No evidence for motor recovery-related cortical reorganization after stroke using resting-state fMRI

Cortical reorganization has been suggested as mechanism for recovery after stroke. It has been proposed that a form of cortical reorganization (changes in functional connectivity between brain areas) can be assessed with resting-state fMRI. Here we report the largest longitudinal data-set in terms of overall sessions in 19 patients with subcortical stroke and 11 controls. Patients were imaged up to 5 times over one year. We found no evidence for post-stroke cortical reorganization despite substantial behavioral recovery. These results could be construed as questioning the value of resting-state imaging. Here we argue instead that they are consistent with other emerging reasons to challenge the idea of motor recovery-related cortical reorganization post-stroke when conceived as changes in connectivity between cortical areas.


Introduction 39
Spontaneous neurological recovery occurs in almost all stroke patients within the first 40 months after the insult. While the underlying physiological changes that accompany spontaneous 41 motor recovery in humans remain largely unknown, data from animal models have been 42 interpreted as showing that cortical reorganization is a potential key mechanism mediating 43 recovery (Dancause & Nudo, 2011;Grefkes & Ward, 2014;Nudo, 2006). 44 In the literature, the term cortical reorganization has been loosely defined and used to 45 refer to any number of structural and physiological changes that follow injury. These changes 46 can span the micro-, meso-and macro-scale, including synaptogenesis, axonal sprouting, and 47 changes in cortical activation maps. We have argued elsewhere that the term functional 48 reorganization should be reserved for those changes, including new cortico-cortical connections, 49 that are causally related to or at least correlated with motor recovery (Krakauer & Carmichael, 50 2017). It should be added that reorganization has also been taken as a qualitative event, 51 exemplified by the idea that one cortical area "takes over" another, which implies a change in the 52 tuning of neurons, for example, when touching the face activates the hand area of sensory cortex 53 in amputees. We argue elsewhere that a qualitative change in cortical representation need not be 54 invoked to explain this result (Krakauer & Carmichael, 2017), but we will not use this definition 55 here. 56 Evidence for functional reorganization after stroke comes primarily from studies of 57 axonal sprouting. For example, Overman and colleagues (2012), in a mouse cortical stroke 58 model, generated sprouting of axonal connections within ipsilesional motor, premotor and 59 prefrontal areas by blocking of an axonal growth inhibitor (epinephrine A5). Similar results were 60 reported for the neuronal growth factor GDF10 (Li et al., 2015). Critically, however, in both 61 studies no direct test of the relevance of axonal sprouting for motor improvement was performed, 62 indeed not even a correlation with the degree of sprouting and behaviour was examined. In 63 addition, most studies describing axonal sprouting after stroke found that it was cortico-64 subcortical instead of cortico-cortical connectivity changes that were linked to motor recovery 65 (see e.g. Lee, 2004;Wahl et al., 2014). Other studies that argue for a role of cortico-cortical 66 connectivity changes underlying stroke recovery are limited by cross-sectional approaches or do 67 not report behavior at all (Dancause et al., 2005;Frost et al., 2003;Liu & Rouiller, 1999; 68 Napieralski et al., 1996). 69 Despite the weak evidence for behaviorally-relevant new cortical connections in animal 70 models post-stroke, these models have nevertheless led to widespread interest in identifying 71 similar processes of functional reorganization in the human brain. One prominent non-invasive 72 method is to measure inter-regional connectivity with resting-state fMRI (rs-fMRI; Biswal et al., 73 1995;Fox & Raichle, 2007). This method relies on correlations between time-series of fMRI 74 activity recorded while the subject is lying in the scanner without performing a task. Most often 75 these correlations are computed between a set of pre-defined regions of interest (ROIs). The 76 underlying assumption is that regions with connected neuronal processing show stronger 77 statistical dependency of their spontaneous neuronal fluctuations. These correlations are 78 commonly regarded as a measure of "functional connectivity", which has been closely linked to 79 structural connectivity (Friston, 2011;van der Heuvel et al., 2009). In the context of stroke 80 recovery, it has been suggested that reorganization can be detected as a change in such 81 correlations/functional connectivity patterns (van Meer et al., 2010). Specifically, for post-stroke 82 recovery of hemiparesis, the advantage of task-free resting-state over task-based fMRI is that it 83 avoids the performance confound (Krakauer, 2004(Krakauer, , 2007; the connectivity measures are not 84 biased by the inability of patients to match control performance due to motor impairment. 85 To date, results from rs-fMRI studies of functional connectivity changes after stroke have 86 been mixed. Although, rs-fMRI studies have frequently found changes in interhemispheric 87 connectivity patterns after stroke (Carter et al., 2010;Chen & Schlaug, 2013;Golestani et al., 88 2013), the direction of these changes and their correlations with behavior have been inconsistent. 89 One study found a positive correlation between motor function and increased functional 90 connectivity between the lesioned M1 and contralateral heterologous cortical areas (Park et al., 91 2011), another study reported that interhemispheric homologous connectivity was associated 92 with lower degrees of motor impairment but only for infratentorial strokes (Lee et al., 2018). Yet 93 another study showed that an increase in M1-M1 connectivity correlated negatively with motor 94 function (Wang et al., 2014). 95 There are many potential reasons for these inconsistencies in rs-fMRI findings. 96 If patients with cortical lesions are included in the study design, it is possible to confuse changes 97 in connectivity measures as a direct consequence of the lesion (e.g. the damaged area becomes 98 disconnected from the brain) with changes associated with true reorganization. Additionally, 99 most studies use different analysis protocols and measures to quantify changes in connectivity, 100 making integration of evidence across studies difficult. Third, the majority of currently available 101 studies have been cross-sectional but it is essential to evaluate changes in connectivity across the 102 time-course of recovery. 103 To address these issues, we here report the results of a longitudinal rs-FMRI study of 104 stroke recovery in patients with hemiparesis after subcortical stroke. Only patients with 105 subcortical lesion locations were included in this study so that any changes in cortical 106 connectivity could not be attributed to the presence of the lesion itself. We provide a detailed 107 The main goal of this study was to determine whether motor impairment recovery following 116 stroke was associated with systematic changes in cortical connectivity. Our two main questions 117 were: 1) Is there a mean difference in the connectivity pattern between five motor regions (S1, 118 M1, PMv, PMd, SMA) when comparing patients and age-matched controls at any time-point 119 during stroke recovery? 2) Is there a change in patients' connectivity patterns over time that is 120 related to motor impairment? 121 We analyzed data from 19 patients with subcortical stroke and 11 healthy controls. 122 Behavioral assessments and resting-state images were obtained at five different time-points over 123 one year. Each patient completed on average 4.5 ±0.7 sessions, with the overall experimental 124 data being 89.5% complete (see also supplementary material for inter-and intrahemispheric connections, Figure S2). An unbalanced 157 mixed-effects ANOVA (see Methods) showed that the intrasession reliability was not 158 significantly different between groups (χ 2 (1)=1.0782, p=0.2991) and showed no changes over 159 time (controls: χ 2 (4)=6.174, p=0.187; patients: χ 2 (4)=1. 922 We also confirmed that the connectivity pattern for controls reflected known anatomical 166 connectivity (Damoiseaux & Greicius, 2009). Within one hemisphere, the highest correlations 167 were found between S1-M1 (0.91 ±0.47, Fisher-Z transformed), while the weakest correlation 168 was found between M1-PMv (0.58 ±0.39). Between hemispheres, S1 right -S1 left demonstrated the 169 highest correlation (0.9 ±0.43), while M1 right -PmV left showed a weaker correlation (0.59 ±0.37). 170 For correlations between hemispheres, homologous ROIs (e.g. M1-M1 or S1-S1) showed higher 171 correlations of the BOLD time series compared to heterologous ROI-ROI connectivity weights 172 (e.g. M1 right -Pmv left or S1 left -Pmd right ) as expected from interhemispheric neural-recordings 173 If disruption of the cortical projections through subcortical stroke leads to an acute 179 reorganization of cortical circuits, one would expect that (on average) acute connectivity patterns 180 of patients and controls would be different. Connectivity patterns for patients and controls were 181 highly correlated in the early period after stroke (acute stage: R=0.69, p=0.0002; see connectivity 182 matrices in Figure 2A and also Figure S4). To statistically test for significant differences 183 between connectivity patterns, we used the Euclidian distance between the two groups' mean 184 patterns and compared it to a null-distribution obtained by a permutation test (Figure 2c) Even though the averaged connectivity patterns for patients and controls were 211 indistinguishable at the acute stage, the heterogeneity in lesion locations for different patients 212 might result in idiosyncratic shifts in connectivity patterns that in the whole group would be 213 reflected as higher variability in patterns. To measure this within-group variability, we calculated 214 the average Euclidian distance of each patient's pattern to the patient group mean pattern and did 215 likewise for controls. The average within-patient distance was 2.955, whereas the average 216 within-control distance was 2.813, resulting in a difference of 0.142 (Δvariability). We compared 217 this value to a null distribution of Δvariability generated with permutation testing. We found that 218 resting-state connectivity patterns of patients showed a higher idiosyncratic, non-systematic 219 variability compared to controls: The difference between the variability lay outside the 2.5% -220 97.5% confidence interval generated by permutation testing (CI 0.018-0.051, Figure 3). Note 221 that the confidence interval was not symmetric around zero, as the N for controls was smaller 222 than for patients. 223 The difference in variability for intrahemispheric lesioned and interhemispheric 224 connections was also higher for patients. For intrahemispheric non-lesioned connections, we 225 found higher variability in controls (intrahemispheric lesioned: Δvariability=0.091, CI 0.002-226 0.023; non-lesioned: Δvariability=-0.01, CI -0.003-0.015; interhemispheric: Δvariability=0.1, CI 227 0.008-0.05, Figure 3 Even though there were no systematic differences between connectivity patterns of 238 patients and controls at the acute stage, we might expect to find changes in patient connectivity 239 patterns over time as they recover from impairment. 240 We therefore quantified Euclidean distances between the average connectivity patterns at 241 the acute stage as reference versus all other weeks (Δweek). Surprisingly, patients showed no 242 increase in Euclidian distances between the acute stage and consecutive weeks (Figure 4 and 243  By examining Euclidian distances between the individual connectivity patterns to the 268 average connectivity pattern, we found a greater non-systematic variability in patients than in 269 controls at the acute stage. However, the idiosyncratic variability of patients themselves did not 270 change from the acute stage compared to the following time-points (Table 3). 271 272 In summary, we found no evidence for a mean difference of connectivity patterns between 276 patients within one year. More importantly, patients did not show any significant longitudinal 277 change in connectivity patterns either systematically or regarding their group variability. 278

Comparison between alternative metrics for M1-M1 connectivity 279
Above we looked at the entire connectivity pattern between five sensorimotor areas 280 within and across hemispheres and found no changes for patients either longitudinally or when 281 compared to controls. In contrast, some previous studies have focused on individual ROI-to-ROI 282 connections and have reported changes after stroke (Thiel & Vahdat, 2015). To test this finding, we investigated changes of interhemispheric M1-M1 connectivity 286 weights over time and between patients and controls in our data set. The analysis showed a 287 significant difference between patients and controls, with patients having a slightly lower 288 Our results also contrast with another published finding that used an alternative metric of 293 connectivity to assess changes in functional connectivity after stroke. Golestani  Similarly, our patients had a lower RelCon for SM1-SM1 compared to controls at all time-299 points. Using a mixed-model, we found a significant difference between the groups 300 (χ 2 (1)=5.2457, p=0.022). However, consistent with our results reported above, we did not find a 301 change over time for RelCon SM1-SM1 in neither controls (χ 2 (4)=2.8087, p=0.5903) nor in 302 patients (χ 2 (4)=8.2243, p=0.0837; Figure 5b). 303 304 Figure 5 Here we report, that there were no longitudinal changes in resting-state functional 313 connectivity (rsFC) between cortical motor areas despite substantial motor recovery over the 314 same time period in a cohort of patients with subcortical stroke. In addition, at no stage of 315 recovery were rsFC patterns different from healthy, age-matched controls. 316 Whenever results are negative, concerns will be raised about the power of the study (addressed 317 below) and the biological validity of the method in general. 318 There have been more than 500 rs-fMRI studies of brain connectivity (Buckner et al., 2013). behavioral, brain stimulation, and rs-fMRI data, they demonstrated that changes in performance 327 after training on a five-digit sequence task led to reliable changes in corticostriatal functional 328 connectivity. When motor memory formation after training was disrupted using rTMS, changes 329 in functional connectivity predicted the modification of memory recall on the next day. 330 Given such results, why were we not able to detect rsFC changes in the setting of stroke 331 recovery? Injury ostensibly triggers functional reorganization, which arguably should be a more 332 dramatic cause of connectivity change as it is associated with structural alterations, e.g. 333 sprouting, and not just learning-related changes in pre-existing connections. There are two 334 potential answers to this question, one is the possibility that the idea that changes in cortico-335 cortical connections promote motor recovery after stroke is ill-conceived, the second is that there 336 are methodological limitations to rs-fMRI. We shall discuss both of these concerns. 337 338 A large number of animal studies, in rodents and non-human primates, have described numerous 339 structural and physiological changes in cortical areas around and beyond the infarct core. These 340 changes have collectively been called reorganization, but in only a small subset of cases have 341 they been correlated with motor recovery, which suggests that most are likely just reactive 342 (Carmichael, 2016). We reasoned that as spontaneous biological recovery is similar for cortical The question must now be asked why it was ever conjectured that changes in connections 355 between cortical regions would enhance recovery from hemiparesis, which is caused by 356 interruption of descending pathways out of a particular region(s). One could rephrase this to ask 357 why would there be a "horizontal" solution to a "vertical" problem? This question is related to 358 the increasing awareness of the questionable relevance of cortical map changes to recovery 359 In the present study, we investigated longitudinal changes in functional connectivity after 413 subcortical stroke. Despite substantial recovery from motor impairment over one year, we found 414 no differences in functional connectivity between patients and controls, nor any changes over 415 time. Assuming that rs-fMRI is an adequate method to capture connectivity changes between 416 cortical regions after brain injury, the results presented here, provide reason to doubt that post-417 stroke cortical reorganization, conceived as changes in cortico-cortical connectivity, is the 418 relevant mechanism for promoting motor recovery after stroke. We suggest instead that it is 419 facilitation of residual cortical descending pathways that are likely to be more causally relevant. 420 It is perhaps time for the field to change its emphasis from changes in "horizontal" connections 421 to changes in "vertical" ones.  Table S1; more detailed information 440 about lesion distribution is shown in Figure S1. 441 Additionally, 11 healthy age-matched control participants (4 females; mean age 65 ±8 442 years; all right-handed), were tested at the same time-points. 443 The study was carried out in accordance with the Declaration of Helsinki and approved 444 by the respective local ethics committee of the participating recruiting centers of SMARTS Rs-fMRI has a relatively low signal-to-noise ratio. Non-neuronal processes, such as sensor noise, 478 head motion, cardiac phase, and breathing, account for a considerable part of the variance of the 479 raw signal (Birn, 2012). It has been argued that markers for the reliability of the sampled rs-480 fMRI data are missing and that the choice of preprocessing steps is often not justified (Bennett & 481 Miller, 2010; Zuo & Xing, 2014). We therefore conducted two different procedures for noise 482 reduction and then compared split-half reliability for the whole connectivity pattern in controls to 483 determine which steps provided higher reliability (see supplementary material). 484

Lesion definition 485
Lesion boundaries were defined as an intensity increase of ≥30% on DWI images, and in a 486 second step manually modified by a neuroradiologist and a neurologist using RoiEditor, see 487 We chose five motor areas (S1=primary somatosensory cortex, M1=primary motor cortex, 491 PMd=dorsal premotor cortex, PMv=ventral premotor cortex, SMA=supplementary motor area) 492 as regions of interest that have been widely accepted as being associated with motor function and 493 motor recovery (Miyai et al., 1999(Miyai et al., , 2002Rehme et al., 2012). Individual T1-images were used 494 to delineate pial-grey matter and grey matter-white matter boundaries using FreeSurfer software 495 (Dale et al., 1999). The cortical surfaces were aligned across participants based on the sulcal-496 depth and local curvature maps. Probabilistic cyto-architectonic maps (Fischl et al., 2008) 497 aligned to the group average surface were then used to define ROIs first on the individual 498 surface, and then back-projected into the subject-native space. 499 The ROIs were defined as follows, M1: surface nodes with the highest probability for 500 Brodmann area (BA) 4. To increase specificity for processes related to recovery of hand 501 function, this ROI was limited to 2cm above and below the hand-knob (Yousry, 1997). S1: nodes 502 in the hand-region in S1 were isolated using BA 3a, 3b, 1 and 2.2cm above and below the hand 503 knob. PMd: nodes with highest probability in BA6, above middle frontal sulcus, but on the 504 lateral surface of the hemisphere. PMv: nodes with the highest probability in BA6, above middle 505 frontal sulcus. SMA: nodes with the highest probability in BA6 on the medial surface of the 506 brain. This ROI therefore includes SMA and preSMA (Picard & Strick, 1996). 507 508

Functional connectivity analysis 509
For each ROI, the time series for all voxels within the ROI were extracted and averaged, 510 resulting in a single BOLD time-course vector for each of the 10 ROIs across the two 511 hemispheres (left-S1, left-M1, left-PMd, left-PMv, left-SMA, right-S1, right-M1, right-PMd, 512 right-PMv, right-SMA). Pairwise correlations between averaged BOLD time-course vectors for 513 the different ROIs were computed and Fisher-Z transformed to conform better to a normal 514 distribution, resulting in a 10×10 matrix of connectivity weights (Figure 2). The matrix thus 515 represents the connectivity weights between all possible ROIs for a patient: 10 intrahemispheric 516 ROI pairs, each within the lesioned and non-lesioned hemispheres, respectively, and 25 517 interhemispheric ROI pairs between the lesioned and non-lesioned hemispheres (overall 45 518 connectivity weights for all ROI pairs). For the rest of this manuscript, this vectorized, Fisher-Z 519 transformed correlation matrix will be referred to as the full connectivity pattern, while the 520 corresponding intra-and interhemispheric subsets of the matrix will be referred to as the 521 intrahemispheric non-lesioned (1×10 vector), intrahemispheric lesioned (1×10 vector), and 522 interhemispheric connectivity patterns respectively (1×25 vector). These connectivity patterns 523 were estimated independently for each session and patient. Connectivity patterns for controls 524 were estimated similarly, with the exception that intrahemispheric connectivity patterns were 525 averaged across both hemispheres. 526 527 4.6. Changes in connectivity patterns in the acute recovery period 528 In the early acute recovery period (week 1-2), stroke-related damage could alter connectivity 529 patterns in patients in two distinct ways: 1) the connectivity pattern could remain the same but 530 overall connection strengths might be increased or decreased, resulting in connectivity patterns 531 in patients DC-shifted but otherwise identical to control patterns. This would indicate that a 532 canonical pattern of connectivity between motor ROIs in healthy people is simply up or down-533 regulated post-stroke either due to maladaptation or compensation for damage. 2) stroke-related 534 damage might alter connectivity weights among only a few select ROIs, e.g. either between 535 ROIs within one hemisphere or across hemispheres. This would alter the shape of the 536 connectivity patterns in patients in comparison to controls. Since we wanted to be sensitive to 537 both kinds of connectivity pattern change, the appropriate statistical test would be a MANOVA 538 between patient and control connectivity patterns. However, due to insufficient degrees of 539 freedom in performing such an analysis (the number of connectivity weights exceeds the number 540 of patients and controls), we instead opted for a permutation test with Euclidean distance as a 541 measure of dissimilarity between patient and control connectivity patterns as it is sensitive to 542 shape and scaling changes of connectivity patterns (for details see supplementary material). Since patients in our cohort demonstrated substantial improvements of upper extremity deficits 559 in the year after stroke (Figure 1), we were interested to see whether there were concomitant 560 longitudinal changes in connectivity patterns. To determine this, we performed two separate but 561 related analyses. First, we independently compared differences in patient connectivity patterns 562 from the acute stage to all consecutive weeks (Δweek from acute to week 4, week 12, week 24, 563 and week 52) to determine how far connectivity patterns diverged over the year from the pattern 564 in the acute post-stroke stage. The same was done for control connectivity patterns to establish 565 intersession reliability. Second, we compared patient's connectivity patterns for all five 566 measurement sessions against the control connectivity patterns to determine how patient patterns 567 changed longitudinally in reference to controls (Δpattern for acute, week 4, week 12, week 24 568 and week 52). Both these analyses were performed using Euclidean distance and permutation 569 testing in the same way as for estimating differences in connectivity patterns at the acute 570 Because changes in functional connectivity between the two primary motor cortices have been 579 reported more consistently than other connectivity changes after stroke, we also explicitly looked 580 at changes of M1-M1 connectivity weights. 581

582
We additionally analyzed our dataset using a metric of functional connectivity that was proposed 583 in the to-date largest longitudinal resting-state stroke study with cortical and subcortical lesion 584 location, which reported changes of M1 interhemispheric connectivity. The metric has been 585 called Relative connectivity (RelCon) and is claimed to have low sensitivity to the temporal 586 signal-to-noise ratio and signal amplitude fluctuations while maintaining a high sensitivity to 587 meaningful signal changes, therefore offering an advantage e.g. in the analysis of data sets 588 acquired with different scanners (Golestani & Goodyear, 2011). for each single week and individual patient/control, as well as looking at the averaged split-half 604 correlation for all weeks together. Reliability between groups was compared using a mixed-605 effects ANOVA, with Group (patients vs. controls) and Week (acute -W52) as fixed, and 606 Subject as a random factor. This was done for all connections, as well as subsets only including 607 interhemispheric, intrahemispheric lesioned or non-lesioned ROIs. 608 Changes of interhemispheric M1-M1 connectivity weights over time between patients and 609 controls were analyzed using a mixed-effects ANOVA, with Group (patients vs. controls) and 610 Week (acute -W52) as fixed, and Subject as a random factor, alternative metrics reported in 611 The complete data set will be openly available in a public repository upon publication. All 617 analysis was performed using built-in and custom-written MATLAB and R scripts that will be 618 made publicly available upon publication. 619

Acknowledgement 620
We like to thank Susumu Mori & Andreia Faria from the Department of Neuroradiology, Johns 621 Hopkins University for their support regarding imaging analysis. 622

Competing interests 623
The authors report no competing interests. 624

Permutation test and Bootstrapping 628
To perform a permutation test, we first identified patients and controls that had estimates of 629 connectivity patterns within the first two weeks after stroke. We estimated Δpattern as the 630 Euclidean distance between the average connectivity pattern for patients and the average 631 connectivity pattern for controls. We then shuffled group assignment labels for connectivity 632 patterns 10,000 times, randomly assigning connectivity patterns to "controls" or "patients". From 633 the shuffled data, we again calculated the Euclidean distance between the average connectivity 634 pattern for patients and controls based on this new assignment. By repeatedly shuffling and 635 computing Euclidean distances, we obtained an estimate of the empirical null distribution of 636 Δpatterne.g. the expected distribution if there was no real difference between the two groups. 637 The measured Δpattern was then compared against this null distribution, and the relative 638 proportion of simulations that showed a larger distance was used as a p-value -the probability 639 that the distance between the mean control and patient pattern would be equal or larger than the 640 measured distance by pure chance. This analysis was carried out independently for the full, 641  m  right  left  58  56  5  2365  53  f  right  right  0  57  4  2395  65  m  right  right  30  21  4  2450  66  m  right  right  66  56  3  2531  66  f  right  right  60  55  5  2565  71  m  right  right  4  0  3  2652  46  m  left  left  4  0  4  2654  46  m  right  right  49  52  5  2663  67  f  right  left  16  2  4  2789  56  m  right  right  64  57  4  2925  59  f  right  left  60  57  5  3176  64  m  left  right  63  57  4  3239  74  m  left  left  5  0  5  3240  80  f  right  left  9  56  5  3241  64  f  right  right  58  39  5  3243  22  m  right  left  63  56  5  3246  53  m  left  left  30  39  5  3247  54  m  right  right  59  57  5  3248  58  m  right  right  61  56  4 658

Data reliability and Preprocessing comparison 667
To estimate the reliability of our measurements within sessions, connectivity patterns were 668 computed as described above for the first 100 volumes and the second 100 volumes 669 independently and correlated with each other to calculate split-half reliabilities. 670 As seen for overall connectivity, intra-and interhemispheric split-half reliabilities were highly 671 reliable for controls and patients (controls: intrahemispheric: r = 0.67 (95% Confidence Interval, condition. Patients always had slightly higher intrasession reliability than although this 684 difference was not significant and was possibly driven by outlier in the control group. 685 The reliability measurement also allowed us to compared two different pre-processing 686 procedures: 687 688 Preprocessing procedure (P1): We removed the first 10 volumes of the functional data, then 689 performed correction for the timing of slice acquisition, motion correction, brain extraction, 690 linear trend removal, and temporal filtering (band pass, 0.01-0.08 Hz) using FSL (FMRIB 691 Software Library (FSL), Oxford University, Oxford, UK). Our analysis was carried out in the 692 native space, and no spatial smoothing was applied. Linear regression was used to remove signal 693 correlated with the global mean signal, and the average time series in the cerebral white matter 694 and cerebrospinal fluid (Fox et al., 2006). 695 Preprocessing procedure (P2): Here, we used an independent component analysis (ICA) 696 approach using FSL MELODIC for artifact reduction (Smith et al., 2004). Again, we removed 697 the first 10 volumes of the functional data. We applied motion correction and brain extraction. 698 Probabilistic independent component analysis was conducted to denoise individual data by 699 removing components such as head motion, scanner artifacts, and physiological noise. Noise 700 components were classified using FMRIB's ICA-based Xnoiseifier (Salimi-Khorshidi et al., 701 2014), which attempts to auto-classify ICA components into "good" vs. "bad" components. The 702 "bad" components were then removed from the functional data. 703 To determine which procedure would provide a more stable result, we calculated the 704 split-half reliability of the ROI-ROI connectivity weights for the whole connectivity pattern over 705 time in controls only. 706 Both procedures lead to good intrasession reliability on average (P1 = 0.64, CI 0.60-0.66; 707 P2 = 0.62, CI 0.57-0.66) but showed no significant difference (χ 2 (1) = 1.231, p = 0.267), while 708 no consistent change over time was found for either procedure by itself (P1: χ 2 (4) = 2.834, p = 709 0.684; P2: χ 2 (4) = 3.007, p = 0.557). Because of the nominal higher intrasession reliability we 710 conducted all subsequent analyses after noise correction using the P1 procedure. Homo-versus Heterologous ROI connectivity 729 We examined the differences in connectivity between homo-and heterologous ROI connection. 730 Homologous connectivity was significantly higher than connections between heterologous ROIs 731 (χ 2 (1) = 108.38, p<0.001) and this effect showed no changes over time χ 2 (4) = 5.8993, p = 732 0.207), Figure S1. Furthermore, we found no differences for this effect between patients and 733 controls (type*group χ 2 (1) = 2.2701, p = 0.132 and type*week*group χ 2 (1) = 2.2187, p = 0.136), 734 Figure S3. 735 737 Figure S3: Average connectivity of interhemispheric homologous versus heterologous ROI-ROI 738 connectivity weights. Homologous regions (e.g. M1-M1, S1-S1, blue line) were higher 739 correlated than heterologous regions (e.g. M1-PmV, PmV-S1, green line). This did not change 740 over the course of a year and no systematic difference was found between both groups (patients 741 right panel, control left panel). 742

Correlations between patient and control connectivity patterns 743
Connectivity patterns for patients and controls were highly correlated in the early period after