Single-Trial Classification of EEG data in non-stationary environments

Brain-Computer Interface (BCI) systems aim to translate the intent of a subject measured from brain signals e.g. EEG into control commands. A popular paradigm for BCI communication is motor imagery i.e. subjects perform the imagination of movements with their feet or hands, the imagined movements are detected and translated into computer commands. A major challenge in BCI research are non-stationarities in the EEG signal which are variations of the signal properties within and across experimental sessions and subjects. They can lead to poor performances with error rates larger than 30% as most machine learning algorithms implicitly assume stationary data. The focus of my research is to understand non-stationarities in EEG data and develop methods and representations which are invariant. Common Spatial Pattern (CSP) is a method which is often used in BCI systems to project the high-dimensional EEG signal into a lower dimensional subspace. Since CSP can be negatively affected by artefacts in the data, it often only works well after an additional data cleaning step which may involve a human expert. In order to overcome this drawback and with the assumption that the non-stationary components of the EEG signal e.g. visual activity or muscle artefacts are not relevant to the BCI task, we proposed a robust extension called stationary CSP [Woj2011a]. This method penalizes non-stationary projection directions, thus produces more robust filters and significantly improves the classification accuracy. As a by-product it also provides physiologically meaningful patterns showing the sources of nonstationarities. We could obtain further performance gain by combining sCSP with unsupervised adaptation [Woj2011b]. In future research I will consider other invariant representations, especially subspaces created by the stationary subspace analysis (SSA) method.