Activity Number:
|
63
- Statistical Methods for Brain Connectivity and Network Analysis
|
Type:
|
Topic Contributed
|
Date/Time:
|
Sunday, July 30, 2017 : 4:00 PM to 5:50 PM
|
Sponsor:
|
Section on Bayesian Statistical Science
|
Abstract #323929
|
View Presentation
|
Title:
|
Use of Brain Connectivity Information in Regression Parameter Estimation via Generalized Regularization
|
Author(s):
|
Damian Brzyski* and Marta Karas and Joaquin Goni and Beau Ances and Timothy Randolph and Jaroslaw Harezlak
|
Companies:
|
Indiana University and Indiana University and Purdue University and Washington University School of Medicine and Fred Hutchinson Cancer Research Center and Indiana University School of Public Health
|
Keywords:
|
Brain imaging ;
Structural connectivity ;
Regularization ;
HIV
|
Abstract:
|
External knowledge is not frequently integrated in the study of associations between the correlated high-dimensional predictors and scalar outcomes. In our work, we incorporate prior information quantifying the relationships between the predictors arising from their connectivity, e.g. in a brain imaging study, we can use the structural connectivity information to study the association of cortical properties with a disease phenotype. We introduce a number of approaches extending the classical Tikhonov regularization by defining a penalty term based on the connectivity-derived Laplacian matrix. We address technical and computational issues caused by the Laplacian matrix non-invertibility both theoretically and in the simulation studies. Moreover, we study the information content of the connectivity matrices in the regression context and propose a solution that is adaptive to it. Our work addresses both theoretical aspects and computational issues of the proposed approach. Finally, we apply our method to study the association of cortical markers with HIV-infected individuals' health outcomes.
|
Authors who are presenting talks have a * after their name.