PACE-ECG pairs a patient's own ECGs across time: 41 times fewer parameters, higher external heart-failure AUROC
Synopsis
ECG foundation models are usually compared on in-domain interpretation tasks, leaving their transfer to longitudinal prognosis across healthcare systems uncertain. PACE-ECG treats two recordings from the same patient within 60 days as a contrastive positive pair and adds waveform- and report-derived multitask supervision, with an encoder of 2.21 million parameters. On MIMIC-IV and two external HEEDB cohorts, its frozen representations predict post-myocardial-infarction all-cause mortality and incident heart failure with higher AUROC than DeepECG-SSL, with the largest gains in external heart failure.
Interpretation
The design of the ECG pretraining objective itself changes downstream transfer, not only model and data scale. Earlier contrastive work mostly used two augmented views of the same ECG as the positive pair, so patient identity or temporal proximity did not define pairing. Within a shared multitask pretraining framework, replacing same-ECG views with two distinct same-patient recordings acquired within 60 days raised external heart-failure and mortality AUROC by 0.0676 and 0.0494 on HEEDB-Boston and by 0.0680 and 0.0591 on HEEDB-Atlanta.
Frozen PACE-ECG representations outperform DeepECG-SSL for post-infarction mortality and incident heart failure, with wider gaps externally. DeepECG-SSL was a strong prior contrastive benchmark across external healthcare datasets, with a 90.4-million-parameter encoder pretrained on more than 1.4 million ECGs. With a 2.21-million-parameter encoder pretrained on 800,035 ECGs, PACE-ECG reached mortality AUROC 0.788 versus 0.773 and heart-failure 0.731 versus 0.703 on MIMIC-IV, 0.769 versus 0.728 and 0.779 versus 0.706 on HEEDB-Boston, and 0.732 versus 0.685 and 0.787 versus 0.709 on HEEDB-Atlanta.
Pretrained representations transfer to external populations better than supervised training from scratch. The scratch control used the same ResNet architecture with random initialization and was trained jointly with the downstream head on MIMIC-IV labels. Every pretrained representation exceeded the scratch model on both outcomes and all three cohorts, with the largest gaps in external heart-failure prediction, for example 0.779 for PACE-ECG versus 0.619 for scratch on HEEDB-Boston.
Perspective
Evaluation covers two outcomes, post-infarction all-cause mortality and incident heart failure, across MIMIC-IV and the HEEDB-Boston and HEEDB-Atlanta cohorts, plus one external foundation-model family; downstream heads are trained only on MIMIC-IV, and neither HEEDB cohort is used for adaptation or model selection.
The 60-day window is a fixed design choice and was not compared with other intervals or unrestricted same-patient pairing; same-patient alignment may retain stable identity, demographic or acquisition-related characteristics as well as clinically relevant physiology, and the evaluation does not separate these mechanisms; both pairing schemes keep auxiliary supervision, so their comparison does not quantify the independent contribution of those tasks.