Motion regression induces global signal related bias in functional connectivity estimates

2024 

Abstract Submission 

Yixiang Mao¹, Conan Chen², Truong Nguyen², Thomas Liu²

¹University of California San Diego, La Jolla, CA, ²University of California San Diego, La Jolla, CA

Yixiang Mao  
University of California San Diego
La Jolla, CA

Conan Chen  
University of California San Diego
La Jolla, CA

Truong Nguyen  
University of California San Diego
La Jolla, CA

Thomas Liu  
University of California San Diego
La Jolla, CA

Regressing out motion parameters estimated from volume registration is a common step in the preprocessing of resting-state fMRI (rsfMRI) data. A prior study has shown that rsfMRI motion parameters can exhibit BOLD-weighted bias associated with the resting-state global activity of the brain, as characterized by the global signal (GS) [1]. In this work, we demonstrate that regression with biased motion estimates can negatively bias resting-state functional connectivity (rsFC) estimates and reduce rsFC differences between young and old subjects.

We used an open source multi-echo fMRI dataset [2], consisting of two rsfMRI scans per subject from a sample of 181 younger and 120 older adults. Motion parameters (3 translation, 3 rotation) were separately estimated for the 1st and 2nd echo data (denoted as e1 and e2 with TE = 14 and 30 ms, respectively) using AFNI 3dvolreg. The GS was computed from the percent change e2 data as the average signal over brain voxels. rsFC estimates were computed with correlation analysis after motion censoring (framewise displacement FD > 0. 2mm) and regression with either the e1 or e2 motion parameters. We used 7 ROIs, including 4 ROIs in the default mode network (DMN) and 3 ROIs in the task positive network (TPN). The primary metric of interest was the difference Δr = Δr_e2 - Δr_e1 between the rsFC estimates obtained by regressing out either the e1 or e2 motion parameters. Because of the greater GS-related bias in the e2 motion parameters [1], the difference Δr reflects the effect of this bias on the rsFC estimates. The corresponding differences Δz in z-scores were also computed. Runs were sorted into one of four groups (low/high motion; low/high GS amplitude) where the respective median values were used to define the boundary between low and high groups, and Δr and Δz were examined within each group. In addition, we looked at the differences in GS amplitude and rsFC estimates between the younger and older subject groups.

Figure 1 shows that there is a negative bias (Δr < 0 in upper triangle; Δz < 0 in lower triangle) in the rsFC estimates between all pairs of ROIs for each of the four groups, with the most negative values observed in the low motion and high GS amplitude (aGS) group. This finding indicates that regression using e2 motion parameters containing GS-related bias can introduce a negative bias in the rsFC estimates, with the magnitude of this bias increasing with higher rsfMRI global activity. Figure 2(a-c) shows the differences in aGS, FD, and Δz (e2-e1) between young and old subjects. The young subjects show significantly higher aGS and lower FD than the older subjects and exhibit more negative Δz values, consistent with the trends observed in Figure 1. Figure 2(d,e) shows the age-related rsFC differences (young-old) as Δz (upper triangle) and effect size d (lower triangle) when regression is performed with either e1 motion parameters with minimal BOLD bias (panel d) or GS-biased e2 motion parameters (panel e). With e1 motion regression, the young subjects showed significantly higher rsFC for all but one ROI pair. Regression with e2 motion parameters reduced the rsFC difference across all ROI pairs, with 5 ROI pairs no longer exhibiting significant differences.

Because motion parameters estimated from rsfMRI data acquired at typical echo times (e.g. 30 ms) can exhibit GS-related bias, regression with these parameters will tend to reduce rsFC values, similar to the effect seen with global signal regression [3]. This negative rsFC bias will tend to be greater in groups with higher GS amplitudes, and can thus alter group differences in rsFC between groups that have different mean GS amplitudes. Thus caution must be used when interpreting rsFC differences obtained when using motion regression as part of the preprocessing pipeline.

Motion Correction and Preprocessing ²

Task-Independent and Resting-State Analysis ¹

FUNCTIONAL MRI

Other - global signal; functional connectivity; motion regression