Residual Relationships Between Motion and Bold Activity Remain after Preprocessing

Background: Head motion is known to influence the BOLD fMRI signal, but its effects are not yet fully understood. Certain techniques, such as functional connectivity MRI (fcMRI), are particularly sensitive to motion, which has been shown to spuriously affect conclusions in studies comparing groups that differ in movement characteristics (Power et al., 2012; Deen & Pelphrey, 2012; Tyszka et al., 2014), such as controls and individuals with ASD. Many data cleanup practices (e.g. censoring/”scrubbing”) require considerable loss of data, and might not fully eliminate residual motion influences (Burgess et al., 2016). Thus a further characterization of the influences of motion on the BOLD signal is needed.

Objectives: To develop new methods to quantify relationships between head movement and BOLD activity, and assess whether current cleanup methods adequately account for movement-related effects on the BOLD signal.

Methods: We analyzed two datasets. (1) Two 16-minute resting state fMRI scans (TR=813ms) from 29 adolescent and adult controls and 24 matched individuals with ASD collected at Indiana University. (2) Two 14-minute resting state scans (TR=720ms) from 75 unrelated Human Connectome Project participants (Van Essen et al., 2013). Both datasets were preprocessed nearly identically using state-of-the-art cleanup methods (ICA-FIX; Salimi-Khorshidi et al., 2014; Smith et al., 2013). Dataset 1 was also alternatively preprocessed using conventional motion and nuisance regression (with and without global signal). We developed a new analytic method for assessing residual movement-linked BOLD artifact, both at the individual and group level, by quantifying the relationship between movement severity and subsequent BOLD activity.

Results: We found that movements are systematically linked with structured and prolonged changes in the BOLD signal that depend on the severity of the preceding motion. This relationship was not limited to high movement epochs; in fact, remarkably small movements were linked with structured BOLD changes occurring considerably later in time. Nearly all motion magnitudes (including those well below typical censoring thresholds) were associated with structured BOLD changes extending as far as 30s later; for some larger motions, over 50s later. Effect sizes of motion-linked BOLD changes were largest at approximately 6s and 20s following motion. These patterns were replicated in 4 independent sessions from two different scanners and persisted robustly across multiple preprocessing methods, but were not observed in four different null models. Note that scrubbing procedures cannot eliminate these temporally distant BOLD changes, as they persist much later than typical temporal masking (e.g. over 5-10 TRs post-motion).

Conclusions:  We provide a novel description of systematic and temporally far-reaching influences of motion on the BOLD signal, supporting previous case reports (Power, 2016; 2014). These influences are not yet adequately handled by state-of-the-art preprocessing methods, and have the potential to artifactually influence results of fcMRI studies that compare groups differing in motion characteristics. Characterizing this pattern is a critical first step in developing new methods that can address it appropriately. Our results suggest caution in interpreting the meaning of different patterns of functional connectivity between ASD and controls until we better understand the interaction between motion and the BOLD signal.