AI Finds Hidden Heart Signal in Routine EKGs That May Flag Sudden Cardiac Death Risk

Most people who die of sudden cardiac arrest outside a hospital are gone within minutes. Roughly 90% do not survive. Many of them never had a reason to suspect anything was wrong, and that is the problem a team at UC Berkeley set out to attack, using a tool already sitting in hospital records: the routine electrocardiogram.

The researchers trained an artificial intelligence system on more than 440,000 EKGs collected in Sweden. They linked those scans to death certificates, which let the system compare the electrical patterns of healthy people, patients with known heart risks, and people who later died of sudden cardiac causes. The training data covered six years of recordings from Sweden's unified healthcare system. The work was published in Nature.

To see whether the patterns held up elsewhere, the team tested the trained system on two further sources: two years of anonymized EKG records from a San Diego hospital system, and a separate set of patient data from Taipei. The aim was to check that what the AI had learned was not peculiar to the Swedish population.

The comparison with current practice is the heart of the result. Doctors now rely largely on a measure of how much blood the heart pumps with each contraction, which helps decide who might benefit from an implantable cardioverter defibrillator, a device that shocks the heart out of a dangerous rhythm. That measure identifies a group with an annual sudden cardiac death rate of 4.6%. The AI identified a group with a rate of 7%. It also flagged a broader population at elevated risk, including thousands of people whom existing guidelines would mostly have classed as low risk.

The current approach has costs on both sides. Getting the pumping measurement generally requires a more extensive medical evaluation, and many people who go on to arrest never receive one. Among patients identified as high risk, about two-thirds of implanted defibrillators never need to deliver a shock, so many people undergo expensive, invasive procedures for an emergency that never comes. Meanwhile, thousands die unexpectedly each year without ever being recognized as at risk. Ziad Obermeyer, the study's lead author, an emergency physician and associate professor at UC Berkeley's School of Public Health, put it this way: "If you knew you were one of the people who was going to drop dead, you would go to a cardiologist and you'd get a defibrillator implanted. The problem is that doctors can't figure out who needs one before it's too late."

The data took about 10 years to assemble, with help from two organizations Obermeyer helped establish, Dandelion Health and Nightingale Open Science. He said data like this is incredibly hard to access, which he cited as a big reason so little clinical AI is in use today. The system works on ordinary EKG images, so it could improve screening without a new kind of test.

Obermeyer is now working with medical systems in Sweden, Taiwan and the United States to run the algorithm across hospital EKG databases. Under the proposed approach, physicians would contact patients whose scans come back high risk and offer them a patch that continuously records the heart's electrical activity. Those longer recordings might reveal the biology behind the signal, and for some patients could support a recommendation for a defibrillator. He has also set up a website where people can leave basic information and an email address to be contacted when EKG analysis becomes more widely available.

One piece remains unexplained: the researchers do not yet fully understand what changes in the heart produce the pattern the AI has learned to recognize. The patch recordings are the next step toward answering that.