Residual Attention-Based Denoising Convolutional Neural Network for Motion Artifact Removal from Single Channel EEG Signals

Motion artifacts in Electroencephalogram (EEG) recordings occur when electrode movement disrupts the electrode-skin interface, causing abrupt fluctuations that distort the recorded brain activity. Thus, removing these artifacts is essential for obtaining accurate and reliable representations of neural dynamics. The absence of spatial information in single channel EEG makes the removal of the artifacts considerably more difficult than in multichannel EEG systems. This work proposes a Residual Attention-Based Denoising Convolutional Neural Network (RADCNN) for effectively removing motion artifacts from single channel EEG signals. By integrating residual learning with an attention mechanism in a deep convolutional architecture, the framework enables robust feature extraction, enhances key signal components, and efficiently suppresses motion-induced noise. The publicly available PhysioNet motion artifact-contaminated EEG dataset was used to train and evaluate the proposed model. The RADCNN attains a Relative Root Mean Square Error (RRMSE) of 0.164 in the temporal domain, 0.315 in the spectral domain, and a Correlation Coefficient (CC) of 0.873, outperforming all other state-of-the-art approaches while preserving the temporal and spectral integrity of the original EEG signals.