Bayesian Analysis of fMRI data with Spatially-Varying Autoregressive Orders

TL;DR

This paper introduces a Bayesian hierarchical model for fMRI data that automatically determines spatially-varying autoregressive noise orders, improving modeling accuracy of the temporal dependence in brain imaging data.

Contribution

The paper presents a novel hierarchical Bayesian approach that allows for spatially-varying autoregressive orders in fMRI noise modeling, addressing limitations of fixed AR order assumptions.

Findings

Enhanced accuracy in modeling fMRI noise with the new method

Effective automatic selection of spatially-varying AR orders

Improved fit to real fMRI data in case studies

Abstract

Statistical modeling of fMRI data is challenging as the data are both spatially and temporally correlated. Spatially, measurements are taken at thousands of contiguous regions, called voxels, and temporally measurements are taken at hundreds of time points at each voxel. Recent advances in Bayesian hierarchical modeling have addressed the challenges of spatiotemproal structure in fMRI data with models incorporating both spatial and temporal priors for signal and noise. While there has been extensive research on modeling the fMRI signal (i.e., the covolution of the experimental design with the functional choice for the hemodynamic response function) and its spatial variability, less attention has been paid to realistic modeling of the temporal dependence that typically exists within the fMRI noise, where a low order autoregressive process is typically adopted. Furthermore, the AR order is held constant across voxels (e.g. AR(1) at each voxel). Motivated by an event-related fMRI experiment, we propose a novel hierarchical Bayesian model with automatic selection of the autoregressive orders of the noise process that vary spatially over the brain. With simulation studies we show that our model has improved accuracy and apply it to our motivating example.