Machine Learning Methods for Brain-Computer Interfaces
Thomas Navin Lal
PhD thesis, Max Plank Institute for Biological Cybernetics --- TU Darmstadt.
Based on an imagery task, we introduce a novel BCI that utilizes Recursive Feature Elimination (RFE) (Guyon et al., 2002), a fea ture selection method which was adapted to the special case of brain imaging data, and a regularized Support Vector Machine (SVM) (e.g. Schoelkopf and Smola, 2002). This BCI was designed and tested based on EEG-recordings (Lal et al., 2004). Together with our collaborators from the department Medical Psychology and Behavioral Neurobi-
ology at the university of TÄubingen (Prof. Dr. N. Birbaumer), we investigated whether the invasive recording technique CoG might be a viable alternative for the completely paralyzed. As a first step our BCI was successfully tested in epilepsy patients who had temporary electrode implants (Lal et al., 2005d). After the proof of concept" was made, the first completely paralyzed patient suffering from ALS was implanted. Unfortunately, we cannot not yet use a BCI to communicate with the patient (Hill et al., 2005b).
The second part of the thesis follows an approach based on the recording technique magnetoencephalography (MEG) to provide a BCI to paralyzed patients. Although an MEG device is currently too expensive and too large to be installed at a patient's home, its compared to EEG| superior signal-to-noise ratio might help to design a patient BCI. Due to the better data quality, patients whose brain signals are not classi¯able using EEG might be classifiable using MEG. In this case a classifying function can be inferred from the MEG-recordings. This function could then be used to provide feedback which encourages the patient to produce more separable brain
signals. One hypothesis is that the brain signals of a patient who was trained in this way later become classifiable using a portable and affordable EEG system. Since MEG has never been used before for feedback experiments, the first hallenge was to get real-time access to MEG data. We were then ready to run an experiment with ten healthy subjects of whom 4 wrote a short name using imagination tasks only (Lal et al., 2005a). It turned out that our hypothesis
was supported by the fact that shorter training times are required as compared to an EEG-based approach.