Example VCG for a single ECG with multiple heartbeats.
Self-Supervised Contrastive Learning Enables Robust ECG-Based Cardiac Classification
We developed a contrastive self-supervised learning framework approach from large collections of unlabeled ECG recordings. We utilized three datasets in this study: a left ventricular ejection fraction-labeled ECG dataset, a large-scale unlabeled ECG dataset, and a serum potassium-labeled ECG dataset. When we applied these models to clinical tasks such as identifying reduced left ventricular ejection fraction and detecting high potassium levels, we found that the pretrained models consistently outperformed untrained models, especially when only small amounts of labeled data were available. Fine-tuning further improved performance, and explicitly modeling relationships between ECG leads helped the models generalize across different tasks. Overall, our results show that learning from large amounts of unlabeled ECG data can produce accurate and flexible tools for heart-related prediction tasks, even in settings where labeled data are scarce. A Key outcome from this study was the use of the vector-cardiogram as a universal framework for performing relevant modifications of ECG signals.
Deekshith Dade, Jake A. Bergquist, Rob S. MacLeod, Benjamin A. Steinberg, Tolga Tasdizen, Self-Supervised Contrastive Learning Enables Robust ECG-Based Cardiac Classification, Heart Rhythm O2, 2026, ISSN 2666-5018, https://doi.org/10.1016/j.hroo.2026.01.016.
Example ECG leads before (blue) and after (red) VCG modification and partial lead blanking as part of ECG modification during training.