AI Finds a Hidden Heart Signal That Could Predict Sudden Death

Key Takeaways

  • Researchers at the University of California, Berkeley, working with colleagues in Sweden, built an AI system that detects a previously unrecognised electrical signal in routine heart recordings.
  • The study, published in the journal Nature on 24 June 2026, used more than 440,000 electrocardiograms (ECGs) from about 180,000 patients in Sweden, matched to national death certificates.
  • The AI identified a small high-risk group, about 2.2 per cent of patients, whose rate of sudden cardiac death within a year was 7 per cent, compared with 4.6 per cent among patients flagged by the standard ultrasound test.
  • The researchers then trained a second AI model to show what the first one was seeing. It revealed a subtle slurring near the end of the R wave in one ECG lead, called aVL, a feature never before described in the medical literature.
  • The team suggests the signal may be linked to myocardial fibrosis, the build-up of collagen between heart muscle cells, but that is a hypothesis, not a confirmed mechanism, and it has not been confirmed with biopsies.
  • The technology is not yet a standard part of medical care. It is a diagnostic aid being prepared for clinical use, not a standalone treatment, and it does not replace evaluation by a cardiologist.

What the Researchers Did

Sudden cardiac death is one of the most feared events in medicine, because it often strikes without warning. A person’s heart suddenly stops beating effectively, usually because of a dangerous disturbance in its electrical rhythm, and without immediate treatment the outcome is fatal. Doctors have long wished for a way to identify, in advance, the people most at risk, so that they could be monitored or given a preventive treatment such as an implantable defibrillator.

The problem is that the people at risk often look healthy. Their hearts may appear normal on standard tests, and the existing screening methods miss many of them. The standard approach uses an ultrasound test to measure how well the heart pumps, and flags people whose pumping function is reduced. But many people who go on to suffer sudden cardiac death have normal pumping function, and they are not flagged at all.

A team led by Ziad Obermeyer, an associate professor at the University of California, Berkeley, asked a different question. Instead of looking for a better version of the existing test, they asked whether a routine electrocardiogram, the familiar ECG trace that records the heart’s electrical activity, already contains information about future risk that human eyes cannot see.

An ECG is cheap, fast and nearly universal. It is a ten-second recording of the heart’s electrical activity, taken with electrodes on the skin. Cardiologists read ECGs every day, looking for patterns that indicate disease. But the trace contains far more information than a human reader can extract by eye, and the Berkeley team wanted to know whether a machine could find something in it that no one had noticed.

To answer that question, they assembled one of the first population-scale datasets of its kind. They used more than 440,000 ECGs from roughly 180,000 patients in Sweden, collected over six years through the Swedish health care system, and matched them to national death certificates. That linkage is what makes the study powerful. It is not enough to show that an AI can find a pattern in an ECG. You need to know what happened to the people afterwards, and the death certificates provided that answer.

The work was published in Nature on 24 June 2026, under the title “An ECG biomarker for sudden cardiac death discovered with deep learning”. The authors are Ziad Obermeyer, Alexander Schubert, James Ross, Sendhil Mullainathan and Markus Lingman. The study’s reference number, its digital object identifier, is 10.1038/s41586-026-10674-6.

What the AI Found

The team trained a neural network to predict the risk of sudden cardiac death from a ten-second ECG. The model they chose was a 64-layer residual neural network, known as a ResNet, a type of deep learning model widely used for pattern recognition. It was not designed specifically for cardiology. It was a general-purpose model, applied to a medical question.

The principal analysis examined 113,072 ECGs from 35,417 people younger than 80 who did not already have defibrillators. That is the group for whom the question matters most: people who have not already been identified as high risk and given a device.

The model identified a high-risk group amounting to about 2.2 per cent of the ECG observations. Within that small group, the observed rate of sudden cardiac death within one year was 7.0 per cent, with a 95 per cent confidence interval of 4.9 to 9.5 per cent. That is the range within which the true rate is likely to fall, given the size of the sample.

That figure is worth comparing with the standard test. Among patients flagged by the usual ultrasound assessment, the rate of sudden cardiac death was 4.6 per cent. The AI’s high-risk group had a higher rate than that, and it included many people who were not already flagged by reduced cardiac pumping function. In other words, the AI found people the existing test missed, and the people it found were at greater risk than the people the existing test catches.

The signal held up when the team tested the model on separate datasets from the United States and Taiwan, suggesting it was not a quirk of Sweden’s population or ECG equipment. Reporting on the study puts the model’s accuracy at 87.2 per cent in the primary validation cohort, 82.2 per cent in the US cohort and 76.7 per cent in the Taiwan cohort, although those specific accuracy figures come from secondary reporting rather than from the study’s own summary, and we have not treated them as the study’s headline result. The figures that are firmly established are the 7 per cent and 4.6 per cent rates, and the 2.2 per cent size of the high-risk group.

It is worth pausing on what those numbers mean, because they are easy to misread. A 7 per cent annual rate of sudden cardiac death is high in medical terms, but it also means that 93 per cent of the people in the high-risk group did not suffer sudden cardiac death within a year. The AI does not predict that a particular person will die. It identifies a group in which the risk is substantially higher than average, so that doctors can pay closer attention to the people in it. That distinction matters, and it is one the researchers themselves have been careful to make.

The Second Model, and the Visible Clue

Finding a pattern is one thing. Understanding it is another, and the second part of this study is what makes it unusual.

Most AI models in medicine are black boxes. They make predictions, but they cannot explain how. A doctor who is told that a patient is at high risk, without being told why, has a difficult decision to make. Do they trust a number they cannot check? That problem, known as the explainability problem, has limited the use of AI in clinical practice, because doctors are rightly reluctant to act on predictions they cannot understand or verify.

Obermeyer’s team tried a different approach. They trained a second neural network, a generative model, whose job was to reveal what the first model was seeing. The generative model’s task was to produce ECG waveforms that looked high risk to the first model, and to show how a waveform changed as predicted risk rose.

The method worked like this. The generative model took a real ECG from a low-risk patient and reworked it step by step, morphing it smoothly into a high-risk version of the same trace. Because the starting point was a real patient’s recording, and the changes were made gradually, the researchers could watch which features of the waveform changed as the risk score increased. That gave them something a single prediction could not: a visible feature they could point to and investigate.

One feature stood out. It had never been described in the medical literature: a subtle slurring near the end of the R wave in one ECG lead, called aVL, which stands for augmented vector left. A lead is a measurement from a particular set of electrodes, giving a particular view of the heart’s electrical activity. Lead aVL looks at the heart from the left side, and the slurring in that view suggested that the heart’s electrical signal was fragmenting as it moved through the muscle.

The researchers quantified the shape by measuring how voltage changed from the peak of the R wave to the end of the QRS complex, the part of the ECG that represents the main electrical impulse through the heart’s lower chambers. That measurement gave them a concrete, checkable feature, one a cardiologist could learn to spot, rather than an abstract risk score.

It is important to be precise about what the visible feature is and is not. The researchers were clear that the slurring in lead aVL represented part of the model’s signal, rather than a substitute for the full model. The complete pattern the AI detects is more complex than a single waveform feature, and the visible clue is a window into it, not a replacement for it. A cardiologist who learned to spot the slurring would catch some of the signal, but not all of it.

What Might Cause the Signal

Finding a pattern is not the same as understanding its cause, and the researchers have been careful about the gap between the two. The underlying physiological mechanism behind the signal remains unknown, and the team has proposed a hypothesis rather than a conclusion.

Their suggestion is myocardial fibrosis, the deposition of collagen, a structural protein, between heart muscle cells. Fibrosis creates small, widely distributed obstacles to electrical conduction, and those obstacles could cause the advancing electrical wave to scatter and change direction, producing the observed shape in the ECG. It is a plausible explanation, because fibrosis is known to interfere with the heart’s electrical system, and because the fragmented, slurred waveform is consistent with a signal that has been disrupted as it travels.

But the link is preliminary. Obermeyer has cautioned that the fibrosis hypothesis has yet to be confirmed with biopsies, which would involve examining actual heart tissue to look for the collagen deposits the hypothesis predicts. Until that work is done, fibrosis remains a testable idea, not an established cause.

The distinction matters for a practical reason. If the signal is caused by fibrosis, then it might be possible to treat the underlying condition, or at least to understand which patients are most likely to benefit from a defibrillator. If it is caused by something else, the treatment implications could be different. The researchers’ next steps include the kind of work that could answer that question, but the answer is not yet in hand.

There is also a subtle point about how the waveform work was done. The exploratory waveform and MRI work used the original training and validation datasets, rather than the untouched dataset used for the final predictive evaluation. That is a methodological detail, but it matters, because it means the visible feature was derived from data the model had already seen, not from a fully independent test. The researchers disclosed this, and it does not invalidate the finding, but it is the kind of detail that careful readers should know about.

Why This Study Stands Out

There are many AI studies in medicine, and most of them follow a familiar pattern: a model is trained on a dataset, it performs well, and the authors suggest it could help doctors. This study is different in a way that has attracted attention from researchers who were not involved in it.

Changxin Lai, a biomedical engineer who wrote an accompanying analysis in Nature and peer-reviewed the study, has said that this is why the work stands out. The two-model setup points to a larger ambition for AI in medicine: getting a machine to surface a fresh clue that human experts can then see and check for themselves. Instead of asking doctors to trust a black box, the study gives them something visible, a feature on an ordinary ECG that a cardiologist could learn to spot.

That ambition matters because it addresses the central obstacle to AI in clinical practice. Doctors do not want tools they cannot understand, and patients do not want decisions made by systems no one can explain. A model that produces a checkable finding is far more useful than a model that produces only a number, however accurate the number might be.

The study also stands out for the scale and quality of its data. More than 440,000 ECGs from about 180,000 patients, matched to death certificates, is an unusually strong foundation, and the fact that the signal held up in datasets from the United States and Taiwan gives it a robustness that single-country studies lack.

The Limits, and What It Does Not Mean

The study’s findings are significant, but they are not a clinical tool yet, and it is important to say plainly what the research does not establish.

First, the technology is at the stage of clinical integration, not standard care. The researchers are working on implementing the algorithm in clinical settings in different countries, but it is not yet part of routine practice, and there is no date for when it might be.

Second, the study does not prove that treatment selected using the model would improve outcomes. Identifying people at higher risk is not the same as showing that acting on that information saves lives. The authors’ analysis suggests a potential benefit, but the step from prediction to improved outcome requires a different kind of study, one that tests whether using the model to guide treatment actually helps patients. That study has not been done.

Third, the fibrosis link is a hypothesis, not a finding. Until it is confirmed with tissue samples, it should be treated as an idea to be tested, not as an explanation.

Fourth, the accuracy figures reported in secondary coverage, including 87.2 per cent in the primary validation cohort, should be treated with care. Accuracy in this context depends on how the threshold is set, and a single percentage does not capture the trade-off between catching people at risk and falsely flagging people who are not. The more robust figures are the ones we have emphasised: the 7 per cent annual rate in the high-risk group, the 4.6 per cent rate among those flagged by the standard test, and the 2.2 per cent size of the group.

Fifth, the study population was people younger than 80 without defibrillators, in a specific set of countries. How the model performs in other populations, in older patients, or in people with different patterns of heart disease, is not yet established.

None of these limits diminish the finding. They define it. A study that identifies a new risk signal, makes it visible, and proposes a testable explanation is a significant contribution, even if it is not yet a clinical tool. The honest way to report it is to say what it found, and to say what it has not yet shown.

What This Means for Patients and Doctors

For now, the practical implications are limited, and it would be misleading to suggest otherwise. No patient should expect their next ECG to be run through this system, and no doctor should change a treatment decision based on it. The work is at the research stage, and the path from research to routine care involves validation, regulatory review and clinical trials that take years.

But the direction of travel is clear, and it is worth understanding. If the signal holds up in further testing, and if the fibrosis hypothesis is confirmed, then a routine, cheap, widely available test could identify people at high risk of sudden cardiac death who are currently missed. Those people could then be monitored more closely, or considered for a defibrillator, which is a device implanted in the chest that can restart the heart if it stops.

That possibility is why the study has attracted attention beyond the research community. Sudden cardiac death is a major cause of death worldwide, and a tool that could identify even a fraction of the people at risk, using a test that is already performed millions of times a year, would be a significant advance.

There is also a broader lesson. The ECG is one of the oldest tools in medicine, and it has been read the same way for decades. This study suggests that it contains information no one knew was there, and that AI can find it. That raises a question about what other familiar tests might be hiding, and whether the next advances in medicine will come not from new machines, but from looking more carefully at the data the old ones already produce.

For readers interested in how AI is being used in health care more broadly, our earlier coverage provides useful context. Our guide to Heidi AI for doctors explains one of the AI tools already in clinical use, our article on AI scribe errors in health care sets out what can go wrong when AI is used in medical settings, and our piece on Vera Health’s AI assistant covers the funding and adoption of AI tools for clinicians.

What We Could Not Verify

  • The study itself was published in Nature on 24 June 2026, and we have verified its title, authors, journal and digital object identifier. The renewed attention to it in November 2026, including reporting by Science Daily, is the reason for this article, but we could not verify the exact date or content of the Science Daily report from the sources we reviewed, and we have not stated it as fact.
  • The accuracy figures of 87.2 per cent, 82.2 per cent and 76.7 per cent for the primary, US and Taiwan cohorts come from secondary reporting, not from the study’s own published summary as we reviewed it. We have reported them with that attribution and have not treated them as the study’s headline result.
  • The claim that using the model to select patients for treatment would reduce sudden cardiac death by a specific amount is based on the authors’ modelling, not on a clinical trial. We have not stated a specific reduction figure, because the context of the reported figures was not clear enough to reproduce accurately.
  • We could not verify the current status of the researchers’ work to implement the algorithm in clinical settings, or the timetable for that work. We have described it as ongoing, without giving a date.
  • We could not verify the fibrosis hypothesis, which the researchers themselves describe as preliminary and unconfirmed by biopsy. We present it as a hypothesis throughout.
  • The study’s findings apply to the population studied, people younger than 80 without defibrillators, in Sweden, the United States and Taiwan. We have not generalised beyond that population.

Sources

Similar Posts