Title: Estimating joint sparse graphical models using fMRI brain imaging data with different coarseness levels
Authors: Eugen Pircalabelu - Université catholique de Louvain (Belgium) [presenting]
Abstract: Brain networks are estimated from fMRI datasets that do not all contain measurements on the same set of regions. For certain datasets, some of the regions have been split in smaller subregions, while others have not been split. This gives rise to mixed scale measurements and the purpose is to estimate sparse undirected graphical models. To overcome the problem of mixed coarseness levels, the expand and the reduce algebraic operators are used. To estimate sparse undirected graphs, the ADMM algorithm is used and to ensure similarity of graphs across coarseness levels, the procedure uses the fused and group lasso penalties for certain block submatrices and a lasso penalty for the remaining submatrices. The method estimates edges for each subject and coarseness level, referred to as within level edges, and identifies possible connections between a large region and its subregions, referred to as between level edges which offer insight into whether a certain large region is constructed by aggregating homogeneous or heterogeneous parts of the brain. It also avoids the tedious task of selecting one coarseness level for carrying out the analysis and produces interpretable results at all available levels. Empirical and theoretical evaluations illustrate the usefulness of the method.