Neighborhood based EEG compression method on P300 speller systems

摘要

Brain computer interfaces (BCI) are used through event-related potentials. An example of BCI systems is speller systems, which aims to identify the letter that focused through visual and auditory stimuli according to the response of the brain after 250-400ms. In the use of spelling systems, very large data sizes are encountered due to the measurements made with stimuli repetition and 64 electrodes so that the result can be more accurately detected. This situation can cause difficulties both in storing the data and transmitting it in online systems. Various techniques is used to compress EEG signals. In this study, neighborhood based data compression method is proposed for compressing EEG signals including P300 data. After the EEG data is divided into epochs, with purpose of the data distributions of the resulting columns are calculated. The comparison is made with the threshold obtained with averages the differences of the variances of each neighboring column. The variance differences of neighboring columns is compared with the threshold, one of the neighboring columns was removed if the difference smaller than the threshold. If the neighborhood variance difference is above the threshold level, both columns are kept in the data block. This process is applied to all neighboring columns pairs. Finally, the data matrix is merged to obtain a compressed EEG signal. In this study, Cz (11), Pz (51) and Poz (58) channels were selected for 15th target letter from the training data belonging to the Subject A selected at random from the data set recorded for the 3rd Brain Computer Interface (BCI) competition. In order to examine the effects on compression of the segmentation time, the signals are divided into 250 - 500 and 1000ms segments. At the result of the study, average 33.056% to 44.583% compression is obtained on the basis of segmentation time, and the analysis of the EEG signal and analysis of the time-domain features is seen that the variance neighborhood based compression process can be performed with 250ms segmentation on the raw signal.