Assessing functional connectivity differences and work-related fatigue in surviving COVID-negative patients

The Coronavirus Disease 2019 (COVID-19) has affected all aspects of life around the world. Neuroimaging evidence suggests the novel coronavirus can attack the central nervous system (CNS), causing cerebro-vascular abnormalities in the brain. This can lead to focal changes in cerebral blood flow and metabolic oxygen consumption rate in the brain. However, the extent and spatial locations of brain alterations in COVID-19 survivors are largely unknown. In this study, we have assessed brain functional connectivity (FC) using resting-state functional MRI (RS-fMRI) in 38 (25 males) COVID patients two weeks after hospital discharge, when PCR negative and 31 (24 males) healthy subjects. FC was estimated using independent component analysis (ICA) and dual regression. When compared to the healthy group, the COVID group demonstrated significantly enhanced FC in the basal ganglia and precuneus networks (family wise error (fwe) corrected, pfwe < 0.05), while, on the other hand, reduced FC in the language network (pfwe < 0.05). The COVID group also experienced higher fatigue levels during work, compared to the healthy group (p < 0.001). Moreover, within the precuneus network, we noticed a significant negative correlation between FC and fatigue scores across groups (Spearman’s ρ = −0.47, p = 0.001, r2 = 0.22). Interestingly, this relationship was found to be significantly stronger among COVID survivors within the left parietal lobe, which is known to be structurally and functionally associated with fatigue in other neurological disorders.

155 fatigue in patients with Chronic Fatigue Syndrome (CFS) and Fibromyalgia (Natelson, 156 2019). To avoid confounding effects from comorbidities, we recruited subjects that were, 157 otherwise, in excellent health conditions prior to hospitalization for COVID-19. For 158 example, 16/47 (34.04%) of our patients were young adults who had no prior record of 159 any comorbidities that could confound the COVID-19 effects. We only had 7/47 (14.89%) 160 subjects with age between 50-54 years (capped at < 55 years as recruitment criteria to 161 avoid aging-related comorbidities), some of whom had reported to have marginal 162 diabetes. The rest (51.07%) in between also did not have any record of comorbidities in 163 the hospital report.
164 Table 1 summarizes the clinical information from the 47 patients included in the current 165 study. Of these 47, 36.17% (17/47) patients were reported to be 'mild', 8.51% (4/47) to 166 be 'moderate' and 36.17% (17/47) to be 'severe'. Information from the rest of the 19.15% 167 (9/47) was not provided from the hospital because those patients did not give consent to 168 sharing their medical symptoms. Please note, we present percentages as a ratio of 169 affected patients with both the available sample with information ('% Out of Avail.' in Table   170 1) and the total sample of patients including those patients who did not give consent to 171 publicly share their clinical data ('% Out of Total' in Table 1). The percentages of 172 'moderate-severe' patients who were administered medications, e.g., Remdesivir, 173 Dexamethasone, Ceftriaxone, Clexane and other antibiotic regimes are provided in Table   174 1. On average these 47 patients stayed in the hospital for approximately 11 ± 3.30 [SD] 175 days.
176 Table 2 summarizes the participant demographics based on age, sex and fatigue. The 177 average age in the HC group was 33.50 ± 9.74 [SD] years and that in the COVID group . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. ; https://doi.org/10.1101/2022.02.01.478677 doi: bioRxiv preprint 222 (FMRIB Analysis Group, Oxford, UK) and AFNI (http://afni.nimh.nih.gov/afni) (Cox, 1996) 223 for housekeeping, visual inspection and quality control purposes. At the beginning, first 224 five time points were excluded from each subject to account for magnetic stabilization.
225 The functional images were motion corrected for head movement using a least squared 226 approach and 6 parameters (rigid body) spatial transformation with respect to the mean 227 image of the scan. The subjects with excessive head motion were identified using 228 framewise displacement (FWD) (Power et al., 2012). Additionally, time frames with high 229 FWD crossing a threshold of 0.5 mm (Power et al., 2012) were identified along with the 230 previous and the next two frames and added as regressors (Yan et al., 2016) during 231 temporal regression of nuisance signals. If more than 50% of the time series data were 232 affected due to regression of high motion frames the participant was removed from the 233 analysis. Moreover, any participant with the maximum framewise translation or rotation 234 exceeding 2 mm was removed from further analysis. Anatomical image from each subject 235 was coregistered to the mean functional image obtained from the motion correction step.
236 T1-weighted image from each subject was segmented into gray matter (GM), white matter 237 (WM), and cerebrospinal fluid (CSF) tissue probability maps and an average template 238 including all participants was generated using DARTEL (Ashburner, 2007). The subject 239 specific tissue maps were non-linearly warped to this template and spatially normalized 240 to the MNI space. These affine transformations were applied to the functional images to 241 normalize all volumes to the MNI space and resampled to isotropic voxel size of 3 mm x 242 3 mm x 3 mm. Time series, from brain compartments with high physiological noise signals 243 such as, CSF and WM was extracted by thresholding the probability maps from the 244 segmentation stage above the 99 th percentile, and first 5 principial components were . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. ; https://doi.org/10.1101/2022.02.01.478677 doi: bioRxiv preprint 245 obtained using a COMPCOR based (Behzadi et al., 2007) principal component analysis 246 (PCA) from both tissues. These 10 components along with Friston's 24-parameter model 247 (6 head motion parameters + 6 previous time point motion parameters + 12 corresponding 248 quadratic parameters) (Friston et al., 1996) and time frames with high FWD (> 0.5 mm) 249 were added as regressors in a multiple linear regression model to remove unwanted 250 signals voxel-wise. The residuals from the regression step were then bandpass filtered 251 between 0.01 to 0.1 Hz and finally, spatial smoothing was performed using a Gaussian 252 kernel of 6 mm full width at half maximum (FWHM).

Head Motion Assessment
254 We performed in-scanner head movement assessment using mean Framewise 255 Displacement (FWD) based on the methods depicted in (Power et al., 2012). A two-tailed 256 two-sample student's t-test revealed no significant differences in mean FWD between the 257 two groups (t = -1.57, p = 0.12, α = 0.05).

ICA and Dual Regression
259 Group level resting state networks were obtained by applying the 'gica' option of the 260 'melodic' module from FSL toolbox (FMRIB Analysis Group, Oxford, UK). All subjects' 4D 261 functional images after pre-processing were temporally concatenated into a 2D matrix of 262 'space' x 'time' as delineated in (C.F. ) and 25 spatial maps were 263 obtained. Resting State Networks (RSNs) were identified by matching ICs with the 1000 264 functional connectome project maps (Biswal et al., 2010) using Dice's coefficient and 265 spatial correlations obtained from AFNI's '3dMatch' program (Taylor & Saad, 2013).
266 Further visual inspection was performed to make sure all network regions aligned with . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. ;

Statistical Analysis
274 To investigate differences in participant demographics, we performed a two-sample t-test 275 on age. Since sex is a categorical variable, we performed a chi-squared test to identify 276 any sex related differences between the groups. Since the fatigue scores deviated from 277 normality (Shapiro-Wilk, p < 0.05), we performed a non-parametric Wilcoxon-Ranksum 278 test on the fatigue scores to identify group level differences in fatigue scores.
279 To investigate FC differences between COVID and HC groups, we performed an unpaired 280 two sample t-test between standardized subject-specific RSN maps from the two groups.
281 To account for confounding effects that may explain some of the variance in the data, 282 age, sex and a regressor to account for two different scanning lengths were also added 283 as covariates of no interest. Cluster-based thresholding was applied at a height threshold 284 of p unc < 0.01, with family wise error (FWE) correction at p FWE < 0.05 for multiple 285 comparisons. The cluster extent threshold (k E ) obtained from this step was used to 286 threshold and generate corrected statistical maps for the contrasts with significant effects.
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. ; https://doi.org/10.1101/2022.02.01.478677 doi: bioRxiv preprint 287 We further wanted to evaluate if the PRN demonstrates correlation with self-reported 288 fatigue among the COVID individuals. We incorporated a multiple linear regression 289 approach where the FC at each voxel was the response variable (Y), and the self-reported 290 fatigue score was the explanatory variable (X). We also added age and sex as covariates 291 of no interest. Significant clusters were obtained in the same manner as described earlier 292 at the end of the previous paragraph for group level differences in FC. For visual 293 representation of the significant relationship between the two variables, the average FC 294 within the significant cluster was obtained from each subject. These average FC values 295 were then linearly regressed against the fatigue scores and visualized within a scatter 296 plot and a line of best fit with 95% confidence interval. Age and sex were regressed out 297 during the linear regression step. The correlation analysis and the graphical plotting was 298 done using 'inhouse' scripts prepared in RStudio (RStudio, 2021).

301
We will present results on participant demographics first and then group level voxel-wise 302 results will be reported. There was no significant differences in age between the two 303 groups (p > 0.05). A chi-squared test on sex revealed no significant effects were observed 304 between the two groups (p > 0.05). The Wilcoxon-Ranksum test revealed significantly 305 higher fatigue levels in the COVID group compared to the HC group (T = 1093, p = 2.86e-306 07). 307 We identified twenty-two large-scale resting state networks (RSNs) (see Figure 1) from 308 the group ICA analysis. Group level statistical analysis was run for five networks of . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. ; 309 interest using standardized subject specific RSN maps obtained from the dual regression 310 step. Significant differences in FC was observed between the COVID and HC groups in 311 particularly three out of the five networks -the basal ganglia (BGN), precuneus (PRN) 312 and language (LANG) networks. Figure 2 shows all significant clusters from obtained from 313 the t-test and the corresponding group level networks where these alterations occur.  Table 3. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. The results from this study support our hypothesis that COVID survivors demonstrate 342 altered FC when compared to HCs, even two weeks after discharge from the hospital. 389 Moreover, these brain regions are also known to be involved in attention processing, 390 therefore, enhanced FC in these regions may indicate possible compensatory 391 mechanisms of attention related symptoms that recovering patients may experience. This 392 is significant, because about 37% of the COVID survivors in our study reported lack of 393 attention and all these 37% of the participants also reported a work-related fatigue score 394 of 2 or higher on a scale of 5. Therefore, further investigations are necessary to 395 understand these processes better, especially, from a clinical perspective.   Table 4 for 442 cluster information). Interestingly, our recent investigation using the same group of 443 survivors revealed stronger positive correlation between GMV and self-reported fatigue 444 within the precuneus and SPL regions, when compared to HCs (Hafiz et al., 2022). This . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023.  (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. ;

491
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. The peak t-score of the cluster was, t peak = 4.40, and corrected for multiple comparisons by controlling false discovery rates, at p fdr < 0.05. The graph on the right shows the linear relationship between FC within the significant cluster and self-reported fatigue scores across all groups. The x-axis represents the residuals plus the average FC (z-scores) across groups from the cluster and the y-axis represents the fatigue scores. The light pink dots represent the COVID group, and the cyan dots represent the HC group. The shaded gray area represents the 95% confidence interval. The blue line represents the least squares regression line of best fit.
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. ; https://doi.org/10.1101/2022.02.01.478677 doi: bioRxiv preprint Table 1. Clinical information detailed by symptom severity and medical treatment of COVID participants. The table shows the number and percentages of participants based on two clinical categories. The first three rows show number and percentage of participants by symptom severity and the last five rows are based on the type of medication administered and the requirement of O 2 support. Particularly, information from 9 participants could not be obtained as these participants did not agree to share the symptom publicly/anonymously. Keys: No. Aff. = number of affected patients, No Info. = no information available because these patients did not give consent to share symptom information, Data Avail. = number of subjects with clinical assessment data available, No. Tot. = total number of patients including those with no information available, % Out of Avail. = proportion of patients affected vs patients with clinical assessment data available (n = 38) in percentages, % Out of Total = proportion of patients affected vs. total number of patients (n = 47) in percentages, O 2 = oxygen supplied to support breathing, BiPap = bilevel positive air pressure, Mix Medications* = a combination of medications -Dexamethasone, Ceftriaxone and Clexane injections . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023.  . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. ; https://doi.org/10.1101/2022.02.01.478677 doi: bioRxiv preprint Table 3. List of spatial regions from significant clusters obtained from the contrast -COVID > HC. The regions from three RSNs -BGN, PRN and LANG which demonstrated significant differences are presented with peak MNI coordinates (X Y Z) and corresponding peak t-score values for each cluster. Keys: ∆FC = Direction of change in Functional Connectivity; Cl. = Cluster; Cl. No. = Number of Clusters; Cl. Size = Cluster Size; Peak t = peak t-score; p FWE = family wise error corrected p-value.
. CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. . CC-BY 4.0 International license available under a (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted January 28, 2023. ; https://doi.org/10.1101/2022.02.01.478677 doi: bioRxiv preprint