Modelling Subject Variability in the Spatial and Temporal Characteristics of Functional Modes

Abstract

Recent work has highlighted the scale and ubiquity of subject variability in observa ons from func onal MRI data (fMRI). Furthermore, it is highly likely that errors in the es ma on of either the spa al presenta on of, or the coupling between, func onal regions can confound cross-subject analyses, making accurate and unbiased representa ons of func onal data essen al for interpre ng any downstream analyses.Here, we extend the framework of probabilis c func onal modes (PFMs) [Harrison et al. ] to capture cross-subject variability not only in the mode spa al maps, but also in the func onal coupling between modes and in mode amplitudes. A new implementa on of the inference now also allows for the analysis of modern, large-scale data sets, and the combined inference and analysis package, PROFUMO, is available from git.fmrib.ox.ac.uk/samh/profumo. Using res ng-state data from , subjects collected as part of the Human Connectome Project [Van Essen et al. ], and an analysis of subjects in a variety of con nuous task-states [Kieliba et al.], we demonstrate how PFMs are able to capture, within a single model, a rich descrip on of how the spa o-temporal structure of res ng-state fMRI ac vity varies across subjects.We also compare the new PFM model to the well established independent component analysis with dual regression (ICA-DR) pipeline. This reveals that, under PFM assump ons, much more of the (behaviorally relevant) cross-subject variability in fMRI ac vity should be a ributed to the variability in spa al maps. This has fundamental implica ons for the interpreta on of cross-sec onal studies of func onal connec vity that do not capture cross-subject variability to the same extent as PFMs.

The relationship between spatial configuration and functional connectivity of brain regions revisited

“…In our previous work, we showed that simulated data containing *only* interindividual variation in node spatial maps resulted in a substantial amount of interindividual information in temporal network matrices estimated with ICA-DR (Bijsterbosch et al, 2018). There, the spatial information in the simulated data (i.e., the simulation ‘ground truth’) were subject-specific PFM maps, which are known to contain spatial overlap (Harrison et al, 2019; Harrison et al, 2015). Therefore, the theory above provides a clean (and mathematical) explanation for how the ‘ground truth’ spatial correlations present between PFM spatial maps can contaminate temporal correlations estimated from traditional ICA-DR.…”

“…An alternative parcellation method, designed to avoid the spatial independence constraint, is PROFUMO. This adopts a hierarchical Bayesian framework that includes spatial priors (for map sparsity and group map regularisation) and temporal priors (consistent with the hemodynamic response function) (Harrison et al, 2019; Harrison et al, 2015). The Probabilistic Functional Mode (PFM) maps obtained from PROFUMO commonly show relatively extensive amounts of spatial overlap (and hence spatial correlation) between nodes.…”

The developing Human Connectome Project (dHCP) automated resting-state functional processing framework for newborn infants

“…This forms a complete probabilistic model for the data, and the group-and subject-level information is inferred together via a variational Bayesian inversion scheme. This is covered in more detail in Harrison et al (2019).…”

Functional and diffusion MRI reveal the functional and structural basis of infants’ noxious-evoked brain activity

“…In the same cohort of 18 infants, nine resting-state networks were robustly identified from separate resting-state scans using probabilistic functional mode analysis 17,18 (Figure 2). These included three sensory and motor networks (two visual, two auditory, and two somatomotor networks) and three cognitive networks (default mode, dorsal attention, and executive control networks).…”

“…The resting-state network analysis performed on the 18-infant nociception-paradigm dataset was closely matched to that described for the dHCP dataset 24 . In brief, probabilistic functional mode (PFM) analysis using FSL's PROFUMO 17,18 was run on both datasets with a pre-specified dimensionality of 25, and using the infant double-gamma HRF 23,41 as the temporal prior. PROFUMO's Bayesian model complexity penalties can eliminate modes, thus returning a number of group-level modes that can be less than the pre-specified dimensionality.…”