Determination of the common electrodes for users and increasing the classification accuracy of motor imagery EEG

In recent studies, it is observed that the success rates for the classification of the EEG (electroencephalogram) data have increased by using deep neural networks. However, the portability and practicability of the brain computer interface systems are still not good since too many electrodes are used, which makes the system expensive, time-consuming to setup and uncomfortable for the users. There are studies that reduce the number of electrodes to increase the classification accuracy and to reduce the computing load. For this purpose, the number of electrodes is reduced by performing some operations on the existing set of electrodes, and a subset is determined for each subject. In this study, a set of five electrodes is selected to classify the EEG signals. Electrodes that would be common to all subjects are investigated by using the Rayleigh coefficient map and divergence measure. After the common electrodes are determined, the data belonging to the other electrodes are removed from the datasets. For the classification of the EEG signals a divergence-based deep neural network (DivFE), which had previously shown high performances, is employed. The preprocesses such as filtering, continuous wavelet transform, common spatial patterns, short-time Fourier transform and independent components analysis are examined to improve the classification accuracy of the DivFE. The 2005 BCI III-3a, 2008 BCI IV-2a and an artificial EEG dataset are used for the training and test processes. Classification success rates of 80.3, 65.1 and 77.1% are obtained for four classes in the 2005 BCI III-3a dataset, for four classes in the 2008 BCI IV-2a dataset and for ten classes in the artificial EEG dataset, respectively. It is observed that the classification accuracies obtained in the literature can be achieved by using only five common electrodes.