Skip to content

Can a Routine Heart Test Reveal Who Is at Risk of Sudden Cardiac Death?

In Simple Terms

A quick, everyday heart test may hold clues that someone could be at risk of dying suddenly from a heart problem—even when standard checks look reassuring. Researchers trained a computer to spot patterns in hundreds of thousands of test recordings. The tool might help doctors decide who needs closer follow-up, but it cannot predict an individual’s future with certainty or replace a doctor.

Hidden Clues in an Everyday Test

An electrocardiogram, usually called an ECG or EKG, records the electrical signals that coordinate the heart’s contractions and relaxation. The result appears as waves and peaks, but it is more than a collection of separate lines. Subtle relationships and changes within the recording may contain information about the heart that is easy to miss when a clinician reads it in the usual way.

For this study, researchers trained a computer program to look for patterns in those signals that were associated with sudden cardiac death later on. Rather than simply comparing people who appeared healthy with patients who had a known heart diagnosis, they linked ECG recordings to death-certificate information. This allowed the system to learn from patterns found in people who later died suddenly from cardiac causes.

The research was led by Ziad Obermeyer, an associate professor at the University of California, Berkeley’s School of Public Health. An emergency physician, Obermeyer studies the intersection of medicine, artificial intelligence, and health policy. The findings were published in Nature. Obermeyer is also part of the Computational Precision Health program, a joint initiative of UC Berkeley and the University of California, San Francisco.

The signal the system learned to recognize does not necessarily depend on a clear abnormality that a doctor could readily spot on an ECG. Instead, the program considers many electrical features together and may identify a pattern in how they combine that points to elevated risk. Researchers do not yet know exactly which biological changes produce that pattern.

Cardiac Arrest Is Not a Heart Attack

To understand why prediction matters, it helps to distinguish a heart attack from sudden cardiac arrest. A heart attack happens when blood flow to part of the heart muscle is reduced or blocked. Cardiac arrest, by contrast, occurs when the electrical system coordinating the heartbeat malfunctions, leaving the heart unable to pump blood effectively.

Cardiac arrest can happen without an obvious warning, making the moments before it difficult to observe and study. According to information in the study, about 90% of people who have sudden cardiac arrest outside a hospital die within minutes. Cardiopulmonary resuscitation and an external defibrillator can sometimes restore a normal rhythm, but the opportunity to intervene shrinks sharply when the danger is unexpected or treatment is not available quickly.

Sudden death also often leaves researchers unable to follow what is happening in the heart just before it stops. An autopsy may reveal a blocked artery, scarring, or other structural changes, but it cannot reconstruct the heart’s electrical activity in the minutes before death. Finding early warning signs in people who are still alive therefore remains an important medical challenge.

Other factors make prediction difficult, too. Cases may differ in their causes and features, and people who die may never have received a specialist heart assessment. Yet routine records, when analyzed at scale, may offer a chance to detect signals that were already present in medical data but had not been the focus of a targeted search.

What Standard Assessments May Miss

To estimate some patients’ risk of sudden cardiac death, doctors measure how much blood the heart pumps with each contraction. This helps assess the heart’s pumping ability and informs decisions about whether some patients might be candidates for an implantable cardioverter-defibrillator, or ICD. The implanted device can deliver an electric shock if it detects a dangerous heart rhythm.

Assessing the heart’s pumping ability usually requires a broader medical evaluation than a routine ECG. Many people may never receive that kind of testing, particularly if they have no symptoms or known reasons for a doctor to suspect a problem. As a result, important clues could go undetected in people who never enter the pathway for specialist assessment.

At the same time, identifying someone as high-risk does not mean an implanted device will necessarily be used. According to the research, about two-thirds of patients who received defibrillators never needed the device to deliver a shock. Implantation is an expensive, invasive procedure, so more accurately identifying those most likely to benefit could help avoid unnecessary treatment without overlooking people who need it.

This creates a screening dilemma. If the criteria are too narrow, people at risk may be missed. If the pool of potential treatment candidates is widened on the basis of uncertain indicators, others may undergo a procedure they do not need. The value of AI, then, is not in making a final decision. It may instead improve the process of sorting patients and identifying who merits further assessment.

From Hundreds of Thousands of ECGs to an International Test

The system was first trained on ECG records collected over six years within Sweden’s unified health care system. The training database contained more than 440,000 recordings. By linking them to death-certificate data, the researchers could compare electrical patterns across different groups, including people who later died suddenly from cardiac causes.

The program learned from thousands of details in the recordings, including the shapes and sequences of waves and the relationships among their components. This approach does not require researchers to decide in advance that any one feature should count as a warning sign. Instead, the system can search for combinations of characteristics that recur alongside the outcome it is designed to predict.

After training, the team tested the tool on independent data from outside Sweden, an important step in checking whether its performance would extend beyond the patient population it learned from. The tests included thousands of additional records from the United States and Taiwan. The U.S. data included ECGs collected over two years from a hospital system in San Diego, as well as a separate set of patient records from Taipei.

Testing across countries and health care systems helps researchers assess whether results may generalize, but it does not prove that performance will be the same in every population or patient group. Differences in the people being studied, the way records are collected, and the care available could all affect how a model performs. That is why testing in multiple settings matters before considering widespread clinical use.

According to Obermeyer, collecting and organizing the data took nearly ten years. Two organizations he helped establish, Dandelion Health and Nightingale Open Science, contributed to the effort. The timeline highlights that developing a medical algorithm involves more than writing software: making records available, linking them reliably, and preparing them for research can itself be a lengthy undertaking.

What Does a 7% Rate Mean?

When the researchers compared the results, the group identified by traditional methods had an annual rate of sudden cardiac death of 4.6%. In the group the AI system identified as being at higher risk, the annual rate was 7%. In other words, the system identified a group in which the recorded risk was higher than in the group flagged by standard assessment methods.

These figures describe annual rates across groups; they do not mean the system can determine what will happen to any one person. The difference between 4.6% and 7% does not make the prediction certain, nor does it prove that everyone in the higher-risk group should automatically receive an implanted device. Rather, it suggests that ranking people by risk could become more accurate, helping a doctor decide what to do next.

One notable finding was that the algorithm also flagged a broader group of people at elevated risk, including thousands who would not ordinarily be considered particularly concerning under current medical guidelines. Most of them would have been classified as relatively low-risk using traditional methods. This raises the possibility that some risk is hidden in everyday records that do not stand out when analyzed in the usual way.

But broadening the pool of people identified as potentially at risk should not automatically mean more people receive treatment. Further testing, continued monitoring, and the decision to implant a device are separate steps, each requiring stronger evidence. The tool’s first practical role may therefore be to direct some patients toward closer observation—not to send them straight from an algorithm’s result to surgery.

Turning a Warning into Follow-Up Care

Obermeyer and his team are working with health systems in Sweden, Taiwan, and the United States to integrate the algorithm into hospital ECG databases. The proposed approach is to screen recordings that already exist for indicators worth investigating, then contact patients whose estimated risk is high and offer them additional follow-up.

One possibility is to ask some patients to wear a small patch that continuously monitors the heart’s electrical activity. Unlike an ECG that records activity for a limited period, longer monitoring could capture intermittent rhythm problems or changes that might not show up in a brief recording. Such data could give doctors and researchers a clearer picture of the signal the AI detected.

If extended monitoring reveals relevant patterns, it could help researchers understand the electrical or biological conditions that make some hearts more vulnerable to a fatal rhythm disturbance. Once enough evidence has been collected, the findings might inform medical decisions about preventive treatments, including an implanted defibrillator for people shown to be most likely to benefit.

These steps are still part of developing a clinical use for the system. They are not an announcement that the test is broadly available or that a treatment has been shown to be right for everyone the program places in its higher-risk group. Moving from a prediction tool to care that improves outcomes will require verifying the accuracy of its predictions, deciding what should happen when it flags someone, and measuring the real-world benefits and risks for patients.

The Mystery the AI Uncovered

The study’s importance goes beyond sorting patients by risk. The algorithm detected electrical patterns that researchers believe are linked to danger, but their biological meaning is not yet fully understood. That opens up a scientific question: what is happening in the heart that leaves a detectable trace on a routine ECG before a fatal rhythm disturbance occurs?

Understanding these patterns could reveal previously unknown changes or improve how the heart is assessed. If researchers can connect a particular signal to a specific mechanism, they could move from observing that some ECGs predict risk to explaining why those patterns appear—and perhaps identifying new opportunities for prevention.

This possibility points to a different role for AI in medicine. It can do more than speed up the reading of a test or automate a routine task. It may help suggest complex relationships within large volumes of data, which scientists can then examine through experiments and clinical observations. Still, an algorithm is a tool for finding patterns; a medical explanation does not become fact simply because a program finds an association with a health outcome.

For this approach to be reliable, doctors need to understand what a result means for the patient and have clear, proportionate steps to follow. Researchers must also confirm that the system works well across diverse groups and that relying on it does not create false reassurance, needless anxiety, or unnecessary testing.

The Bottom Line

A study from the University of California, Berkeley, suggests that a routine ECG may contain subtle clues about a person’s risk of sudden cardiac death. An AI system learned from more than 440,000 ECGs collected in Sweden over six years and was then tested on thousands of additional records from the United States and Taiwan. It identified a group with an annual death rate of 7%, compared with 4.6% in a group flagged by traditional methods, and also highlighted people standard assessments would often consider to be at lower risk.

The findings do not determine what will happen to any individual, and they do not mean that everyone the program flags should receive an implanted defibrillator. But the approach could help doctors decide who needs closer follow-up, such as longer-term monitoring with a heart-recording patch. If future testing confirms that it is useful and safe, the method could help prevent some deaths while also revealing more about the signals that precede a fatal heart rhythm problem.