Deep learning for suppressing EMG and motion artifacts in armband ECG R-peak detection

Continuous cardiac monitoring through wearable electrocardiogram (ECG) devices faces significant challenges from electromyogram (EMG) contamination and motion artifacts during daily activities. This paper introduces a novel deep learning framework employing a convolutional neural network accentuating encoder-decoder (CNNAED) architecture for robust R-peak identification in artifact-corrupted armband ECG signals. The study enrolled 10 healthy participants who underwent simultaneous 24 h monitoring using both experimental armband systems and clinical-grade Holter monitors. Time-frequency spectrograms derived from 10 s ECG segments provided input features for the CNNAED architecture. The network training objective focused on selective R-peak enhancement through amplitude accentuation while suppressing interfering signal components. Model validation employed subject-independent testing protocols emphasizing daytime recordings with increased artifact prevalence. Statistical analysis adopted paired t tests across 11,598 validated segments to assess performance improvements. Heart rate estimation accuracy showed substantial improvement, with mean absolute error decreasing from 13.26 to 9.37 beats/min following CNNAED processing, indicating 29.3% improvement with high statistical significance (p < 0.001). Root mean square of successive differences (RMSSD) showed 6.1% improvement (p < 0.001), indicating preserved cardiac timing characteristics essential for heart rate variability analysis. The proposed methodology provides a statistically robust solution for EMG artifact mitigation in wearable cardiac monitoring, enabling improved signal quality and expanded usable data availability during ambulatory conditions.