Parallel Computing Sparse Wavelet Feature Extraction for P300 Speller BCI.

Abstract

This work is intended to increase the classification accuracy of single EEG epoch, reduce the number of repeated stimuli, and improve the information transfer rate (ITR) of P300 Speller. Target EEG epochs and nontarget EEG ones are both mapped to a space by Wavelet. In this space, Fisher Criterion is used to measure the difference between target and nontarget ones. Only a few Daubechies wavelet bases corresponding to big differences are selected to construct a matrix, by which EEG epochs are transformed to feature vectors. To ensure the online experiments, the computation tasks are distributed to several computers that are managed and integrated by Storm so that they could be parallelly carried out. The proposed feature extraction was compared with the typical methods by testing its performance of classifying single EEG epoch and detecting characters. Our method achieved higher accuracies of classification and detection. The ITRs also reflected the superiority of our method. The parallel computing scheme of our method was deployed on a small scale Storm cluster containing three desktop computers. The average feedback time for one round of EEG epochs was 1.57 ms. The proposed method can improve the performance of P300 Speller BCI. Its parallel computing scheme is able to support fast feedback required by online experiments. The number of repeated stimuli can be significantly reduced by our method. The parallel computing scheme not only supports our wavelet feature extraction but also provides a framework for other algorithms developed for P300 Speller.