Improved P300 speller performance using electrocorticography, spectral features, and natural language processing

Abstract

ObjectiveThe P300 speller is a system designed to restore communication to patients with advanced neuromuscular disorders. This study was designed to explore the potential improvement from using electrocorticography (ECoG) compared to the more traditional usage of electroencephalography (EEG). MethodsWe tested the P300 speller on two epilepsy patients with temporary subdural electrode arrays over the occipital and temporal lobes respectively. We then performed offline analysis to determine the accuracy and bit rate of the system and integrated spectral features into the classifier and used a natural language processing (NLP) algorithm to further improve the results. ResultsThe subject with the occipital grid achieved an accuracy of 82.77% and a bit rate of 41.02, which improved to 96.31% and 49.47 respectively using a language model and spectral features. The temporal grid patient achieved an accuracy of 59.03% and a bit rate of 18.26 with an improvement to 75.81% and 27.05 respectively using a language model and spectral features. Spatial analysis of the individual electrodes showed best performance using signals generated and recorded near the occipital pole. ConclusionsUsing ECoG and integrating language information and spectral features can improve the bit rate of a P300 speller system. This improvement is sensitive to the electrode placement and likely depends on visually evoked potentials. SignificanceThis study shows that there can be an improvement in BCI performance when using ECoG, but that it is sensitive to the electrode location.