Learning general representation of 12-lead ECG with a joint-embedding predictive architecture

Area of Science:

• Cardiology
• Artificial Intelligence
• Machine Learning

Background:

• Electrocardiogram (ECG) data is crucial for diagnosing heart conditions.
• Supervised learning for ECG analysis is limited by the scarcity of labeled data.
• Self-supervised learning (SSL) offers a way to learn from unlabeled ECG data.

Purpose of the Study:

• To introduce ECG-JEPA, a novel SSL model for 12-lead ECG analysis.
• To leverage masked modeling in the latent space for learning semantic ECG representations.
• To improve ECG analysis by avoiding reconstruction of raw signals and noise.

Main Methods:

• Developed ECG-JEPA, an SSL model utilizing masked modeling in the latent space.
• Introduced Cross-Pattern Attention (CroPA), a masked attention mechanism for 12-lead ECG.
• Trained ECG-JEPA on a large dataset of approximately 180,000 unlabeled ECG samples.

Main Results:

• ECG-JEPA learns effective semantic representations of ECG data.
• The model bypasses raw signal reconstruction, avoiding noise and L2 loss limitations.
• Achieved state-of-the-art performance on diagnostic classification, feature extraction, and segmentation tasks.

Conclusions:

• Latent space masked modeling is a powerful SSL approach for ECG analysis.
• ECG-JEPA offers advantages over existing SSL methods in the ECG domain.
• The proposed method enhances the utility of unlabeled ECG data for various clinical applications.