Dimension reduction of dynamic functional connectivity network based on clustering for ASD diagnosis

Abstract

Brain functional connectivity (FC) network has been widely used to identify the biomarkers of neuropsychiatric disorders, especially dynamic FC network (D-FCN) based on sliding window strategy that can describe the characteristics of FC over time. However, many existing D-FCNs have dimensional disasters in varying degrees, which not only increases the computation complexity, but also damages the generalization performance of the network model. In this work, we design a network reduction scheme based on clustering strategy by grouping all correlation time series (i.e., dynamic FC series) in D-FCN into different clusters, then calculating the mean dynamic FC series in each cluster to estimate the reduction network. Specifically, the whole RS-fMRI time series is divided into several overlapped sub-series segments by sliding window strategy, and construct a temporal low-order FC network (t-FCN) and a temporal high-order FC network (t-HFCN) for each sub-segment. The set of t-FCNs is a low-order dynamic network (Lo-D-FCN), while the set of t-HFCNs is a high-order dynamic network (Ho-D-FCN). And then we use the clustering algorithm to group the time series in the network into several clusters to find the internal similar patterns of the dynamic FC series in the same cluster, then calculating the mean dynamic FC series in each cluster and stacking them in parallel to estimate the reduction network. Finally, we used the mean values of all the elements in each mean dynamic FC series as a recognition feature, and input the feature vectors into SVM for classification. The reduction network greatly reduces the dimension of feature recognition and by fusing different types of reduction network classification results, the final classification result reached 83.7%.