Title: Clustering longitudinal biomarker data using Dirichlet process mixtures
Authors: Wesley Johnson - UC Irvine (United States) [presenting]
Abstract: The aim is to jointly model longitudinal diagnostic outcome data (ELISA (continuous) and Fecal Culture (dichotomous)) for individuals that are not infected at the beginning of time, but where some individuals become infected over the course of the study. Individuals that remain uninfected throughout the study have continuous serologic responses that continue to vary about their own baseline level. Responses after infection ultimately increase and then tend to plateau at a higher baseline level. Infection times are modeled using change points, and curve shapes for infected individuals (after unknown time of infection) are modeled using a Dirichlet process mixture of 5 parameter sigmoid curves. This permits clustering of curve shapes, thus allowing for some individuals to a more rapid increase and or shape in response curve than others. All of this is helpful for estimating receiver operating characteristic curves for use of the biomarker in determining cutoffs for detection of infection for individuals in different groups, and for understanding the behavior of testing procedures in different populations where individuals may be known to have been infected longer or shorter periods of time.