Classification of Frequency and Phase Encoded Steady State Visual Evoked Potentials for Brain Computer Interface Speller Applications using Convolutional Neural Networks

Over the past decade there have been substantial improvements in vision based Brain-Computer Interface (BCI) spellers for quadriplegic patient populations. This thesis contains a review of the numerous bio-signals available to BCI researchers, as well as a brief chronology of foremost decoding methodologies used to date. Recent advances in classification accuracy and information transfer rate can be primarily attributed to time consuming patient specific parameter optimization procedures. The aim of the current study was to develop analysis software with potential ‘plug-in-and-play’ functionality. To this end, convolutional neural networks, presently established as state of the art analytical techniques for image processing, were utilized. The thesis herein defines deep convolutional neural network architecture for the offline classification of phase and frequency encoded SSVEP bio-signals. Networks were trained using an extensive 35 participant open source Electroencephalographic (EEG) benchmark dataset (Department of Bio-medical Engineering, Tsinghua University, Beijing). Average classification accuracies of 82.24% and information transfer rates of 22.22 bpm were achieved on a BCI naïve participant dataset for a 40 target alphanumeric display, in absence of any patient specific parameter optimization.