Researchers have successfully developed artificial intelligence models capable of predicting the risk of sudden cardiac arrest in the general population by analyzing electronic health records and electrocardiogram data. Published in JACC: Advances, the study marks a significant advancement in forecasting a medical emergency that accounts for over 400,000 deaths annually in the United States, with a survival rate of merely 10 percent. Dr. Neal Chatterjee, the lead investigator from the University of Washington School of Medicine, stated that the findings demonstrate that predicting cardiac arrest in the general public is feasible using AI applications and health data. The research team analyzed a dataset comprising approximately 1.7 million patients from a large U.S. healthcare system. They developed three distinct AI models: one utilizing only electrocardiogram (EKG) data, one using only electronic health record (EHR) data which weighed 156 clinical features, and a combined model integrating both datasets. When validated against a real-world cohort, the combined EHR-EKG model correctly identified 153 out of 228 high-risk individuals who subsequently experienced cardiac arrest. Dr. Chatterjee highlighted that this technology improves risk stratification from approximately one in 1,000 to one in 100, a shift that makes the theoretical risk concrete and actionable for patients and providers. Notably, the AI-enhanced analysis of standard 12-lead EKGs alone showed strong predictive ability, performing only modestly worse than models incorporating broader EHR data. This suggests that the EKG, a widely available and low-cost tool, could serve as an effective screening method for cardiac arrest risk in diverse communities globally. Beyond traditional cardiovascular markers, the AI models identified unique risk contributors, including electrolyte disorders, substance use, and medication interactions. These findings point to potentially modifiable risk factors that could be addressed through medication reviews or lifestyle changes once a patient is flagged. Despite these promising results, the study acknowledges important limitations. All data originated from a single healthcare system, leaving the generalizability of the models to populations with different demographics or care patterns uncertain. Furthermore, the analysis was limited to individuals who received an EKG, which may not represent the broader population. There is also a potential for bias within the AI-enhanced EKG representations linked to specific demographics and healthcare access patterns. Dr. Chatterjee emphasized that while the prediction capability is established, further research is required to determine the optimal clinical response when a model flags a high-risk patient. Questions remain regarding the specific follow-up studies, screening protocols, surveillance methods, and interventions necessary to manage this information effectively. The study was conducted in collaboration with co-senior authors from Massachusetts General Hospital and the Broad Institute of MIT and Harvard.