Improving Classification Performance of BCIs by Using Stationary Common Spatial Patterns and Unsupervised Bias Adaptation

Non-stationarities in EEG signals coming from electrode artefacts, muscular activity or changes of task involvement can negatively affect the classification accuracy of Brain-Computer Interface (BCI) systems. In this paper we investigate three methods to alleviate this: (1) Regularization of Common Spatial Patterns (CSP) towards stationary subspaces in order to reduce the influence of artefacts. (2) Unsupervised adaptation of the classifier bias with the goal to account for systematic shifts of the features occurring for example in the transition from calibration to feedback session or with increasing fatigue of the subject. (3) Decomposition of the CSP projection matrix into a whitening and a rotation part and adaptation of the whitening matrix in order to reduce the influence of non-task related changes. We study all three approaches on a data set of 80 subjects and show that stationary features with bias adaptation significantly outperforms the other combinations.