Learning Lead-Invariant ECG Representations via Cross-View Contrastive Learning

Electrocardiograms (ECGs) are inherently multi-view signals, with cardiac electrical activity observed through multiple leads that differ in anatomical orientation and sensitivity to underlying cardiac sources. Although contrastive learning has been previously applied to ECG signals, its effects on the geometry of learned representations remain poorly understood. Here, we study how crossview contrastive learning between single-lead and full 12-lead ECGs reshapes representation structure. Using the PTB-XL dataset, we train a self-supervised model to align the embeddings of randomly sampled single-lead views and the corresponding 12-lead signals. We find that contrastive learning induces near complete linear alignment across leads, a structure absent in raw waveform space, and that embeddings of full 12-lead ECGs are also largely linearly recoverable from single-lead embeddings. We further show that this geometric alignment is accompanied by substantially improved diagnostic decodability across leads using linear probes, while preserving physiologically structured lead differences. Together, these results provide a mechanistic explanation for the robustness of contrastively learned ECG representations.