Abstract
Electrocardiogram (ECG) captures the heart's electrical signals, offeringvaluable information for diagnosing cardiac conditions. However, the scarcityof labeled data makes it challenging to fully leverage supervised learning inmedical domain. Self-supervised learning (SSL) offers a promising solution,enabling models to learn from unlabeled data and uncover meaningful patterns.In this paper, we show that masked modeling in the latent space can be apowerful alternative to existing self-supervised methods in the ECG domain. Weintroduce ECG-JEPA, a SSL model for 12-lead ECG analysis that learns semanticrepresentations of ECG data by predicting in the hidden latent space, bypassingthe need to reconstruct raw signals. This approach offers several advantages inthe ECG domain: (1) it avoids producing unnecessary details, such as noise,which is common in ECG; and (2) it addresses the limitations of na\"ive L2 lossbetween raw signals. Another key contribution is the introduction ofCross-Pattern Attention (CroPA), a specialized masked attention mechanismtailored for 12-lead ECG data. ECG-JEPA is trained on the union of several openECG datasets, totaling approximately 180,000 samples, and achievesstate-of-the-art performance in various downstream tasks including ECGclassification and feature prediction. Our code is openly available athttps://github.com/sehunfromdaegu/ECG_JEPA.