U-Net Based Decomposition of Respiratory Surface Electromyography for Removal of Cardiac Artifacts

Diaphragmatic surface electromyography (dEMG) has emerged as a promising tool for continuous respiratory monitoring, with broad clinical applications. However, real-time analysis of dEMG is hindered by strong electrocardiogram (ECG) contamination, limiting its utility in clinical settings. To address this, we propose a U-Net-based model for dEMG decomposition, enabling separation of ECG and EMG components. Unlike traditional methods, our approach does not rely on R-peak detection and effectively handles non-linear signal decomposition, overcoming key limitations of gating and wavelet-based techniques. The model was trained on synthetic mixed signals, combining PTB-XL ECG recordings with simulated EMG signals generated using the BioMime model, and random noise. We demonstrate that our method significantly improves dEMG separation, achieving a superior SNR improvement (SNRi) of 14.03 dB, compared to 0.18 dB and 4.46 dB for conventional gating and wavelet-based approaches, respectively.Clinical Relevance-This work represents a key component of a broader analysis pipeline required to improve non-invasive respiratory monitoring. Enhancing dEMG signal quality, our approach could significantly advance wearable respiratory monitoring devices, facilitating continuous assessment of lung function for asthma and other respiratory diseases in both hospital and home settings.