Event-Related Potential Classification Based on EEG Data Using xDWAN with MDM and KNN

Abstract

A way of measuring a brain response during cognitive, sensory, or motor event is called event related potential (ERP). Electroencephalography (EEG) data during ERP can be used to cognitive operations. Recently, to classify ERP from EEG signal has begun to apply a compact convolutional neural network (EEGNet) but its performance depends on the data size.To discriminate between correct and incorrect ERP responses of a subject here a classification algorithm with smoothing of xDAWAN spatial filtering method is proposed. A spatial filtering xDAWAN method extracted only true ERP signal from EEG signal by increasing signal to noise ratio and discarding noise. We have reduced the dimension of the true ERP data by applying Riemannian geometry. Riemannian geometry calculated covariance matrix from true ERP data and all covariance data mapped into a euclidean space called tangent space and produced Riemannian feature vector. Finally, feed this Riemannian feature vector by a machine-learning algorithm to predict ERP response. As a machine-learning algorithm, we used here specially K-nearest neighbor (KNN) and Minimum Distance to Mean (MDM) as a classifier. To evaluate our model, here used EEG data, which contains data for left ear auditory, right ear auditory stimulation and left visualization, right visualization stimulation. xDAWAN is effective due to EEG is a narrow and noisy signal. Performance of Riemannian geometry also not depend on noisy or data size. Our model achieved better performance than other did specially EEGNet. We believe our model will be considered as a great invention in this research domain.