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

Background. Electrocardiogram (ECG) foundation models are proposed for clinical use and promoted as fair, yet such claims are rarely tested with subgroup-stratified evaluation. We measured demographic disparities across four ECG foundation models, tested whether mitigation removes them, and asked whether the dominant disparity originates in the models or in the data and the clinical task. Methods. We probed frozen embeddings of four architecturally distinct ECG foundation models (ECGFounder, xECG, ECG-FM, DeepECG-SSL) in thirteen cohorts on four continents (5.4 million analysed ECGs, 31 endpoints, six held-out cohorts), stratified by age, sex and race. We examined embedding geometry, tested post-hoc correction, larger labelled training sets, comorbidity exclusion and low-rank fine-tuning, and compared each network with an untrained copy of itself and with logistic regression on classical ECG measurements. Findings. Disparities concentrated on age, not sex or race (median worst-group AUROC gap 0.028 vs 0.004 and 0.016; age exceeded sex in 97% of results) and replicated externally (ED-BWH 30-day MACE AUROC 0.81-0.90 under 40 years, 0.64-0.71 at >=75). Age formed a smooth gradient in the embeddings, whereas sex, though most decodable, carried no gap. No remedy closed it (low-rank adaptation: -0.007 to -0.004). Untrained copies (r=0.81, 102 pairs) and classical measurements (rho=0.90) reproduced the endpoint-specific gaps, and pretraining raised AUROC (+0.067) without changing them. The largest race effect, higher potassium error in Black patients, vanished at matched true potassium. Interpretation. ECG foundation models carry an age-structured disparity that replicates externally and survives mitigation. Because untrained networks and a classical model reproduce it, the results indicate that it is inherited from the age structure of the ECG and its labels rather than created by the models, so switching or adapting models will not remove it. Age-stratified reporting, untrained and classical baselines, and age-specific thresholds and recalibration are therefore needed. Funding. Austrian Research Promotion Agency (FFG; CARDISENSE project, grant FO999922422).