Temporally Dynamic Resting-State Functional Connectivity Networks for Early MCI Identification

Abstract

Resting-state functional Magnetic Resonance Imaging (R-fMRI) scan provides a rich characterization of the dynamic changes or temporal variabilities caused by neural interactions that may happen within the scan duration. Multiple functional connectivity networks can be estimated from R-fMRI time series to effectively capture subtle yet short neural connectivity changes induced by disease pathologies. To effectively extract the temporally dynamic information, we utilize a sliding window approach to generate multiple shorter, yet overlapping sub-series from a full R-fMRI time series. Whole-brain sliding window correlations are computed based on these sub-series to generate a series of temporal networks, characterize the neural interactions between brain regions at different time scales. Individual estimation of these temporal networks overlooks the intrinsic temporal smoothness between successive overlapping R-fMRI sub-series. To handle this problem, we suggest to jointly estimate temporal networks by maximizing a penalized log likelihood via a fused lasso regularization: 1) l ₁-norm penalty ensures a sparse solution; 2) fused regularization preserves the temporal smoothness while allows correlation variability. The estimated temporal networks were applied for early Mild Cognitive Impairment (eMCI) identification, and our results demonstrate the importance of including temporally dynamic R-fMRI scan information for accurate diagnosis of eMCI.