Key result
ECGFounder diagnostic accuracy drops significantly on athlete ECGs, achieving a ~0.84 AUROC.
Why the study?
Before deploying automated ECG foundation models trained on general clinical populations for large-scale athlete cardiac screening, it was necessary to verify whether they are applicable for diagnosing athlete-specific ECG findings.
Do ECG foundation models trained and fine-tuned on general clinical populations accurately generalize to diagnose athlete-specific ECG findings in young competitive athletes?
Observational (n=43)
No
Do ECG foundation models trained and fine-tuned on general clinical populations accurately generalize to diagnose athlete-specific ECG findings in young competitive athletes?
Clinically trained ECG foundation models show partial transferability to competitive athletes—reliably detecting sinus bradycardia and sinus arrhythmia—but suffer significant sensitivity drops due to physiological and semantic domain shifts, entirely missing labels like bundle branch block.
Answers are grounded in this paper and cite their sources.
Limited generalization of clinical ECG models to athletes warrants caution in screening; leaves open the value of domain-specific retraining.
AI-powered cardiac screening could evolve sports medicine by boosting accuracy, efficiency, and cost-effectiveness, with automated ECG interpretation based on foundation models becoming key for large-scale athlete screening. Before deployment, it is crucial to verify whether these clinically developed models are applicable to this population. We evaluated whether ECG foundation models trained on general clinical populations can diagnose athlete-specific ECG findings. We fine-tuned three models (ECG-FM, HuBERT-ECG, ECGFounder) on PhysioNet/CinC 2021, 87,662 samples, and tested them on 43 expert-annotated ECGs from highly trained/national-level athletes. Performance was strong on PhysioNet/CinC 2021 (AUROC 0.97, AUPRC 0.57, Sensitivity 0.89) but dropped on athletes’ data, confirming domain shift. ECGFounder performed best (AUROC 0.84, AUPRC 0.72, Sensitivity 0.57) and detected the most frequent findings, Sinus Bradycardia (AUROC 0.87) and Sinus Arrhythmia (AUROC 0.95), well. Per-label estimates for rare classes must be interpreted cautiously, given the small cohort. This work contributes to developing athlete-focused AI frameworks for reliable ECG-based screening.
No takes yet. Share an insight, caveat, or question.
Goelz et al. (2026) conducted an observational in Cardiovascular screening / athlete-specific ECG findings in competitive athletes (n=43). ECG foundation models (ECGFounder, ECG-FM, HuBERT-ECG) fine-tuned on PhysioNet/CinC 2021 vs. Expert-annotated ground truth ECG diagnoses and in-domain clinical test performance (PhysioNet/CinC 2021) was evaluated on Diagnostic classification performance (Macro-AUROC) of best-performing model (ECGFounder) on athlete ECGs (95% CI 0.74-0.92, p=0.0122). ECGFounder achieved a Macro-AUROC of 0.84 and Macro-Sensitivity of 0.57 on athlete ECGs, dropping from in-domain clinical performance.
Synapse has enriched 10 related papers. Consider them for comparative context: