CFE 2020: Start Registration
View Submission - CMStatistics
Title: Adaptive regularization for multi-modal brain imaging Authors:  Jaroslaw Harezlak - Indiana University School of Public Health-Bloomington (United States) [presenting]
Damian Brzyski - Wroclaw University of Science and Technology (Poland)
Kewin Paczek - Jagiellonian University (Poland)
Timothy Randolph - Fred Hutchinson Cancer Research Center (United States)
Joaquin Goni - Purdue University (United States)
Abstract: The problem of adaptive incorporation of multi-modal brain imagining data in the multiple linear regression setting is addressed. We assume the model of the form $E[Y|X,Z] = X*beta + Z*b$, where the response variable $Y$ corresponds to a neuropsychological outcome, $X$ are the possible confounders, and $Z$ are the explanatory variables (e.g. cortical thickness or area) for which the functional and structural connectivity information exists. The connectivity information is used to build the adaptive penalty terms in the regularized regression problem. The general idea of incorporating connectivity information in regularization approach via linear mixed model representation has been recently established in our prior work: ridgified Partially Empirical Eigenvectors for Regression (riPEER). We incorporate multiple sources of information, e.g. functional and structural connectivity network structure, and estimate the regression parameters with multiple penalty terms via a riPEER extension called AIMER (Adaptive Information Merging Estimator for Regression). We present an extensive simulation study testing various realistic scenarios and apply msPEER to data arising from the Human Connectome Project (HCP) study.