AI Tool Predicts Risk of Out-of-Hospital Cardiac Arrest

Sudden cardiac arrest is a lethal event that often occurs without warning, causing more than 400,000 deaths in the U.S. each year and a survival rate of about 10%. Clinicians struggle to identify who in the general population is at greatest risk before collapse occurs. Forecasting risk accurately could enable earlier surveillance and preventive care. To help address this challenge, researchers have developed artificial intelligence (AI) models that scan electronic health records and 12-lead electrocardiograms to flag individuals at elevated risk.

Investigators at the University of Washington School of Medicine (Seattle, WA, USA) developed three AI models to estimate future risk of out-of-hospital cardiac arrest. The models were built separately on electrocardiogram (EKG)-only data, electronic health record (EHR)-only data that weighed 156 clinical features, and a combined EHR–EKG approach. Co-senior authors were from Massachusetts General Hospital and the Broad Institute of MIT and Harvard, and the findings were published in JACC: Advances on May 12, 2026.

Model development and validation used three patient cohorts drawn from a large U.S. health system. The training cohort included 993 people with out-of-hospital cardiac arrest from 2013 to 2021 and 5,479 age- and sex-matched controls without arrest. A separate testing cohort of 463 arrest cases from 2022–2023 and 2,979 controls verified that risk associations were consistent with training. A real-world cohort of 39,911 individuals who received an EKG in 2021 was then analyzed over the next two years to assess performance in routine care.

In the real-world cohort, the combined EHR–EKG model correctly predicted 153 of 228 people who were identified as high-risk and subsequently experienced cardiac arrest. An EKG-only model showed strong predictive ability that was only modestly lower than models using EHR data. The analysis also surfaced risk features not typically emphasized in cardiovascular disease, including electrolyte disorders, substance use, and medication interactions.

The authors noted important limitations. All data originated from a single health system, limiting generalizability to other populations. The real-world cohort included only individuals who underwent EKG testing, and representations learned from EKGs could reflect demographic and care-pattern biases. Additional work is needed to determine the most appropriate clinical actions when a patient is flagged as high-risk.

“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.

“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?” said Dr. Chatterjee.

Related Links
University of Washington School of Medicine