armamentarium.”

In the coming years, Friedman foresees the information derived from ECGs broadening still further, taking the place of multiple different tasks or tests. So far, he said, “we know and have experience with it being used as a screen for left ventricular dysfunction, aortic stenosis, hypertrophic cardiomyopathy, amyloid heart disease, hyperkaliemia, hypokalemia, [and] cirrhosis of the liver,” he said. “It won't be that the clinician has to choose. It would be the clinician elects to order an ECG and gets a broad screening test. At least that's how I'd like to see this unfold.”

Vest, too, said the “real value” for SAMDs like these is to be able to increase efficiency within a health system and identify “patients who would benefit from more-specialized resources that they have not otherwise been directed toward.”

Ultimately, change will come, but gradually, Psotka predicted. “Artificial intelligence will change the way we practice medicine. It is a slow, slow process that has a lot of steps that are hidden under the surface.”