Sudden cardiac death. It’s the grim headline that hits 300,001 people in the U.S annually.
Implantable defibrill exist. They stop lethal rhythms. The tech works fine. The real headache? We have no idea who needs them until it is often too late.
That changes slightly today.
A team led by Ziad Obermeya, an associate professor at UC Berkeley, just dropped a study in Nature. They trained a neural network to spot danger on a ten second electrocardiogram (ECG). Then, crucially, they built a second network to show how the first one did it.
Why bother with the second model?
Because medicine hates black boxes. We need human experts to see the clue for themselves.
Right now cardiologists rely on left ventricular ejection fraction (LVEF). It measures how much blood the heart pumps per beat via ultrasound.
It is far from perfect.
“A lot of people who suddenly die… had it [ultrasound] and the results were normal” says Obermeyer. Meanwhile many high-risk folks never got scanned. On the flip side most people flagged as high-risk never actually trigger their defibrillators. We are overtreating some. Undertreating others.
Obermeyer’s team wanted better.
They turned to ECGs. Cheap. Ubiquitous. But for a century cardiologists saw only familiar waveforms. No hidden code for sudden death. So they let deep learning dig.
The engine was a 64-layer ResNet. Boring? Sure.
“It’s kind of a workhouse model everyone uses.”
The magic was the fuel. They fed it over 440,090 ECGs from 180k patients in Sweden linked to death records. A massive dataset. The AI flagged roughly 2.2% as high-risk.
Does it hold up elsewhere?
Yes. They tested on U.S. and Taiwan data. The signal stayed sharp. Within that tiny 2.2 group the annual death rate hit 7%. Standard ultrasound tests miss this. Over 86% of the AI’s picks wouldn’t have raised alarms with traditional tools. These patients likely walked home without protection.
“But what is the machine seeing?” Obermeyer asked.
Standard explainability tools highlight pixels. Not helpful to a human doctor. You need a pattern. A sketchable wave.
So they built a generative model. Its job: morph a safe ECG into a risky one step-by-step until the first network screamed danger.
Most changes were known. Expected.
But one feature? New. Never described.
A subtle slurring in lead aVL. Like the electrical signal fragments as it hits muscle.
“We extracted new knowledge from artificial intelligence,” says Changxin Lai at Johns Hopkins, who was not part of the study but reviewed the work.
Sounds scary?
Maybe. For some high-risk patients MRIs showed diffuse fibrosis. Scarring that matches the weird waves the AI generated. A physical cause for the digital warning. Obermeyer admits the biopsy confirmation is missing. It’s preliminary.
Don’t panic yet.
“This is not ready to guide treatment” warns Sumeet Chugh from Cedars-Sinai. He sees a lot more research before this identifies defibrillator candidates in practice.
But think about the logistics. MRIs are expensive. Rare for screening.
ECGs? Anywhere. An Apple Watch can record them. Yes the AI prefers hospital grade data but the gap is narrowing. The drop in quality from consumer tech is minor according to the team.
The beauty of this approach? You don’t have to trust the robot.
“You can just use it to target additional,” Obermeyer explains. It flags who gets the deeper dive. Not who gets surgery today.
It leaves a question hanging in the air though.
If an AI spots a fracture in the signal we ignored for a hundred years, how many other ghosts are hiding in the plain sight of our charts?
We might never know unless we listen to them.
