Scaling ECG Foundation Models and Identifying a Threshold for Effective Representation Learning

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Foundation models are increasingly used in healthcare.
• Electrocardiogram (ECG) analysis is crucial for diagnosing heart conditions.
• The impact of unlabeled pretraining data volume on ECG foundation models is not well understood.

Purpose of the Study:

• To evaluate the effect of unlabeled pretraining data volume on the performance of ECG foundation models.
• To determine the minimum ECG data threshold required for effective self-supervised pretraining.

Main Methods:

• Utilized one-dimensional vision transformer (1D-ViT) masked autoencoders for pretraining.
• Pretrained models on progressively larger ECG datasets.
• Fine-tuned pretrained models for various downstream tasks: rhythm, morphology, diagnostic, and structural heart disease.
• Compared performance against control groups without self-supervised pretraining.

Main Results:

• Models pretrained with ≤400,000 ECGs did not consistently outperform controls without self-supervised pretraining.
• Pretraining with 600,000-800,000 ECGs significantly improved Area Under the Receiver Operating Characteristic curve (AUROC) across all evaluated tasks.
• A minimum data volume appears necessary for effective ECG representation learning.

Conclusions:

• Unlabeled pretraining data volume is a critical factor for ECG foundation model performance.
• A minimum threshold of 600,000-800,000 ECGs is suggested for effective self-supervised pretraining.
• These findings have implications for developing robust AI tools for ECG analysis.