Enhancing Performance of Classification Techniques for EEG Based Brain–Computer Interface

Abstract

Mental task (MT) Electroencephalography (EEG) is EEG recorded during performance of non-motor general mental tasks. MT based brain-computer interface (BCI) paradigm using non-motor mental tasks can be viewed as generalization of motor imagery (MI) based BCI paradigm. MT based BCI paradigm shows better potential than MI based BCI paradigm in enhancing the quality of life of the physically disabled persons. Consequently, the focus of this current research lies in MT based BCI systems. Classification of such non-motor cognitive Electroencephalography (EEG) signals produced during performance of mental task is the central challenge in developing such type of EEG based non-invasive BCI systems. The human brain shows extremely complex nonlinear and non-stationary spatiotemporal patterns of EEG signals that vary over multiple temporal scales. We believe that accurate representation of such type of complex and subtle patterns contained in the signal dynamics holds the key in enhancing the real time performance of a BCI system. But the neurological control signals driving the MT based BCI paradigm contain a number of different types of sophisticated spatiotemporal patterns which have not been identified yet. Current feature extraction algorithms relying on prior assumptions about the patterns may discard meaningful information contained in the data. Due to this, their ability to accurately identify new type of patterns is limited. The prime motivation of our work stems from this need for discovering new patterns through nonlinear signal processing or from the geometrical and topological properties of the RPS. In this dissertation, we aim at enhancing the classification performance in mental task (MT) based BCI paradigm. We investigated the following feature extraction approaches: 1. Introducing the largest singular value of the phase space matrix in the multivariate empirical mode decomposition( MEMD) domain as feature for classifying mental task in MT based brain-computer interface. 2. Introducing Eigen values of the covariance matrix of the coefficient matrix of the multivariate autoregressive( MVAR) model in the MEMD domain. 3. Proposing multivariate multi scale entropy values as EEG features for classifying non-motor mental task EEG. 4. Singular values in the phase space of original signal as features. In the first approach, we employed singular value decomposition (SVD) based phase space analysis of the multivariate intrinsic mode functions (IMFs) and extracted largest singular values from the phase space matrices of the sensitive IMFs for constructing the feature vectors. With these new feature vectors, we achieved highest classification accuracy of 83.33% for binary classification between mental arithmetic and mental letter composing. Our second approach is based on deriving multivariate autoregressive (MVAR) models of the set of relevant multivariate intrinsic mode functions (IMFs) generated from the Multivariate empirical mode decomposition (MEMD) of the multi-channel EEG signals. In this approach, the set of statistically significant Eigen values computed from the derived multivariate AR models of the set of relevant IMFs were used for constructing the feature vectors. Finally, we classified the constructed feature vectors by employing LS-SVM classifier with three different kernel functions. We achieved highest average classification accuracy of 94.3% for binary and 77.7% for three class classification. In the third approach, we proposed multivariate multi scale entropy based complexity measures as EEG features for classifying EEG signals in MT based BCI paradigm. These entropy values computed over selected scales have been employed for constructing the feature vectors. We achieved highest classification accuracy of 100% for binary classification of the two pairs mental tasks. We tested all our approaches on a bench mark EEG data set and evaluated the results. The accuracy, speed and consistency of the test results show efficacy of the proposed features. In this way, this thesis presents several novel results in the broad area of brain signal classification using EEG recordings which further leads to better understanding of cognitive brain dynamics and improved performance of next generation of noninvasive BCI systems.