Self-supervised contrastive learning enables robust electrocardiogram-based cardiac classification

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Machine Learning

Background:

• Self-supervised contrastive learning is a powerful method for learning from unlabeled data.
• In electrocardiogram (ECG) analysis, pre-training enhances classification, particularly with scarce labeled data.

Purpose of the Study:

• To investigate and improve contrastive self-supervised learning techniques for ECGs.
• Systematically combine advances in augmentation, contrastive loss, and encoder architectures for ECG analysis.

Main Methods:

• Implemented a contrastive pre-training framework using vectorcardiography (VCG)-based augmentations, interlead/intersegment contrastive loss, and patient-aware sampling.
• Developed a dual-stream TemporalNet architecture processing grouped ECG leads independently.
• Pre-trained on ~1 million unlabeled ECGs, evaluated on low left ventricular ejection fraction (LVEF) and high potassium chloride tasks across various supervision levels (1%-100%).

Main Results:

• Contrastive pre-training consistently improved performance across all supervision levels.
• Achieved 3-4% higher AUC for LVEF and 5-7% higher AUC for potassium chloride tasks in low-label settings (1-10% supervision) compared to baseline.
• Performance advantage for pre-trained models persisted even with increased labeled data.

Conclusions:

• Contrastive pre-training substantially enhances ECG classification, especially in low-data regimes.
• A scalable framework trained on 1 million ECGs offers practical guidance and architectural innovations for ECG foundation models.
• The developed methods are applicable to a broad range of clinical prediction tasks using ECG data.