His Heart Was Pumping at 10 Percent. The Real Question Is Why It Took an Algorithm to Notice.

A 45-year-old man walks into an emergency room. He can’t catch his breath, he’s been coughing for weeks, and he’s a smoker. The doctors hand him a bronchodilator and a course of steroids, tell him it’s probably reactive airway disease, and send him home.

His heart was pumping at 10 percent of normal.

That number stopped me cold. A healthy left ventricle ejects somewhere north of 55 percent of its blood with each beat. Ten percent is the kind of failing heart that ends on a transplant list, which is exactly where this man ended up, after a genetic test turned up a lamin mutation driving a dilated cardiomyopathy. He got a new heart. And the reason we know his story at all is that an AI model named EchoNext caught what the ER waved through, in a case report published in Nature Medicine that became a New York Times feature with a familiar shape: the machine saw what your doctor couldn’t.

That’s the version everyone ran with. The case file says something less flattering, and the calendar says something else again.

Let me be honest, because I came into this one ready to be charmed. The biology underneath EchoNext had me leaning in. It reads a standard 12-lead ECG, the same squiggle your GP has been printing out for decades, and pulls from it a signature of structural heart disease that trained cardiologists staring at the same tracing simply cannot see. How? The model was trained on more than 1.2 million ECG-echocardiogram pairs from over 230,000 patients, learning to tie faint, distributed waveform patterns to the chamber sizes and valve problems an echo would later confirm. Put it head to head against 13 attending cardiologists reading ECGs alone and EchoNext caught 77 percent of structural heart problems against their 64 percent. In a silent real-world deployment across nearly 85,000 patients with no prior echo, it flagged about 9 percent as high-risk, and roughly 73 percent of those who actually got the follow-up scan turned out to have disease, twice the hit rate of people sent for a first echo the ordinary way.

That part holds up. An ECG carries more information than the human eye was ever going to extract from it, and a machine trained on a million heartbeats can read the fine print. I’m not going to pretend otherwise, and the reflex that waves off every AI-in-medicine claim as vaporware would be wrong here.

Then my enthusiasm hit a wall, the same wall oncologist Vinay Prasad walked into when he picked the story apart this week. Of all the cases EchoNext’s makers could have put in the showcase window, they chose the one that argues against them.

Go back to the man in the ER. He’s a smoker. He’s short of breath. By the case report’s own account he had elevated cardiac biomarkers and an abnormal ECG. As Prasad reads it, you don’t need a neural network to know what comes next: a 45-year-old smoker with dyspnea, a bad ECG, and raised cardiac markers gets an echocardiogram, not an inhaler and a ride home. That isn’t frontier medicine. It’s the workup a competent intern is supposed to finish before the patient leaves the department. The story sold as AI seeing the invisible is, in Prasad’s telling, a standard evaluation that didn’t happen, rebranded as a triumph.

What’s missing from the writeup sharpens his point. The case report never says which biomarker was elevated or how high, the one number that would tell a clinician how loud this presentation actually was. It skims the physical exam, the very place a heart at 10 percent function tends to announce itself: distended neck veins, crackles at the lung bases, swelling in the legs. Prasad says he’s frankly suspicious the note is incomplete on exactly this point, because either those signs were there and went unacted-on, or the exam was never properly done. Neither version flatters the machine. Both are a human miss dressed in the language of innovation, and Prasad’s read is that no working clinician seems to have touched the framing before Nature Medicine and the Times ran it.

Now look at the calendar, the part the feel-good version skips. The case report dropped this week. The next day, Pathway Labs, the commercial spinout built around EchoNext and co-founded by the model’s lead developer, announced the FDA had cleared the tool to detect six forms of structural heart disease, alongside a deal to license it to OpenEvidence, the clinical search engine hundreds of thousands of clinicians already keep open in a browser tab. Columbia holds the patent. Seen against that, a transplant case report and a sympathetic newspaper feature aren’t only science communication. They’re the warm-up act for a product launch, and a heart transplant the algorithm gets the credit for is a hell of a thing to carry into an FDA-cleared rollout.

So two things are true at once. The tool looks good, and the idea that an ECG hides a readable signature of a failing heart is the kind of biology that makes this job fun. But the case picked to sell it is the weakest possible advertisement for it, because it documents a basic miss rather than a hidden catch, and it arrived gift-wrapped exactly when a newly cleared company needed a headline.

Here’s what I’d actually watch, and it isn’t the transplant anecdote, which tells you more about that one emergency department than about the algorithm. Watch whether EchoNext, now that it’s commercial and wired into the tools doctors already use, catches disease earlier in a real population: more hidden cases found, fewer missed, without burying clinicians in false positives and a flood of needless echoes. That’s the trial that earns the hype. A man rescued from a workup his doctors skipped is not it, and the people telling you it is have a product to move.

Sources

1. Vinay Prasad – “AI will beat your doctor but a NYTimes story and NatMed paper ain’t it”
2. Nature Medicine – case report: AI-enhanced diagnostics leading to heart transplantation (2026)
3. Nature – Elias et al., “Detecting structural heart disease from electrocardiograms using AI” (2025)
4. Columbia University Irving Medical Center – “Can AI Detect Hidden Heart Disease?”
5. STAT – Pathway Labs, EchoNext, and the OpenEvidence licensing deal
6. Medical Economics – “FDA-cleared AI tool flags hidden heart disease from a standard ECG”
7. TCTMD – “AI for Heart Failure Care Is Evolving Rapidly, THT 2026 Makes Clear”