EEG Functional Connection Analysis Based on the Weight Distribution of Convolutional Neural Network

Functional connections are commonly used when exploring the human brain, especially in brain data analysis. However, most of the studies concentrate on traditional statistical analysis. In this paper, we innovatively combined the functional connection with the deep learning algorithms and analysed the matrices after the weight distribution of each layer of the convolutional neural network (CNN) to obtain the connections that play a vital role in the classification. The electroencephalogram (EEG) data used in this paper was acquired through a visual mismatch negativity (MMN) experiment. When dealing with this data, each electrode was regarded as a node in the network, and the phase lag index (PLI) was calculated to construct the functional connection matrices, which were used as inputs for the CNN classification and feature extraction. The matrices after the weight distribution were further analysed by means of graph theory. In this paper, the classification accuracy for deviation and standard stimuli are over 95%, and the theta band achieved the highest accuracy. Through the distributed matrices, we found that there are two regions that obtained larger weights from the convolutional layers, i.e., the temporal lobe and the occipital region. The occipital region is related to our visual experiment, and the temporal lobe region is connected with MMN mechanism. We also considered the strategy of the three-layer CNN according to weight distribution processing.