Motor imagery (MI) brain-computer interface (BCI) performance is highly influenced by non-stationarity and artifact contamination of electroencephalogram (EEG) signals. This paper presents a framework for overcoming EEG uncertainties in real-time multiclass MI BCI. An artifact rejected multiclass extension of common spatial pattern (CSP) by using joint approximate diagonalization (JAD) is proposed for feature extraction. Artifactual trials are excluded in spatial filters calculation that results in more informative features. In order to cope with non-stationarities, an adaptive resonance theory (ART) based neuro-fuzzy classifier, named self-regulated supervised Gaussian fuzzy adaptive system Art (SRSG-FasArt) is implemented for multiclass applications. The proposed framework is evaluated based on a standard dataset of BCI competition IV. Applying the system in real-time performance shows significant improvement in multiclass classification accuracy compared to state of the art methods.
