The electrocardiogram is more than a century old, and cardiologists have studied what it shows for generations. It’s hard to imagine much was left to find.
A team led by Ziad Obermeyer at the University of California, Berkeley, thinks there was. They trained an AI system to read routine ECGs and flag people at elevated risk of sudden cardiac death, which often arrives without warning, then used a second AI model to figure out what the first one was picking up on.
It surfaced a small pattern in the heart’s electrical trace that, as far as the researchers know, hadn’t been described before.
There’s a big limit, though, and the study’s lead is the first to say so: the findings aren’t ready to guide treatment decisions. A separate research group, backed by new funding, is tackling the same problem from a different angle.
The Problem Is Knowing Who to Protect
The treatment exists; the hard part is deciding who should get it.
Sudden cardiac death kills about 300,000 people in the U.S. each year, and it’s not the same thing as a heart attack. A heart attack is a circulation problem caused by a blocked artery. Cardiac arrest is electrical: the heart’s rhythm goes chaotic, and it stops pumping blood.
The standard way to pick defibrillator candidates leans heavily on LVEF, an ultrasound measure of how well the heart pumps. Obermeyer says many people who die suddenly never had that ultrasound, or had it and got normal results.
It also misfires in the other direction, since many people who get a defibrillator on that basis never have it fire, though that doesn’t mean their implants were unnecessary.
The Algorithm Found an Overlooked Group
The headline numbers sound modest until you see how little they overlap with the old method.
The study, published in Nature, trained a neural network on more than 440,000 ECGs from roughly 180,000 patients in Sweden, each matched to national death certificates. It was then tested on separate datasets from the United States and Taiwan, and the signal held up.
The model flagged about 2.2% of patients as high risk. Within that group, the annual rate of sudden cardiac death reached 7%, compared with 4.6% among patients flagged by the standard ultrasound. More than 86% of the people the algorithm picked were not flagged by the ultrasound at all.
Worth keeping in mind: the 10 seconds is the recording length, and the 7% is a risk over a full year, so nothing here predicts an event minutes away. The 2.2% is simply the share who got flagged, and it says nothing about the model’s overall accuracy.
A Second AI Showed Its Work
Rather than stopping at a risk score, the team built another model to show what the first one had noticed.
The second model was generative. Starting with a real low-risk patient’s ECG and guided by the first model’s risk score, it nudged the trace step by step toward a high-risk version so researchers could see which changes mattered.
Many of the features it highlighted were already familiar to cardiologists. One stood out: subtle slurring in the aVL lead, suggesting the heart’s electrical signal fragments as it travels through muscle.
Changxin Lai, a biomedical engineer who wrote an accompanying analysis in Nature and peer-reviewed the study, put it this way: “We extracted new knowledge from an artificial intelligence model.”
A possible explanation comes from cardiac MRIs of some high-risk patients, which showed subtle, diffuse fibrosis, a kind of scarring tied to arrhythmias. It fits the electrical changes, but Obermeyer cautions that the link is preliminary and no biopsies have confirmed it.
Related: These 7 Habits Increase Your Risk of Early Death
Why the Researchers Are Urging Caution
The people who built the tool are the first to slow things down.
Save this article
Enter your email address and we'll send it straight to your inbox.
Obermeyer sees the ECG as a practical screening tool, far easier to use than expensive, inconvenient imaging like MRI. But he doesn’t want a high-risk result treated as a verdict.
Sumeet Chugh, who directs the Center for Cardiac Arrest Prevention at Cedars-Sinai and wasn’t involved in the study, calls this an important area of research. But from a patient care standpoint, he says, much more work is needed before findings like these can help identify candidates for a preventive defibrillator.
Wearables are a separate question. The model was trained on medical-grade ECGs and did slightly worse with lower-quality signals from consumer devices, a gap the researchers described as minor. That’s a research finding, not a product.
Another Team Is Betting on Fitness Trackers
Cedars-Sinai’s newly funded project is separate work that looks at risk over the long haul and in the hours before an event.
Announced on October 19 during Sudden Cardiac Arrest Awareness Month, the project has more than $5 million in new funding from the National Institutes of Health and the American Heart Association. The investigators are Chugh, who weighed in on the Berkeley study from the outside, and colleague Kyndaron Reinier, and they’re building on two decades of joint research.
The long-term half draws on de-identified health records from about 400,000 Los Angeles County residents treated at Cedars-Sinai. It will test how well a risk tool called VFRisk, which weighs 13 variables including ECG measures and clinical history, predicts events over time, and whether tracking changes in heart health improves it.
The short-term half is where the fitness trackers come in. Over three years, 600 high-risk patients will wear devices that record heart rate, heart-rate variability, oxygen saturation, and activity. At the same time, researchers scan rolling 48-hour windows for departures from each person’s baseline.
That window is for analysis, not a promised warning time, and the project has not announced any results yet.
Why the Clock Matters So Much
Prediction matters because cardiac arrest rarely gives anyone time to react.
Reinier says cardiac arrest is fatal in about 90% of cases, and for up to half of patients it’s the first sign of underlying heart disease. Without immediate CPR and defibrillation, death follows within minutes.
Investigators also plan to compare AI-based approaches with conventional ones. If the wearable work pans out, Reinier describes one eventual payoff: clearer guidance on when to seek urgent care when tracker readings and symptoms point to rising short-term risk.
Her instructions for the moment itself haven’t changed: if you witness a cardiac arrest, call 911, start CPR, and use an automated external defibrillator if one is available.
If an AI flagged your routine ECG and your doctor suggested extra testing, what’s the first question you’d ask?
