A genetic algorithm for single-trial P300 detection with a low-cost EEG headset

Abstract

Brain machine interface (BMI) devices facilitate communication and control of computers using signals measured from within the brain of the operators. These signals are detected using electroencephalography (EEG) devices. Research in this field aims to enable victims of `locked-in syndrome' as a result of amyotrophic lateral sclerosis, spinal injury, cerebral palsy, muscular dystrophies, or multiple sclerosis. BMI systems also increase diversity in human computer interaction methods. One of the BMI target signals, known as the P300, is an involuntary reaction to a desired visual stimulus. BMI systems capable of detecting P300 signals allow direct brain-device interaction, without the need for muscle excitation. Because EEG P300 signal suffers low signal to noise ratios, classification of user intent can be difficult. Typically P300 systems use repeated visually evoked potentials (VEPs) to increase classifier accuracy; however this results in lower information transfer rates. To improve single-trial P300 detection we use a genetic algorithm (GA) in combination with both a neural network and linear discriminant analysis classifiers. The GA improved feature selection for training the classifiers. We explore the results of those features found influential on P300 classification and suggest direction for future research in single-trial P300 detection.