Improving BCI performance by modified common spatial patterns with robustly averaged covariance matrices

EEG single-trial analysis requires methods that are robust against noise and disturbance. In this contribution, based on the framework of robust statistics, we propose a simple modification of Common Spatial Patterns by the robust calculation of covariance estimators against outlying trials caused, for example, by artifacts. We tested the proposed robust filters with EEG recordings from 80 subjects and obtained, not only a significant improvement in performance, but for some subjects, also better neuro-physiologically interpretable filters.