PACE-ECG: Patient-Aware Contrastive ECG Pretraining; Leveraging Longitudinal Structure for Transferable Representations
Organizations: King’s College London, St Thomas’ Hospital, London, UK
Abstract
Foundation models for physiological signals are commonly evaluated on in-domain interpretation tasks, but their ability to transfer to longitudinal clinical prognosis across healthcare systems remains uncertain. We developed PACE-ECG, an ECG pretraining framework that combines contrastive learning across temporally neighbouring recordings from the same patient with clinically guided multitask supervision. We evaluated frozen ECG representations for post-myocardial-infarction all-cause mortality and incident heart failure prediction across MIMIC-IV and two external HEEDB cohorts, comparing the proposed representation with DeepECG-SSL, a state-of-the-art model, and two control baselines. Across both downstream tasks, the proposed model outperformed DeepECG-SSL in-domain (AUROC +1.9% for mortality and +4.0% for heart failure) and on external populations in HEEDB-Boston (+5.6% for mortality +10.3% for heart failure) and HEEDB-Atlanta (+6.9% for mortality +11.0% for heart failure) while using an encoder 41 times smaller and a fraction of the pretraining data. A matched ablation showed that patient-temporal pairing improved external mortality and heart failure AUROC over generic self-supervision by 7.0% and 9.4% on HEEDB-Boston and 8.9% and 9.3% on HEEDB-Atlanta. These findings indicate that pretraining-objective design greatly influences downstream transfer alongside model and data scale, and that in-domain performance alone is insufficient to establish the clinical generalisability of ECG foundation representations.
Figures & tables
| Model | MIMIC-IV | HEEDB Boston | HEEDB Atlanta |
|---|---|---|---|
| PACE-ECG | 0.788 (0.786–0.789) | 0.769 (0.767–0.770) | 0.732 (0.728–0.715) |
| SSL model | 0.749 (0.747–0.752) | 0.719 (0.715–0.722) | 0.672 (0.668–0.676) |
| DeepECG-SSL | 0.773 (0.772–0.774) | 0.728 (0.724–0.732) | 0.685 (0.684–0.686) |
| Trained from scratch | 0.747 (0.730–0.759) | 0.650 (0.628–0.690) | 0.659 (0.629–0.685) |
| Model | MIMIC-IV | HEEDB Boston | HEEDB Atlanta |
|---|---|---|---|
| PACE-ECG | 0.731 (0.725–0.734) | 0.779 (0.777–0.781) | 0.787 (0.784–0.790) |
| SSL model | 0.683 (0.679–0.687) | 0.712 (0.708–0.715) | 0.720 (0.716–0.723) |
| DeepECG-SSL | 0.703 (0.702–0.705) | 0.706 (0.703–0.708) | 0.709 (0.702–0.715) |
| Trained from scratch | 0.651 (0.642–0.661) | 0.619 (0.606–0.631) | 0.668 (0.660–0.675) |
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
| Characteristic | MIMIC-IV MI cohort | HEEDB Boston | HEEDB Atlanta |
|---|---|---|---|
| Patients | 8,494 | 10,222 | 1,855 |
| Mortality events | 3,074 | 1,393 | 329 |
| HF events | 2,407 | 1,644 | 664 |
| Age at MI, years, median (IQR) | 71.0 (61.0–81.0), | 65.8 (57.0–73.1) | 65.1 (54.9–73.0) |
| Female, n (%) | 3,289 (38.8%) | 3,659 (35.8%) | 730 (39.4%) |