A Novel Approach to Classify Motor-Imagery EEG with Convolutional Neural Network Using Network Measures

Electroencephalogram (EEG) signal recorded throughout motor imaging (MI) tasks has been wide applied in brain-computer interface (BCI) applications as a communication approach. To improve the classification success rate of MI EEG classification tasks, this paper proposes a completely unique input form based on brain network connectivity measures for the datasets from BCI Competition IV. First, using connectivity patterns between brain regions during MI task, six more frequent network features were selected and their maps were generated in 2D format; then a simple yet powerful convolutional neural network (CNN) with one convolutional layer was deployed for binary classification of MI tasks (left-hand, right-hand, both feet and tongue movements). The discrimination ability of these features was compared with each other. Our results demonstrate that CNN fed with path length feature map can further improve classification performance in most binary problems. While all classification results are better than 86%, the best accuracy using brain network features is 96.69% in right-tongue separation. The present study shows that the proposed method is efficient to classify MI tasks, and provides a practical method for classification of non-invasive EEG signals in BCI applications.