Physiology-guided beat-level arrhythmia classification from ECG using a CNN-transformer hybrid neural network

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Accurate electrocardiogram (ECG)-based arrhythmia classification is crucial for widespread screening and continuous patient monitoring.
• Current methods face challenges due to the complex interplay of local waveform morphology and global rhythm context in diagnostic cues.
• The need for robust and efficient arrhythmia detection systems is paramount for timely clinical intervention.

Purpose of the Study:

• To develop and evaluate TransECG-Net, a novel physiology-guided hybrid CNN-Transformer network for five-class heartbeat classification aligned with AAMI standards.
• To assess the performance, noise robustness, and edge-deployability of the proposed TransECG-Net model.
• To compare TransECG-Net against existing deep learning models for ECG arrhythmia classification.

Main Methods:

• Developed TransECG-Net, integrating a CNN branch for local ECG waveform features (P-QRS-T morphology, QRS width, amplitude) and a Transformer branch for global temporal dependencies.
• Employed a learnable dimension-wise gated fusion module to combine CNN and Transformer representations.
• Utilized public ECG recordings, segmented into fixed-length windows, and stratified into training (70%), validation (15%), and testing (15%) sets for comprehensive evaluation.

Main Results:

• TransECG-Net achieved a high accuracy of 99.52% on the testing set, correctly classifying 4,976 out of 5,000 samples.
• Demonstrated strong class-wise performance with F1-scores ranging from 99.12% to 99.90%, yielding a macro-averaged F1-score of 99.52%.
• Outperformed established models like DeepECG-Net (98.30%) and Hybrid CNN-BLSTM (94.20%), while maintaining low latency (35ms) and memory footprint (28MB).

Conclusions:

• TransECG-Net offers a highly accurate, physiology-guided, and noise-tolerant solution for ECG arrhythmia screening.
• The model's efficiency and small footprint make it suitable for edge deployment in wearable devices and clinical monitoring.
• This hybrid approach advances the capabilities of automated ECG analysis for improved cardiovascular health management.