Meta-analysis of functional neuroimaging data using Bayesian nonparametric binary regression

TL;DR

This paper introduces a Bayesian nonparametric binary regression approach for neuroimaging meta-analysis, allowing spatially adaptive probability modeling of brain activation peaks across multiple studies.

Contribution

It develops a novel Bayesian method that adaptively models voxel activation probabilities, improving over traditional kernel-based meta-analyses.

Findings

Method outperforms standard kernel-based approaches in simulations

Provides voxel-specific miscoding probabilities for quality control

Successfully applied to emotion neuroimaging data

Abstract

In this work we perform a meta-analysis of neuroimaging data, consisting of locations of peak activations identified in 162 separate studies on emotion. Neuroimaging meta-analyses are typically performed using kernel-based methods. However, these methods require the width of the kernel to be set a priori and to be constant across the brain. To address these issues, we propose a fully Bayesian nonparametric binary regression method to perform neuroimaging meta-analyses. In our method, each location (or voxel) has a probability of being a peak activation, and the corresponding probability function is based on a spatially adaptive Gaussian Markov random field (GMRF). We also include parameters in the model to robustify the procedure against miscoding of the voxel response. Posterior inference is implemented using efficient MCMC algorithms extended from those introduced in Holmes and Held [Bayesian Anal. 1 (2006) 145--168]. Our method allows the probability function to be locally adaptive with respect to the covariates, that is, to be smooth in one region of the covariate space and wiggly or even discontinuous in another. Posterior miscoding probabilities for each of the identified voxels can also be obtained, identifying voxels that may have been falsely classified as being activated. Simulation studies and application to the emotion neuroimaging data indicate that our method is superior to standard kernel-based methods.