Key result
Retrained neural readout cuts false-positive abnormal ECG calls in athletes by ~35 percentage points.
Why the study?
Whether a publicly released ECG foundation model trained on hospital ECGs respects the International Recommendations for ECG interpretation in athletes, and whether its errors can be corrected without athlete data, needed to be tested.
Does retraining the readout layer of an ECG foundation model using athlete-criteria labels on non-athlete data reduce false-positive abnormal ECG calls in healthy athletes while maintaining sensitivity for screening-relevant abnormalities?
Observational (n=78)
Yes
Does retraining the readout layer of an ECG foundation model using athlete-criteria labels on non-athlete data reduce false-positive abnormal ECG calls in healthy athletes while maintaining sensitivity for screening-relevant abnormalities?
An off-the-shelf ECG foundation model misclassifies a third to half of normal athlete ECGs as abnormal due to benign training-related adaptations like sinus bradycardia, but retraining a small readout layer using International Criteria labels on non-athlete ECGs eliminates two-thirds of these false alarms without sacrificing sensitivity.
Answers are grounded in this paper and cite their sources.
May reduce unnecessary athlete ECG referrals; hypothesis-generating for foundation-model adaptation to specialized populations.
Objective To test whether a publicly released ECG foundation model respects the International Recommendations for ECG interpretation in athletes, and whether its errors can be corrected without athlete data. Methods We applied ECGFounder, trained on about 10 million hospital ECGs, to non-athletes with only training-related (“green”) findings (PTB-XL, Chapman/Ningbo) and to two athlete groups: 26 Norwegian endurance athletes read as normal by a cardiologist and 156 green-only recordings from 52 Spanish footballers. In a sequence of pre-registered steps we tested four fixes, anchoring sensitivity on screening-relevant abnormal ECGs, and confirmed the best fix with thresholds set on data it had never seen. Results Green-only ECGs were called abnormal in 33.7% of PTB-XL and 17.9% of Chapman/Ningbo cases, and the model called abnormal 34.6% of normal endurance athletes and 47.4% of footballer recordings, sometimes labelling infarction. Rules applied over the model’s own named findings, with or without measured intervals, made things worse. A small readout retrained on the model’s features with athlete-criteria labels, using non-athletes only, called 17.6% (95% CI 10.3% to 25.4%) of normal athletes abnormal against 52.8% for the model’s own output, at 87.5% against 89.4% sensitivity for screening-relevant abnormal ECGs, and held across three training seeds. Conclusion The model’s representation already separates athletic adaptation from disease; its default output does not. Relabelling the readout against the International Criteria removes about two-thirds of the false alarms.
No takes yet. Share an insight, caveat, or question.
Chukwudi Samuel Eke (2026) conducted an observational in Athletic cardiac adaptation (normal athlete ECGs with training-related 'green' findings) (n=78). Neural network readout retrained on ECGFounder features with athlete-criteria labels vs. Default ECGFounder foundation model output was evaluated on False-alarm rate (proportion of normal athletes called abnormal) at the 90% screening-red sensitivity threshold (ARR 35.2%, 95% CI 26.8-44.0). A neural readout retrained with athlete-criteria labels reduced false-positive abnormal ECG calls in normal athletes from 52.8% to 17.6% vs default ECGFounder output.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: