Additive and Interaction Models for Nonparametric Regression of Biomedical Imaging Data, with Applicaton to Ophthalmological Multi-Level Data on the Sphere

• Activity Number: 107
• Type: Invited
• Date/Time: Monday, August 4, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Jeffrey S. Morris*+ and Veera Baladandayuthapani
• Companies: MD Anderson Cancer Center and MD Anderson Cancer Center
• Keywords: Image Analysis ; Functional Data Analysis ; Multi-level Modeling ; Bayesian Analysis ; Mixed Models ; Big Data

Abstract

In biomedical research, numerous assays produce complex, high dimensional image data. A common analysis goal is regression to either identify which image regions differ across groups or to classify subjects based on these image data. Efficient methods must be flexible enough to capture their rich internal structure, account for between-image correlation from the design, and scale up to enormous data sizes. We demonstrate how Bayesian functional mixed models can accomplish these goals, and can accommodate nonparametric effects for continuous predictors and their interactions that borrow strength and vary freely across the image, and include functional growth curve components to capture longitudinal correlations that also vary across the image. We apply these methods to a glaucoma study of pressure-induced scleral strain yielding longitudinal measurements of densely sampled image data on the scleral surface of the eye to nonparametrically estimate the effect of age on strain across the scleral surface and its derivative. Our analysis demonstrates strain is greatest near the optic nerve, where there is clearly a decreasing trend with age that accelerates around age 65.