Title: Dynamic functional connectivity: A sparse group fused lasso approach
Authors: David Degras - University of Massachusetts Boston (United States) [presenting]
Abstract: A novel approach is presented to assess dynamic functional connectivity (DFC) in neuroimaging data. Modeling brain signals as piecewise structural vector autoregressive (SVAR) processes, the method decomposes the time range of the data into intervals of homogeneous functional connectivity (FC). The piecewise SVAR model can be rapidly fitted to data (typically in minutes) thanks to an efficient implementation of sparse group fused lasso (SGFL). To handle the high dimension of the parameter space, SGFL makes two sparsity assumptions: (i) at each time point, there are only few nonzero regression coefficients (i.e. the number of functional connections between pairs of brain regions is small), and (ii) regression coefficients only change infrequently over time, i.e. regime changes in FC are relatively rare. Enforcing these assumptions in model fitting produces a challenging problem of nonsmooth convex optimization, which we solve with a novel hybrid algorithm that combines block coordinate descent, forward-backward algorithms, iterative soft thresholding, and subgradient methods at different levels of optimization. A numerical comparison of the hybrid approach with state-of-the-art optimization procedures is presented, as well as applications of SGFL to resting-state fMRI and EEG data.