Title: A Bayesian clustering approach for the analysis of repeated cognitive tests, imaging, genetic data and time-to-dementia
Authors: Sylvia Richardson - MRC Biostatistics - Cambridge (United Kingdom)
Paul Kirk - University of Cambridge (United Kingdom)
Brian Tom - MRC Biostatistics Unit (United Kingdom)
Anais Rouanet - Bordeaux Population Health Center (France) [presenting]
Abstract: The aim is to present an outcome-guided Bayesian Dirichlet Process Mixture Model developed to identify patterns of cognitive decline associated with differential dementia risk, as well as specific profiles of baseline socio-demographic, imaging and genetic data. Our model links a longitudinal outcome, a time-to-event and a set of correlated variables, called profile variables, to help identify clusters of patients. Adopting a Bayesian approach with a Dirichlet Process prior upon the mixture distribution allows us to quantify the uncertainty in both the number of clusters and their characteristic profiles. Given the cluster allocations, the longitudinal outcome and the profile variables are described by mixed-effects model and multinomial (or multivariate Gaussian) distributions, respectively. The model is estimated by MCMC and uncertainty of the final partition is assessed through model-averaging techniques. We present results from the North American Alzheimer's Disease NeuroImaging (ADNI) study, which demonstrate the utility of our approach to refine the stratification of subjects with high risk of dementia.