A novel joint sparse partial correlation method for estimating group functional networks.

Advances in graph theory have provided a powerful tool to characterize brain networks. In particular, functional networks at group-level have great appeal to gain further insight into complex brain function, and to assess changes across disease conditions. These group networks, however, often have two main limitations. First, they are popularly estimated by directly averaging individual networks that are compromised by confounding variations. Secondly, functional networks have been estimated mainly through Pearson cross-correlation, without taking into account the influence of other regions. In this study, we propose a sparse group partial correlation method for robust estimation of functional networks based on a joint graphical models approach. To circumvent the issue of choosing the optimal regularization parameters, a stability selection method is employed to extract networks. The proposed method is, therefore, denoted as JGMSS. By applying JGMSS across simulated datasets, the resulting networks show consistently higher accuracy and sensitivity than those estimated using an alternative approach (the elastic-net regularization with stability selection, ENSS). The robustness of the JGMSS is evidenced by the independence of the estimated networks to choices of the initial set of regularization parameters. The performance of JGMSS in estimating group networks is further demonstrated with in vivo fMRI data (ASL and BOLD), which show that JGMSS can more robustly estimate brain hub regions at group-level and can better control intersubject variability than it is achieved using ENSS.