A new AI system trained on more than 440,000 electrocardiograms can identify patients at high risk of sudden cardiac death better than the standard clinical test, researchers report today in Nature. The model flagged a high-risk group with a 7% annual rate of the fatal event, compared to 4.6% for the current ejection fraction measurement - a difference that could translate to thousands of preventable deaths each year in the U.S. alone.
Sudden cardiac arrest kills over 300,000 people in the U.S. each year. The heart's electrical system malfunctions without warning, and while internal defibrillators can save lives, doctors lack a reliable way to decide who needs one. The most common test measures how much blood the heart pumps out with each contraction. If that ejection fraction falls below a certain threshold, a patient may qualify for an implantable defibrillator.
Why current detection falls short
The ejection fraction test requires a more involved evaluation that most victims never receive. Two-thirds of implanted defibrillators never end up firing, meaning patients undergo invasive procedures for a risk they may never face. Meanwhile, many people who die suddenly never knew they were at risk. "The problem is that doctors can't figure out who needs one before it's too late," said Ziad Obermeyer, associate professor at UC Berkeley's School of Public Health and the study's lead author.
Training on global EKG data
Obermeyer's team used six years of electrocardiograms from Sweden's unified health system, matched with death certificates, to train an AI model. They then validated the algorithm on deidentified EKGs from a hospital system in San Diego and a dataset from Taipei. The model learned to spot waveform patterns associated with later sudden cardiac death. "Good AI starts with good data," Obermeyer said. "Unfortunately, data like the ones we used for this study are incredibly hard to access. It's a big part of why there's so little clinical AI in use today."
The AI's risk prediction outperformed the ejection fraction benchmark, isolating a larger high-risk pool that would have been missed by current standards. The findings rely on scans that are widely available in medical centers around the world.
A physiological signal hiding in plain sight
The model may have uncovered something new about heart physiology. Because most sudden cardiac deaths happen abruptly, what goes wrong inside the heart before it stops has remained largely a black box. By detecting subtleties in the EKG waveform, the AI points to a signal that appears to correlate with the heart's fatal misfiring - a finding that could spur new research into the underlying mechanism. "Medical decisions are really hard, and I think that's why AI is so exciting for me," Obermeyer said. "We can not only make better decisions, but also start to understand what's actually going on with these patients before their heart stops."
Obermeyer sees this as the start of AI for Science & Research opening up new investigative paths. "There is also going to be a new way of doing science that comes out of these tools," he said, "and it's fun to think about how that starts happening."
Next steps for deployment
The team is already working with health systems in Sweden, Taiwan, and the U.S. to deploy the algorithm on hospital EKG databases. Patients flagged as high-risk would be notified and given the option to wear a patch that continuously monitors their heart. That data could help researchers pin down the physiological mechanism and eventually lead to a life-saving internal defibrillator. Obermeyer also built a website where people can submit basic information to be contacted for EKG analysis once the tool is more widely available.
Why this matters for science and research professionals
This study shows how an AI model trained on routine clinical data can extract a latent risk signal that outperforms a decades-old standard. For researchers, it demonstrates the value of large, linked datasets - the project took about a decade to compile - and the potential of AI to generate testable hypotheses about disease mechanisms. The method does not require new imaging or invasive tests; it works with widely available electrocardiograms. That accessibility, combined with the model's performance, suggests a scalable approach to studying other conditions where the physiological precursors remain poorly understood.
Your membership also unlocks:
