Title: Incorporation of brain spatial and connectivity-based information in statistical regularization
Authors: Jaroslaw Harezlak - Indiana University School of Public Health-Bloomington (United States) [presenting]
Aleksandra Steiner - University of Wroclaw (Poland)
Damian Brzyski - Wroclaw University of Science and Technology (Poland)
Timothy Randolph - Fred Hutchinson Cancer Research Center (United States)
Joaquin Goni - Purdue University (United States)
Abstract: Prior information use in a principled manner can improve the quality of the regression coefficient estimation. Our proposal incorporates structural connectivity derived from Diffusion Weighted Brain Imaging and cortical spatial distance in the penalized approach. Extending previously developed methods informing the estimation of the regression coefficients, we incorporate such information via a Laplacian matrix based on the proximity measures. The penalty term is constructed as a weighted sum of structural connectivity and proximity between cortical areas. Simulation studies show improved estimation accuracy. We apply our approach to the data collected in the Human Connectome Project, where the cortical properties of the left hemisphere are found to be associated with vocabulary comprehension.