Abstract

The human brain is the most intriguing and complex signal processing unit ever known to us.
A unique characteristic of our brain is its plasticity property, i.e., the ability of neurons to modify
their behavior (structure and functionality) in response to environmental diversity. The plasticity
property of brain has motivated design of brain-computer interfaces (BCI) to develop an alternative
form of communication channel between brain signals and the external world. The BCI systems
have several therapeutic applications of significant importance including but not limited to rehabilitation/
assistive systems, rehabilitation robotics, and neuro-prosthesis control. Despite recent
advancements in BCIs, such systems are still far from being reliably incorporated within humanmachine
inference networks. In this regard, the thesis focuses on Motor Imagery (MI)-based BCI
systems with the objective of tackling some key challenges observed in existing solutions. The
MI is defined as a cognitive process in which a person imagines performing a movement without
peripheral (muscle) activation. At one hand, the thesis focuses on feature extraction, which is
one of the most crucial steps for the development of an effective BCI system. In this regard, the
thesis proposes a subject-specific filtering framework, referred to as the regularized double-band
Bayesian (R-B2B) spectral filtering. The proposed R-B2B framework couples three main feature
extraction categories, namely filter-bank solutions, regularized techniques, and optimized Bayesian mechanisms to enhance the overall classification accuracy of the BCI. To further evaluate the effects
of deploying optimized subject-specific spectra-spatial filters, it is vital to examine and investigate
different aspects of data collection and in particular, effects of the stimuli provided to subjects to
trigger MI tasks. The second main initiative of the thesis is to propose an element of experimental design dealing with MI-based BCI systems. In this regard, we have implemented an EEG-based
BCI system and constructed a benchmark dataset associated with 10 healthy subjects performing
actual movement and MI tasks. To investigate effects of stimulus on the overall achievable performance,
four different protocols are designed and implemented via introduction of visual and voice
stimuli. Finally, the work investigates effects of adaptive trimming of EEG epochs resulting in an
adaptive and subject-specific solution.