General linear model (GLM) analysis
The task-based GLM-analysis was performed by means of SPM12 Version 7487 (
https://www.fil.ion.ucl.ac.uk/spm/) and the SPM toolbox TFCE (r201 from 2020 to 04-21) in Matlab R2018a (9.4.0.949201 Update 6, MathWorks Inc., Natick, Massachusetts). The first 13 images (10 s) of each session were removed to account for T1-equilibration effects that go beyond the initial dummy scans removed by Siemens for fast fMRI protocols. The images were realigned to the first one of each scanning session and were then stereotactically normalized into the standard anatomical space defined by the Montreal Neurological Institute (MNI) template by means of the DARTEL algorithm including geodesic shooting using an existing MNI-template (
http://nist.mni.mcgill.ca/?p=904) through the use of the CAT12 toolbox (version 1450)
(Ashburner
2007). Therefore, the stereotactic coordinates in this paper refer to the MNI coordinate system. The normalized images were smoothed with a three-dimensional isotropic 4 mm Gaussian kernel and the realignment parameters and a high-pass filter (128 s) were integrated into the design matrix. The effect of the different stimulation conditions on regional BOLD responses was estimated according to the general linear model including the realignment parameters (Friston et al.
1995b). The conditions (GVS right, GVS left, GNS right, GNS left) were modelled as blocks.
Statistical parametric maps (SPMs) were generated on a voxel-by-voxel basis with a hemodynamic model of the stimulation periods present during the session (Friston et al.
1995a). To analyse differences in activations during both stimulations in general, we defined the contrasts to include the main effects for GVS applied on the left and right mastoid and for both the left and the right GNS experiments. These results are referred to as “vestibular stimulation” and “nociceptive stimulation” in the following sections.
Single subject t-contrasts were computed for each stimulation condition compared to the rest condition of each session and entered into a second-level statistical analysis to test for effects on a between-subject basis. Paired t-tests were performed between the GVS and GNS contrast, a conjunction analysis to test for areas significantly activated by both GSN and GVS and a correlation analysis including the pain scale and pain sensitivity questionnaire.
Statistical significance was determined using TFCE, with the default parameters after 10,000 permutations using a threshold of
p < 0.05 corrected for multiple comparisons via false discovery rate (FDR) (Smith and Nichols
2009). When applicable and available, the cytoarchitectonic maps of the occipital and temporal lobe, the insular gyri and the parietal operculum were used to calculate the respective overlay of our results (Eickhoff et al.
2005). Results were localized and visualized using the anatomy toolbox (Eickhoff et al.
2005), the ANL atlas (Edlow et al.
2012) and MRIcroGL by Chris Rhorden (
https://www.mccauslandcenter.sc.edu/mricrogl/).
Functional network analysis
After data quality control assessment via MRIQC (Esteban et al.
2017) to detect banding artefacts from multi-band imaging and excessive head movements, preprocessing for functional connectivity analysis was performed using fMRIPprep 1.2.5 (Esteban et al.
2019), based on Nipype 1.1.6 (Gorgolewski et al.
2011). T1 images were bias field corrected and skull stripped. Spatial normalisation was performed to the ICBM 152 Nonlinear Asymmetrical template version 2009c (Fonov et al.
2009) using nonlinear registration (see specifics in the online appendix) and brain tissue was segmented into cerebrospinal fluid, white matter and grey matter. BOLD images were registered to the normalised T1 image. Head motions parameters were estimated with six rotation and translation parameters. No slice timing correction was performed. BOLD times-series were resampled, corrected for head-motion and susceptibility distortions, and normalised to MNI152NLin2009cAsym space. Framewise displacement (FD) and DVARS were calculated and three region-wise global signals were extracted within the CSF, the WM, and the whole-brain masks. For detailed methods, see Online Appendix.
Fmriprep and MRIQC summary outputs were also used for quality control. Because functional connectivity data are particularly susceptible for motion, we used a strict inclusion criterion of a mean framewise displacement of FD > 0.2 as an output in MRQC in any run performed, or BOLD signal extinction in cortical brain areas after fmriprep preprocessing. For the within-group comparison applying these criteria resulted in a dataset of fifteen participants.
For further signal extraction and correction, CONN 18.b was used (Whitfield-Gabrieli and Nieto-Castanon
2012). Extraction was performed separately for the GVS and GNS data applying the same parameters. The reoriented and normalised functional data were used for signal extraction from 100 ROIs (7 Network parcellation), as defined by Schaefer et al. (
2017). Data were despiked, detrended and filtered with a band-pass filter of 0.008–1 Hz to obtain a signal in the standard frequency range used for resting-state analysis. After filtering, regression was performed. For the stimulation sessions, we used a finite impulse response regressor to control for the influence of the mean event responses on functional connectivity values, as suggested by Cole et al. (
2019). Further regressors included motion, CSF and WM signal as determined by fmriprep (raw signal as well as first-order derivative). High motion frames were also accounted for by creating a scrubbing regressor, which included all frames with a framewise displacement above 0.9 mm or BOLD signal changes above five standard deviations. Pearson correlation was calculated for the extracted and denoised signals and adjacency matrices were created for each participant and each condition. Each participant contributed to the analysis with six adjacency matrices in total: three from the GVS experiments (resting-state and GVS stimulation) and three from the GNS experiment (resting state, GNS stimulation left and GNS stimulation right). All further analysis steps were based on these correlation matrices.
General whole brain network changes associated with vestibular stimulation were determined using a within-participant design for the stimulation sessions (GVS and GNS) and the resting-state sessions from the two different experiments.
Two types of functional network analyses were conducted. The first analysis was performed using network-based statistics (NBS), which focuses on differences in individual connections within the network. The second analysis was focused on differences in modularity of the network, i.e. whether functionally related regions (i.e. groups of nodes) maintain or change their affiliation during different conditions. As a control, the two resting-state sessions of the different experiments within the same participant were compared, no changes in network architecture were expected there.
Changes in network connectivity: The NBS toolbox by Zalesky et al. (
2010) was used to determine changes at the level of graph connections. In NBS, statistical tests are performed at every connection—only connections surpassing a primary threshold are further used to identify topological clusters. Considering the arbitrary nature of selecting the primary threshold, we used a range of primary thresholds (from 2 to 3.5 in steps of 0.3). For each component a FWER-corrected
p value is determined with permutation testing at 10 000 permutations using the method of Freedman and Lane (
1983). We only considered a component to be significant, if the
p value was below 0.1 consistently across all primary thresholds tested. Both component extent and component intensity were investigated. Weak effects that include many connections tend to become significant with component extent, whereas testing for component intensity is better for detecting strong, focal connections.
Changes in network modularity: To determine how nodes differ in terms of their functional network participation during the GVS and the GNS sessions, i.e. whether nodes interacted with the same nodes throughout the conditions or whether they changed in terms of their interactions, a consensus modularity analysis as described in Castrillon et al. (
2020) was conducted using custom made Matlab and R scripts (4.0.2 within RStudio 1.3.1056). The analysis was only marginally modified from Castrillon et al. (
2020). For each participant in each of the four conditions, classification was performed using the Louvain algorithm with a gamma of 1.3 (i.e. larger than the default value of 1 to detect smaller modules) and no pre-defined module affiliation. The parameter for consensus modularity analysis was left at tau = 0.4 (Castrillon et al.
2020; Lancichinetti and Fortunato
2012). The result of this analysis was the classification consistency (
z) and diversity (
h) for each node in each of the four conditions (i.e. both resting-state sessions and both stimulation sessions (GVS and GNS). Classification consistency was based on the within-module degree z-score [a within-module version of degree centrality (Rubinov and Sporns
2010)], classification diversity was based on participation coefficient, a measure of diversity of intermodular connections of individual nodes. Functions from the Brain Connectivity Toolbox (Rubinov and Sporns
2010; Bullmore and Sporns
2009) were used to calculate these graph measures. To determine significant differences in classification consistency and diversity between the four conditions, Kruskal–Wallis tests were performed.