Novel functional brain network methods based on CNN with an application in proficiency evaluation

Abstract

With the advancements in the research about the EEG signal processing, there are a large number of different analytical methods include the functional brain network (FBN) methods, most of them are unsupervised and need multi handcrafted steps until classification. It is expected to propose a general-purpose method based on deep learning methods to improve the situation. Therefore, in this paper, we proposed two methods based on the convolutional neural network (CNN). One was analyzing adjacent matrixes computed by the phase locking value (PLV) to generate features based on CNN, which was equivalent to the graph-theoretic (GT) indexes functionally. The other method was a novel end-to-end method, brain connection based on CNN (BCCNN), which used a factorized 1-D CNN to filter the temporal part of the raw EEG, computed the correlation coefficients among the electrodes to build a new kind of FBNs and then extracted features from the FBNs like the last method. Then we performed a working memory (WM) experiment to verify the validity of the two methods. Those methods were used to detect the proficiency of the subjects in a WM task. In the results, the accuracy of the first method was 99.33%, which was as good as that of the GT indexes (99.35%). The accuracy of the second methods was 96.53%, which was lower than the performance of the PLV but higher than that of two conventional CNNs (94.37%, 90.83%).