A cross-sectional case–control study on the structural connectome in recovered hospitalized COVID-19 patients

COVID-19 can induce neurological sequelae, negatively affecting the quality of life. Unravelling this illness's impact on structural brain connectivity, white-matter microstructure (WMM), and cognitive performance may help elucidate its implications. This cross-sectional study aimed to investigate differences in these factors between former hospitalised COVID-19 patients (COV) and healthy controls. Group differences in structural brain connectivity were explored using Welch-two sample t-tests and two-sample Mann–Whitney U tests. Multivariate linear models were constructed (one per region) to examine fixel-based group differences. Differences in cognitive performance between groups were investigated using Wilcoxon Rank Sum tests. Possible effects of bundle-specific FD measures on cognitive performance were explored using a two-group path model. No differences in whole-brain structural organisation were found. Bundle-specific metrics showed reduced fiber density (p = 0.012, Hedges’ g = 0.884) and fiber density cross-section (p = 0.007, Hedges’ g = 0.945) in the motor segment of the corpus callosum in COV compared to healthy controls. Cognitive performance on the motor praxis and digit symbol substitution tests was worse in COV than healthy controls (p < 0.001, r = 0.688; p = 0.013, r = 422, respectively). Associations between the cognitive performance and bundle-specific FD measures differed significantly between groups. WMM and cognitive performance differences were observed between COV and healthy controls.


Participants' recruitment and consent
Twenty patients with clinical signs of COVID-19 pneumonia were included in this study.Patients were eligible to participate if they had a positive reverse transcriptase-polymerase chain reaction test (RT-PCR) and were hospitalised at the UZ Brussel.A radiology resident performed patient recruitment in collaboration with the department of infectious diseases.The Intensive Care Unit and Infectiology Department created a list of all hospitalised COVID-19 patients at the UZ Brussel to facilitate the recruitment.An infectious disease specialist contacted the patients on the list by phone or asked about the patient's willingness to participate during a follow-up consultation.Patients who expressed their willingness to participate were contacted by the radiology resident to further explain the study protocol.In addition to the cohort of patients with COVID-19, a control group of 18 individuals (i.e.healthy controls) was included and was age-matched to the patient cohort.Individuals who never had a symptomatic COVID-19 infection were eligible for participation.The recruitment of these individuals was conducted by means of convenience sampling and through the network of the co-authors.All participants provided their written informed consent after being informed both verbally and in writing regarding the study protocol.

Protocol
All tests and measurements were conducted at the department of Radiology-Magnetic Resonance (UZ Brussel).Participants underwent a Magnetic Resonance Imaging (MRI) brain scan and cognitive test battery.The patients with COVID-19 underwent the MRI and cognitive test battery upon hospital discharge.The MRI brain scan was acquired in a supine position using a 3 T MRI Ingenia scanner with a 32-channel head coil (Philips Medical Systems, Best, The Netherlands).The protocol for both cohorts consisted, among others, of an 3D-T1 weighted spin-echo images, dMRI (48 directions at a b-value of 3000s/mm 2 ).Characteristics of the different brain imaging techniques applied are available in the paper by Tassignon et al. (2023)  18 .

White matter microstructure-whole-brain fixel based analysis
Single-shell two tissue FBA followed the steps described by Raffelt et al. 2017 31 , and detailed in the MRtrix3 user guide 39 .Briefly, fiber orientation distribution (FOD) maps were calculated for each subject based on group averaged response functions for anisotropic tissue (white matter) and isotropic tissue (grey matter and cerebrospinal fluid).Subject FOD maps were then normalized over the two tissue types and used to generate population averaged (study-specific template) FOD images.Fixels were generated from the template FOD maps and for each individual subject's FOD maps after warping to template space.Individual subject fixels were assigned to the template fixels, then the CSD-derived metrics FD, FC, and FDC were calculated for each subject.A whole brain tractogram with 20 million streamlines was generated from the template FOD maps and filtered with SIFT to 2 million streamlines.This tractogram was used to define a fixel-to-fixel connectivity matrix, and to define a fixel mask with at least 150 streamlines per fixel.All fixel maps were smoothed based on a sparse fixel-to-fixel connectivity matrix.Finally, statistical analysis compared FD, FC (log scaled), and FDC measures between COVID-19 patients and healthy controls with fixelcfestats 30 .Fixelcfestats controls for familywise error rate using permutation testing to control for multiple testing 30 .Total intracranial volumes calculated by FreeSurfer were used as regressors of noninterest while comparing FC and FDC measures between the two groups.

White matter microstructure-bundle-specific CSD-derived metrics
Bundle specific tractography was carried out using the KUL_FWT pipeline 40 in native subject space, and resulting tractograms were warped to the fixel template space, to sample the individual subject mean FD, FC, and FDC within the traversed fixels.

Statistical analysis
All statistical analyses were performed using R (version 4.1.2;R Core Team, 2022) 42 .A p-value below 0.05 was considered statistically significant.To investigate the first aim, the difference between COVID-19 patients and healthy controls in structural brain connectivity was assessed by means of the Welch-two sample t-test or, when the assumptions were not met, by the non-parametric Wilcoxon rank sum exact test.In particular, group differences in terms of the clustering coefficient and global and local efficiency were inspected parametrically.The characteristic path length was examined non-parametrically because of non-normality (healthy controls: Shapiro-Wilk W = 0.864, p = 0.018; COVID-19 patients: Shapiro-Wilk W = 0.741, p < 0.001).No corrections for multiple testing were applied.
For the second aim, whole-brain FBA was conducted to examine differences between COVID-19 patients and healthy controls by means of whole-brain FD, FC and FDC (Fig. 2).Subsequently, tract-specific FBA was performed to test the differences in motor-related tracts between groups terms of mean FD, FC and FDC, respectively.Hence, three multivariate analysis of variance (MANOVA) models were constructed per bundle to avoid multiple testing, using the lme4 43 and lmerTest 44 packages in R. All assumptions were checked and fulfilled.First, multivariate results were examined, after which univariate parameter estimates were inspected.Hedge's g were calculated as it is a standardized mean difference measure that adjusts for potential bias in the estimation of the effect size due to small sample sizes.For the third aim, the assumptions for performing a MANOVA were checked but not fulfilled (i.e., violation of both normality and homoscedasticy).Therefore, Wilcoxon rank sum exact tests were used to analyse group differences in cognitive performance on the motor praxis, digit symbol substitution and psychomotor vigilance tests.Wilcoxon test effect sizes (r values) were calculated and are provided in the results section.No corrections for multiple testing were applied.
For the fourth aim, two-group structural equation modelling was used to unravel the associations between cognitive performance and white matter microstructure and whether these associations differ between the healthy controls and the COVID patients.Specifically, a two-group path model, grouping on participant type (healthy controls, COVID-19 patients), was constructed using the lavaan package (version 0.06-11) in R 45 .Multicollinearity between the different variables was evaluated.The correlations between the left whole pyramidal tract and left corticospinal tract (r = 0.94) and between the corresponding right measures (r = 0.94), were almost perfect, signifying redundant variables.Consequently, the left and right corticospinal tracts were dropped from the model.The robust maximum likelihood estimator was used.A stepwise model building approach was adopted.In the first model, all hypothesized paths were modelled, and paths with p-values > 0.10 in both groups were removed, resulting in a good fitting model.We manually defined parameters to test whether the parameter estimates differ significantly between both groups using Wald tests.Before interpreting the parameter estimates, we evaluated how well the proposed model fits the data using the following cut-offs.First, if the Chi-square test is non-significant, the model fit is considered acceptable as the observed covariance matric is deemed similar to the model implied covariance matric.It is advised that the Comparative Fit Index exceeds 0.90 or, preferably, 0.95 46 .For the Tucker Lewis Index, a value between 0.90 and 0.95 is considered as a marginal fit, and values exceeding 0.95 represent a good fit 47 .Concerning the Root Mean-Square Error of Approximation a value below 0.04 describes a good fit and below 0.08 a moderate fit 48 .Values of the Standardized Root Mean Square Residual exceeding 0.10 are indicative of a poor fit 48 .Statistical analyses and graphical representations were made using several R-packages 45,[48][49][50][51][52][53][54] .

Participants
Table 1 provides the participants characteristics.A total of 20 COVID-19 patients and 18 healthy controls were included in this study.Within the group of COVID-19 patients, one male patient dropped out due to personal reasons and one male participant did not undergo the MRI brain scan due to claustrophobia.Within the group of healthy controls, one female participant refused to undergo an MRI brain scan due to claustrophobia.Accordingly, the data of 18 COVID-19 patients and 17 healthy controls were analysed.There were no baseline agedifferences between groups (p-value = 0.146).

Aim 1: structural brain connectivity
Figure 1 shows boxplots for the normalized whole-brain structural graph metrics of the COVID-19 patients and healthy controls for each graph theory measure separately.No significant differences between COVID-19 patients and healthy controls are detected (all p > 0.24), as can be seen in Table 2. Whole-brain FBA Figure 2 visualises the differences in whole-brain FD, FC and FDC between healthy controls and COVID-19 patients.A threshold of α = 0.10 was used as results with a significance value of 0.10 > p (Free Water Elimination) < 0.05 were considered borderline significant.FD showed minor differences in the corpus callosum and subcortical white matter of the left precentral gyrus (medially), FC showed minimal fixel differences in the left cerebral peduncle, and FDC showed more prominent differences in the left pyramidal tract, corpus callosum (parietal segment) and subcortical white matter (precentral gyrus).

Bundle-specific CSD-derived metrics
Figure 3 visualises FD of the motor segment of the corpus callosum, parietal segment of the corpus callosum, premotor and supplementary motor segment of the corpus callosum, bilateral medial lemnisci, bilateral whole pyramidal tracts and corticospinal tracts in COVID-19 patients and healthy controls.A significant difference between COVID-19 patients and healthy controls was detected for the motor segment of the corpus callosum (Table 3).The multivariate ANOVA model showed a significant association between group and the corpus callosum (F(3,31) = 0.771, p = 0.042).Univariate inspection revealed only a significant association between group membership and the motor segment of the corpus callosum with a large effect size ( β group = −0.043,p = 0.012, Hedges' s g = 0.884).No other statistically significant differences were found between COVID-19 patients and healthy controls for the other regions.
Figure 4 shows the boxplots of FC of the motor segment of the corpus callosum, parietal segment of the corpus callosum, premotor and supplementary motor segment of the corpus callosum, bilateral medial lemnisci, bilateral whole pyramidal tracts and corticospinal tracts in COVID-19 patients and healthy controls.No clear differences were visually observed between both groups.
Yet, the statistical modelling showed a marginal significant effect of FC in the left corticospinal tract in COVID-19 patients compared to healthy controls ( β group = 2.929, p = 0.096, Hedges' s g = 0.566).No other www.nature.com/scientificreports/statistical differences were found between COVID-19 patients and healthy controls.Results of multivariate models per brain region of interest of the differences in fiber cross-section between groups are provided in Table 4. Figure 5 visualises FDC of the motor segment of the corpus callosum, parietal segment of the corpus callosum, premotor and supplementary motor segment of the corpus callosum, bilateral medial lemnisci, bilateral whole pyramidal tracts and corticospinal tracts in COVID-19 patients and healthy controls.A clear difference between COVID-19 patients and healthy controls was detected for the motor segment of the corpus callosum.This finding was confirmed by statistical testing, as can be seen in Table 5. Modelling showed a lower FDC in the motor segment of the corpus callosum in COVID-19 patients compared to healthy controls ( β group = −0.081,p = 0.007, Hedges' s g = 0.945).No other statistical differences were found between COVID-19 patients and healthy controls.www.nature.com/scientificreports/Wilcoxon rank sum exact tests were subsequently used to statistically assess the difference in cognitive performance between COVID-19 patients and healthy controls.Median reaction times on the motor praxis test and digit symbol substitution test were significantly worse in COVID-19 patients compared to healthy controls (W = 28, p < 0.001, r = 0.688; W = 73, p = 0.013, r = 0.422, respectively).Despite the small sample size, medium to large effect sizes were found on the difference in cognitive performance for the motor praxis test and digit symbol substation test.Performance on the psychomotor vigilance test did not differ between COVID-19 patients and healthy controls (W = 136, p = 1, r = 0).

Aim 4: associations between structural brain MRI measures and cognitive performance
Table 6 provides an overview of the two-group path model, grouping on patient type.This model looks at the associations between cognitive performance and the FD of the motor segment of the corpus callosum, parietal segment of the corpus callosum, premotor and supplementary motor segment of the corpus callosum, bilateral medial lemnisci, corticospinal tracts and whole pyramidal tracts.The outcome variables were the test score on the motor praxis test, digit symbol substitution test and psychomotor vigilance test, while the predictor variables comprised the aforementioned fiber bundles.The constructed model converged normally after 345 iterations and fits the data well according to the robust fit indices ( χ 2 8 = 1.029, p = 0.490; CFI = 1.00,TLI = 1.00,RMSEA = 0.000, SRMR = 0.033).New parameters were constructed.Here, a positive estimated difference denotes that the estimate of the COVID-19 patients was lower than that of the healthy controls.A negative estimated difference implies that the estimate of the COVID-19 patients score was higher than that of the healthy controls.
A visualisation of the associations between cognitive performance and FD is provided in Figs.7 and 8, respectively for COVID-19 patients and controls.

Discussion
This study explored the differences in structural whole-brain organisation, local white matter brain microstructure and cognitive performance between COVID-19 patients and healthy controls.Furthermore, we investigated whether differences in cognitive performance are associated with white-matter brain microstructure for www.nature.com/scientificreports/COVID-19 patients and healthy controls.No differences in the structural whole-brain organisation were found between COVID-19 patients and healthy controls.Whole-brain FBA showed marginally significant differences in the corpus callosum and the subcortical white matter of the medial left precentral gyrus.Whole-brain FC showed marginally significant differences in the left cerebral peduncle and whole-brain FDC in the left pyramidal tract, the parietal segment of the corpus callosum, and the precentral gyrus of the subcortical white matter.On a tract-specific level, we found a reduction of FD and FDC in the motor segment of the corpus callosum and a marginal reduction in FC in the left corticospinal tract among COVID-19 patients indicating differences in intra-axonal volume (e.g.axonal loss) and macroscopic cross-sectional axonal size compared to healthy controls, respectively 31 .The motor segment of the corpus callosum is the main interhemispheric connection between the primary motor cortices, whereas the corticospinal tract transmits motor-related impulses from the cerebral cortex to the spinal tract 40 .Consequently, the differences in FD, FC and FDC could indicate a changed ability to relay information impacting cognitive performance 40 .Regarding cognitive performance, we found that performance on the motor praxis and digit symbol substitution tests was worse in COVID-19 patients than the healthy controls, indicating reduced sensory-motor speed and problems with complex tracking and visual scanning 41 .
The associations between the microstructural brain structure (i.e.fiber density) and performance on the cognitive tasks were examined using structural equation modelling.Significant differences in associations were found between the performance on the motor praxis test, and the motor segment of the corpus callosum, the left and right whole-pyramidal tracts and the right medial lemniscus in COVID-19 patients compared to the group of healthy controls (Table 6).The associations between the digit symbol substitution test and the left wholepyramidal tract, and between the psychomotor vigilance test and parietal segment of the corpus callosum, the right medial lemniscus and the left and right whole pyramidal tracts also significantly differed between COVID-19 patients and the healthy controls (Table 6).These results suggest that COVID-19 may induce structural white-matter brain changes that are likely to induce cognitive problems.
Bispo et al. found a reduction of fibre density in recovered COVID-19 patients compared to healthy controls in the posterior genu and rostral body of the corpus callosum, the arcuate fasciculus, the cingulum, the fornix, the inferior frontal-occipital fasciculus, the inferior and superior longitudinal fasciculus, the uncinate fasciculus, the corona radiata, and corticospinal tracts 33 .These reductions in fibre density were also correlated with a worse outcome for COVID-19 patients on reaction time and visual memory tests 33 .These results are consistent with our results.However, caution is needed when comparing our results to those of Bispo et al.Their results might be less specific to the white matter microstructure because their acquisition protocol constituted only 32 directions www.nature.com/scientificreports/with a b-value of 800 s/mm 233 .This is below the advised b-value > 2000s/mm2 for CSD, resulting in lower angular resolution and less specificity for restricted diffusion within the axons 55 .The population difference between the two studies must also be pointed out.Bispo et al. (2022) mainly included non-hospitalised individuals tested three months after their COVID-19 recovery, while the present study included hospitalised patients tested one month after recovery 33 .This prompts the question about the influence of hospitalisation length and related disease severity on the effect of COVID-19 on brain structure and cognitive performance 56 .Consequently, more research is needed to determine whether experiencing a COVID-19 infection affects the brain structure and cognitive performance while accounting for hospitalisation-related factors such as the degree of hospitalisation (e.g.residing on an intensive care unit, use of artificial breathing machines) and hospitalisation length.A strength of this study is that both groups were homogeneous regarding vaccination status, as no vaccination was possible at the time of the data collection, thus limiting the selection and information bias.Hence, vaccination status has no confounding effect [57][58][59][60][61][62][63] .Future COVID-19 studies should account for the influence of vaccination status as a possible confounding factor.A study limitation is our small sample size.Therefore, we attached more significance to the visualisations (i.e.data distributions) and effect sizes than to the p-values.In our analyses, we also neutralised this limitation using statistical tests and models designed to control for a small sample size.Additionally, the small sample size limits performing a covariates analysis to correct for possible baseline differences, despite the study groups being matched by age.Future research should implement such analysis to acquire a more comprehensive understanding of the groups.Furthermore, caution is needed when interpreting the results of the white-matter microstructure.Due to the limited sample size, we could not control for brain-size scaling effects.Therefore, we decided not to include the FC and FDC in our last hypothesis (i.e.associations), because of to their intrinsic relation to intracranial volume 30,31 .Inclusion of these measures in future studies might shed a broader light on the effects of COVID-19 on white-matter macro-structure.A last limitation comprises our cross-sectional study design.This design limits us from examining the differences between COVID-19 patients and healthy controls at one given point in time.Future research should include the longitudinal follow-up of brain structure and cognitive performance in COVID-19 patients compared to healthy controls to unravel interactions between brain structure and cognitive performance over time 64 .

Figure 6
Figure6shows the boxplots of the cognitive performance, expressed as reaction time, on the motor praxis, digit symbol substitution and psychomotor vigilance tests in COVID-19 patients and healthy controls.This visualisation indicate a difference between groups on the Motor Praxis test and Digit Symbol substitution test.

Figure 2 .
Figure 2. Whole-brain fixel-based analysis results for the Healthy controls > COVID-19 patients contrast showed fixel differences at p (Free Water Elimination) < 0.1 significance threshold.R right, L left.

Figure 3 .
Figure 3. Boxplots of fiber density of the motor segment of the corpus callosum, parietal segment of the corpus callosum, premotor and supplementary motor segment of the corpus callosum, bilateral medial lemnisci, bilateral whole pyramidal tracts and corticospinal tracts in COVID-19 patients (COV) and healthy controls (HC).

Figure 4 .
Figure 4. Boxplots for fiber density of the motor segment of the corpus callosum, parietal segment of the corpus callosum, premotor and supplementary motor segment of the corpus callosum, bilateral medial lemnisci, bilateral whole pyramidal tracts and corticospinal tracts in COVID-19 patients (COV) and healthy controls (HC).

Figure 5 .
Figure 5. Fiber density cross-section of the motor segment of the corpus callosum, parietal segment of the corpus callosum, premotor and supplementary motor segment of the corpus callosum, bilateral medial lemnisci, bilateral whole pyramidal tracts and corticospinal tracts in COVID-19 patients (COV) and healthy controls (HC).Each facet presents the median and the IQR (box), Q1 − 1.5* IQR and Q3 + 1.5*IQR (whiskers), and individual observations (dots).

Figure 8 .
Figure 8. Associations in healthy controls between the cognitive performance on the motor praxis test (mpt), digit symbol substitution test (dss) and psychomotor vigilance test (pvt) and the fiber density of the motor segment of the corpus callosum (mot cc), parietal segment of the corpus callosum (par cc), premotor and supplementary motor segment of the corpus callosum (p&s cc), left and right medial lemnisci (med L & med R, respectively) and left and right whole pyramidal tracts (pyr L & pyr R, respectively).Positive estimates are indicated with green arrows, negative estimates with red arrows.The stronger the effect, the thicker the line of the arrows will be.Estimates are standardized.Single headed arrows are regressions, double headed arrows are covariances.
41gnitive performanceThe computerized cognitive test battery "Cognition" (Joggle® Research, Seattle, WA, USA) was conducted using an iPad.The test battery has an average duration of approximately 18 min, is sensitive to multiple domains at high-level cognitive performance and has been proven to engage specific brain regions, evidenced by functional neuroimaging41.The test battery is compiled out of 7 tests, including the motor praxis test, visual object learning test, abstract matching, line orientation test, digit symbol substitution test, balloon analogue risk test, NBACK and psychomotor vigilance test.These tests measure sensorimotor speed, spatial learning and memory, abstraction, spatial orientation, complex scanning and visual tracking, risk decision-making, working memory, and vigilant attention.Participants practised each cognitive test once to mitigate learning effects.For this study the primary outcome of interest was the median reaction time on the motor praxis, digit symbol substitution and psychomotor vigilance tests.

Table 1 .
Patient characteristics.NA not applicable at the time of the experimental session.

Table 2 .
Statistical analysis of differences in graph theory measures between COVID-19 patients (n = 18) and healthy controls (n = 17).HC healthy controls, COV COVID-19 patients, SD standard deviation, NA not applicable due to non-normality.*Wilcoxon rank sum exact test used due to outliers and non-normality.+ Unless stated otherwise.++ Wilcoxon W-value.

Table 3 .
Results of analysis of variance models per brain region of interest of the differences in fiber density between COVID-19 patients (n = 18) and healthy controls (n = 17).Group = COVID-19 patients, SMA supplementary motor area, SE standard error, CI confidence interval.Significant values are in bold.

Table 4 .
Results of multivariate models per brain region of interest of the differences in fiber crosssection between COVID-19 patients (n = 18) and healthy controls (n = 17).Group = COVID-19 patients, SMA supplementary motor area, CI confidence interval, NA not applicable, SE standard error.*Univariate test used instead of multivariate test due to assumptions that were not fulfilled.+ F-value.++ Mean squared error.

Table 6 .
Results of the two-group path model on the associations between cognitive performance and fiber density for COVID-19 patients (n = 18) and healthy controls (n = 17).mpt motor praxis test, dss digit symbol substitution test, pvt psychomotor vigilance test, mot cc motor segment of the corpus callosum, par cc parietal segment of the corpus callosum, p&s cc premotor and supplementary motor segment of the corpus callosum, med L left medial lemniscus, med R right medial lemniscus, pyr L left whole pyramidal tract, pyr R right whole pyramidal tract, SE standard error.Significant values are in bold.