Classification of 3D interpolated EEG signals using hybrid R-3DCNN

Abstract

Preserving both, the spatial and temporal features of electroencephalogram (EEG) signals, have been the most sought after objectives of researchers to extract better features. To maintain spatial features, each electrode of an EEG cap, located on the scalp of a subject, has to be represented such that the spatial locations of the electrodes are preserved with respect to each other. We propose to project the three-dimensional (3D) coordinates of the electrodes of the EEG cap into the 3D space and interpolate the power spectral density (PSD) values to obtain 3D images, from which features can be extracted using a 3D Convolutional Neural Network (3D-CNN). These features are passed to a Recurrent Neural Network (RNN) to extract the temporal properties of the EEG signals. The proposed approach of using a hybrid Recurrent 3D-CNN (R-3DCNN) supersedes the accuracy of previously adopted practices which use two-dimensional (2D) interpolation techniques of representing PSD values in a similar setting. The proposed approach has been demonstrated successfully on a motor imagery classification task.