Robust common spatial filters with a maxmin approach

Electroencephalographic signals are known to be non-stationary and easily affected by artifacts, therefore their analysis requires methods that can deal with noise. In this work we present two ways of calculating robust common spatial patterns under a maxmin approach. The worst-case objective function is optimized within prefixed sets of the covariance matrices that are defined either very simply as identity matrices or in a data driven way using PCA. We test common spatial filters derived with these two approaches with real world brain-computer interface (BCI) data sets in which we expect substantial "day-to-day" fluctuations (session transfer problem). We compare our results with the classical common spatial filters and show that both can improve the performance of the latter.