Content area
Electroencephalography (EEG)-based brain-computer interfaces (BCIs) are devices that allow users to control computers or robotic devices using signals recorded directly from their brains. Since these devices bypass the need for muscle or speech activation, they have the potential to replace or restore motor functions for motor-impaired patients. BCIs may also improve the lives of the general population by providing a direct line of communication with their personal devices and the Internet of Things. EEG signals are recorded non-invasively from outside of the brain, making them a safe option for BCI systems, particularly for users who are not candidates for invasive surgery. However, EEG signals also have relatively low signal-to-noise ratios, poor spatial resolution, and high variability across subjects and sessions, which has so far limited the performance and applications of these devices compared to invasive BCI methods. This thesis aims to improve the performance and reliability of EEG-based BCIs by addressing three of the main components of BCI systems: the control paradigm, the signal processing algorithms, and the end application. Specifically, the results of the three studies included in this work show that integrating several control paradigms can produce multiple EEG feature sets simultaneously, that online deep learning-based decoding can improve performance in continuous control tasks, and that the resulting system can be used for complex tasks involving physical robotic devices. As a culmination of this work, we demonstrate that the proposed EEG BCI system using real-time deep learning-based decoding allows both able-bodied and motor-impaired users to continuously control a robotic arm to pick up, move, and place cups around a set of shelves using only their EEG signals. These studies provide a contribution towards the advancement of EEG-based BCIs and show the potential for these systems to move towards real-world and clinical applications.
