Sparse Common Spatial Pattern for EEG Channel Reduction in Brain-Computer Interfaces

Abstract

Common spatial pattern (CSP) is widely used in brain-computer interfaces (BCIs) to extract features from the multichannel EEG signals. However, the CSP method easily overfits to the data of small training sets and its performance can be degraded by highly noisy and interference channels. Furthermore, more recording channels imply more processing and computation time in practical applications. To overcome these drawbacks, in this paper we propose a novel sparse CSP algorithm by introducing sparsity into spatial filters. The proposed method adopts $l_{1}$ norm as sparsity metric and constrains the ratio between variances of spatially filtered EEG signals of two classes larger than a specified threshold. To improve computational efficiency, an iterative approach based on general eigenvalue decomposition is further developed. The experimental results on 9 subjects from BCI competition datasets publicly available demonstrate that the proposed algorithm can achieve comparable classification accuracy even when the number of channels is small.