A system based on genetic algorithms for on-line single-trial P300 detection

Abstract

Brain-machine interface (BMI) systems collect and classify electroencephalogram (EEG) data to predict the desired command of the user. The P300 EEG signal is passively produced when a user observes or hears a desired stimulus. The P300 can be used with a visual display to allow a BMI user to select commands from an array of selections. The visual stimuli are often repeated and averaged to increase classification accuracy. In this paper we explored classification of single-epoch P300 signals. An EEG BMI system was constructed to allow offline training and live testing. Using a genetic algorithm to select features of data we achieved 78.3% signal detection accuracy using a Support Vector Machine classifier. Using this classifier we constructed a simulated mobile robot steering system, which could be controlled with little training and achieved up to 7.5 commands/minute.