AI spots hidden heart disease from routine ECGs

Colonoscopies screen for colon cancer and mammograms catch breast tumors, but structural heart disease—affecting millions worldwide—has never had a cheap, routine screening tool.

For decades, doctors were taught in medical school that you simply cannot diagnose structural valve defects or cardiomyopathy from a standard electrocardiogram (ECG). If a patient needed answers, they had to be sent for an expensive echocardiogram ultrasound.

Now, researchers from Columbia University and NewYork-Presbyterian have broken that rule. In a study published in Nature, the team unveiled EchoNext, an AI model that reads standard, low-cost ECG traces to flag hidden structural heart problems before symptoms spiral.

Beating cardiologists at reading the wire

Because structural heart issues often show few early symptoms, patients usually discover them only after severe damage has already set in. Nobody orders a costly cardiac ultrasound just in case.

EchoNext flips the economics of the workflow. The team trained a deep learning network on over 1.2 million ECG–ultrasound pairs from 230,000 patients, teaching the AI to catch the faintest electrical ripples caused by physical muscle thickening or failing valves.

In a direct benchmark across 3,200 ECGs, EchoNext went head-to-head with 13 cardiologists. The human specialists managed a 64% accuracy rate. EchoNext scored 77%, outperforming even doctors who had access to traditional AI interpretation software.

Finding 2,000 missed patients in the real world

Lab benchmarks are fine, but clinical utility is what matters. To test the system in the wild, researchers quietly ran EchoNext on nearly 85,000 hospital patients who had received an ECG but never an ultrasound. Doctors were blinded to the predictions.

The algorithm flagged roughly 7,500 people—about 9% of the cohort—as high-risk for undiagnosed disease. Over the next year, 55% of those flagged patients happened to get an echocardiogram ordered through standard care.

The result: nearly three-quarters of them actually had structural heart disease. That is twice the hit rate seen in standard patient populations undergoing their first ultrasound. The team estimates that if every flagged patient had received an immediate follow-up scan, around 2,000 serious, hidden heart conditions would have been caught a year earlier.

The researchers have already released a de-identified dataset for the broader community and started a clinical trial across eight emergency departments.

Why it matters

We think EchoNext is one of the smartest applications of clinical AI to date, simply because it upgrades infrastructure that already exists.

Around 400 million ECGs are performed globally every year. They are cheap, non-invasive, and available at virtually every clinic. Instead of inventing a proprietary scanner or selling an expensive test, EchoNext turns an existing $20 routine test into a population-scale screening funnel. For patients, that means potentially catching progressive heart failure years before it lands them in the ICU, without spending an extra dime out of pocket.

For health systems, it sets a template for how diagnostic AI should work. The biggest bottleneck in healthcare is not interpreting high-end MRIs; it is triaging who actually needs the scarce, high-end equipment in the first place.

The catch, of course, is capacity. Flagging 9% of all routine ECGs as high-risk will flood ultrasound suites with referrals. If hospitals cannot scale up cardiac ultrasound imaging to match that demand, a warning from an algorithm is just another stress test on an already stretched healthcare system.