A new UC Berkeley-led study discovered a physiological signal to predict sudden cardiac arrest with the help of AI, potentially saving thousands of lives.
About 350,000 people in the U.S. experience sudden cardiac arrest annually, and nearly 90% of them die. Cardiac death is, in theory, preventable with a medical device called a defibrillator. Nevertheless, patients die from the condition each year because clinicians fail to predict those at risk.
In the new study, researchers trained an AI model using more than 440,000 electrocardiograms — a medical test that records the heart’s electrical activity — along with death certificates and electronic health records. In the process, they discovered a new physiological sign associated with the risk of sudden cardiac arrest.
According to the study, the standard measure for identifying sudden cardiac arrest is prone to error: It misses most sudden cardiac deaths and can flag low-risk patients, leading to unnecessary and expensive defibrillator installations. Only about one-third of implanted defibrillators ever deliver therapy.
Researchers tested the AI model on thousands of patient files from San Diego and Taiwan, and found that the algorithm isolated a higher-risk group more than the standard test. In fact, more than 80% of the high-risk patients flagged by the AI model were not flagged by the standard clinical test.
Among the population of high-risk patients that the AI model flags, only about 7% are predicted to experience sudden cardiac arrest. According to UC Berkeley Ph.D. student Alexander Schubert, who worked on the study, while most flagged patients wouldn’t need immediate defibrillator implantation, it can help doctors monitor these patients more closely.
Flagging high-risk patients ahead of time can also open the door to more research examining what the heart looks like before sudden cardiac arrest — something not shown in an autopsy, according to the study.
“These patients are hard to study because they look fine and then they drop dead,” Schubert said. “It’s a very hard-to-study population compared to, for instance, cancer patients, who you observe while they go through chronic disease.”
Markus Lingman, senior author in the study and chief medical intelligence officer at Sahlgrenska University Hospital, said in an email he hopes the study can “inspire others to use AI to identify risk and support thinking about causality.”
Researchers are now using the algorithm to scan hospital EKG databases in Sweden, Taiwan and the United States. Patients identified as high-risk are offered an EKG patch that collects more data, helping researchers better understand the newly discovered physiological mechanism — and potentially leading to defibrillator placement.
UC Berkeley public health professor Ziad Obermeyer, lead researcher on the study, also built a website where individuals who want to assess their risk for sudden cardiac arrest can join an interest list to be scanned by the AI system once it becomes publicly available.