Abstracts Category : Other

Motor Imagery Based Brain Computer Interfaces

by Syed Salman Ali

Brain Computer Interface (BCI) is an emerging technology which enables humansto communicate with external devices using their brain signals. Motor imageryclassification is one such area in BCI systems where the goal is to detect and classifythe motor related intentions of humans using their brain signals. The desired accuracyof BCI systems cannot be achieved without the preprocessing and the extractionof discriminant features from raw brain signals.To this end, this thesis first develops a novel extension of multivariate empiricalmode decomposition (MEMD) for the preprocessing of the raw brain signals. MEMDis a mathematical tool that is used to decompose multivariate time signals into a setof basis functions called intrinsic mode functions (IMFs). In order to extract IMFs,MEMD is required to estimate the local mean of multivariate signal by taking theprojections of input signal on a dense uniformly sampled hyper-sphere. A novelnon-uniform sampling scheme is proposed to estimate the local mean of multivariatesignals. The non-uniform samples are generated by linearly transforming the hypersphereinto an N dimensional ellipsoid using singular value decomposition (SVD). Anumber of experiments on synthetic and real-world signals were conducted to showthat the non-uniform sampling scheme is helpful to generate meaningful IMFs whenthe input multichannel signals are highly correlated. In addition, the performance ofproposed algorithm was also evaluated in motor imagery based BCI systems.The second part of this thesis talks about the common spatial pattern (CSP) which is commonly used to extract discriminant features from the raw brain signalsfor motor imagery based BCI systems. The performance of the CSP algorithm relieson the estimation of the covariance matrix which becomes an ill posed problem whenthe number of training samples is limited. In this thesis, a novel extension of theCSP algorithm is proposed to address the limited sample size problem. The proposedmethodology is evaluated on publically available BCI competition III dataset. Theresults show improved performance of proposed method when compared with thetraditional CSP algorithm especially when the number of training samples is limited.keywords: MEMD, non-uniform sampling scheme, common spatial pattern,maximum entropy, motor imagery, BCI competitionAdvisors/Committee Members: Zhang, Lei (advisor), Bais, Abdul (committeemember).