Speeding up classification of multi-channel Brain-Computer Interfaces: Common spatial patterns for slow cortical potentials

During the last years interest has been growing to find an effective communication channel which translates human intentions into control signals for a computer, the so called Brain-Computer Interface (BCI). One main goal of research is to help patients with severe neuromuscular disabilities by substituting normal motor outputs. Various cortical processes were identified which are suitable for implementing such a system on basis of scalp recorded electroencephalogram (EEG), e.g., slow cortical potentials (SCP) and event-related desynchronisation (ERD) of 10–20 Hz brain rhythms. Until quite recently BCI systems used only few EEG channels but by use of advanced machine learning techniques it became possible to exploit the spatial information provided by multi-channel EEG. While the use of such high density spatial sampling increases the accuracy of the system it may—depending on the computational effort of the signal processing— pose a problem for the implementation of the feedback in real-time. Here we propose a method that offers a substantial speed-up for classification of SCP features as used in the Berlin Brain computer interface (BBCI) [1]. Instead of applying the time consuming low-pass filtering to all, say 120, EEG channels a suitable spatial projection extracts only 2 or 4 new channels which can be used without any loss of classification accuracy in our experiments. Our approach is based on the technique of common spatial patterns (CSP) which were suggested in [2] to extract ERD features from EEG. While in its original form CSP is only applicable to oscillatory features we present a new variant which allows to use CSP for SCP features without regularisation even in case of large channel numbers or few training samples. Keywords—multi-channel electroencephalogram, brain-computer interface, slow cortical potentials, movement related potential, Bereitschaftspotential, common spatial patterns, single-trial classification.