Skip to main content
Science Catalogs
Health Sciences

Deep learning finds an ECG marker of sudden cardiac death risk missed by standard tests

Researchers at the University of California, Berkeley have used deep learning to find an electrical signature in routine electrocardiograms (ECGs) that is linked to sudden cardiac death. In the study, published in Nature, the algorithm picked out a group of patients with a higher annual rate of sudden cardiac death than the group identified by the standard measure of cardiac pumping function. It also flagged many people who would usually be considered low risk.

The clinical problem

Sudden cardiac arrest, which stems from failure of the heart's electrical system rather than a blocked artery, causes more than 300,000 deaths a year in the United States. Roughly 90% of people who have an arrest outside hospital die. Implantable cardioverter defibrillators can prevent some of these deaths, but selecting recipients is difficult.

Current practice relies heavily on a measure of how much blood the heart ejects with each beat. Obtaining it usually requires more extensive cardiac assessment, which many people who later suffer an arrest never receive. Among high-risk patients who do receive a device, about two-thirds never need a shock. Meanwhile, many deaths occur in people who were never identified as at risk.

How the model was built and tested

The team, led by Ziad Obermeyer of Berkeley's School of Public Health, trained a neural network on more than 440,000 ECGs from Sweden's unified healthcare system, covering six years. The recordings were linked to death certificates, allowing the system to learn differences between healthy people, patients with known cardiac risk, and those who later died of sudden cardiac causes.

The model was then evaluated on two further datasets: two years of anonymised ECG records from a hospital system in San Diego, and patient data from Taipei. The authors on the paper are Obermeyer, Alexander Schubert, James Ross, Sendhil Mullainathan and Markus Lingman.

What the results show

Using the conventional pumping measure, clinicians can identify a group with an annual sudden cardiac death rate of 4.6%. The AI identified a group with a rate of 7%. It also recognised a larger population at elevated risk, including thousands of people who would not normally attract concern under current guidelines and who would mostly have been rated low risk by standard methods.

Because the model works from ordinary ECG images, which hospitals already collect widely, it could in principle be applied to existing records without a new type of test. The source material does not give detailed performance metrics beyond the death rates above, and the results come from retrospective datasets rather than a prospective trial.

Open questions and next steps

The biological meaning of the signal is not yet understood. The algorithm has learned patterns that researchers cannot yet link to specific changes in the heart, which makes the finding a prompt for physiological work as well as a possible screening tool.

Obermeyer is working with health systems in Sweden, Taiwan and the United States to run the algorithm over existing ECG databases. Under the proposed approach, patients with high-risk readings would be contacted and offered a wearable patch that records heart activity continuously. Those recordings could shed light on the mechanisms behind the signal and, for some patients, support a decision to implant a defibrillator.

The data behind the study took about ten years to assemble, with help from two organisations Obermeyer helped establish, Dandelion Health and Nightingale Open Science. He points to limited access to high-quality clinical data as a major reason so little clinical AI is currently in use.

Key questions remain: whether the model's flagged patients benefit from monitoring or devices, how it performs across more diverse populations, and what electrical or structural changes it detects. Prospective studies will be needed to answer them before the tool could influence guidelines.

Source: AI discovers a hidden heart signal that could predict sudden death (ScienceDaily: health and medicine, 15 November 2026).

Image: AI-generated illustration.

Working on this topic? Find a journal for your research or join the discussion on Scholars Post.