Abstract
Deep Auto-Encoders (DAE) have been widely used in dimensionality reduction and feature extraction of brain functional networks. However, the features in aggregation matrices of functional networks obtained by DAE dimensionality reduction might lose part of time-varying information, so that DAE cannot learn the distribution of original features well. To solve these problems, we extracted and analyzed brain functional statuses for early Mild Cognitive Impairment (eMCI) based on Variational Auto-Encoder (VAE). The high dimensions of features in aggregation matrices of dynamic functional networks were reduced by VAE to obtain the corresponding hidden variable matrices. Gaussian Mixture Model (GMM) was used to cluster the features in the hidden variable matrices to form several Common Functional Networks (CFNs) representing different functional statuses. We analyzed the similarities and differences of functional statuses between eMCI subjects and normal subjects in different sub-segments, as well as the switching of functional statuses in the entire time series. The experimental results show that there are similarities between the most frequent functional statuses of the two types of subjects and differences between the least frequent functional statuses. The proposed method can more significantly reveal the similarity and difference of functional statuses between eMCI subjects and normal subjects than the comparable methods, and the switching rule of functional statuses can help better understand the dynamic characteristics of brain functional networks for eMCI patients.