External validation of ECG artificial intelligence for emergency and cardiac assessment across a large-scale U.S. healthcare system.

An ECG-based artificial intelligence (AI) model was previously developed to generate ten digital biomarkers for emergency and cardiac assessment and is currently deployed in clinical practice in Korea (ECG Buddy, ARPI Inc.). Its external validity within U.S. healthcare settings has not been established. This study evaluated model performance using a large-scale, multi-center U.S. dataset via the Mayo Clinic Platform (MCP) Discover. Two validation cohorts were assessed: an emergency-diagnosis cohort (mortality, AMI, STEMI and equivalents, hyperkalemia, pulmonary edema) using initial ED ECGs, and a cardiac-function cohort (left and right ventricular systolic dysfunction, pulmonary hypertension, hemodynamically significant pericardial effusion) anchored to echocardiography or right-heart catheterization. The primary objective was non-inferiority of the Area Under the Receiver Operating Characteristic Curve (AUC) relative to prespecified benchmarks. Across ten target conditions, AUCs ranged from 0.883 to 0.949; all met non-inferiority criteria. Performance was consistent across sex and age strata. In a paired subset (N = 1368), ECG-AI outperformed initial troponin T for AMI (AUC 0.920 vs. 0.878) and STEMI equivalents (AUC 0.932 vs. 0.736; all P < 0.001). These findings support the model's potential for ECG-based screening and triage, and provide a foundation for prospective evaluation of calibration, clinical integration, and impact across diverse populations.