Key result
Bi-stream attention model achieves ~94% accuracy for ECG arrhythmia classification under inter-record evaluation.
Why the study?
Accurate classification of ECG beats remains difficult when minority arrhythmias exhibit marked morphological variation across recordings.
Does an attention-based bi-stream deep learning architecture improve ECG arrhythmia beat classification accuracy and macro-F1 compared to a temporal mean-pooling control?
Does an attention-based bi-stream deep learning architecture improve ECG arrhythmia beat classification accuracy and macro-F1 compared to a temporal mean-pooling control?
A compact attention-based bi-stream deep learning model achieves high overall ECG beat classification accuracy and noise robustness, though inter-record generalization for minority arrhythmia classes (supraventricular ectopic and fusion beats) remains limited.
Requires external validation before clinical ECG deployment; leaves open whether Bahdanau attention outperforms mean pooling for minority arrhythmias.
Accurate classification of electrocardiogram (ECG) beats remains difficult when minority arrhythmias exhibit marked morphological variation across recordings. A compact bi-stream architecture is presented in which target-beat morphology is encoded by a convolutional stream and short-term rhythm context is represented by a recurrent stream formed from five causal beat tokens. Conventional additive Bahdanau attention is evaluated against an otherwise identical temporal mean-pooling control. Under a secondary heartbeat-random 80/20 protocol, the attention model attains 99.3333% accuracy and 93.1581% macro-F1. Under the more demanding canonical DS1 → → DS2 inter-record protocol, the same model attains 93.7949±1.9588% accuracy and 36.3326±0.5670% macro-F1 over three seeds; a strict sensitivity analysis excluding record 202 gives 93.5877±2.0469% accuracy and 35.6715±0.7045% macro-F1. The divergence between aggregate accuracy and macro-F1 reveals weak transfer of supraventricular ectopic (S) and fusion (F) beats across records. The attention model contains 85,627 parameters. Quantitative evaluation on the MIT-BIH Noise Stress Test Database yields mean SNR improvements of 19.1143, 5.9946, and 4.7341 dB for baseline wander, muscle artifact, and electrode-motion noise, respectively. The results characterize a compact morphology–rhythm formulation while identifying minority-class inter-record generalization as the principal limitation.
No takes yet. Share an insight, caveat, or question.
Rayasam et al. (2026) studied ECG arrhythmias. Attention-based bi-stream deep learning architecture (additive Bahdanau attention) vs. Identical temporal mean-pooling control was evaluated on ECG beat classification accuracy under canonical DS1 to DS2 inter-record protocol. The attention-based bi-stream model achieved 93.79% accuracy and 36.33% macro-F1 under the canonical inter-record protocol for ECG arrhythmia classification.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: