Increasing performance of classifiers for SSVEP-based brain-computer interfaces using extension methods
Ethan Webster · 2020
Brain-computer interfaces (BCI) provide an alternative communication method that does not require standard physical mediums (speech, typing, etc.). These systems have been implemented to provide additional communication and control options for people with certain motor disabilities. Classification is an important part of BCI systems and consists of inferring user commands from brain activity. Supervised classification methods often achieve higher accuracy, but unsupervised classification methods are useful when training is not practical for the user. This thesis focuses on unsupervised classification algorithms used for a BCI speller application and presents extensions for two existing classifiers that improve classification accuracy and thus the potential speed of the system. First, an extension to canonical correlation analysis (CCA) is proposed that uses temporally local covariance. Second, a generalized version of extended multivariate synchronization index (EMSI) is proposed that models brain activity as a dynamical system by utilizing time delay embedding. Both methods are evaluated on a common EEG dataset and both provide improved system speed in the average case.