Nonnegative matrix factorization common spatial pattern in brain machine interface

Abstract

Fast and accurate discrimination of Electroencephalography (EEG) data is necessary for controlling brain machine interface. This paper introduces a novel method to discriminate 2-class motor imagery states (left and right hand) using nonnegative matrix factorization (NMF), common spatial pattern (CSP) and random forest. Conventionally CSP is used after extracting frequency band segment of EEG signal, which is called bandpass-filtered CSP (BPCSP). Especially filter bank CSP (FBCSP) has been extensively used to extract feature vectors from EEG data. However in these methods, the range of frequency band needed to be specified in advance and the performance depends on the selected frequency band. Our new method can decide the frequency band automatically by using NMF (NMFCSP). After the feature vectors were extracted from EEG data, random forests (RF) method was adopted as a classification algorithm. The mean accuracy rate of 2-class classifier using NMFCSP was 78.8±3.27%. This is higher than the accuracy rate of BPCSP (64.4±8.53%) and FBCSP (68.4±6.81%).