AI Spots Death Signal Hiding in ECGs

Tablet with FDA on screen amid medical tools and ECG charts
Photo: ra2 studio / Shutterstock

A hidden flicker in a 100-year-old heart test may reveal who is about to die suddenly, years before it happens.

Quick Take

  • Researchers trained an artificial intelligence (AI) system on more than 440,000 heart tests, called electrocardiograms (ECGs), from about 180,000 Swedish patients.
  • The AI flagged a high-risk group making up 2.2% of patients, who died suddenly from heart causes at a rate of 7.0% per year.
  • Most of those flagged, 86.1%, would have been missed by the standard heart-pumping test doctors use today.
  • The AI spotted a specific squiggle in the heart tracing that no doctor had described in medical literature before.

A Decades-Old Test Yields a Brand New Clue

Doctors have used ECGs since the early 1900s to check heart rhythm. The test has barely changed. But researchers at UC Berkeley and partner institutions ran over 440,000 of these recordings from roughly 180,000 Swedish patients through a deep learning system, linking each test to the patient’s eventual cause of death. That scale of real-world data let the AI hunt for patterns no cardiologist had ever cataloged.

The AI did not just rank risk in general terms. It carved out a specific slice of patients, 2.2% of the entire group, who went on to die suddenly from cardiac causes at a rate of 7.0% every year. That is a steep jump from the overall population and higher even than the rate seen in patients already known to be high-risk through standard testing.

Beating the Old Benchmark for Heart Risk

For decades, doctors have leaned on a measurement called left ventricular ejection fraction, or LVEF, which checks how well the heart’s main pumping chamber squeezes blood. Patients with a weak LVEF reading are considered high risk and often get a defibrillator implanted. In this study, the LVEF-flagged group saw a 4.6% annual sudden death rate, meaningfully lower than the AI’s 7.0% figure.

Even more striking, the AI caught danger that LVEF testing missed entirely. Researchers reported that 86.1% of the patients the AI flagged as high-risk would not have been identified using the standard pumping-strength test. That gap matters because LVEF has long been criticized for casting too wide a net while still missing many people who go on to die suddenly, a problem cardiologists have documented for years.

Researchers Found the Signal Checked Out Elsewhere Too

A single country’s data can be a fluke. So the team tested the model against two outside groups: a health system in the United States and a hospital registry in Taiwan. In the American data, the AI predicted dangerous heart rhythm disturbances tied to sudden death. In the Taiwanese data, it predicted future cardiac arrests caused by those same rhythm problems. Different populations, different healthcare systems, same warning signal showing up again.

The Slurred Wave Nobody Had Named

To figure out what the AI was actually seeing, researchers paired the prediction model with a second AI tool that generates altered versions of a heart tracing, essentially showing what a “safer” heartbeat would look like next to a risky one. The difference centered on a subtle slurring in one section of the tracing, called lead aVL, that had never been written up in medical journals before. It is a quiet, almost invisible wrinkle that a computer caught and a century of human eyes had not.

The study also found a hopeful thread worth watching. Among the AI-flagged high-risk patients who already had a defibrillator implanted, death rates ran 54.4% lower than expected. That is not proof the device caused the improvement, since patients who get defibrillators often differ in other ways too. But it is the kind of early signal that makes researchers want to run a real trial next.

What Still Needs to Happen Before This Reaches Your Doctor

This is a retrospective study, meaning it looked backward at records rather than testing the AI on new patients in real time. No one has yet run a trial proving that acting on the AI’s warning actually prevents deaths. That next step, a randomized study comparing AI-guided care against standard practice, is the test the finding still needs to pass before it changes how your cardiologist reads your chart.

Still, the pattern fits a field that has been building toward this moment for years. Reviews of AI-driven ECG analysis have already shown these systems can outperform older risk scores at spotting sudden cardiac death risk, precisely because standard tools like LVEF were never built to catch the specific electrical glitches that trigger a fatal rhythm. This study adds a large, real-world dataset and a newly named warning sign to that growing case.

For now, the headline takeaway is straightforward and well-earned: a common, cheap, decades-old test may be hiding a death-predicting signal that doctors have simply never been trained to see. Turning that discovery into a bedside tool that saves lives is the work still ahead.

Sources:

pmc.ncbi.nlm.nih.gov, pubmed.ncbi.nlm.nih.gov, scientificamerican.com, news.berkeley.edu