Bayesian Multimodal Local False Discovery Rate in Neuroimaging Studies

Abstract

For exploratory neuroimaging studies comparing disease group and healthy control (HC) group, one of the primary objectives is to identify any differential connectivities that may be potentially associated with the pathophysiology of neurological and psychiatric disorders. As such, thousands of hypotheses are tested simultaneously that is known as the large-scale simultaneous hypothesis testing problem. Most of the hypotheses tested are null, that is, nothing but noise, while a very small number of them may contain true signals. The false discovery rate (FDR) and local FDR (Lfdr) methods have been developed to address this problem. Typical neuroimaging studies usually have small sample size due to the high economic cost, leading to low statistical power and high probability of falsely significant findings. The existing methodologies do not provide a satisfied control of FDR especially for small sample size studies, neither capabilities that allow integration of multimodal neuroimaging data into statistical modeling. The information provided by multimodal imaging techniques can be complementary to each other and thus integrated multimodal analysis enables us to borrow strength from different modalities. A covariate-modulated Lfdr method has been used in genome-wide association studies and proved to be efficient by increasing power. We extend this method to multimodal neuroimaging data, aimed to improve FDR control and sensitivity to detect differential functional connectivity (FC) links between disease group and HC group in a cross-sectional, comparative, multimodal neuroimaging study with small sample size. We implement a Bayesian multimodal Lfdr approach, which utilizes a Bayesian mixture model to leverage structural connectivity (SC) statistics and enhance modeling of the density of FC statistics. The utility of Bayesian multimodal approach is illustrated with extensive simulation study and a neuroimaging study in late-life depression (LLD) in which both FC (using resting-state functional magnetic resonance imaging) and SC (using diffusion tensor imaging) data were measured on each participant. We demonstrate in simulation study, that Bayesian multimodal Lfdr method performs numerically better in terms of FDR control by comparison with the traditional Lfdr method that solely considers FC especially when the sample size is small. In addition, we employed a Bayesian multiple comparison method via a non-parametric Bayesian Dirichlet process mixture model directly on FC data in the LLD neuroimaging study, as a comparison with the results from Bayesian multimodal Lfdr method.