Single-trial classification of ERPS using second-order blind identification (SOBI)

Single-trial classification of EEG signals has received increasing attention in both basic research and for the development of EEG based brain computer interfaces (BCI). Typically, such classification has been performed using signals from a set of selected EEG sensors. Since EEG sensor signals are mixtures of signals from multiple intra- and extra-cranial sources, single-trial sensory and motor evoked potentials can be difficult to detect and classify. In this paper, second-order blind identification (SOBI) was used to preprocess EEG data and extract activity from the left and right primary somatosensory (SI) cortices. Subsequently, classification of event-related potentials (ERPs) evoked by a sequence of randomly mixed left, right, and bilateral median nerve stimulations was performed by back-propagation neural networks, using as inputs the two SOBI-recovered SI components or the two "best sensors". Results from four subjects showed that classification accuracy was significantly higher when SOBI-recovered left and right SI components were used for classification than when the EEG sensor signals were used directly.