Neural signal processing techniques to analyse motor functions in eeg-based brain-computer interfaces
Shenoy Handiru, Vikram
Date of Issue2018-02-12
School of Computer Science and Engineering
The electroencephalogram (EEG)-based Brain-Computer Interface (BCI) provides an alternative pathway to transmit neural information to a computer. Nowadays, EEG-based Brain-Computer Interfaces are becoming popular due to the cost-effectiveness, portability, and the high temporal resolution of EEG. Also, there have been substantial progress in the EEG-based BCI during the past two decades. However, there are still several shortcomings that need to be addressed such as - poor spatial resolution, non-stationary nature of cortical signals, and high variability of BCI performance across subjects. As the recorded EEG signals are usually noisy, special approaches for signal processing need to be used to extract useful information from highly non-stationary EEG signals. To this end, several research challenges concerned with the spatial filtering algorithms are addressed in this thesis by developing a computationally efficient channel selection algorithm to improve the scalp-spatial resolution and also by robustifying the covariance estimation to handle nonstationarities in a small-sample BCI setup. As the EEG sensor-space recordings are affected by the linear mixing of cortical sources and the volume conduction effect, the latter half of the thesis is devoted to the EEG source localization to extract underlying features relevant to motor tasks. The research presented in this thesis explored different datasets related to motor-imagery and motor-execution tasks in a BCI setup and investigated the robustness of spatial filtering algorithms in terms of cross-subject generalization, source-space analysis and also the clinical stroke data. A novel supervised factor analytic approach for EEG source-space feature extraction was also developed that allows for a better interpretation of task-related cortical sources and particularly useful in multi-class BCIs. After evaluating on several types of datasets, it is demonstrated that the EEG source-space features not only helps in obtaining better performance in classifying the multi-class motor tasks but also helps in the better interpretation of neural correlates. Continuing further, the EEG source-space analysis is extended to study the causal effects of transcranial direct current stimulation on the Motor-Imagery BCI performance of stroke patients. With the increasing number of source imaging approaches extending beyond BCI applications to clinical diagnosis, this thesis suggest that the EEG source-space studies will have a significant attention in the coming years.