Efficient EEG frequency band selection techniques for a robust motor imagery based brain-computer interface

Recently, Electroencephalogram (EEG)-based Brain-Computer Interfaces (BCIs) have become a hot topic in the study of neural engineering, rehabilitation and brain science. BCIs translate human intentions into control signals to establish a direct communication channel between the human brain and output devices bypassing brain’s normal output pathway of nerves and muscles. This new approach is a promising communication channel for paralyzed patients to interact with the external world and a new direction in entertainment through BCI games in healthy people as well. In order to decipher the intentions accurately, it is important to obtain distinguishable EEG features. At present, event-related de/synchronization (ERD/ERS) patterns during imagination of motor movements or motor imagery have been extensively applied to design BCI. These patterns are the attenuation and enhancement of EEG rhythmic power during motor imagery. One of the critical elements in the design of any BCI is the extraction of reliable and discriminative features that represent the intended task by the user. The work presented in this thesis focuses on improving the discrimination between features extracted during various motor imagery tasks. More specifically, techniques towards a robust motor imagery based BCI by selecting the relevant frequency components carrying the discriminative information are proposed. The first proposed algorithm in the work is a discriminative filter bank (DFB) based approach to distinguish between motor imagery patterns. The algorithm named as Discriminative Filter bank Common Spatial Pattern (DFBCSP) employs a parent filter bank of twelve bandpass filters to filter the EEG recorded from sensory motor cortex. The Fisher ratio values computed at each filter output determine the discriminative capability of the respective bands. A set of four bandpass filters offering highest Fisher ratio values are selected from the parent filter bank to form DFB. Common Spatial Pattern (CSP) features extracted from the DFB output are used for distinguishing the various motor imagery tasks. Experimental results show that the classification performance of DFBCSP is better than the existing filter bank based method.DOCTOR OF PHILOSOPHY (SCE