Feature selection for brain-computer interfaces

Brain-Computer Interface (BCI) systems are a means of establishing communication for severely paralyzed patients. Based on the brain activity signals during the execution of mental tasks by a user, a computer system translates those signals first into higher-level features and finally into control commands for communication interfaces. This involves a number of algorithmic steps that have to be optimized individually for each patient in order to attain high performance of the BCI. One of these steps is the choice of a suitable set of brain signal features. This set is supposed to provide good discriminability of the mental tasks, allow for introspection, simplify the experimental effort and thus increase acceptance in patients. In terms of EEG electrodes, a smaller feature set entails a reduction of the immense effort of the daily setting-up of electrodes prior to the start of an experiment. The problem of feature selection is hard to tackle as optimal features may vary between subjects and even between sessions with the same subject. This thesis proposes a new signal processing framework for BCI that incorporates a quick and fully algorithmic feature selection step combined with an SVM classification in embedded form. For the evaluation of this new methodology, the results of own online and offline studies with electroencephalogram (EEG), electrocorticogram (ECoG) and magnetoencephalogram (MEG) will be presented, including the first ever implementation of a motor-imagery BCI system in MEG. In addition, the framework has been evaluated against several existing filter and wrapper approaches for feature selection. According to these results, the new method is capable of adapting to the changing signal characteristics of BCI users, can be used without prior neurophysiological knowledge about underlying mental tasks, reduces the number of features from several hundreds to just about 10% of the original features all while remaining highly accurate in terms of classification performance. Furthermore, the results of the feature selection step prove to be plausible in terms of neurophysiology, i.e. chosen EEG channels agree well with the expected underlying cortical activity patterns during the mental tasks. Under restricted conditions, it is shown that the optimized feature sets determined by the new signal processing framework can be transferred across subjects with only a small drop in performance.