AHA Scientific Sessions 2026 Abstract · Nov 2, 2026
Dual-Model Consensus Strategy to Improve Specificity of Cardiac Artificial Intelligence Phenotyping in Elite Athletes
By de Sousa Barzon Serra V, Zanco J, Ibrahim R et al.
Background: Artificial intelligence (AI) algorithms applied to electrocardiograms (ECG) and echocardiography are emerging tools for cardiovascular screening. Athletic cardiac remodeling, however, overlaps substantially with pathologic phenotypes—including hypertrophic cardiomyopathy (HCM) and cardiac amyloidosis (AMY)—and may generate false-positive classifications in elite athletes. Hypothesis: We aimed to evaluate whether a dual-model consensus strategy — defined as concurrent positive classification by two independently trained cardiac AI algorithms — could reduce false-positive screening rates for hypertrophic cardiomyopathy and cardiac amyloidosis in healthy elite athletes. Methods: We retrospectively evaluated 64 elite athletes (mean age 24.8 ± 5.4 years; 85.9% male; 64.1% professional baseball, 15.6% soccer) at a single tertiary cardiovascular center. All participants were free of structural heart disease on echocardiography, defining the cohort as disease-negative for specificity analysis. Two independently developed AI algorithms—Model A (ECG plus echocardiography) and Model B (ECG alone)—generated probability outputs for HCM and AMY using published thresholds. A dual-model consensus positive was defined a priori as concurrent flagging by both algorithms. Specificity with 95% Clopper-Pearson confidence intervals was calculated; between-model agreement was assessed with McNemar test. Results: Mean left ventricular ejection fraction (LVEF) was 59.7 ± 3.9%. Model A flagged 8 athletes as false-positive (specificity 87.5%, 95% CI 76.8-94.4) and Model B flagged 7 (specificity 89.1%, 95% CI 78.8-95.5; p=0.80). False-positive classifications were entirely non-concordant: no athlete was flagged by both algorithms simultaneously. Dual-model consensus reclassified all 15 false positives as negative, achieving a specificity of 100% (95% CI 94.4-100). Conclusions: Two independently developed cardiac AI algorithms produced entirely non-overlapping false-positive classifications of HCM and AMY in elite athletes. A dual-model consensus rule achieved 100% specificity and represents a computationally simple strategy to reduce diagnostic errors in AI-based cardiovascular preparticipation screening. Prospective validation in disease-positive cohorts is warranted.