Optimizing the EEG Electrode Configuration for Signal Acquisition in P300 Speller Systems

HE P300 speller is a brain-computer interface (BCI) system that restores communication ability to “locked-in” patients by detecting evoked responses to visual stimuli in the user’s electroencephalogram (EEG) signal [1]. Choosing the optimal placement and number of electrodes is essential in any EEG system as it balances the amount of available data against the set-up time and the amount of computation work required for signal classification. While several different layouts have been tested [2], no systematic study has been performed to determine the optimal electrode placement. The goal of this project was to optimize EEG electrode placement for P300 studies using an offline Gibbs sampling method. This study used a data set of 15 healthy subjects who spelled between 8 and 10 five letter words using Donchin’s P300 speller with 32 electrodes in an established pattern [3, 4]. Batch Gibbs Sampling was used to find samples from the distribution of electrodes that can best classify EEG signals. Electrode sets were constructed by finding the groups of electrodes that co-occurred most frequently in the samples. Classification with a naïve Bayes classifier [5] was performed offline using data from each of these electrode sets and results were evaluated using information transfer rate (ITR) [6].