ECG-FM: an open electrocardiogram foundation model

Kaden McKeen; Sameer Masood; Augustin Toma; Barry Rubin; Bo Wang

doi:10.1093/jamiaopen/ooaf122

ECG-FM: an open electrocardiogram foundation model

JAMIA Open. 2025 Oct 16;8(5):ooaf122. doi: 10.1093/jamiaopen/ooaf122. eCollection 2025 Oct.

Authors

Kaden McKeen^{1

2

3

4

5}, Sameer Masood^{1

6}, Augustin Toma^{4

7}, Barry Rubin^{1

2

3}, Bo Wang^{1

2

3

4

5

7

8}

Affiliations

¹ Toronto General Hospital Research Institute, University Health Network, Toronto, M5G 2C4, Canada.
² Peter Munk Cardiac Centre, University Health Network, Toronto, M5G 2N2, Canada.
³ UHN AI Hub, University Health Network, Toronto, M5G 2C4, Canada.
⁴ Vector Institute for Artificial Intelligence, Toronto, M5G 0C6, Canada.
⁵ Department of Laboratory Medicine and Pathobiology, University of Toronto, Toronto, M5S 3K3, Canada.
⁶ Department of Medicine, University of Toronto, Toronto, M5S 3H2, Canada.
⁷ Department of Medical Biophysics, University of Toronto, Toronto, M5G 2C4, Canada.
⁸ Department of Computer Science, University of Toronto, Toronto, M5S 2E4, Canada.

Abstract

Objectives: To develop ECG-FM, an open-weight foundation model for electrocardiogram (ECG) analysis, rigorously evaluate its performance on clinically salient tasks, and openly release it alongside a public benchmark.

Materials and methods: In a study using 1.5 million 12-lead ECGs, we present ECG-FM, a transformer-based foundation model pretrained with hybrid self-supervision that combines masked reconstruction and contrastive learning with ECG-specific augmentation. Downstream, we evaluate multi-label ECG interpretation and prediction of reduced left ventricular ejection fraction (LVEF), introducing an openly available benchmark on the MIMIC-IV-ECG dataset. We assess ECG-FM's capabilities through data scaling experiments, latent-space structure analysis, and attention-based saliency.

Results: Finetuned ECG-FM models outperform task-specific baselines in the small-to-medium-scale data regime, exhibit strong label efficiency and cross-dataset generalizability, and achieve high AUROC on salient labels, including atrial fibrillation (0.996) and LVEF $\leq 40 %$ (0.929). The pretrained encoder showcases competitive linear probing performance, with functionally discriminative embeddings.

Discussion: Findings indicate that ECG-FM is generalizable, label-efficient, and discriminative for screening, risk stratification, and monitoring. Its representations capture low-level morphology and high-order cardiac semantics, and the pretrained encoder serves as a robust feature-set generator. This work mitigates reliance on large labeled datasets, reduces compute and data requirements, and lowers barriers to reproducibility and cross-study comparison.

Conclusion: ECG-FM is an open, rigorously validated ECG foundation model intended to accelerate transparent, comparable research in the ECG analysis subfield. It is designed for rapid integration and evaluation, especially for delivering practical gains in low-label settings. We release our code, model weights, tutorials, and benchmark at https://github.com/bowang-lab/ECG-FM/.

Keywords: deep learning; electrocardiography; foundation model; self-supervised learning; time series analysis.