This title appears in the Scientific Report :
2023
Please use the identifier:
http://dx.doi.org/10.34734/FZJ-2023-03041 in citations.
Voxel- versus ROI-level Nuisance Regression for Functional Connectivity Analyses
Voxel- versus ROI-level Nuisance Regression for Functional Connectivity Analyses
Introduction: Individual-level analysis of functional connectivity (FC) using large datasets demandsconsiderations of computational efficiency. The extraction of FC involves two mainpreparatory steps: 1) cleaning the voxel-level time series from nuisance variables, and 2)aggregation of the voxel-lev...
Saved in:
Personal Name(s): | Muganga, Tobias (Corresponding author) |
---|---|
Sasse, Leonard / Larabi, Daouia / Eickhoff, Simon / Patil, Kaustubh | |
Contributing Institute: |
Gehirn & Verhalten; INM-7 |
Imprint: |
2023
|
DOI: |
10.34734/FZJ-2023-03041 |
Conference: | Organization for Human Brain Mapping (OHBM), Montreal (Canada), 2023-07-22 - 2023-07-26 |
Document Type: |
Poster |
Research Program: |
Brain Dysfunction and Plasticity |
Link: |
OpenAccess |
Publikationsportal JuSER |
Introduction: Individual-level analysis of functional connectivity (FC) using large datasets demandsconsiderations of computational efficiency. The extraction of FC involves two mainpreparatory steps: 1) cleaning the voxel-level time series from nuisance variables, and 2)aggregation of the voxel-level time series into region-level time series. Computationalefficiency can be improved by cleaning the region-level instead of the voxel-level time series.When using mean aggregation, the two approaches should be numerically equivalent(Friston et al., 2006). However, the two approaches have not been systematically comparedfor another popular option of aggregation: the first eigenvariate. While these two aggregationschemes are similar for regions with homogenous voxel-level time series, the firsteigenvariate better captures the signal of heterogeneous regions (Friston et al., 2006). In thecurrent study, we evaluated individual specificity and the capacity to predict cognitive scoresof FC obtained from both cleaning approaches and aggregation schemes.MethodWe used FIX-cleaned resting-state data acquired with RL- and LR-phase encodingdirections (Smith et al., 2013) from two scan sessions (one day apart) from the HCP-YAS1200 release (N = 376; NFemale = 182; Mage = 28.75; SDage = 3.82; Van Essen et al., 2013).The 400 region version of the Schaefer atlas (Schaefer et al., 2018) was used to computeaggregate region-level time series as the mean or first eigenvariate of all voxel time serieswithin each region. Signal cleaning was applied either voxelwise with subsequentaggregation into the region-level time series or aggregation was performed first followed bycleaning. Cleaning steps included linear detrending, bandpass filtering (0.01-0.08 Hz) andlinear nuisance regression. Nuisance variables included mean white matter (WM), meancerebral spinal fluid (CSF), mean gray matter (GM), 24 motion regressors (Friston et al.,1996), and 21 aCompCor regressors from WM, CSF and GM (Muschelli et al., 2014). Next,FC was computed as Pearson’s correlation between all pairs of regional time series and thetwo FCs from the same session acquired with two phase encoding directions wereaveraged. We assessed individual specificity in the resulting FCs by computing identificationaccuracy (Finn et al., 2015) and differential identifiability (Amico & Goni, 2018). Thepredictive capacity of interindividual differences from FC was assessed using kernel ridgeregression in a 5-fold cross-validation scheme with 5 repeats to predict age and fivecognitive scores; cognitive flexibility, working memory, fluid intelligence, vocabulary andreading.Results: We found that moving from voxelwise cleaning to regionwise cleaning led to a 2-folddecrease in computation time per FC (Msaving = 110 sec.). While individual specificity betweenthe two cleaning approaches was similar in the case of mean aggregation, regionwisecleaning resulted in higher individual specificity for first eigenvariate aggregation (Fig. 1). Asimilar yet more diverse picture was found for the prediction analysis. Here, for meanaggregation both cleaning approaches performed similarly for all predicted scores. Whenusing eigenvariate aggregation, both cleaning approaches performed equally for ageprediction. However, regionwise cleaning resulted in better prediction performance of fluidintelligence, vocabulary and reading while voxelwise cleaning led to higher predictionperformance for cognitive flexibility, and working memory scores (Fig. 2).Conclusion: With our findings, we encourage a more efficient default of regionwise cleaning for FCextraction when mean aggregation is used. In the first eigenvariate aggregation, theindividual specificity was better with regionwise cleaning but prediction results were mixed.Thus, further investigation is advised before extending the regionwise cleaning for firsteigenvariate aggregation. |