AI models identify cardiac-arrest risk from patient records and heart readings
Researchers have developed artificial intelligence models that scan electronic health records and electrocardiograms to identify people at elevated risk for sudden cardiac arrest, a condition that kills more than 400,000 Americans annually with only a 10% survival rate.
The study, published May 11 in JACC: Advances, tested three AI models on roughly 1.7 million patients in a large U.S. healthcare system. One model analyzed EKGs alone, another examined electronic health records, and a third combined both data sources.
The combined model correctly identified 153 of 228 high-risk patients who experienced cardiac arrest within two years. That improves the baseline risk estimate from about 1 in 1,000 to 1 in 100.
"With these models, we're able to enrich risk prediction from about 1 in 1,000 down to 1 in 100," said Dr. Neal Chatterjee, the study's lead investigator and a cardiologist at the University of Washington School of Medicine. "If your doctor were to tell you that your risk of cardiac arrest is 1 in 100, that would catch your attention."
How the models were built and tested
Researchers trained the AI models using data from 993 patients who had out-of-hospital cardiac arrests between 2013 and 2021, plus 5,479 matched control patients. They then validated the models on a separate group of 463 cardiac arrest cases from 2022-2023 with 2,979 controls.
The real-world test involved 39,911 individuals who received EKGs in 2021. Researchers tracked which patients experienced cardiac arrest over the next two years and compared those outcomes to the risk predictions the models had made.
EKG analysis alone shows promise
A notable finding: the AI-enhanced EKG analysis alone predicted risk nearly as well as models that incorporated additional health record data. This matters because EKGs are inexpensive and available globally.
"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.
Unexpected risk factors emerged
The models identified cardiac-arrest risk factors beyond traditional cardiovascular disease markers. Electrolyte disorders, substance use, and medication interactions all appeared in the risk profiles.
"We show some relatively low hanging fruit - modifiable risk factors," Chatterjee said. "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."
What comes next remains unclear
Chatterjee emphasized that the study demonstrates feasibility but stops short of clinical guidance. Clinicians still need clarity on how to respond when a model flags elevated 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?" he said.
Study limitations
All data came from a single healthcare system, which limits how well the findings apply to other populations with different demographics or care patterns. The real-world cohort included only patients who received EKGs, a group that may differ from those who didn't undergo testing. The AI models could also reflect biases tied to demographics and healthcare access patterns.
For healthcare professionals exploring AI for Healthcare applications, understanding both the capabilities and constraints of predictive models is essential. The AI Data Analysis techniques used here show how machine learning can surface patterns in clinical data, but implementation requires careful validation and consideration of real-world constraints.
Your membership also unlocks:
