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

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Electrocardiogram (ECG)-based arrhythmia classification is crucial for widespread screening and continuous monitoring.
• Challenges in ECG analysis stem from distributed diagnostic cues in waveform morphology and temporal rhythm.
• Accurate recognition requires sophisticated methods to interpret complex ECG signals.

Purpose of the Study:

• To develop TransECG-Net, a novel physiology-guided CNN-Transformer hybrid network.
• To achieve accurate, AAMI-aligned five-class heartbeat classification.
• To evaluate the network's performance, noise robustness, and edge-deployability.

Main Methods:

• Developed TransECG-Net, integrating CNN for local morphology and Transformer for temporal dependencies.
• Employed a learnable dimension-wise gated fusion module to combine CNN and Transformer representations.
• Utilized public ECG recordings, segmented into heartbeat windows, and stratified into training, validation, and testing sets (70%/15%/15%).

Main Results:

• TransECG-Net achieved 99.52% accuracy on 5,000 testing samples, correctly classifying 4,976.
• Macro-averaged F1-score was 99.52%, with high class-wise F1-scores (e.g., 99.90% for Normal).
• Outperformed existing models like DeepECG-Net (98.30%) and Hybrid CNN-BLSTM (94.20%) with low latency (35ms) and memory footprint (28MB).

Conclusions:

• TransECG-Net offers accurate, physiology-guided ECG arrhythmia screening.
• The model demonstrates noise tolerance and suitability for edge deployment.
• Enables efficient arrhythmia detection for both wearable and clinical monitoring applications.