Multi class EEG artifacts classification and removal using adaptive neural filter

Abstract

Numerous adaptive filtering methods have been developed to enhance the quality of electroencephalogram (EEG) signals and thereby improve the efficacy of artifact removal and classification. The goal of this work is to categorize and remove artificial, eye blink, and muscle motion artifacts. An adaptive neural network (NN) based filter is proposed to simultaneously suppress all motion artifacts. To increase classification accuracy, an extended statistical feature set is employed. Using hybrid principal component analysis (PCA) and quadratic support vector machine (SVM) kernel combinations based on Bayesian optimization improves the prediction accuracy of EEG artifact classification. The results of artifact removal using the proposed adaptive neural filter are presented. The NN filter is trained on the filtered ensemble empirical mode decomposition (EEMD) and canonical correlation analysis (CCA) signal to eliminate motion artifacts. The dataset comprises eye blinks, muscle artifacts, and synthetically generated artifact classes. It is found that quadratic and cubic SVM offer an accuracy of 94.7%, which increases to 100% using the proposed optimization approach. Additionally, it is observed that the NN filter is capable of completely eliminating high-peak eye blink artifacts while maintaining the shape of the true EEG signal based on learning.