Robust multi-command bci systems bci systems based on steady state visual evoked potentials

In this thesis, three multi-stage algorithms applied to the design of asynchronous Robust Multi-command Brain Computer Interface (BCI) based on Steady-state Visual Evoked Potentials (SSVEP) are presented. The presented BCI system permits to the subject to move an object in different directions along a monitor screen. The subject controls the movement of the object by gazing at a set of small reversal checkerboards flickering at different frequencies. Based on fast and reliable estimations of the SSVEP responses, the three presented BCI algorithms permit the system control with high and robust performances. The three methods relied on two Blind Source Separation (BSS) methods, the AMUSE algorithm and a Constrained Independent Component Analysis with references (CICAr), to achieve the robust performance against the presence of artifacts. The AMUSE algorithm, a second order statistic BSS method was applied in two of the algorithms at the artifact rejection stage, permitting an automatic way to attenuate the artifacts of the EEG with low computation cost. While the CICAr method, a high order statistical BSS method, was included in the third proposed algorithm, and it was applied as an feature extraction method, focusing only in the SSVEP signals. The three algorithms were validated for the developed BCI system, although, only two of them based in the AMUSE algorithms, were tested in real-time. The first proposed multistage algorithm, consisted on a sequence of stages, starting with the automatic artifact rejection with AMUSE algorithms. Subsequently, the cleaned EEG signals were filtered using a bank of filters that defined the feature vector. The final decision stage of the BCI system was performed by an ANFIS classifier calibrated for each subject during a short session. The second proposed algorithm relied in the AMUSE algorithm too. However, feature vector was defined by the estimations of the energies of the SSVEP responses obtained by applying a Multi-Input Multi-Output adaptive RLS filter. Instead of using a classifier to map the feature vector to the subject's command, a set of statistical tests were performed over the corrected values of the estimated SSVEP energies. The correction factors of the SSVEP estimations were obtained from the estimations of the spontaneous EEG activity by applying cubic spline approximations for the data spectrum. The BCI system with this algorithm does not require any calibration session that is, a plug and play BCI system. This algorithm reached a high performance with a low computational cost. In order to see the advantage of applying a high order statistic method, the third proposed algorithm relied on the CICARr extraction method. A constrained Independent Component analysis where a set of equalities and inequalities were defined to efficiently extract and compared the elicited SSVEP activities. The CICAr method allowed to solve the scale and order ambiguities present in ICA algorithms. The SSVEP information provided by the CICAR algorithm was contrasted by applying a surrogate method. This algorithm does not require any classifier, and hence it is a plug and play system. This algorithm reached a very high performance for the electrode placed in the occipital area. The asynchronous multi-command BCI systems based on SSVEP could be useful to help paralyzed subjects to control objects in different environments and applications, smart houses, vehicle panels, robot control, etc. In particular, it can be used in the design of smart houses, by distributing the visual flickering stimulators in the house to control different objects.