Probes minimization still maintaining high accuracy to classify target stimuli P300

Abstract

We present a novel approach to detect P300 Signal by using the difference between target trial signal and nontarget trail signal generated by P300 BCI Speller. Conventional P300 BCI Speller uses 8 probes at pre-defined locations on the scalp. It works for any person and could deliver a reasonable performance. In P300 BCI speller, system needs to decipher the event related potential (ERP), called P300. Though it is strong in the parietal region of the brain, location of the strongest signal varies from person to person. If we can adapt the location of probes for an individual, we could eliminate un-necessary probes. At first, we want to eliminate probes that did not generate similar signal for target stimuli. We calculated signal distance between different experiments for every probe. The probes that generated quite different signal in different experiments are unsuitable for classification. Then, we search for the probes that will generate strong P300 signal from the remaining probes. In this work, we concentrated on the difference between target trail signal and non-target trail signal. P300 signal's distinctive property is that its amplitude has a strong negative correlation with the event probability. In P300 BCI Speller, the target trail's probability is 1/6 and the non-target trail's probability is 5/6. Infrequent target trail P300 signal's amplitude is larger than the nontarget trial signal's amplitude. We designed an algorithm to calculate the P300 signal's amplitude and selected the probes with larger difference between target trial P300 signal's amplitude and non-target trail signal's amplitude. We achieved over 81% classification accuracy on average with 3 probes from only one pair of target stimuli signal and non-target stimuli signal.