A user-customized self-paced brain computer interface

Much attention has been directed towards synchronous Brain Computer Interfaces (BCIs). For these BCIs, the user can only operate the system during specific system-defined periods. Self-paced BCIs, however, allow users to operate the system at any time he/she wishes. The classification of Electroencephalography (EEG) signals in self-paced BCIs is extremely challenging, as the BCI system does not have any clue about the start time of a control task. Also, the data contains a large number of periods during which the user has no intention to control the BCI. For sensory motor self-paced BCIs (focus of this thesis), the brain of a user goes through several well-defined internal state changes while performing a mental task. Designing classifiers that exploit such temporal correlations in EEG data can enhance the performance of BCIs. It is also important to customize these BCIs for each user, because the brain characteristics of different people are not the same. In this thesis, we first develop a unified comparison framework to compare the performance of different classifiers in sensory motor BCIs followed by rigorous statistical tests. This study is the largest of its kind as it has been performed on 29 subjects of synchronous and self-paced BCIs. We then develop a Bayesian optimization-based strategy that automatically customizes a synchronous BCI based on the brain characteristics of each individual subject. Our results show that our automated algorithm (which relies on less sophisticated feature extraction and classification methods) yields similar or superior results compared to the best performing designs in the literature. We then propose an algorithm that can capture the time dynamics of the EEG signal for self-paced BCI systems. We show that this algorithm yields better results compared to several well-known algorithms, over 13 self-paced BCI subjects. Finally, we propose a fully automatic, scalable algorithm that customizes a self-paced BCI system based on the brain characteristics of each user and at the same time captures the dynamics of the EEG signal. Our final algorithm is an important step towards transitioning BCIs from research environments to real-life applications, where automatic, scalable and easy to use systems are needed.Applied Science, Faculty ofElectrical and Computer Engineering, Department ofGraduat