Band selection by distance of spatial patterns for brain machine interfacing

This paper has proposed a new method based on a metric between frequency components to design the passband for feature extraction for electroencephalogram In motor imagery based brain machine interfaces (MI-BMI), because the appropriate passband should be tuned for data, this problem is a crucial issue. In the proposed method, observed signals are divided into some frequency bins and covariance matrices in each bin are obtained. And then we measure the distances of the covariance matrices between the bins by the common spatial pattern method. The near bins on this metric are selected as the passband. By adopting this metric, we can select the passbands being useful for classification problem in MI-BMI. We conducted the experiment with a dataset of MI-BMI. The proposed method outperforms some conventional methods in the classification accuracy.