Predicting Treatment Efficacy via Quantitative MRI: A Bayesian Joint Model

• Activity Number: 149
• Type: Invited
• Date/Time: Monday, August 10, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Timothy D. Johnson* and Jincao Wu
• Companies: University of Michigan and FDA
• Keywords: quantitative MRI ; Bayesian analysis ; joint model ; image analysis

Abstract

We present a joint, two-stage Bayesian model of treatment efficacy, as determined by one-year survival status, using temporal changes in quantitative MRI (qMRI) measures and assess its predictive performance. In stage I, we smooth the qMRI data with a multivariate spatio-temporal pairwise difference prior model. We propose four summary statistics from the qMRI data in the first stage. In stage II, the statistics enter a generalized non-linear model (GNLM) as predictors of one-year survival status. Covariates enter the systematic component via a multivariate adaptive regression spline basis. The systematic component is linked to the outcome with a probit link. The number of bases is assumed unknown and is estimated by reversible jump MCMC. We apply our model to a pilot study designed to assess the ability of qMRI to predict one-year survival status after three weeks of therapy (currently, assessment is determined five months post treatment initiation; typically too late for second line therapies). By modeling correlation in the images and allowing a flexible decision boundary, our model achieves higher overall correct classification rates than do simpler models.