Positive-call rates and agreement of two artificial intelligence electrocardiogram models for hypertrophic cardiomyopathy and cardiac amyloidosis in competitive athletes

BACKGROUND: Artificial intelligence-enabled electrocardiography (AI-ECG) models detect hypertrophic cardiomyopathy (HCM) and cardiac amyloidosis in clinical populations; whether two independently developed models agree in athletes is unknown. METHODS: We applied two independently developed AI-ECG models to the screening electrocardiograms (ECGs) of 716 competitive athletes aged 40 years or younger, screened by ECG alone and presumed free of structural heart disease. The RST algorithms return per-condition probabilities dichotomized at published operating points; the AZ model, a four-class ECG-image classifier, assigns one class per ECG. The primary outcome was each model's positive-call rate (proportion of athletes flagged) for HCM and cardiac amyloidosis; agreement was described with Cohen kappa and positive agreement. Because the cohort was presumed disease-free, sensitivity was not assessed. RESULTS: Median age was 22.1 years; 86.0% were men. The RST algorithms flagged HCM in 13 athletes (1.8%; 95% CI, 1.0-3.1) and cardiac amyloidosis in 2 (0.3%; 0.0-1.0); the AZ model flagged HCM in 29 (4.1%; 2.7-5.8) and cardiac amyloidosis in 18 (2.5%; 1.5-3.9). Of 60 athletes flagged by either model, 2 were flagged by both (kappa 0.07 and - 0.01; positive agreement 0.10 and 0.00); continuous outputs were uncorrelated for HCM (Spearman rho 0.00). Of 59 flagged athletes with a stored cardiologist interpretation under the international criteria, 39 had a normal ECG. All 18 AZ amyloidosis calls occurred in men. CONCLUSIONS: The two models flagged different athletes at different rates. Without verified cases these rates describe referral burden, not accuracy; whether either detects disease requires validation against verified diagnoses.