ACL-ECG: Anatomy-Aware Contrastive Learning for Multi-Lead Electrocardiograms.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Deep learning for electrocardiogram (ECG) analysis requires large labeled datasets, which are costly and time-consuming to create.
• Self-supervised learning, particularly contrastive learning, offers a way to learn from unlabeled ECG data.
• Current methods often lack the incorporation of domain-specific knowledge, limiting representation quality.

Purpose of the Study:

• To develop a novel self-supervised method, Anatomy-Aware Contrastive Learning for ECG (ACL-ECG), for automated ECG analysis.
• To leverage cardiac anatomical relationships within a contrastive learning framework to enhance ECG representation learning.
• To reduce the reliance on extensive labeled datasets in ECG analysis.

Main Methods:

• Proposed ACL-ECG, a self-supervised contrastive learning framework incorporating cardiac anatomy.
• Developed a physiology-aware augmentation strategy including random scale cropping, cardiac-cycle masking, and temporal shifting.
• Grouped ECG leads into anatomical regions (anterior, inferior, septal, lateral) and introduced region-level contrastive objectives.

Main Results:

• ACL-ECG outperformed state-of-the-art contrastive baselines in downstream tasks under linear probing, with up to 1.29% AUROC and 3.57% AUPRC improvements.
• Fine-tuning ACL-ECG with 10% of labeled data achieved performance comparable to fully supervised training, reducing annotation needs by 5-8x.
• Ablation studies confirmed the effectiveness of both the physiology-aware augmentation and anatomy-aware contrastive objectives.

Conclusions:

• ACL-ECG enhances ECG representation quality without increasing annotation burden.
• The anatomy-informed approach provides a robust foundation for self-supervised ECG analysis in label-scarce clinical settings.
• ACL-ECG offers a scalable and efficient alternative for automated ECG interpretation.