AI spots a hidden ECG warning sign that could predict sudden cardiac death

What it's about

More than 300,000 people in the US die of sudden cardiac arrest every year. The heart's electrical system suddenly fails, and the heart stops pumping blood. It strikes older people with known heart problems, but also young athletes who seem perfectly healthy. An implantable defibrillator can prevent many of these deaths. It's a small device placed under the skin that delivers a shock when the heart rhythm turns dangerous. The catch: doctors often don't know who needs one.

A team led by Ziad Obermeyer, an emergency physician and associate professor at the School of Public Health at the University of California, Berkeley, trained an AI to tackle that problem. It reads a person's risk from an ordinary ten-second electrocardiogram (ECG). An ECG records the heart's electrical activity. The study, "An ECG biomarker for sudden cardiac death discovered with deep learning", was published in the journal Nature. Alongside Obermeyer, the authors include Alexander Schubert, James Ross, Sendhil Mullainathan and Markus Lingman. The university had already presented the results in June 2026.

Why the current test falls short

Right now, cardiologists rely mainly on left ventricular ejection fraction (LVEF). It's an ultrasound measurement that shows how much blood the heart's left chamber pumps out with each beat. If the value falls below a certain threshold, a defibrillator becomes an option. But the method has two big weaknesses:

• Many people at risk are never tested. The measurement requires a more involved exam, and people without symptoms don't get one. "Many people who die suddenly of cardiac arrest either never had the ultrasound or their results were normal," Obermeyer told Scientific American.
• Many implants are never needed. About two-thirds of defibrillators implanted in high-risk patients never deliver a shock. Those people go through an expensive, invasive procedure to guard against an emergency that may never come.

About 90 percent of people who suffer sudden cardiac arrest outside a hospital don't survive. That's why spotting at-risk people early matters so much. Obermeyer puts it this way:

"What makes this problem so tragic, but also so well suited to AI: we have the cure. If you knew you were one of the people who would drop dead, you'd go to a cardiologist and get a defibrillator. The problem is that doctors can't figure out who needs one until it's too late."

It's important not to confuse this with a heart attack. In a heart attack, a blood vessel is blocked. In cardiac arrest, the electrical control of the heartbeat fails.

How the AI was trained

The model itself is nothing special. The team used a 64-layer ResNet, a "residual" neural network. This widely used architecture for image and signal analysis can go very deep thanks to shortcuts between layers. "It's a workhorse that everyone uses. There's nothing interesting about it," Obermeyer told Scientific American. "What's interesting is the data it learned from."

The training data covers six years of ECG records from Sweden's unified healthcare system: more than 440,000 ECGs from about 180,000 patients, linked to national death certificates. The AI saw ECGs from healthy people, from patients with known risk factors and from people who later died of sudden cardiac death. From that, it learned which combinations of spikes and waves point to elevated risk.

The team then tested the model on two independent datasets: two years of anonymized ECGs from a hospital network in San Diego and a separate collection from Taipei. The goal was to show that the result isn't a quirk of the Swedish population or Swedish equipment. According to Obermeyer, building the datasets alone took about ten years. Two organizations he co-founded were involved: Dandelion Health and Nightingale Open Science. "Good AI starts with good data," he says. But that kind of data is extremely hard to get, he adds, and that's a major reason so little clinical AI is in use today.

The results in numbers

The AI flagged about 2.2 percent of patients as high risk. In that group, the risk of death was markedly higher than in the group the conventional test identifies. According to Scientific American, more than 86 percent of the patients the AI picked out wouldn't have been flagged at all under the standard LVEF criterion. Many of them would have been sent home without a defibrillator. UC Berkeley says this involves thousands of patients a year who count as low risk by today's standards.

One practical advantage: the AI works with completely ordinary ECG images that hospitals around the world already produce every day. No new exam is needed. The model also works with ECGs from consumer devices like an Apple Watch or simple smartphone accessories. It performs somewhat worse there, but according to the researchers only slightly.

A second AI translates what the first one sees

The second part of the study is what really stands out. Common explainability tools like saliency maps only highlight which parts of a signal mattered most to the network. They don't show what a suspicious signal looks like. A cardiologist can sketch an unusual wave; a neural network can't do that on its own.

So the team built a generative AI model, one that creates new data itself. "Its job was to generate ECG traces that look high-risk to the first model," Obermeyer explains. Guided by the first network's risk score, the generative model gradually transformed a real low-risk patient's ECG into a high-risk version of the same trace. That revealed which features change.

Cardiologists already knew many of those features. But one had never been described in the medical literature: a subtle "slurring" in lead aVL (augmented vector left). A lead is a specific measurement direction between electrodes. The pattern suggests that the electrical signal is fragmenting as it travels through the heart muscle.

Biomedical engineer Changxin Lai reviewed the study and wrote an accompanying commentary in Nature. He sees this as the heart of the work: the ECG has been around for more than 100 years, generations of cardiologists have carefully analyzed this kind of data, and now new knowledge has been extracted from an AI model.

For some high-risk patients, cardiac MRI scans were also available. They showed fine, diffuse fibrosis, or scarring of heart tissue, which can disrupt electrical conduction. That would fit the artificially generated traces. But Obermeyer stresses that this link is preliminary and still needs to be confirmed with tissue samples.

What's next

Obermeyer is already working with health systems in Sweden, Taiwan and the US to run the algorithm on hospitals' existing ECG databases. Doctors will contact anyone flagged as high risk. Those patients can then wear a patch that continuously records their heart rhythm. The longer recordings should help uncover the biological mechanism behind the warning sign. In some cases, they could also make the case for a defibrillator. Obermeyer has also set up a website where people can leave basic information and their email address so they can be contacted later for an ECG analysis. The sources don't give the address.

Obermeyer himself warns against jumping to conclusions:

"I wouldn't advise anyone to get a defibrillator just because we say their ECG shows high risk. The beauty is that you don't have to believe the AI. You can simply use it to trigger targeted follow-up tests."

The bigger picture

The study is notable for two reasons. First, it uses a cheap, widely available test to find people who currently slip through the cracks. Second, it doesn't just deliver a "black box" prediction. It produces a concrete, human-visible feature that cardiologists can check and learn to recognize. That sets it apart from many medical AI projects that get the right answers but can't explain why.

Plenty of questions remain. There's no proof yet that using the algorithm actually prevents deaths. Both SciTechDaily and independent experts point this out. Sumeet S. Chugh, director of the Center for Cardiac Arrest Prevention at Cedars-Sinai Medical Center, calls the field important. But from a patient-care standpoint, he says, much more research is needed before such findings can be used to select defibrillator candidates. And even a rate of 7 percent a year means the vast majority of flagged people won't die in any given year. Broad screening would therefore trigger a lot of follow-up testing. The suspected link to fibrosis is only a hint so far. Still, the work shows how AI can go beyond pure prediction and help generate new medical insights. Obermeyer calls it "a new way of doing science."