Researchers have developed artificial intelligence (AI) models that can scrutinize electronic health records (EHR) and electrocardiograms to identify individuals in the general population at elevated risk for sudden cardiac arrest — a condition that causes more than 400,000 U.S. deaths annually and has a survival rate of only 10%.
The finding represents a significant advance in predicting a largely unpredictable medical emergency that often strikes people with no known heart disease.
"Using artificial intelligence applications and health records data, the prediction of cardiac arrest in the general population is feasible," said Dr. Neal Chatterjee , the study's lead investigator and a cardiologist at the University of Washington School of Medicine.
JACC: Advances, a journal of the American College of Cardiology, published the paper today. Other co-senior authors are from Massachusetts General Hospital and the Broad Institute of MIT and Harvard.
The researchers' test population comprised approximately 1.7 million patients in a large U.S. healthcare system. Three AI models were developed with discrete datasets: "EKG-only," "EHR-only" (weighing 156 clinical features of patients' records), and a combined model that integrated EKG and EHR data sets.
The investigators developed and validated the AI model with three distinct patient groups:
Training cohort: Researchers trained the models with data from 993 individuals who had experienced out-of-hospital cardiac arrest between 2013 and 2021 and 5,479 age- and sex-matched control patients who did not. This group taught the AI models to recognize the patients' EHR data entries and EKG readings that were associated with higher cardiac arrest risk.
Testing cohort: To verify that the AI models accurately distinguished factors of high and low risk, the researchers applied them to a separate group of 463 cardiac arrest cases from 2022-2023 and 2,979 control patients. The test showed that the models' risk associations aligned similarly with those established by the training cohort.
Real-world cohort: This group included 39,911 individuals who had received EKGs during 2021, regardless of their health status. Researchers analyzed the records of the subset of those patients who experienced cardiac arrest during the next two years, to discern how well they aligned with the risk profiles established by the AI models, Chatterjee explained.
With the real-world cohort, the combined EHR-EKG model correctly predicted 153 of 228 people who were high-risk and who went on to experience a cardiac arrest.
"With these models, we're able to enrich risk prediction from about 1 in 1,000 down to 1 in 100," Chatterjee said. "If your doctor were to tell you that your risk of cardiac arrest is 1 in 100, that would catch your attention. "We're bringing a theoretical risk into focus."
Another promising finding was that AI-enhanced EKG analysis alone showed strong predictive ability, only modestly lower than the two models that incorporated EHR data.
"The 12-lead EKG is a low-cost tool that might stratify patients' risk for cardiac arrest in any community around the world," Chatterjee said.
The study also identified cardiac-arrest risk features outside of those typically associated with cardiovascular disease. Among these contributors were electrolyte disorders, substance use and medication interactions.
"We show some relatively low hanging fruit … modifiable risk factors," Chatterjee noted. "A model that flags a patient as high-risk might prompt somebody taking care of a patient to review their medical history and their medications."
While the findings demonstrate feasibility of risk prediction, Chatterjee emphasized that more study is needed to determine best clinical responses when a model flags a patient's elevated cardiac-arrest risk.
"We need to figure out which follow-on studies to pursue to understand what we do with this patient information. What screening, what surveillance, what intervention is warranted?"
The study has important limitations. All data came from a single healthcare system, and generalizability to other populations with different demographics and care patterns is unknown. The real-world cohort was limited to individuals who received an EKG, who may differ from those not undergoing EKG evaluation. The AI-enhanced EKG representations could potentially reflect biases linked to demographics and healthcare patterns.
The research was supported with funding from the National Institutes of Health (K23HL169839, R01 HL160003, R01 HL168889, K24 HL153669, R01HL092577, R01HL157635), the American Heart Association (23CDA1050571, 961045), the European Union (MAESTRIA 965286) and from the Foundation Leducq (24CVD01). Chatterjee is supported by philanthropic donation of Kevin and Ann Harrang and the John and Cookie Laughlin Endowed Professorship.
The authors' conflict-of-interest statements are in the published paper, which will be provided to journalists upon request.
