Discovery Among Binary Biomarkers in Heterogeneous Populations

Abstract

Biomarkers have great potential to improve disease diagnosis and treatment. Disease may arise via multiple pathways, however, each associated with distinct complex interactions among multiple biomarkers, and hence patients exhibit considerable heterogeneity in the biomarker-disease association despite sharing the same clinical diagnosis. Thus identification of clinically useful biomarker combinations requires statistical methods that accommodate population heterogeneity and enable discovery of possibly complex interactions among biomarkers that associate with disease. We address jointly modeling binary and continuous disease outcomes when the association between predictors and these outcomes exhibits heterogeneity. In the context of binary biomarkers, we use ideas from logic regression to find Boolean combinations of these biomarkers that predict the binary disease outcome. The associated continuous outcome is modeled as Gaussian. Heterogeneity is cast as unknown subgroups in the population, with the associations between the joint outcome and biomarkers and other covariates varying by subgroup. We adopt a mixture of finite mixtures (MFM) fully Bayesian formulation to simultaneously estimate the number of subgroups, the subgroup membership structure, and the subgroup-specific relationships between outcomes and predictors. We describe how our model incorporates the Boolean relations as parameters arising from the MFM model and our approach to the associated challenges of specifying the prior distribution and estimation using Markov chain Monte Carlo. We illustrate the performance of the methods using simulation and discuss application.