![[2602.23060] RhythmBERT: A Self-Supervised Language Model Based on Latent Representations of ECG Waveforms for Heart Disease Detection](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.23060] RhythmBERT: A Self-Supervised Language Model Based on Latent Representations of ECG Waveforms for Heart Disease Detection
Summary
RhythmBERT is a novel self-supervised language model designed for ECG waveform analysis, enhancing heart disease detection by treating ECG data as a structured language.
Why It Matters
This research addresses significant limitations in current ECG analysis methods by introducing a generative model that captures both rhythm semantics and morphology. By improving diagnostic accuracy with a single lead, it has the potential to transform cardiac care, making advanced analysis more accessible and efficient.
Key Takeaways
- RhythmBERT encodes ECG segments into symbolic tokens for better analysis.
- The model is pretrained on 800,000 ECG recordings, enhancing label efficiency.
- It achieves performance comparable to 12-lead systems using only a single lead.
- The approach aligns ECG analysis with physiological semantics, improving diagnostic capabilities.
- RhythmBERT can generalize across various heart conditions, including complex cases.
Computer Science > Machine Learning arXiv:2602.23060 (cs) [Submitted on 26 Feb 2026] Title:RhythmBERT: A Self-Supervised Language Model Based on Latent Representations of ECG Waveforms for Heart Disease Detection Authors:Xin Wang, Burcu Ozek, Aruna Mohan, Amirhossein Ravari, Or Zilbershot, Fatemeh Afghah View a PDF of the paper titled RhythmBERT: A Self-Supervised Language Model Based on Latent Representations of ECG Waveforms for Heart Disease Detection, by Xin Wang and 5 other authors View PDF HTML (experimental) Abstract:Electrocardiogram (ECG) analysis is crucial for diagnosing heart disease, but most self-supervised learning methods treat ECG as a generic time series, overlooking physiologic semantics and rhythm-level structure. Existing contrastive methods utilize augmentations that distort morphology, whereas generative approaches employ fixed-window segmentation, which misaligns cardiac cycles. To address these limitations, we propose RhythmBERT, a generative ECG language model that considers ECG as a language paradigm by encoding P, QRS, and T segments into symbolic tokens via autoencoder-based latent representations. These discrete tokens capture rhythm semantics, while complementary continuous embeddings retain fine-grained morphology, enabling a unified view of waveform structure and rhythm. RhythmBERT is pretrained on approximately 800,000 unlabeled ECG recordings with a masked prediction objective, allowing it to learn contextual representations in a label-ef...