Spectrogram–scalogram fusion with CNN–BiLSTM– attention for record-derived ECG segment classification
Background Beat segments extracted from recordings labelled as Normal rhythm, atrial fibrillation (AF), or premature ventricular contraction (PVC) may exhibit distinct morphological and transformed-domain characteristics. In this study, AF-associated beats inherit the label of their source AF recording and do not represent independent beat-level diagnosis. Methods A supervised three-class ECG classification study was conducted using the PTB-XL dataset. Following filtering, normalization, and Pan–Tompkins R-peak detection, 2-second, 12-lead ECG segments (1,000 × 12) were extracted. These R-peak-centered segments could contain multiple cardiac cycles and inherited their source recording labels. Spectrograms and scalograms were then generated and classified using the CNN–BiLSTM–attention framework. Results On the independent held-out test set, the proposed CNN–BiLSTM–attention model achieved a reported accuracy of 93.4%, compared with 88.6%, 86.9%, and 91.8% for CNN-only, BiLSTM-only, and CNN–BiLSTM, respectively. The training curves were examined over 300 epochs, but complete convergence could not be established from the recorded validation trajectories. Conclusion The proposed framework effectively classified record-derived Normal-associated, AF-associated, and PVC-associated beat segments. AF-associated classification should not be interpreted as diagnosis of AF from an isolated beat.
Authors
- Taisheng Zeng (ORCID: https://orcid.org/0000-0001-7806-0351)
- 张之东
- Kavimbi Chipusu (ORCID: https://orcid.org/0009-0005-9258-6030)
- Jianlong Huang
- Yunyi Zeng
- Guangfeng Li
Institutions
- Quanzhou Normal University (CN)
- University of Saskatchewan (CA)
- Education University of Hong Kong (HK)
- Henan Provincial People's Hospital (CN)
Publication Details
- Journal
- Frontiers in Physiology
- Published
- 2026-11-18
- DOI
- https://doi.org/10.3389/fphys.2026.1854203
- Primary Topic
- ECG Monitoring and Analysis
- Type
- article
- Field-Weighted Citation Impact
- 0.00