Learning Cardiac Latent Representations in Vectorcardiogram Space

Researchers introduce LVCG, a self-supervised learning framework that represents cardiac electrical activity in vectorcardiogram (VCG) space rather than traditional ECG signal space. By learning unified latent representations instead of lead-specific artifacts, the method reduces redundancy, minimizes spurious correlations, and demonstrates improved generalization across cardiac assessment tasks.

This research addresses a fundamental limitation in cardiac signal processing: the redundancy inherent in twelve-lead ECG systems. Traditional ECG operates in observable signal space, where multiple leads represent projections of identical underlying cardiac electrical activity from different angles. This structural redundancy creates opportunities for spurious correlations and overfitting, degrading model robustness across different clinical settings and patient populations.

The proposed LVCG framework leverages the Frank vectorcardiogram model as theoretical foundation, shifting representation learning into a physically grounded latent VCG space. This approach extracts view-invariant features that capture essential cardiac electrical dynamics while eliminating lead-specific artifacts. The self-supervised learning paradigm enables the model to learn representations without extensive labeled data, a significant practical advantage in healthcare applications where annotated datasets remain scarce.

The framework's performance gains are particularly pronounced in domain shift scenarios, where models trained on one patient population generalize to different cohorts or clinical conditions. This capability directly translates to improved clinical utility, as cardiac assessment systems must perform reliably across diverse patient demographics and healthcare settings. The reduction in redundancy potentially strengthens model interpretability by focusing on medically relevant features rather than signal artifacts.

For the broader AI-healthcare space, this work exemplifies how domain-specific knowledge—here, the geometric principles underlying cardiac electrophysiology—can enhance machine learning architectures. The demonstrated improvement in generalization and robustness suggests similar principles could apply to other physiological signal processing tasks, influencing how neural networks approach medical time-series analysis.

• →LVCG operates in vectorcardiogram space rather than observable ECG space, eliminating redundancy from multiple lead projections
• →Self-supervised representation learning in VCG space reduces spurious correlations and improves model generalization across domains
• →The framework demonstrates enhanced robustness in domain shift settings, critical for real-world clinical deployment
• →View-invariant latent representations capture essential cardiac electrical activity while filtering out lead-specific artifacts
• →Physically grounded latent spaces offer a template for integrating domain knowledge into deep learning architectures