Unmasking Hidden Waves: A Proof-of-Concept Transformer-Based Electrocardiogram Signal Decomposition for Complex Arrhythmias
Abstract Standard automated electrocardiogram (ECG) algorithms rely on sequence classification, frequently failing to provide morphological insight during complex tachycardias where critical physiological events, such as P waves, are buried. Here, we explore the feasibility of shifting the paradigm from proxy-based classification to explicit signal decomposition to unmask hidden cardiac electrical activity. We developed the ECG Signal Decomposition Transformer (ESD-former), a hybrid 1D U-Net architecture with a Transformer bottleneck, designed to continuously disentangle raw ECG signals into explicitly separated P, QRS, and T tracks. Trained on physiologically constrained synthetic overlapping arrhythmias and optimized using a custom ShapeAwareLoss, the model was tested for proof-of-concept real-world feasibility on continuous 5-minute resting ECGs. While baseline models failed on overlapping waveforms, the ESD-former reconstructed hidden features like retrograde P waves in synthetic examples. In an exploratory clinical validation, it achieved a Positive Percent Agreement (PPA) of 92.9% for QRS, 76.0% for T waves, and 82.7% for atrial activity against a reference software, while successfully generating 81 additional deeply buried P wave detections that were subsequently confirmed by manual expert review. This explicit signal decomposition framework provides a foundational proof-of-concept that, following rigorous clinical validation, could potentially assist electrophysiologists in the assessment of complex arrhythmias.
Authors
- Michał Szafarczyk (ORCID: https://orcid.org/0009-0000-2270-5106)
- Peter van Dam
Institutions
- Jagiellonian University (PL)
- Szpital Uniwersytecki w Krakowie (PL)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-17
- DOI
- https://doi.org/10.21203/rs.3.rs-10680222/v1
- Primary Topic
- ECG Monitoring and Analysis
- Type
- preprint