SA-Divider: A Novel Automatic EEG Artifact Removal Method Based on Self-Attention Convolutional Neural Network

Various artifacts, often accompanied during the acquisition of electroencephalograph (EEG), hinder further progress of EEG-based applications. However, hand-craft and semi-automatic-based methods are inaccurate and timeconsuming, and automatic artifact removal remains challenging. In this study, we propose a novel automatic EEG artifact removal method based on self-attention convolutional neural network (CNN), named SA-Divider. SA-Divider consists of a Self-attention module (SA) and a Divider module. The SA is designed to capture more accurate temporal and spatial features of EEG, and the Divider, which comprises an Encoder, a Decoder, and a Decomposer, aims to acquire better performance of artifact removal in EEG with less computation time. Experimental results on two EEG datasets show that SA-Divider achieved significant performance in four metrics and time consumption, indicating that SA-Divider can effectively remove artifacts and retain clean EEG-related information from raw EEG signals. These findings suggest that the proposed SA-Divider provides an efficient method to remove artifacts and extract brain cognitive-related EEG components while reducing computational complexity.