Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OpenECG: Benchmarking ECG Foundation Models with Public 1.2 Million Records

Zhijiang Wan, Qianhao Yu, Jia Mao, Wenfeng Duan, Cheng Ding

TL;DR

OpenECG addresses the need for scalable, generalizable ECG foundation models by aggregating 1.2 million public ECG records from nine centers into a public benchmark. It compares three SSL approaches (SimCLR, BYOL, MAE) with two backbones (ResNet-50 and ViT) and uses leave-one-dataset-out and data-size scaling to assess generalization. Key findings show BYOL and MAE outperform SimCLR, data diversity drives robustness, and performance saturates for BYOL/MAE around 60-70% data, suggesting data efficiency. The work demonstrates that public datasets can rival proprietary data for training robust ECG models and establishes a publicly available OpenECG benchmark to foster cross-site validation and future multimodal extensions.

Abstract

This study introduces OpenECG, a large-scale benchmark of 1.2 million 12-lead ECG recordings from nine centers, to evaluate ECG foundation models (ECG-FMs) trained on public datasets. We investigate three self-supervised learning methods (SimCLR, BYOL, MAE) with ResNet-50 and Vision Transformer architectures, assessing model generalization through leave-one-dataset-out experiments and data scaling analysis. Results show that pre-training on diverse datasets significantly improves generalization, with BYOL and MAE outperforming SimCLR, highlighting the efficacy of feature-consistency and generative learning over contrastive approaches. Data scaling experiments reveal that performance saturates at 60-70% of total data for BYOL and MAE, while SimCLR requires more data. These findings demonstrate that publicly available ECG data can match or surpass proprietary datasets in training robust ECG-FMs, paving the way for scalable, clinically meaningful AI-driven ECG analysis.

OpenECG: Benchmarking ECG Foundation Models with Public 1.2 Million Records

TL;DR

OpenECG addresses the need for scalable, generalizable ECG foundation models by aggregating 1.2 million public ECG records from nine centers into a public benchmark. It compares three SSL approaches (SimCLR, BYOL, MAE) with two backbones (ResNet-50 and ViT) and uses leave-one-dataset-out and data-size scaling to assess generalization. Key findings show BYOL and MAE outperform SimCLR, data diversity drives robustness, and performance saturates for BYOL/MAE around 60-70% data, suggesting data efficiency. The work demonstrates that public datasets can rival proprietary data for training robust ECG models and establishes a publicly available OpenECG benchmark to foster cross-site validation and future multimodal extensions.

Abstract

This study introduces OpenECG, a large-scale benchmark of 1.2 million 12-lead ECG recordings from nine centers, to evaluate ECG foundation models (ECG-FMs) trained on public datasets. We investigate three self-supervised learning methods (SimCLR, BYOL, MAE) with ResNet-50 and Vision Transformer architectures, assessing model generalization through leave-one-dataset-out experiments and data scaling analysis. Results show that pre-training on diverse datasets significantly improves generalization, with BYOL and MAE outperforming SimCLR, highlighting the efficacy of feature-consistency and generative learning over contrastive approaches. Data scaling experiments reveal that performance saturates at 60-70% of total data for BYOL and MAE, while SimCLR requires more data. These findings demonstrate that publicly available ECG data can match or surpass proprietary datasets in training robust ECG-FMs, paving the way for scalable, clinically meaningful AI-driven ECG analysis.

Paper Structure

This paper contains 17 sections, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Architecture design of FMs for ECGs
  4. SSL-based pre-training strategies of ECG-FMs
  5. Materials and Methods
  6. Data preparation
  7. Dataset description
  8. Data pre-processing
  9. Pre-training of ECG-FMs
  10. Self-Supervised Learning Methods
  11. Experiments and Results
  12. Comparison of ECG rhythm classification performance with previous studies
  13. Testing Generalizability via Iterative Dataset Exclusion
  14. Effect of Training Data Size on ECG Foundation Model Performance
  15. Discussion
  16. ...and 2 more sections

Figures (3)

  • Figure 1: The key dataset statistics of previous ECG Foundation Models.
  • Figure 2: Three SSL pre-train methods.
  • Figure 3: Impact of Training Data Size on AUROC Performance Across Different Datasets and Self-Supervised Learning Methods