Single trial analyses of encephalogram data

In this thesis, inspired by the development of the Brain-computer-interface (BCI) technology, we present novel methods for the analysis of macroscopically recorded brain signals. Here the focus is put on improved feature extraction methods, the detection of mental states and the analysis of variability of brain responses. Conditional event-related (de-)synchronization: The fluctuation of signal power in a narrow band induced by an event is conventionally termed event-related (de-)synchronization and is quantified as the relative deviation from the mean baseline activity. We extend the ERD terminology with respect to a generalized reference. To this end, we oppose the time course of the event-related activity against those obtained from single trials without specific stimulus processing. From this generalized approach we derive a method to determine the dependencies of the ERD response on initial cortical states. A comparative study of surrogate and real ERD data validates this approach. Spatio-spectral filters: The common-spatial-pattern algorithm (CSP) determines optimally discriminative spatial filters from multivariate broad-band signals. We extend the conventional algorithm such that it additionally obtains simple frequency filters. This enables adaptation to the individual characteristics of the power spectrum and thus improves feature extraction. An empirical comparison with the conventional CSP method reveals the advantages of our approach in the context of the classification of imaginary unilateral hand movements. Extraction of event-related potentials (ERP): Independent component analysis (ICA) is a tool for statistical data analysis that is able to linearly decompose multivariate signals into their underlying source components. We present an ICA method that uses prior knowledge about the phase-locked property of ERPs for their improved extraction from single trial EEG. The application on artificially generated and real world data validates this approach in terms of an improved signal-to-noise ratio of the extracted ERPs. Adaptive feature combination across time: Lateralized mu-rhythm ERD and lateralized movement-related potentials are the most commonly used discriminative features for the classification of imaginary hand movements. In the context of real time classification we present a method that efficiently combines these temporally differently accentuated features. To this end, we first train weak classifiers for each time instance and each feature separately. Subsequently we combine these weak classifiers in a strictly causal, probabilistic manner. The effectiveness of this approach was proven by its successful application to data from the international BCI competitions in 2003 and 2005.