AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Comparative Diagnostic Specificity of Two Independent Artificial Intelligence Electrocardiogram Algorithms for Pre-Participation Cardiovascular Screening in Athletes
By de Sousa Barzon Serra V, Zanco J, Ibrahim R et al.
Background: Adding a 12-lead electrocardiogram (ECG) to pre-participation cardiovascular screening improves sensitivity for underlying structural heart disease. Artificial intelligence (AI)-ECG algorithms can identify hypertrophic cardiomyopathy (HCM) and cardiac amyloidosis (AMY) in general clinical populations, but their specificity in athletes—where training-induced cardiac remodeling mimics pathologic phenotypes—has not been rigorously compared across independently developed models. Hypothesis: We hypothesized that two independently trained AI-ECG models would differ significantly in diagnostic specificity for HCM and AMY in healthy athletes, and that their false-positive classifications would target largely distinct physiological subsets. Methods: We performed a retrospective cohort study of 738 healthy athletes (mean age 22.4 years; 86.5% male) undergoing pre-participation 12-lead ECG at the Mayo Clinic enterprise (2017-2025). Athletes with concurrent echocardiography were excluded from this study. All participants were free of structural heart disease (disease-negative for specificity analysis). Two AI algorithms analyzed each ECG independently: the RST model and the AZ. Specificity for the composite endpoint (HCM or AMY) was the primary outcome, with 95% Clopper-Pearson confidence intervals. Between-model differences were assessed with McNemar test. Results: The RST model flagged 15 athletes as false-positive for the composite endpoint (specificity 98.0%, 95% CI 96.7-98.9), versus 48 for the AZ model (specificity 93.5%, 95% CI 91.5-95.2; p<0.001). For individual conditions, RST specificity was 98.2% (95% CI 97.0-99.1) for HCM and 99.7% (95% CI 99.0-100.0) for AMY; AZ specificity was 96.1% (95% CI 94.4-97.4) for HCM and 97.4% (95% CI 96.0-98.4) for AMY. Among 61 total false-positive events, only 2 athletes (3.3%) were flagged by both models. AI-predicted ECG age exceeded chronological age by 5.15±6.41 years (Pearson r = 0.571, p<0.001), consistent with athletic electrical remodeling. Conclusions: In a large ECG-only athlete cohort, two independently developed AI-ECG algorithms demonstrated significantly different diagnostic specificity for HCM and AMY, with near-complete non-concordance of false-positive identifications. These findings suggest that a sequential dual-model screening strategy — confirming an initial positive flag with a second independent algorithm before triggering additional workup — may improve specificity and reduce unnecessary cardiovascular evaluations in athletic pre-participation screening. Prospective validation is warranted.