Functional Principal Components Mixture Regression with Application to CT Image Data

• Activity Number: 422
• Type: Contributed
• Date/Time: Tuesday, August 6, 2013 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Lucy Robinson*+ and Sriram Balasubramanian and Silpa Reddy
• Companies: Drexel University and Drexel University and Drexel University
• Keywords: image data ; functional data analysis ; mixture models

Abstract

We propose a novel functional principal components mixture regression model with application to CT image data. As a motivating example, we consider ribcage images of pediatric subjects with thoracic deformities. Image data are used as covariates in a regression model predicting scalar pulmonary function measures. Each rib pair can be considered as a functional data object, and within each subject we may have a mixture of ribs of a normative shape and ribs exhibiting some deformity. Variation across rib pairs within subjects and variation between subjects are described using a multilevel functional principal components analysis. The relationship between the scalar response and the functional principal components is described by a mixture regression model in which the regression function depends on an unobserved latent class variable. Model parameters are estimated in a Bayesian framework using MCMC.