Channel Extension Uniting Information Theoretic Feature Extraction Based Multiclass Common Spatial Pattern Used in BCI

摘要

Common Spatial Pattern (CSP) has been widely used in decoding the spatial patterns of corresponding neuronal activities from electroencephalogram (EEG) signals in Brain-computer Interface (BCI), and has attained good performance in the discrimination of two-class motor imagery. However, for multiple classes of motor imagery, the effect of discrimination is unsatisfactory. Information Theoretic Feature Extraction (ITFE) provides a method to choose Independent Components (ICs) that approximately maximizes mutual information of ICs and eliminates the need for heuristics in multiclass CSP. But the effect of discrimination is also unsatisfactory in the classification of multiclass EEG data with a small number of recording channels. To solve the problem, this paper proposes a method for extending channels, named Channel Extension (CE), through the time delay on the original signal and superimposition of the new signals on the spatial pattern, by which we extend channels without increasing electrodes. We pointed out the relationship between spatial patterns and the classification accuracy, and gave a general rule to choose the spatial patterns. We also tested the best multiple of channel extension. The experimental results showed that the proposed method outperforms the original ITFE based CSP algorithm in terms of classification accuracy.