The design of spatial selection using CUR decomposition to improve common spatial pattern for multi-trial EEG classification

Abstract

The most important factor in EEG signal processing is the determination of relevant features in encoding the meaning of the signal. Obtaining relevant features for EEG can be done using a spatial filter. The Common Spatial Pattern (CSP) is known to produce discriminative features when processing EEG signals. Yet, CSP is also sensitive to noise and is channel-dependent, as it is considered to be a spatial filter. However, the disadvantage of CSP is that channels containing only noise are also considered as active channels. In this paper, the design of a filter for spatial selection is proposed using CUR decomposition to select important channels or the time segment of EEG trials in order to improve CSP performance. CUR decomposition can also be used as a noise rejection technique because CUR can be used in factorizing the given EEG signals. In other words, CUR decomposition rejects the non-active channels, which typically contain noise, before spatially filtering the EEG signals. Once the EEG signal is decomposed based on the importance of the channels, time segmentation, and EEG factorization, the decomposed signal can be used as input to the CSP. In general, three approaches were proposed in this framework: (1) channel selection, i.e., C selection; (2) time segment selection, R; and (3) signal factorization, U. Furthermore, the performance accuracy between the original CSP and CSP in which the input was spatially filtered by the proposed framework was validated using datasets IVa of BCI competition III. The test results show that the CSP with spatial selection using C selection and U factorization offers 12% and 9% improvement compared to the original CSP, respectively. Hence, the proposed method in this study can be used as a spatial filter to improve the CSP performance.