SportsECG-FM: a sport-specific ECG interpretation algorithm enhancing athletes cardiac screening

Background Electrocardiographic (ECG) screening is central to the prevention of sudden cardiac death (SCD) in competitive athletes. Exercise-induced cardiac remodelling produces borderline ECG phenotypes that morphologically overlap with pathological findings, complicating ECG interpretation in athletes. While large-scale artificial intelligence (AI) foundation models (FMs) have demonstrated high accuracy in clinical populations, their capacity to discriminate physiological adaptation from pathological change in athlete cohorts remains limited. Without sports specific domain adaptation, these models may risk false reassurance or unnecessary restriction for athletes, affecting both safety and career outcomes. Purpose To develop and evaluate SportsECG-FM, a domain-adaptation framework designed to differentiate physiological athletic ECG patterns from pathological abnormalities in professional athletes. Methods The study compared two distinct cohorts: (i) a clinical reference group and (ii) a professional football athlete cohort (Pro-Football 12-lead Database, PF12RED). The outcomes of interest were cardiologist-confirmed ECG diagnostic phenotypes (sinus rhythm, sinus bradycardia, incomplete right bundle branch block, and T-wave inversion in different territory). We utilised a clinical ECG model pre-trained on over 10 million clinical ECG recordings, to extract digital phenotypes that characterise cardiac electrical patterns. To account for exercise-induced cardiac remodelling, we implemented a domain adaptation strategy using cardiologist-labelled athlete ECGs. Specifically, we trained a prototypical network on the PF12RED dataset using an 8-way variable-shot approach. Model performance was evaluated in the athlete cohort before and after adaptation, and compared against cardiologist labels to assess discrimination across ECG phenotypes. Results In the athlete cohort, SportsECG-FM improved agreement with cardiologist-confirmed ECG diagnoses. Mean AUROC increased from 0.63 to 0.76 (+20.6% relative) and mean sensitivity from 0.71 to 0.84 (+18.3% relative), while mean specificity was maintained (0.75 vs. 0.76). In the clinical reference cohort, clinical ECG model achieved mean AUROC 0.80 and mean sensitivity 0.83. Furthermore, the adapted model enhanced common physiological athletic ECG patterns identification, including sinus bradycardia and incomplete right bundle branch block, supporting discrimination between physiological and pathological abnormalities. Conclusions These findings provide preliminary evidence that athlete-specific domain adaptation bridges the performance gap between clinical and athlete ECG interpretation, with improvements in screening accuracy.Validation in larger and more diverse athletic populations is required before broader implementation.