WAVE: Wall-Aligned Vector Embedding for Self-Supervised Learning of Electrocardiograms

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Deep learning models for electrocardiogram (ECG) analysis require extensive labeled data, increasing annotation costs.
• Developing cost-effective and label-efficient methods for ECG analysis is crucial for widespread clinical application.

Purpose of the Study:

• To propose wall-aligned vector embedding (WAVE), a self-supervised learning framework for ECG analysis.
• To reduce the reliance on labeled ECG data by extracting prior knowledge from unlabeled ECGs.
• To improve the efficiency and reduce the cost of cardiovascular diagnostics.

Main Methods:

• WAVE framework leverages multi-lead ECGs, considering lead diversity, synergy, and correlations with cardiac walls.
• A multi-branch network captures lead-wise diversity.
• Wall-wise synergy is modeled using concatenation, shared projection, and a dual alignment task for intra- and inter-wall correlations.

Main Results:

• WAVE consistently outperformed existing methods across various evaluation settings.
• The framework demonstrated strong performance with limited labeled ECG data.
• Key components like dual alignment and shared projection significantly improved pretraining quality.

Conclusions:

• WAVE effectively learns informative ECG representations from unlabeled data.
• The proposed method enables low-cost and label-efficient ECG analysis.
• WAVE holds significant potential for real-world cardiovascular diagnostics.