Networks-based brain computer interfaces for decoding neurophysiological dynamics

The brain is an extremely complex organ and probably one of the greatest mysteries of the universe that has baffled scientists and thinkers for centuries. Today, one of the ways in which we understand and investigate the brain is by modelling various neural phenomena with the help of machine learning and sophisticated mathematical algorithms. Technological and scientific advancements go hand in hand. One such area of technology receiving great attention from the scientific community is brain-computer interface (BCI). A BCI is an alternate channel of communication between the brain and a computer, without the involvement of peripheral muscles, via direct transmission of brain signals to a computer. The primary aim of this thesis is to develop and use models of dynamic interactions between brain regions which can be used in different types of applications, from machine learning to classify different types of hand movements to understanding different cognitive styles in autistic adults compared to typically developing (TD) controls. For the first study, electroencephalogram (EEG) data was acquired from seven subjects performing centre-out hand movements in four orthogonal directions in the 2D horizontal plane. A modified form of phase-locking based estimator of inter-channel synchrony, called Wavelet Phase-Locking Values (W-PLV), was proposed to carry out binary classification of hand movements in different directions. It was hypothesized that different parts of the brain would show different connectivity patterns to carry out a fairly complex motor task. Results showed that different spatial and time-frequency patterns of pairwise inter-channel synchrony could be identified for hand movements in different directions using the proposed quantifier, thus providing evidence in support of the above hypothesis. Binary classification using the proposed features achieved satisfactory performance. This study was then extended to investigate the flow of information between different brain regions using a simple independent component analysis (ICA)-based method. The experimental paradigm allowed the subjects a mental preparation phase before the actual execution of hand movements. It was thus demonstrated that the task involved participation of motor as well as non-motor (cognitive) parts of the brain using a data-driven approach. Results in the EEG source space showed that during the visuomotor task, independent components derived from EEG data were not limited to the motor region but were in fact spread out throughout the cortical surface. This observation had been previously reported in other electrophysiological modalities such as electrocorticography (ECoG). Thus, results obtained from this EEG study are in agreement with the expected neurophsyiology of motor action and preparation. Having investigated the dynamic interactions between different brain regions in healthy subjects performing motor tasks, research was then conducted on more sophisticated models of dynamic brain interactions to study human behaviour in a neurodevelopmental disorder called Autism Spectrum Disorder (ASD). TD and high functioning autistic adults participated in a study to identify complex positive and negative facial emotions, using video stimuli of actors, while their EEG data was collected. While the groups performed at par in recognizing complex negative emotions, the autistic group performed significantly poorly compared to the TD controls. The aim of this research was then to obtain a dynamic functional brain networks perspective of the differences observed in the facial emotion recognition (FER) accuracy. Features extracted from the EEG cortical source space showed dynamic cognitive processing during the FER task showing that the brain may have to reorganize its neural resources dynamically. Next, it was also seen that functional brain networks in the autistic group had a closer inclination to random networks than their TD counterparts, especially in lower frequency bands, which are typically attributed to the integration of information between long-range brain regions. A multi-temporal scale, coarse-grained connectome study was then carried out which showed that there seemed to be a spatio-spectral shift in dynamic functional connectivity during complex emotional processing from the lower frequency bands to the gamma and beta sub-bands were in the autistic group compared to TD controls. This spectral shift was seen to occur more drastically for emotions resulting in poorer recognition accuracy. Results of this work showed that psycho-cognitive theories of ASD could potentially have a basis in EEG-based dynamic functional brain networks and provide motivation to further investigate the dynamics of brain networks to study human behaviour. The research presented in this thesis shows that EEG-based dynamic brain interactions can be modelled to decode and investigate motor activity directly from brain data as well as analyze the sub-optimal cognitive performance in autistic adults with evidence supporting widely promoted theories in psychology. With re- cent advances in Deep Learning, future research can use the work presented in this research with BCIs and appropriately designed experiments to decode finer and more complex motor activity or gain better understanding of human behaviour.Doctor of Philosoph