Automatic and Semi-automatic Approaches for Selecting Prominent Spatial Filters of CSP in BCI Applications

Abstract

Common Spatial Patterns (CSP) has become a popular approach for feature extraction in Brain Computer Interface (BCI) research. This is because it can provide a good discrimination between 2 motor imaginary tasks. In theory, the first and the last spatial filters from CSP should exhibit the most discriminate between 2 classes. But in practice, this is not always true, especially if either of these 2 filters emphasizes on a channel that has high variability of variance of sample matrices among trials. Such spatial filter is unstable on a single class, and thus it is not appropriate to use for discrimination. Furthermore, one or both of these 2 spatial filters may not localize the brain area that relates to motor imagery. The desired spatial filters may be at the second or at an even greater order of sorted eigenvalue. In this work, we propose to find an appropriate set of spatial filters of CSP projection matrix, which may provide higher classification accuracy than using just 2 peak spatial filters. We present 2 selection approaches to select the set of prominent spatial filters: the first one is automatic approach; the second one is semi-automatic approach requiring manual analysis by human. We assessed both of our approaches on the data sets from BCI Competition III and IV. The results show that both selection approaches can find the appropriated prominent spatial filters.