A Reduced Rank Regression Framework for Interpretable Image-On-Scalar Regression with Application to Alzheimer's Disease

• Activity Number: 528 - Contributed Poster Presentations: Section on Statistics in Imaging
• Type: Contributed
• Date/Time: Wednesday, July 31, 2019 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Imaging
• Author(s): Tianyu Ding* and Rob Krafty and Dana Tudorascu and Annie Cohen
• Companies: University of Pittsburgh and University of Pittsburgh and University of Pittsburgh and University of Pittsburgh
• Keywords: Reduced rank regression; Image-on-scalar ; Encoder and decode; Penalized regression; ADMM; Down Syndronme

Abstract

A key question to advancing our understanding of the multidimensional nature of Alzheimer's Disease(AD), which is characterized by both neurological and cardiovascular pathologies, is how brain structure is linked to vascular burden. Reduced rank regression is a widely used technique for overcoming the high-dimensional nature of multivariate regression, such as the image-on-scalar regression of brain images onto vectors of vascular burden. Existing penalized extensions allow for the incorporation of smoothness and sparsity on reduced rank subspaces. However, sparsity and smoothness on reduced rank subspace do not directly translate into meaningful structures on image spaces, blurring biological interpretation. In this study, we propose a sparse reduced rank regression framework, which includes a novel fused group lasso penalty, combined with an encoder decoder, that ensures smoothness, sparsity and interpretability on the image space. An algorithm is developed for model fitting that uses alternating direction method of multipliers (ADMM)to perform optimization within low-dimensional subspaces.