SSVEP-based Brain Computer Interface for 2-D cursor control: Comparative study of four classifiers

The present work was devoted to a comparative study of four classification algorithms targeting applications for Brain Computer Interface (BCI). BCI is a communication system that interprets brain signals (frequently, electrical activity) and converts them into control commands for an external environment. BCI can be viewed as a communication pathway for severely disabled individuals. Steady State Visual Evoked Potential (SSVEP) based BCI attract particular interest in today's research because of the possibility of a highly accurate and reliable BCI system with little or no user training.;SSVEP is an oscillatory response elicited by the brain that can be found in EEG and consists of fundamental frequency and number of harmonics corresponding to the frequency of visual stimulus – a flickering light, for instance, a phase reversal checkerboard pattern, or frequency modulated LEDs. The present study concerns a 2-dimensional cursor control application of BCI using SSVEP paradigm. Visual stimuli were implemented on the background of different colors to assess the effect of colors on the classification accuracy.;A spatial filter was designed using bipolar channel selection of EEG to enhance SSVEP. Four different methods were considered for detection and classification of SSVEP. Power spectrum-based approach using both parametric and non-parametric spectral estimators was explored as it is a widely used method for SSVEP-based BCI. Other methods, such as Canonical Correlation Analysis (CCA), Linear Discriminant Analysis (LDA), and Neural Network based detection and classification were considered. A comparison between these methods was performed based on classification accuracy, requirement of machine learning, possibility of high bit-rate, and computational complexity.;The power spectrum-based method was found impractical for designing the BCI system. CCA provides better accuracy than the power-spectrum approach and requires very few customizations for inter-subject variability, however, the accuracy is rather low. Higher accuracy was achieved using LDA classifier, which was found as more computationally efficient. Neural Network based classification provided the highest accuracy and was observed as a robust and reliable approach.