A New Market For A Century-Old Test

Why bring AI into a test that has been around for a hundred years?

Cheap, fast, everywhere; about 300 million ECG tracings are produced worldwide each year on a machine that costs almost nothing to run. Now the tracings that haven’t altered in a century are being interpreted by AI. And the change is real.

Ask Viz.ai, Anumana and AliveCor whether their algorithms are ready to read the 12-lead ECG, and you get a close-to-unified answer: already cleared, already deployed at hundreds of US hospitals and clinics, already producing peer-reviewed evidence. Ask some clinicians, and the answer is different: not fully trusted, yet.

The disagreement is not over whether AI can extract signal from a tracing. Both accept that it can. The disagreement is over what counts as proof, who decides when an algorithm has earned a place at the bedside and what to do in the gap between FDA clearance and national guidelines. I discussed this very question with three AI-ECG CEOs and a renowned ECG educator.

Has the evidence arrived?

Dr. Amal Mattu, who founded one of the field’s first Emergency Cardiology Fellowships at the University of Maryland and has spent nearly 30 years training clinicians to read ECGs, sets an explicit bar. “AI needs to prove its worth in large studies, [then] it needs to be approved for clinical use, [then] it needs to be adopted into national guidelines before cardiology will routinely accept AI interpretations to influence care,” he says. His estimate for when the loop closes: 10 years.

The market is moving fast. Anumana, a joint venture between Mayo Clinic and the AI company nference, develops ECG-based AI algorithms and has framed the work in pharma terms. “We have spent a lot of time, over 100 publications in peer-reviewed journals,” says CEO Maulik Nanavaty. “We follow a very traditional development path to make sure that we’re building out the same type of evidence as a pharmaceutical or a drug device would do.” The flagship is the EAGLE trial in Nature Medicine, a pragmatic cluster-randomized study of 22,641 patients and 358 primary care clinicians that found AI-ECG screening lifted the diagnosis rate of low ejection fraction by about a third over usual care. In March, the FDA cleared Anumana’s algorithm for the early detection of pulmonary hypertension; cardiac amyloidosis followed shortly after.

Viz.ai received the first De Novo clearance from the FDA in 2023 for cardiovascular machine-learning-based notification software, for an algorithm that flags hypertrophic cardiomyopathy on routine ECGs. The platform is now deployed at hundreds of US sites including Mount Sinai, Cleveland Clinic and UCSD. “We’re seeing time to diagnosis for HCM drop from five years to five weeks,” CEO and co-founder Chris Mansi says.

AliveCor makes FDA-cleared personal ECG devices, putting the classic device into a user’s pocket. It received clearance in January for five additional cardiac determinations on its handheld Kardia 12L, bringing the device’s total to 39. “We are building a ‘hospital-in-your-pocket’ that clinicians actually trust for diagnosis,” says CEO and President Priya Abani.

The clearances exist. The trials exist. Mattu disputes neither. He disputes whether either has yet crossed into the territory where a cardiologist treating a patient should accept the model’s read.

Will the cardiologists use it?

Mattu sees the second flashpoint as the steepest. “The interventional cardiologist is the person that does the catheterization; and so if [they] don’t believe in AI, the AI result is almost irrelevant,” he says. He imagines the scenario: “if I’m working a shift and caring for a patient with chest pain and I get an ECG, pretend that the ECG appears to me to be ok, but the AI says it is an acute heart attack. Then I call the cardiologist and say ‘I have an ECG that I’m not worried about, but the AI device says you need to take this patient emergently to the cath. lab!’ In all likelihood, the cardiologist will probably just say ‘goodbye’ and hang up.”

Mansi disputes that framing. “Quite the contrary. Interventional cardiologists are enthusiastic adopters,” he says. “They benefit from AI-ECG because the ECG is often the triage investigation that gets the right patients to them in the first place.” In his telling the friction is upstream of the cath lab, not at it. “The real friction lives upstream, in fragmented workflows and delayed routing.”

Six hundred US sites. Cleveland Clinic among them. Mattu’s broader point still lands: FDA clearance is not the same as guideline adoption, and “the cardiologists I know are unlikely to accept an AI read as something which should, in and of itself, influence care.”

Replace or augment?

All agree the algorithm should not be the decision-maker.

Mattu wants the physician at the wheel. “The physician needs to be the driver here; and AI can, at best, make some suggestions.”

Mansi’s articulation of the same line is more design-specific. “The goal isn’t to replace clinical judgment. It’s to make sure that judgment gets applied to every patient, every time, no matter where they show up.” He pushes the design principle further. “Clinicians need to remain active reasoners, not passive approvers.”

Nanavaty puts it in even more conservative terms. “One cannot aspire to go and change the physician’s practice. What one needs to do is to really provide the right tools that allow them to make their decisions at that point without necessarily adding steps and complexity.”

Three different framings of the same posture: AI as augmentation, physician as driver. The unresolved question is whether the daily workflow actually preserves that posture once enough hospitals deploy. Like any AI augmentation, once a recommendation is given, it is difficult to disagree with—creating a cognitive bias, and a medicolegal dilemma for the practicing physician.

What counts as success?

For Mattu, the missing evidence is patient outcomes. “They are successfully publishing articles in the cardiology literature. But I would like to see them perform randomized studies that show that reliance on AI is not just superior to human diagnosis, but that reliance on AI actually results in lives saved.”

Mansi acknowledges the demand and complicates it. “It’s a fair question. RCTs are the gold standard, and we respect that. But they present a genuine ethical challenge in AI-enabled care pathways. Once you have strong real-world evidence that a tool is getting patients to treatment faster, randomizing patients away from it becomes a challenge.”

That is the deepest line of disagreement. Mattu wants RCTs of mortality. Mansi argues the field is past the point where withholding the tool from a control arm passes the ethics board. Both can be true. The next five years will be spent deciding which evidence practicing clinicians agree to count.

Where they agree

Adoption still has a long way to go, although Medicare’s 2025 outpatient rule added codes for AI-enabled ECG analysis, reimbursing hospitals per use. Last July, Philips launched the first commercial AI-ECG marketplace, with Anumana as its founding algorithm partner.

As a physician who has trained in reading ECGs for decades, the integration of AI-ECG marks an exciting opportunity for better care. At the same time, I wonder: will future physicians lose this fundamental skill altogether? Will they become QA specialists, reviewing the logic of a recommendation rather than producing it themselves? It is a question our profession is starting to share with software engineers in the era of Claude Code.