Researchers have developed an artificial intelligence (AI) model that can scrutinize electronic health records (EHRs) and electrocardiograms to identify individuals in the general population at high risk of sudden cardiac arrest. Sudden cardiac arrest kills more than 400,000 people a year in the United States, with a survival rate of only 10%.
The discovery represents a major advance in predicting the largely unpredictable medical emergencies that often strike people with no known heart disease.
Artificial intelligence applications and health record data can be used to predict cardiac arrest in the general population. ”
Dr. Neil Chatterjee, Principal Investigator and Cardiologist, University of Washington School of Medicine
JACC: ProgressThe Journal of the American College of Cardiology published the paper today. Other co-senior authors are from Massachusetts General Hospital, the Broad Institute of MIT and Harvard University.
The researchers’ test population consisted of approximately 1.7 million patients from a large U.S. health system. The three AI models were developed using separate datasets. “EKG only,” “EHR only” (weighting 156 clinical features of the patient record), and a combined model that integrates EKG and EHR datasets.
The researchers developed and validated the AI model using three different patient groups:
Training cohort: Researchers trained the model using data from 993 people who experienced out-of-hospital cardiac arrest between 2013 and 2021 and 5,479 age- and gender-matched control patients who did not. The group taught an AI model to recognize patient EHR data inputs and electrocardiogram readings that are associated with a higher risk of cardiac arrest.
Testing cohort: To verify that the AI model accurately differentiates between high- and low-risk factors, researchers applied the AI model to a separate group of 463 cardiac arrest cases and 2,979 control patients from 2022 to 2023. This test showed that the model’s risk associations matched similarly to those established by the training cohort.
Real-world cohort: This group included 39,911 people who underwent an electrocardiogram during 2021, regardless of their health status. The researchers analyzed records of a subset of patients who experienced cardiac arrest over the next two years to see how well they matched the risk profile established by the AI model, Chatterjee explained.
In a real-world cohort, the combined EHR-EKG model accurately predicted 153 of 228 high-risk cardiac arrest patients.
“With these models, we can improve risk prediction by about 1 in 1,000 to 1 in 100,” Chatterjee said. “If a doctor tells you there’s a 1 in 100 risk of cardiac arrest, that’s going to get a lot of attention. We’re focusing on the theoretical risk.”
Another promising finding was that AI-enhanced ECG analysis alone showed strong predictive ability, but only slightly lower than two models that incorporated EHR data.
“The 12-lead ECG is a low-cost tool that has the potential to stratify the risk of cardiac arrest in patients anywhere in the world,” Chatterjee said.
The study also identified characteristics of cardiac arrest risk beyond those typically associated with cardiovascular disease. Some of these factors include electrolyte disturbances, substance use, and drug interactions.
“We have some modifiable risk factors that are relatively easy to implement,” Chatterjee said. “Models that flag a patient as high risk could prompt those caring for the patient to review their medical history and medications.”
Professor Chatterjee stressed that while the study results demonstrate the feasibility of risk prediction, further research is needed to determine the best clinical response if the model indicates an increased risk of cardiac arrest in a patient.
“To understand what to do with this patient information, we need to determine what follow-on research should proceed. What screening, what surveillance, what interventions are warranted?”
This study has important limitations. All data were obtained from a single health system, and generalizability to other populations with different demographics and care patterns is unknown. The actual cohort is limited to individuals who underwent ECG testing and may differ from those who did not undergo ECG evaluation. AI-enhanced ECG representations may reflect biases related to demographics and medical patterns.
sauce:
University of Washington School of Medicine
Reference magazines:
Sharma, S. Others. (2026). Artificial intelligence-enhanced electrocardiography and health records to predict cardiac arrest. Jack: Progress. DOI: 10.1016/j.jacadv.2026.102787https://www.sciencedirect.com/science/article/pii/S2772963X26002085
