An Optimized SWCSP Technique for Feature Extraction in EEG-based BCI System

Brain-computer interface (BCI) is an evolving technology having huge potential for rehabilitation of patients suffering from disorders of the nervous system, besides  many other nonmedical applications. Multichannel electroencephalography (EEG) is widely used to provide input signals to a BCI system. Significant research in methodology employed to implement different stages of BCI system, has led to discovery of new issues and challenges. The raw EEG data includes artifacts from environmental and physiological sources, which is eliminated in preprocessing phase of BCI system. It is then followed by a feature extraction stage to isolate a few relevant features for further classification to a particular motor imagery (MI) activity. A feature extraction approach based on spectrally weighted common spatial pattern (SWCSP) is proposed in this paper to improve overall accuracy of a BCI system. The reported literature uses SWCSP for feature extraction, as it has outperformed other techniques. The proposed approach enhances its performance by optimizing its parameters. The independent component analysis (ICA) method is used for detection and removal of irrelevant data, while linear discriminant analysis (LDA) is used as a classifier. The proposed approach is executed on benchmark data-set 2a of BCI competition IV. It yielded classification accuracy of 70.6% across nine subjects, which is higher than all the reported approaches.