LVCG: Learning Cardiac Representations in Vectorcardiogram Space
A new self-supervised learning framework called LVCG operates in vectorcardiogram (VCG) space rather than standard ECG signal space. Standard twelve-lead ECG records multiple projections of the same cardiac activity, introducing redundancy and risk of overfitting. LVCG, motivated by the Frank VCG model, learns a unified latent representation directly in VCG space. This is the first general self-supervised representation learning framework designed for this physically grounded space. The approach aims to improve tasks like disease diagnosis and clinical report generation.
Key facts
- ECG is a cornerstone of cardiac assessment
- Standard twelve-lead ECG represents multiple projections of the same cardiac activity
- Representation learning in ECG space introduces redundancy and risk of overfitting
- LVCG is motivated by the Frank vectorcardiogram (VCG) model
- LVCG learns a unified latent representation in VCG space
- LVCG is the first general self-supervised representation learning framework for VCG space
- The framework is designed for tasks like disease diagnosis and clinical report generation
- The paper is available on arXiv with ID 2605.31249
Entities
Institutions
- arXiv