Resting-state fMRI (IV)

In Study IV, we analysed the rsfMRI data of the prospective REVIS trial. Structural and functional images were acquired from 16 chronic occipital stroke patients participating in the Helsinki arm of the trial and 14 healthy control subjects who gave their informed consent. The study was approved by the Ethics Committee of the Helsinki and Uusimaa Hospital District (No. 49/13/03/01/13, date 13/03/2013 and no.

HUS/576/2017, date 06/03/2017). The control subjects were examined by a neurologist who assessed their visual fields with confrontation testing. The flowchart of the study is depicted in Figure 8.

Figure 8. Flowchart of Study IV. Modified from Publication IV. Permission to reproduce granted by publishing terms by SAGE Publishing. HUH, Helsinki University Hospital; VFD, visual field defect; rtACS, repetitive transorbital alternating current stimulation; fMRI, functional magnetic resonance imaging.

Occipital stroke

n = 18

Healthy control n = 14

rtACS n = 9

Sham n = 9

rtACS

n = 8 Sham

n = 8 Healthy control

n = 12

Excluded n = 2 reduced vigilance

Analysed Post treatment fMRI

Follow-up fMRI

Baseline fMRI 2 weeks

2 months Excluded

n = 1 MRI incompatible

Randomisation

Excluded n = 1 claustrophobia - Age 18-75

- Homonymous VFD due to occipital infarct - Lesion age > 6 months

- Stable VFD in baseline perimetry - Corrected visual aquity ≥ 0.4 - No diseases interfering with the study

- Age- and sex-matched to occipital stroke patients

- No neurological or ophthalmological diseases

- No VFD in confrontation perimetry - No contraindication for MRI Stroke patients treated in HUH Healthy population

50

The fMRI measurements were performed with 3T Siemens Magnetom Skyra scanner (Siemens, Erlangen, Germany) at the Advanced Magnetic Imaging Centre (Aalto University, Espoo, Finland) with a 30-channel coil (modified from 32-channel Siemens head coil). An fMRI session comprised anatomical images (high- and/or low-resolution T1, T2, FLAIR) and two 6-min rsfMRI runs. During the resting-state imaging, the subjects were advised to keep their mind empty of thoughts while they fixated on a cross on a back-projection screen. The fMRI sequences covered the occipital, parietal, and most of the temporal and frontal lobes, whereas parts of the inferior frontal and anterior temporal lobe and the cerebellum were excluded to optimise the images for visual cortical areas [1,42].

Functional and anatomical preprocessing were performed with tools from FSL [334,335], AFNI [336], and ITK-SNAP [337], integrated in a Nipype pipeline [338].

Preprocessing steps included skull stripping, deletion of the first six volumes to reach stable magnetisation, normalisation, slice timing correction, and transforming all images to the MNI152 standard space. In addition, we removed variance associated with several nuisance regressors, including temporal filtering and 24 motion parameters. Head movement was measured as framewise displacement [339]; its average magnitude in each run was calculated, and the higher of the two values was included in statistical analyses as a covariate to describe head motion.

We analysed the fMRI data with a method introduced by Craddock et al. [340].

The method is based on a multivariate regression connectivity model that uses a support vector regression analysis to calculate a predictive model for activity of a region of interest (ROI) based on time series of voxels outside the ROI. An advantage of the model is that it does not require a preselection of certain ROIs before analyses.

For the connectivity analysis, we divided the brain into 74 cortical ROIs according to the Harvard-Oxford atlas [341]. The volume covered by infarcted tissue was excluded from the analysis with lesion masks. The support vector regression analysis was run consecutively for all ROIs. First it created a predictive linear model for the time series of the studied ROI based on the averaged time series data of all voxels outside the ROI from the first run of a rsfMRI session. Then it tested the model with data from the second run, and vice versa. Both runs produced one vector of prediction weight parameters. The analysis yielded correlation coefficients prediction accuracy and reproducibility for each ROI: the former between the predicted and observed time series and the latter between the two model vectors. The model is described in more detail in Figure 9.

51

Figure 9. Schematic model of the multivariate regression method (modified from Craddock et al. [340]). Modified from the supplemental material of Publication IV. Permission to reproduce granted by publishing terms by SAGE Publishing. Data 1 and 2 depict two resting-state fMRI runs acquired during one session. Both data are divided in 74 regions of interest (ROI), which are examined one at a time. First, support vector regeression (SVR) analysis is applied to Data 1 to create a linear model (Model 1) for the activity of the ROI based on time series of all voxels outside the ROI (blue arrows). Next, Model 1 is applied to Data 2 to create a prediction of time series of the studied ROI in Data 2 by multiplying time series of the voxels outside the ROI with model weights (red arrows). The predicted time series of the ROI is then compared to the observed time series of the ROI from Data 2, and a correlation coefficient between these two is called Prediction accuracy. The same procedure is repeated for all ROIs and for both Data 1 and 2, resulting in two prediction accuracies for each ROI per session. For analyses, Prediction accuracy 1 and 2 are averaged. Reproducibility is the correlation coefficient between model vectors from Model 1 and 2. In the end, the analysis has yielded 74 Prediction accuracy and Reproducibility values per subject per session.

Furthermore, the model weights from the support vector regression analysis were used in a network analysis according to the graph theory [342]. In a directed and weighted graph, the ROIs represented nodes and the model weights edges between them. From the graph, we calculated four network parameters: centrality degree, centrality eigenvector, average shortest path, and clustering. Centrality degree is the sum of the edge weights that connect to a node [343], whereas centrality eigenvector sums not only the weights of a node but also the weights of the nearest nodes [344]. Average shortest path is the average length of the shortest paths between all possible pairs of nodes in the network, and clustering (clustering coefficient) reports the fraction of a node’s nearest neighbours that are also neighbours to each other [343].

Data 1 outside the studied ROI PREDICTION

52