Tracing the Heart's Pathways: ECG Representation Learning from a Cardiac Conduction Perspective

TL;DR

The paper addresses limitations of ECG self-supervised learning that neglect conduction-driven heartbeat and lead-specific variation and the stepwise diagnostic workflow. It proposes CLEAR-HUG, a two-stage framework: CLEAR pretraining with conduction-guided and view-guided reconstruction using sparse attention, followed by HUG finetuning with hierarchical lead grouping that mirrors clinical diagnosis. It introduces a reconstruction objective $L = ||D(E(X_{masked})) - X||_2^2$ and a sparse attention design to isolate heartbeat-specific and lead-specific information, plus a seven-head lead fusion strategy. Across six datasets, CLEAR-HUG achieves substantial gains, including strong few-shot performance, demonstrating improved physiological relevance, interpretability, and robustness for downstream ECG tasks.

Abstract

The multi-lead electrocardiogram (ECG) stands as a cornerstone of cardiac diagnosis. Recent strides in electrocardiogram self-supervised learning (eSSL) have brightened prospects for enhancing representation learning without relying on high-quality annotations. Yet earlier eSSL methods suffer a key limitation: they focus on consistent patterns across leads and beats, overlooking the inherent differences in heartbeats rooted in cardiac conduction processes, while subtle but significant variations carry unique physiological signatures. Moreover, representation learning for ECG analysis should align with ECG diagnostic guidelines, which progress from individual heartbeats to single leads and ultimately to lead combinations. This sequential logic, however, is often neglected when applying pre-trained models to downstream tasks. To address these gaps, we propose CLEAR-HUG, a two-stage framework designed to capture subtle variations in cardiac conduction across leads while adhering to ECG diagnostic guidelines. In the first stage, we introduce an eSSL model termed Conduction-LEAd Reconstructor (CLEAR), which captures both specific variations and general commonalities across heartbeats. Treating each heartbeat as a distinct entity, CLEAR employs a simple yet effective sparse attention mechanism to reconstruct signals without interference from other heartbeats. In the second stage, we implement a Hierarchical lead-Unified Group head (HUG) for disease diagnosis, mirroring clinical workflow. Experimental results across six tasks show a 6.84% improvement, validating the effectiveness of CLEAR-HUG. This highlights its ability to enhance representations of cardiac conduction and align patterns with expert diagnostic guidelines.

Figures (8)

• Figure 1: Cardiac conduction and the relationship between the 12 ECG leads: (1) Electrical activity propagates through the heart (a) and reflects on the 12 leads in the same time window (b). (2) The 12 leads capture the heart's electrical activity from different views (c).
• Figure 2: Illustration of proposed CLEAR-HUG framework. CLEAR-HUG is composed of two stages: (1) CLEAR Pre-training stage to learn specified representations of 12 leads, and (2) CLEAR-HUG Finetuning to integrate the lead feature from the pretrained encoder and simulate the clinical diagnosis procedure to provide predictions for downstream tasks. In which MHA layer stands for multi-head attention layer vaswani2017attention.
• Figure 3: Illustration of proposed HUG head. The HUG head integrates the three ECG lead groups via a three-stage hierarchical framework.
• Figure 4: Visualization of reconstructed tokens for CLEAR variants w/o components: 12 leads in different colors, with poor reconstructions marked by red boxes.
• Figure 5: The visualization of activation ratios on 7 group combinations from HUG on different diseases.