An Attention-UNet Model to Remove Artifacts from Single-Channel EEG

Electroencephalography (EEG) signals are often unavoidably contaminated by physiological artifacts during acquisition. Reducing the effects of these artifacts is important for various EEG-based interpretation tasks. However, the existing deep learning EEG denoising methods currently have some shortcomings: information lost caused by multiple downsampling and upsampling, artifact-related regions not considered separately, etc. In this study, a novel deep learning network named Attention-Unet (AUNet) is proposed for EEG denoising. AUNet uses skip connections to connect the decoder and encoder for feature information fusion, while the attention gate focuses on activating artifact-related area information. The developed network is evaluated on a semi-simulated dataset including ocular electrocardiography (EOG), electromyography (EMG), and electrocardiography (ECG). The experimental results show that the presented AUNet outperforms existing methods in metrics including relative root mean square error (RRMSE) and signal-to-noise ratio improvement (ΔSNR).