Capturing Cardiac Cyclicity through Phase-Equivariant Self-Supervised Learning

TL;DR - Winder is a self-supervised architecture that explicitly encodes cardiac-cycle symmetry by separating phase-invariant features from phase-rotating harmonic representations. It preserves diagnostically useful ECG information with interpretable latent geometry and roughly 1 million parameters.

• Introduces a phase-equivariant objective tailored to cyclic physiological signals.
• Uses a fixed, closed-form transport operator derived from cardiac-cycle geometry, adding no learned parameters.
• Achieves PTB-XL linear-probe diagnostic accuracy within the reported range of state-of-the-art self-supervised methods.
• Produces phase-equivariant representations tied directly to a measurable physiological quantity.