The Artifact Removal Algorithm for EEG Based on Contrastive Learning

Electroencephalogram (EEG) is easily contaminated by various artifacts, and many studies use the decomposition and blind source separation method to identify and remove artifact-related components. However, most studies using the threshold-based decomposition-blind source separation methods, leading to excessive suppression of brain activity and may not be able to remove specific artifact. Recently, self-supervised learning based on contrastive learning is developing rapidly and has achieved great success in the field of computer vision. At the same time, recent research shows that contrast learning can learn better representation with more generalization ability by comparing the pairwise distance within the feature space, and can achieve better results than supervised learning models after fine-tuning. In this study, an adaptive few-channel EEG denoising algorithm based on outlier detection was proposed. The potential components are extracted by blind source separation method, and the anomaly detection algorithm is used to identify the abnormal components according to the distribution of the components in the feature space. The entire process is completely data-driven and adaptive, so it is possible to remove a variety of different artifacts at the same time in a single session without over-inhibiting brain activity.