Disruption of the Atrophy-based Functional Network in Multiple Sclerosis Is Associated with Clinical Disability: Validation of a Meta-Analytic Model in Resting-State Functional MRI
In multiple sclerosis (MS), gray matter (GM) atrophy exhibits a specific pattern, which correlates strongly with clinical disability. However, the mechanism of regional specificity in GM atrophy remains largely unknown. Recently, the network degeneration hypothesis (NDH) was quantitatively defined (using coordinate-based meta-analysis) as the atrophy-based functional network (AFN) model, which posits that localized GM atrophy in MS is mediated by functional networks.
To test the NDH in MS in a data-driven manner using the AFN model to direct analyses in an independent test sample.
Materials and Methods
Model fit testing was conducted with structural equation modeling, which is based on the computation of semipartial correlations. Model verification was performed in coordinate-based data of healthy control participants from the BrainMap database (https://www.brainmap.org). Model validation was conducted in prospectively acquired resting-state functional MRI in participants with relapsing-remitting MS who were recruited between September 2018 and January 2019. Correlation analyses of model fit indices and volumetric measures with Expanded Disability Status Scale (EDSS) scores and disease duration were performed.
Model verification of healthy control participants included 80 194 coordinates from 9035 experiments. Model verification in healthy control data resulted in excellent model fit (root mean square error of approximation, 0.037; 90% CI: 0.036, 0.039). Twenty participants (mean age, 36 years ± 9 [standard deviation]; 12 women) with relapsing-remitting MS were evaluated. Model validation in resting-state functional MRI in participants with MS resulted in deviation from optimal model fit (root mean square error of approximation, 0.071; 90% CI: 0.070, 0.072), which correlated with EDSS scores (r = 0.68; P = .002).
The atrophy-based functional network model predicts functional network disruption in multiple sclerosis (MS), thereby supporting the network degeneration hypothesis. On resting-state functional MRI scans, reduced functional network integrity in participants with MS had a strong positive correlation with clinical disability.
© RSNA, 2021
In multiple sclerosis, disruption of the atrophy-based functional network is strongly associated with clinical disability, thereby providing support for the network degeneration hypothesis of localized gray matter atrophy.
■ The atrophy-based functional network (AFN) model predicted functional network disruption in a prospective study of 20 study participants with relapsing-remitting multiple sclerosis (MS).
■ By applying the AFN model in resting-state functional MRI, deviation from optimal model fit was detected, which indicates reduced functional network integrity (root mean square error of approximation, 0.071).
■ Expanded Disability Status Scale scores for participants with MS had a strong positive correlation with AFN-predicted functional network alteration (r = 0.68; P = .002).
Routine clinical MRI enables highly sensitive qualitative evaluation of white matter lesions to guide diagnosis and treatment of patients with multiple sclerosis (MS) (1). However, the use of white matter lesions as a marker of disease progression is problematic given its weak correlation with clinical status (2). Recently, increased attention has been placed on the role of neurodegeneration in MS, particularly in gray matter (GM). Current literature reports the presence of GM atrophy throughout the disease course and across all MS subtypes, an indication that it is likely independent of white matter inflammation (3). Additionally, a clinically relevant pattern of GM atrophy has been described that affects localized brain regions, including the basal ganglia, thalamus, and sensorimotor cortex (4–6). Although it is evident that GM atrophy is regionally specific, there has yet to be consensus on the pathophysiologic mechanism.
From observations of localized GM atrophy in cognitive and motor neurodegenerative disorders, the network degeneration hypothesis (NDH) has emerged as a plausible concept (7). The NDH describes co-atrophy that occurs in a network-based manner (ie, normally functionally connected brain regions are prone to co-atrophy in disease) (7). In MS, network mediation in the context of localized GM atrophy has also been proposed, where “hubs” (ie, regions of the brain that have dense functional connectivity) are preferentially susceptible to GM atrophy (8). Thus, the NDH could be extended to the disease pattern in MS to help explain regional selectivity of GM atrophy.
A practical approach to define and test a data-driven hypothesis based on the NDH is to use meta-analytic coactivation modeling. This method uses a big-data approach to create a functional connectivity model with a task-based functional literature sample, which can help predict resting-state functional connectivity in prospective data (9,10). Blood oxygen level–dependent functional MRI helps provide an indirect measure of neuronal activity and can be performed in the resting state without task stimulation (ie, resting-state functional MRI) (11). Using resting-state functional MRI, functional connectivity can be derived statistically and is defined as the temporal correlation of neurophysiologic events between spatially distinct brain regions (12). By taking this approach, a meta-analytic node-and-edge atrophy-based functional network (AFN) model in MS was recently constructed (4). In this two-part model, nodes represent atrophy-prone brain regions in MS and were computed using anatomic likelihood estimation; the edges represent interregional functional covariances in healthy control participants and were computed using functional meta-analytic connectivity modeling (4,13). This meta-analytic AFN model could then be subsequently used to organize resting-state functional connectivity analyses for characterizing functional network alterations in MS (Fig 1).
We hypothesized that the AFN model would help detect functional network disruption in MS that is associated with clinical disability. The objective of our study was to test the NDH in a data-driven manner by using the literature-based AFN model to direct analyses in a prospective resting-state functional MRI sample of participants with MS. To our knowledge, this is the first study to investigate functional network disruption in MS by testing a model-based hypothesis. Furthermore, this is, to our knowledge, the first study to develop a model using quantitative meta-analysis with primary data model validation. The results would inform further development of the AFN model as a quantitative functional imaging tool for individualized assessment.
Materials and Methods
This prospective study was approved by the institutional review board and was compliant with Health Insurance Portability and Accountability Act guidelines. Written informed consent was obtained from all participants.
Twenty-two study participants with relapsing-remitting MS were prospectively recruited from an MS clinic between September 2018 and January 2019, with participants forming a random series. Inclusion criteria of participants with MS were age 18–50 years and fulfillment of 2017 revised McDonald criteria for relapsing-remitting MS (1). Exclusion criteria were screening failures, clinical relapse or use of intravenous steroid medications within the past month, structural brain disease, previous brain surgery, excessive image artifact, uncontrolled psychiatric condition, and claustrophobia or other contraindications to MRI. It has been shown that intravenous steroids can affect scoring of clinical disability and quantification of MRI measures (14).
Whole-brain MRI scans were obtained with a 3.0-T scanner (Tim Trio; Siemens Medical Solutions) using a standard 12-channel head coil as the radiofrequency receiver and the integrated circularly polarized body coil as the radiofrequency transmitter. Functional T2*-weighted (blood oxygen level–dependent) MRI was performed by using a multiband gradient-echo echo-planar imaging sequence with agreement from the University of Minnesota (15). A total of 700 volumes were acquired with the following parameters: repetition time msec/echo time msec, 1400/30; flip angle, 52°; field of view, 211 × 211 mm; base resolution, 88 × 88; multiband acceleration factor, 3; and isotropic voxel size, 2.4 mm. Participants were given instructions to remain awake with their eyes closed and to let their minds wander. Dummy scans were obtained to establish steady-state magnetizations. To correct for distortions, a gradient-echo field map was acquired with the same shimming and acquisition matrix. Three-dimensional T1-weighted images were obtained by using the magnetization-prepared rapid acquisition gradient-echo pulse sequence with the following parameters: 1900/2.26/900 (repetition time msec/echo time msec/inversion time msec); flip angle, 9°; field of view, 256 × 256; in-plane image matrix, 256; and isotropic voxel size, 1 mm. T2-weighted fluid-attenuated inversion recovery images were obtained with the three-dimensional turbo spin-echo sequence to characterize lesion volume. Acquisition parameters were as follows: 5000/335/1800; echo train duration, 673 msec; echo spacing, 3.1 msec; turbo factor, 221; sagittal sections, 160; field of view, 256 × 256; in-plane image matrix, 256 × 256; and isotropic voxel size, 1 mm.
All acquired images were visually inspected for quality control before image analysis. Functional data preprocessing was performed with FSL software (version 5.0.11; FMRIB Software Library, https://fsl.fmrib.ox.ac.uk/fsl/fslwiki) (16). The FEAT software package within FSL was used for data preprocessing, which included linear registration to T1-weighted images and nonlinear registration to the Montreal Neurological Institute 152 space template, motion correction, field map unwarping, section timing correction, brain extraction, spatial smoothing with 5-mm full width at half maximum, and temporal filtering with a high-pass filter of 100 seconds (17). Additional noise reduction steps included removal of motion-related artifacts by using Independent Component Analysis–based Automatic Removal of Motion Artifacts, or ICA-AROMA, and nuisance regression by using the mean signal of white matter and cerebrospinal fluid as regressors (18).
The workflow in applying the AFN model in resting-state functional MRI is shown in Figure 2. Regions of interest (ROIs) (radius, 5 mm) were defined by nodes in the AFN model. Binarized ROIs were transformed from standard (Montreal Neurological Institute) to the native echo-planar imaging space of each participant with nonlinear registration. The ROIs were rebinarized to minimize effects of interpolation. Then, each ROI was used to sample the preprocessed resting-state functional MRI data by extracting the mean time series values across the four-dimensional resting-state functional MRI data.
White matter lesions were segmented by the lesion growth algorithm (19) as implemented in the Lesion Segmentation Tool (version 2.0.15; https://www.statistical-modelling.de/lst.html) for Statistical Parametric Mapping (version 12; Wellcome Center for Human Neuroimaging; https://www.fil.ion.ucl.ac.uk/spm/). The algorithm first segments the T1-weighted images into the three main tissue classes (cerebrospinal fluid, GM, and white matter). This information is then combined with the coregistered fluid-attenuated inversion recovery intensities to calculate lesion belief maps. By thresholding these maps with a prechosen initial threshold (κ, 0.3) an initial binary lesion map is obtained, which is subsequently grown along voxels that appear hyperintense on the fluid-attenuated inversion recovery image. The optimal initial threshold was determined by two board-certified neuroradiologists (F.F.Y, with 9 years of experience, and B.T., with 20 years of experience), who were provided only with thresholded binary lesion maps. Visual inspection was performed independently, without disagreement in evaluation. The resulting lesion probability map was thresholded to obtain a binary lesion segmentation. The total lesion volume was normalized for head size, resulting in normalized lesion volume.
To minimize the effect of T1 hypointensities on brain volumetric measurements, T1-weighted images were preprocessed by using the “lesion_filling” tool in FSL software (20). Subsequently, SIENAX, a software package in FSL, was used to obtain global brain tissue volumes, normalized for head size (20).
Coordinate-based Healthy Control Data from BrainMap
A data set of coordinate-based results from the literature was generated by using the BrainMap neuroimaging database (21). The inclusion search filters “activations-only” and “healthy control” were applied in Sleuth on July 2, 2020 (version 2.4; http://www.brainmap.org/sleuth/); no records were excluded, following best practices (22). ROIs defined by the AFN model were used to sample the modeled activation values in standard (Montreal Neurological Institute) space (Fig 2). These experiment-level results data are precursors to anatomic likelihood estimation values, which indicate the probability of a convergent effect in the sampled literature (23).
Path analysis in structural equation modeling (SEM) was used to assess the model fit of AFN. All SEM analyses were performed with Amos (version 25.0; SPSS) and are based on the computation of standardized semipartial regression coefficients using the maximum likelihood estimator. Compared with traditional correlation analyses, SEM is a rigorous statistical method that allows for simultaneous analysis of the structural relations involving the entire system of variables while correcting for measurement error (24). In SEM, pairwise associations are tested while controlling for variables in the rest of the model; this enables statistical computation of associations that are conditionally independent. In this study, structural relations (ie, modeled covariances) represent functional connectivity specified by the AFN model. Based on path estimates, model fit can be computed to determine the extent to which the model predicts the data (24).
The SEM path diagram derived from the AFN model included 14 region-to-region paths. Observed variables and paths were specified by the nodes and edges in the AFN model, respectively (Fig 3). A standard recursive SEM model was constructed, so only the stronger path in bidirectional AFN coactivations was retained in the path diagram (24,25). The root mean square error of approximation (RMSEA) was selected as the primary model fit criterion given that it is a widely used fit statistic and is relatively insensitive to the effects of sample size (26,27). RMSEA of 0.05 or 0.08 indicates a reasonably good fit to the data (28). The Tucker-Lewis Index was also computed (29). A sample size requirement of 200 samples per observed variable is typical in SEM, which was exceeded in this study (30).
The AFN model was quantified and formalized with SEM in the present study. As part of model construction, the model verification step provides model fit statistics, which were not previously computed. The use of disease-specific nodes and semipartial correlations follow current best practices (31). Because the functional connectivity predicted by the AFN model is based on meta-analytic results of healthy controls, model fit verification was first performed using healthy control data from the BrainMap database. This verification step serves to define optimal model fit of the AFN. To perform model validation, the AFN model was directly tested (without model optimization) in prospective resting-state functional MRI data of participants with MS by computing both groupwise and individual model fit. To better represent temporal effects in functional MRI time series data, a unified SEM approach with multivariable autoregressive modeling was taken (Appendix E1 [online]) (32).
Correlation analyses were performed with software (version 25.0; SPSS). Partial correlation coefficients were computed between per-participant imaging measures (RMSEA, Tucker-Lewis Index, normalized lesion volume, and other brain volumetric measurements listed in Table 1) and clinical data (Expanded Disability Status Scale [EDSS] scores and disease duration). Correlations with EDSS were adjusted for disease duration, age, and sex. Correlations with disease duration were adjusted for age and sex.
Among 22 recruited study participants with relapsing-remitting MS, two participants were withdrawn because of screening failures (n = 1) and excessive image artifact (n = 1). Thus, 20 participants (mean age, 36 years ± 9 [standard deviation]; 12 women) with relapsing-remitting MS were included in the study (Fig 4). Healthy control data generated by using the BrainMap neuroimaging database included 80 194 coordinates, 9035 experiments, and 2449 papers. Demographic information for the 20 participants with MS included in the study can be found in Table 1.
Verification and Validation of the AFN Model
Results from verification and testing of the AFN model demonstrated model fit deviation in participants with MS from optimal model fit in healthy control participants (Table 2). Model verification in coordinate-based data of healthy control participants resulted in excellent model fit. Thus, model optimization was not performed before testing the model in participants with MS. In prospective resting-state functional MRI data of participants with MS, application of the AFN model resulted in borderline acceptable model fit, which deviated from optimal model fit results in healthy control data.
Associations between Participant-level Imaging and Clinical Measures
Partial correlations between participant-level imaging findings and clinical data were computed (Fig 5, Table 3). Per-participant RMSEA and EDSS had a strong correlation (r = 0.68; P = .002). Per-participant Tucker-Lewis Index and EDSS also had a strong correlation (r = −0.53; P = .03). Remaining correlations between imaging and clinical measures were not significant. The clinical associations of model fit indices and normalized lesion volume are displayed as a scatterplots (Fig 6).
Recently, the network degeneration hypothesis (NDH) was quantitatively defined (using coordinate-based meta-analysis) as the atrophy-based functional network (AFN) model, positing that localized gray matter atrophy in multiple sclerosis (MS) is mediated by functional networks. In this study, the NDH was tested in participants with MS by applying the AFN model to prospectively acquired resting-state functional MRI data. Two predictions were confirmed: (a) the AFN model detected functional network disruption on resting-state functional MRI scans of participants with MS (root mean square error of approximation, 0.071); and (b) the degree of functional network disruption had a strong correlation with clinical disability (r = 0.68; P = .002). On the other hand, no correlation existed between normalized lesion volume and Expanded Disability Status Scale scores (r = −0.06; P = .82). Our results indicate that it is feasible to apply the AFN model in resting-state functional MRI to assess clinically relevant functional network alteration in MS.
The model testing strategy we used was multistep and without sample overlap or analytical asymmetry, using distinctly different data sets at each stage of model building. Previously, functional connectivity predicted by the AFN model was determined meta-analytically by using task-activation functional MRI and PET coordinate-based data of healthy control patients (4). In our study, the AFN model was first verified in healthy control coordinate-based data. Then, given satisfactory results from model fit verification, the AFN model was subsequently tested—without finetuning or adjustment—in an independent resting-state functional MRI test sample of participants with MS.
At the group level, the AFN model helped detect a reduction in functional network integrity on resting-state functional MRI scans of participants with MS. Although the RMSEA in MS indicated a model fit that was borderline acceptable, the model fit deviated considerably from the meta-analytically verified results in healthy control BrainMap data. These observations suggest that the AFN model (based on healthy functional connectivity involving MS-specific atrophy-prone GM regions) could provide baseline measures for identifying functional network alterations in MS. Current literature describing functional connectivity changes in MS has been somewhat heterogeneous with reports of both increase and decrease in functional connectivity (33). This phenomenon may be due to the multiphasic nature of functional connectivity changes (34). Further, functional connectivity changes within a network are considered nonuniform across brain regions and may be influenced by involvement of hubs (5,35). Therefore, a network-based approach may be able to more fully capture functional connectivity alteration in MS (33,36,37). Although our results were limited by the cross-sectional design of the study, we observed that functional network disruption could be detectable at the group level by applying the AFN model in a prospective sample.
In addition to groupwise assessment, the AFN model was applied at the single-participant level. This was an important step to include to account for intraparticipant variation (38). Subsequent correlation analysis between per-participant model fit indices and EDSS scores demonstrated increasing deviation from optimal model fit as clinical disability worsened. This finding suggests that there is increasingly compromised integrity of the predicted functional network as disease progresses. Given that edges in the AFN model were defined using healthy control coordinate-based data, it is reasonable that deviation in model fit occurs as clinical disability worsens in patients with MS. The associations between model fit measures and disease duration were relatively weaker, which could be related to differences in medication regimen, compliance, and treatment response of the participants with MS. However, correlation analysis using normalized lesion volume confirmed no association with disease severity.
Clinically, results from this study demonstrate the feasibility of using the AFN model to assess clinical progression of MS by using resting-state functional MRI. The AFN model describes a functional network likely to be consistently affected in MS, but additional connections may coexist. Thus, in consideration of clinical utility, surveillance of the AFN in MS would be important in enabling widespread use of functional network measures. From a practical perspective, evaluation of neurovascular dynamics by means of resting-state functional MRI is a sensitive method that could provide early detection of disease-related changes (39). In contrast, assessment of atrophy may have more limited clinical use because atrophy often indicates the presence of irreversible brain tissue loss and permanent structural damage (34). In particular, resting-state functional MRI (ie, task-free functional MRI) is useful in the clinical setting (as opposed to task-based functional MRI) given its nondependence on task performance. Overall, the AFN model has helped provide replicable neuroimaging features that could be used to develop sensitive imaging tools for individualized patient assessment.
Our study had limitations. First, the AFN model was only tested in the relapsing-remitting subtype of MS. Second, to show the biologic basis of these functional network changes, there is a need to move beyond the use of composite scores (eg, EDSS scores) to associate specific clinical symptoms to functional network disruptions. Third, our sample size of 20 participants was small.
In conclusion, this study showcased a stepwise approach for targeted imaging biomarker discovery. By applying a meta-analytically derived atrophy-based functional network (AFN) model in resting-state functional MRI of participants with multiple sclerosis (MS), data-driven testing of the network degeneration hypothesis was accomplished. Present findings confirm that network mediation may explain localized gray matter atrophy in MS and that the AFN model could be used to examine clinically relevant functional network changes. Additional work is needed to disentangle path-level (or region-specific) contributions to these per-participant network-level changes, which may offer insights regarding diagnostic utility. Further model development as a quantitative functional imaging tool for evaluation of disease progression is encouraged.Disclosures of Conflicts of Interest: F.L.C. Activities related to the present article: institution received grants from National Institutes of Health and Radiological Society of North America. Activities not related to the present article: disclosed no relevant relationships. Other relationships: filed patent with U.S. Patent Office. M.F. disclosed no relevant relationships. R.S.R. Activities related to the present article: disclosed no relevant relationships. Activities not related to the present article: is a consultant for BSM, Viela Bio, and Genentech. Other relationships: disclosed no relevant relationships. L.P. disclosed no relevant relationships. C.G.F. disclosed no relevant relationships. S.D. disclosed no relevant relationships. J.P.G. disclosed no relevant relationships. F.F.Y. disclosed no relevant relationships. B.T. disclosed no relevant relationships. S.Y.H. Activities related to the present article: institution received grant from National Institutes of Health (NIH). Activities not related to the present article: institution has grants/grants pending with Siemens Healthineers; has received payment for lectures including service on speakers bureaus from Siemens Healthineers. Other relationships: disclosed no relevant relationships. P.T.F. Activities related to the present article: institution received grants from NIH and National Institute of Mental Health. Activities not related to the present article: is a consultant for NIH Center for Scientific Review; receives payment from Wiley as editor in chief of Human Brain Mapping; has grants/grants pending with NIH, U.S. Department of Defense, U.S. Department of Veterans Affairs, and Michael J. Fox Foundation; filed patent with U.S. Patent Office; has received reimbursement from various institutions for travel, accommodations, and meeting expenses. Other relationships: receives payment as a licensee of intellectual property.
The authors thank Radhika Mohan, BS, Matthew Deng, MD, Matthew Lu, MD, Brian Seegmiller, BS, and Brian Doss, BS, for their assistance in data acquisition.
Author contributions: Guarantors of integrity of entire study, F.L.C., P.T.F.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, F.L.C., P.T.F.; clinical studies, F.L.C., M.F., B.T., S.D., F.F.Y., R.S.R; statistical analysis, F.L.C., L.P., C.G.F., J.P.G., P.T.F.; and manuscript editing, F.L.C., R.S.R., L.P., C.G.F., S.D., J.P.G., F.F.Y., B.T., S.Y.H., P.T.F.
Supported by the National Institutes of Health (grant nos. R01MH074457, R25EB16631, P41EB030006, and K23NS096056) and Radiological Society of North America (grant no. RR1826).
- 1. . Diagnosis of multiple sclerosis: 2017 revisions of the McDonald criteria. Lancet Neurol 2018;17(2):162–173. Crossref, Medline, Google Scholar
- 2. . The clinico-radiological paradox in multiple sclerosis revisited. Curr Opin Neurol 2002;15(3):239–245. Crossref, Medline, Google Scholar
- 3. . Assessing brain atrophy rates in a large population of untreated multiple sclerosis subtypes. Neurology 2010;74(23):1868–1876. Crossref, Medline, Google Scholar
- 4. . Localised grey matter atrophy in multiple sclerosis is network-based: a coordinate-based meta-analysis. Clin Radiol 2019;74(10):816.e19–816.e28. Crossref, Google Scholar
- 5. . Cortical atrophy patterns in multiple sclerosis are non-random and clinically relevant. Brain 2016;139(Pt 1):115–126. Crossref, Medline, Google Scholar
- 6. . Localized grey matter atrophy in multiple sclerosis: a meta-analysis of voxel-based morphometry studies and associations with functional disability. Neurosci Biobehav Rev 2013;37(5):819–830. Crossref, Medline, Google Scholar
- 7. . Neurodegenerative diseases target large-scale human brain networks. Neuron 2009;62(1):42–52. Crossref, Medline, Google Scholar
- 8. . What lies beneath grey matter atrophy in multiple sclerosis? Brain 2016;139(Pt 1):7–10. Crossref, Medline, Google Scholar
- 9. . Correspondence of the brain’s functional architecture during activation and rest. Proc Natl Acad Sci U S A 2009;106(31):13040–13045. Crossref, Medline, Google Scholar
- 10. . Cognitive relevance of the community structure of the human brain functional coactivation network. Proc Natl Acad Sci U S A 2013;110(28):11583–11588 [Published correction appears in Proc Natl Acad Sci U S A 2013;110(38):15502.]. Crossref, Medline, Google Scholar
- 11. . A default mode of brain function. Proc Natl Acad Sci U S A 2001;98(2):676–682. Crossref, Medline, Google Scholar
- 12. . Functional and effective connectivity in neuroimaging: A synthesis. Hum Brain Mapp 1994;2(1-2):56–78. Crossref, Google Scholar
- 13. . Beyond the single study: function/location metanalysis in cognitive neuroimaging. Curr Opin Neurobiol 1998;8(2):178–187. Crossref, Medline, Google Scholar
- 14. . Effects of IV methylprednisolone on brain atrophy in relapsing-remitting MS. Neurology 2001;57(7):1239–1247. Crossref, Medline, Google Scholar
- 15. . Evaluation of slice accelerations using multiband echo planar imaging at 3 T. Neuroimage 2013;83:991–1001. Crossref, Medline, Google Scholar
- 16. . Advances in functional and structural MR image analysis and implementation as FSL. Neuroimage 2004;23(Suppl 1):S208–S219. Crossref, Medline, Google Scholar
- 17. . Temporal autocorrelation in univariate linear modeling of FMRI data. Neuroimage 2001;14(6):1370–1386. Crossref, Medline, Google Scholar
- 18. . ICA-AROMA: A robust ICA-based strategy for removing motion artifacts from fMRI data. Neuroimage 2015;112:267–277. Crossref, Medline, Google Scholar
- 19. . An automated tool for detection of FLAIR-hyperintense white-matter lesions in Multiple Sclerosis. Neuroimage 2012;59(4):3774–3783. Crossref, Medline, Google Scholar
- 20. . Evaluating and reducing the impact of white matter lesions on brain volume measurements. Hum Brain Mapp 2012;33(9):2062–2071. Crossref, Medline, Google Scholar
- 21. . Opinion: Mapping context and content: the BrainMap model. Nat Rev Neurosci 2002;3(4):319–321. Crossref, Medline, Google Scholar
- 22. . The hippocampal network model: A transdiagnostic metaconnectomic approach. Neuroimage Clin 2018;18:115–129. Crossref, Medline, Google Scholar
- 23. . The functional connectivity of the human caudate: an application of meta-analytic connectivity modeling with behavioral filtering. Neuroimage 2012;60(1):117–129. Crossref, Medline, Google Scholar
- 24. . Structural Equation Modeling with Amos: Basic Concepts, Applications, and Programming. 3rd ed. New York, NY: Routledge, 2016https://doi.org/10.4324/9781315757421. Crossref, Google Scholar
- 25. . Progressive bidirectional age-related changes in default mode network effective connectivity across six decades. Front Aging Neurosci 2016;8:137. Crossref, Medline, Google Scholar
- 26. . The Performance of RMSEA in Models With Small Degrees of Freedom. Sociol Methods Res 2015;44(3):486–507. Crossref, Google Scholar
- 27. . Comparison of model fit indices used in structural equation modeling under multivariate normality. J Mod Appl Stat Methods 2015;14(1):152–167. Crossref, Google Scholar
- 28. . Alternative ways of assessing model fit. In: Bollen KA, Long JS, eds. Testing Structural Equation Models. Beverly Hills, Calif: Sage, 1993; 136–162. Google Scholar
- 29. . In: Little TD, ed. Principles and Practice of Structural Equation Modeling. 4th ed. New York, NY: Guilford, 2016. Google Scholar
- 30. . Nonconvergence, improoper solutions, and starting values in LISREL maximum likelihood estimationo. Psychometrika 1985;50(2):229–242. Crossref, Google Scholar
- 31. . Network modelling methods for FMRI. Neuroimage 2011;54(2):875–891. Crossref, Medline, Google Scholar
- 32. . Unified structural equation modeling approach for the analysis of multisubject, multivariate functional MRI data. Hum Brain Mapp 2007;28(2):85–93. Crossref, Medline, Google Scholar
- 33. . Functional connectivity in multiple sclerosis: Recent findings and future directions. Front Neurol 2018;9:828https://doi.org/10.3389/fneur.2018.00828. Crossref, Medline, Google Scholar
- 34. . Detecting neurodegenerative pathology in multiple sclerosis before irreversible brain tissue loss sets in. Transl Neurodegener 2019;8(1):37 [Published correction appears in Transl Neurodegener 2020;9:3.]. Crossref, Medline, Google Scholar
- 35. . Explaining the heterogeneity of functional connectivity findings in multiple sclerosis: An empirically informed modeling study. Hum Brain Mapp 2018;39(6):2541–2548. Crossref, Medline, Google Scholar
- 36. . Reduced network dynamics on functional MRI signals cognitive impairment in multiple sclerosis. Radiology 2019;292(2):449–457. Link, Google Scholar
- 37. . Resting-state functional MR imaging: a new window to the brain. Radiology 2014;272(1):29–49. Link, Google Scholar
- 38. . A Manifesto on Psychology as Idiographic Science: Bringing the Person Back Into Scientific Psychology, This Time Forever. Measurement 2004;2(4):201–218. Google Scholar
- 39. . Cognitive function and fMRI in patients with multiple sclerosis: evidence for compensatory cortical activation during an attention task. Brain 2002;125(Pt 6):1275–1282. Crossref, Medline, Google Scholar
Article HistoryReceived: Aug 19 2020
Revision requested: Oct 23 2020
Revision received: Nov 8 2020
Accepted: Nov 23 2020
Published online: Feb 2 2021
Published in print: Apr 2021