Data-driven multivariate and multiscale methods for brain computer interface

Abstract

This thesis focuses on the development of data-driven multivariate and multiscale methods for brain computer interface (BCI) systems. The electroencephalogram (EEG), the most convenient means to measure neurophysiological activity due to its noninvasive nature, is mainly considered. The nonlinearity and nonstationarity inherent in EEG and its multichannel recording nature require a new set of data-driven multivariate techniques to estimate more accurately features for enhanced BCI operation. Also, a long term goal is to enable an alternative EEG recording strategy for achieving long-term and portable monitoring. Empirical mode decomposition (EMD) and local mean decomposition (LMD), fully data-driven adaptive tools, are considered to decompose the nonlinear and nonstationary EEG signal into a set of components which are highly localised in time and frequency. It is shown that the complex and multivariate extensions of EMD, which can exploit common oscillatory modes within multivariate (multichannel) data, can be used to accurately estimate and compare the amplitude and phase information among multiple sources, a key for the feature extraction of BCI system. A complex extension of local mean decomposition is also introduced and its operation is illustrated on two channel neuronal spike streams. Common spatial pattern (CSP), a standard feature extraction technique for BCI application, is also extended to complex domain using the augmented complex statistics. Depending on the circularity/noncircularity of a complex signal, one of the complex CSP algorithms can be chosen to produce the best classification performance between two different EEG classes. Using these complex and multivariate algorithms, two cognitive brain studies are investigated for more natural and intuitive design of advanced BCI systems. Firstly, a Yarbus-style auditory selective attention experiment is introduced to measure the user attention to a sound source among a mixture of sound stimuli, which is aimed at improving the usefulness of hearing instruments such as hearing aid. Secondly, emotion experiments elicited by taste and taste recall are examined to determine the pleasure and displeasure of a food for the implementation of affective computing. The separation between two emotional responses is examined using real and complex-valued common spatial pattern methods. Finally, we introduce a novel approach to brain monitoring based on EEG recordings from within the ear canal, embedded on a custom made hearing aid earplug. The new platform promises the possibility of both short- and long-term continuous use for standard brain monitoring and interfacing applications.