EOG Artifacts Reduction from EEG Based on Deep Network and Recursive Least Squares Adaptive Filter

Abstract

We developed a cascade of deep network and recursive least squares adaptive filter (DN-RLS) for electrooculogram (EOG) artifacts removal. The proposed method can be divided into offline stage and online stage. During the offline stage, EOG signals are used to train an DN to learn features of EOG signals. During the online stage, the learned DN is used to extract EOG artifacts from electroencephalogram (EEG), then a RLS filter is used to further remove EOG artifacts. The proposed method not only just needs few number of EEG channels in removal process, but also doesn’t need additional EOG recordings during online stage. We compared the proposed method to the independent component analysis (ICA) technique and a shallow network combined with RLS method. Experimental results show that the DN-RLS can learn features of EOG artifacts better and result in higher classification accuracy.