ECG foundation models share an age-related performance disparity that originates in the data, not the models

Age-stratified performance disparities in ECG foundation model embeddings persist across four models and six held-out cohorts, with a median worst-group AUROC gap of 0.028.

• Why it matters: Fairness claims for ECG foundation models are rarely tested by subgroup, and it is unclear whether age-related gaps come from the models or from the ECG data and clinical labels.
• What they did: Frozen embeddings from ECGFounder, xECG, ECG-FM, and DeepECG-SSL were probed across 13 cohorts (5.4 million ECGs, 31 endpoints), with age/sex/race stratification and tests including post-hoc correction, larger labeled sets, comorbidity exclusion, and low-rank fine-tuning.
• The result: The age gap replicated externally (ED-BWH 30-day MACE AUROC 0.81–0.90 under 40 vs 0.64–0.71 at ≥75) and survived mitigation (low-rank adaptation −0.007 to −0.004), while untrained copies and logistic regression on classical ECG measurements reproduced endpoint-specific gaps.

AI summary of the abstract · may contain errors · not medical advice · read the paper.