Evidential multi-band common spatial pattern in brain computer interface

In electroencephalography (EEG)-based brain computer interfaces (BCI) one way to extract discriminative command-related patterns is to use common spatial pattern (CSP). Speech-imagery BCIs on the other hand, are a more recent type of BCIs in which imagination of a vowel, sound or word is detected from EEG measurements. Since the performance of CSP highly depends on the spectral filters which are previously applied to the signal, finding the appropriate frequency band seems to be a major concern. In order to overcome this problem, the novel evidential multi-band common spatial pattern (EMBCSP) is proposed. In EMBCSP, features are extracted using CSP after multiple band-pass filtering. The preliminary classification results are later given to a fusion system and then the features are labeled regarding the larger combined probability value. The performance of this method has been assessed using speech-imagery-based EEG measurements and has also been compared with some prevalent methods. The results suggest the outperforming of EMBCSP with accuracies between 75 to 96.67 percent and give rise to opportunities to apply such a method to other common types of BCIs.