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

Accurate ECG-based arrhythmia classification is essential for large-scale screening and continuous monitoring, but recognition remains challenging because diagnostic cues are distributed across local waveform morphology and temporal rhythm context.

We developed TransECG-Net, a physiology-guided CNN-Transformer hybrid network for AAMI-aligned five-class heartbeat classification. The CNN branch extracts local P-QRS-T morphology, QRS-width variation, and amplitude-shape features, while the Transformer branch models global temporal dependencies using positional encoding and multi-head self-attention. Both representations are combined through a learnable dimension-wise gated fusion module. Public ECG recordings were segmented into fixed-length heartbeat windows and split into training, validation, and testing subsets using a stratified 70%/15%/15% protocol. Performance was assessed using accuracy, precision, recall, specificity, and F1-score, with additional noise-robustness and edge-device evaluation.

TransECG-Net correctly classified 4,976 of 5,000 testing samples, achieving 99.52% accuracy. Class-wise F1-scores were 99.90% for N, 99.43% for L, 99.56% for R, 99.61% for A, and 99.12% for V, with a macro-averaged F1-score of 99.52%. It outperformed DeepECG-Net (98.30%) and Hybrid CNN-BLSTM (94.20%) while maintaining 35 ms latency and 28 MB memory footprint.

TransECG-Net supports accurate, physiology-guided, noise-tolerant, and edge-deployable ECG arrhythmia screening for wearable and clinical monitoring.