AI finds hidden ECG signals that may point to risk of sudden cardiac arrest

Researchers at the University of California, Berkeley (UC Berkeley) have developed an artificial intelligence system that uses ordinary electrocardiograms (EKGs), which record the heart's electrical activity, to identify people at elevated risk of sudden cardiac arrest. The findings were published in Nature. More than 300,000 people in the United States experience sudden cardiac arrest each year, when the heart's electrical system suddenly fails and the heart stops pumping blood.

How the system was trained

The algorithm was trained on more than 440,000 EKG recordings from Sweden's unified healthcare system, covering six years. The scans were linked to death certificates so the system could distinguish healthy people, patients with known heart risks and those who later died of sudden cardiac causes. It was then tested on two years of anonymized EKG records from a San Diego hospital system and a separate patient dataset from Taipei.

Comparison with current practice

Doctors currently rely mainly on how much blood the heart pumps with each beat when deciding whether a patient needs an implantable defibrillator. This approach identifies a group with an annual sudden cardiac death rate of 4.6%, while the AI identified a group with a rate of 7%. The algorithm also flagged a broader group at elevated risk, including thousands of people who would be considered low risk under current guidelines. The system works with standard EKG images that hospitals already collect.

The present approach has drawbacks too: about two-thirds of implanted defibrillators in high-risk patients never need to deliver a shock. Meanwhile, roughly 90% of people who suffer cardiac arrest outside a hospital die within minutes.

Next steps

Lead author Ziad Obermeyer, an associate professor at UC Berkeley's School of Public Health and an emergency physician, is working with medical systems in Sweden, Taiwan and the United States to introduce the algorithm into hospital EKG databases. Patients with a high-risk result could be offered a patch that continuously records the heart's electrical activity. That data could help explain which changes in the heart produce the signals the AI recognizes. For some patients, the extra information might support a recommendation to receive a defibrillator.