Cross-Sectional and Longitudinal Brain Scans Reveal Accelerated Brain Aging in Multiple Sclerosis

Multiple sclerosis (MS) is an inflammatory disorder of the central nervous system. By combining longitudinal MRI-based brain morphometry and brain age estimation using machine learning, we tested the hypothesis that MS patients have higher brain age relative to chronological age than healthy controls (HC) and that longitudinal rate of brain aging in MS patients is associated with clinical course. Seventy-six MS patients, 71 % females and mean age 34.8 years (range 21-49) at inclusion, were examined with brain MRI at three time points with a mean total follow up period of 4.4 years. A machine learning model was applied on an independent training set of 3208 HC, estimating individual brain age and calculating the difference between estimated brain age and chronological age, termed brain age gap (BAG). We also assessed the longitudinal change rate in BAG in MS individuals. We used additional cross-sectional MRI data from 235 HC for case-control comparison. MS patients showed increased BAG (4.4 ±6.6 years) compared to HC (Cohen’s D = 0.69, p = 4.0 × 10−6). Longitudinal estimates of BAG in MS patients suggested an accelerated rate of brain aging corresponding to an annual increase of 0.41 (±1.23) years compared to chronological aging for the MS patients (p = 0.008). On average, patients with MS have significantly higher BAG compared to HC and accelerated rate of brain aging compared to chronological aging. Brain age estimation represents a promising method for evaluation of brain changes in MS, with potential for predicting future outcome and guide treatment.

After the initial diagnosis, which is usually established in early adulthood, patients experience different disease activity and longitudinal accumulation of neurological damage.
Development of robust brain imaging markers that can parse this between-subject heterogeneity of the clinical trajectories, predict future progression of disability, and monitor the effects of treatment is a major aim with important clinical implications 4 5 . Current imaging markers with relevance for MS are associated with disease activity and progression, and include, among other features, number or volume of hyperintense brain lesions visible on T2-weighted MRI images, contrast-enhancing T1 lesions, increased annual brain volume loss and T1-hypointense "black holes" 4 6 7 . Increased rate of brain volume loss, which is best captured using longitudinal designs 8 , reflects accelerated neurodegeneration 9 . Tuning the analysis towards specific brain regions may boost the correlations between estimated brain atrophy and disability 4 . However, identifying robust associations between clinical outcomes and MRI measures has been challenging 10 . This clinico-radiological paradox in MS is likely explained by a combination of lack of sensitivity and specificity both in the clinical and imaging domain. Indeed, expanded disability status scale (EDSS) and relapse rate, which are frequently used as clinical measures in MS phase III trials, are not sensitive to the full clinical spectrum in MS 11 12 . Ongoing efforts are made to develop and validate novel multidimensional measures to capture subtle changes in MS activity and progression 12 . The concept of "No evidence of disease activity" (NEDA) has emerged in the last decade as one such promising multidimensional measure of MS disease activity 13 . On the imaging side, recent technological advances for improved acquisition and analyses are likely to overcome some of the current challenges related to brain scan reproducibility and across-site harmonization and contribute towards a predictive MRI marker for disability and the effect of treatment in patients with MS in a clinical setting 4 .
Essentially, brain age estimation uses machine learning on a large training set of MRI data from healthy controls (HC) to develop a model that can accurately predict the individual age from brain imaging data [14][15][16] . Utilizing sensitive measures of MRI-based brain morphometry, brain age estimation provides a robust imaging-based biomarker with potential to yield novel insights into similarities and differences of disease pathophysiology across brain disorders 15 17 . Such imaging-based brain age has been shown to be reliable both within and between MRI scanners, and is a candidate biomarker of an individual's brain health and integrity 14 15 17 . Different approaches to brain age estimation utilize information from a variety of brain regions (e.g. hippocampus, subcortical, grey matter and white matter 18 ) or MRI sequences (e.g. T1, T2, diffusion tensor imaging and functional MRI) to inform the estimation model 17 . An older appearing brain, which is related to advanced physiological and cognitive ageing and mortality 17 , has been found to be an imaging-based hallmark across several brain disorders. To our knowledge, only one preprint (n = 254 MS patients) and one abstract (n = 17 MS patients) have reported brain age estimations in MS 15 19 , and both reported older appearing brains in patients with MS compared to HC.
Here, combining cross-sectional and sensitive measures of MRI-based regional and global brain morphometry in MS and HC (cross-sectional only), we tested the hypothesis that MS patients have higher estimates of brain age than HC in a cross-sectional design. Next, using longitudinal MRI data in MS patients we tested the hypothesis that brain aging accelerates in MS and that the rate of acceleration is associated with a more severe clinical outcome, investigating the associations between the rate of brain aging, disease-modifying treatment (DMT) and clinical trajectories.

Participants
We included 76 MS patients recruited at Oslo University Hospital 20 21  fingolimod, natalizumab or alemtuzumab.
The HC group was recruited through newspaper ads or after a stratified random selection drawn from the Norwegian National Population Registry to two parallel studies 18 23 .
Exclusion criteria included estimated intelligence quotient (IQ) <70, history of neurologic or psychiatric disease and current medication significantly affecting the nervous system 24 .
The project was approved by the local ethics committee and in line with the Declaration of Helsinki. All participants received oral and written information and gave their written informed consent.

MRI acquisition
All MS patients were scanned at up to three time points between January 2012 and August 2017, using the same 1.5 T scanner (Avanto, Siemens Medical Solutions; Erlangen, Germany) equipped with a 12-channel head coil. Structural MRI data were collected using a 3D T1- Several processing steps, such as skull stripping, Talairach transforms, atlas registration as well as spherical surface maps and parcellations were then initialized with common information from the within-subject template, increasing reliability and power 26 .
Manual quality control of the MRI scans from patients was performed by trained research personnel to identify and edit segmentation errors where possible (n = 43 MRI scans) and exclude data of insufficient quality (n = 6 MRI scans). In addition, eight brain scans were removed due to missing sequences of the 263 MRI scans from MS patients. Lesion filling was performed utilizing automatically generated lesion masks from Cascade 28 with the lesion filling tool (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/lesion_filling) in FSL 29 . The lesion masks were assessed by a trained neuroradiologist. For a probabilistic representation of the lesions, the lesion masks were normalized into MNI-standard space using FSL FLIRT 30

Brain age estimation model
The training set for the brain age estimation included data from 3208 HC >12 years (54 % women, mean age 47.5 (±19.8), age range 12-95) obtained from several publicly available datasets ( Supplementary Fig. 2). We trained one "xgboost" machine learning model for each sex to predict brain age derived from a total of 1118 cortical and subcortical brain imaging features 15 31 32 .
Based on a recent implementation 15 , brain age estimations were performed both using global and regional features as input. The brain regions reported were based on lobe labels from Freesurfer 25 . The brain age estimation model was carefully evaluated using 10-fold cross-validation within the training set and showed good performance and generalizability ( Supplementary Fig. 3, r = 0.91). For all patients and HC, we calculated the brain age gap (BAG, defined as the difference between chronological age and imaging-based brain age).
Using linear regressions, we removed any common variance with age, age 2 and sex to account for confounding factors before submitting the residualized version of BAG to further analyses 33 . When pooling estimates of BAG from the 1.5 T and 3 T scanners, we adjusted BAG for scanner effect on BAG estimates by extracting the scanner coefficient from a linear mixed effects (LME) model for global and all brain regions. When comparing BAG between patients and matched HCs we report the actual adjusted difference in BAG between these two groups.

Statistical analyses
We used R (R Core Team, Vienna, 2018) for statistical analyses. All LME models accounted for age, age 2 , sex and scanner 34 .
We estimated an annualized rate of change in BAG by dividing the total change in BAG by the time interval between the time points. We utilized the longest time interval between time points and excluded MS patients lacking longitudinal data (n = 8). The unit used for brain aging is "change in BAG per year". Here, 0 would indicate that the rate of brain aging corresponds to chronological aging, and positive and negative values correspond to accelerated and decelerated brain aging compared to chronological aging, respectively. For all brain regions we tested for significant change in the rate of brain aging by performing a onesample t-test on BAG with 0 set as test value. To assess brain aging within groups stratified by DMT, we used the coefficients from linear models with the individual slope in BAG as dependent variable and age, age 2 , sex and DMT group as independent variables. We estimated the annual global brain atrophy by comparing estimated total brain volume from the Freesurfer output (BrainSegVolNotVent) between time points.
To assess reliability of brain age across time we computed the intraclass correlation coefficient (ICC) using the R package "irr" (https://CRAN.R-project.org/package=irr).

Data availability
The data are not publicly available due to local restrictions, since they contain information that may compromise the privacy of research participants. The code needed to reproduce our results is available from the authors upon request.

Cross-sectional case-control analyses (3 T)
At time point 3 we found significantly higher BAG for the MS cohort compared to matched HC for all brain regions except the temporal region ( Fig. 1; Table 2). The most prominent differences in BAG were 4.4 years for global BAG (Cohen's D = 0.69) and 6.2 years for subcortical and cerebellar brain regions (Cohen's D = 0.72).
At time point 3, 58 MS patients underwent one MRI scanning in the 1.5 T and one in the 3 T scanner with two days apart. Whereas absolute estimates of brain age varied between scanners for all brain regions except insula (BAG scanner difference -6.08 to 10.60 years, see Supplementary Table 2; Supplementary Fig. 4), brain age estimates from the two scanners were highly correlated for global BAG and all brain regions (r = 0.67-0.86, p<0.001).

Figure 1. Cross-sectional comparison of brain age gap between MS patients and healthy
controls. The distribution of brain age gaps across brain regions based on the cross-sectional 3 T MRI data from matched HC and MS patients at time point 3. We found increased brain age gaps for all brain regions except from the temporal brain region. Brain age gaps are residualized for age, age 2 and sex. Cohen's D effect sizes for the brain age gap between HC and MS patients are depicted using the colour bar.

Longitudinal MS sample (1.5 T)
Fig. 2 shows the estimated brain age and rate of brain aging within the longitudinal MS cohort. The correlations between chronological age and global brain age were r = 0.  Table 2 and   Supplementary table 3).
We found a significant annual increase in global BAG of 0.41 (±1.23) years (p = 0.008) in patients with MS ( Fig. 3; Supplementary Table 4 and 5). None of the regional measures showed significantly increased annual change in BAG (Supplementary Table 5).
We found no significant difference in BAG between the raw and the lesion filled MRI ICCs for all brain regions across all time points varied from 0.79-0.94 for residualized brain age gap and 0.78-0.95 for raw predicted age. The regional feature of cerebellar and subcortical brain regions showed highest reliability with an ICC of 0.94 for BAG and 0.95 for predicted age (Supplementary Table 8).
Mean annualized estimated change in global brain volume from all three time points.
from Freesurfer was -0.30 % (±0.53 %). ICC for global brain volume was 0.97-0.99. Mean annualized change in WMLL was 504 mm 3 (±1299 mm 3 ). ICC for WMLL was 0.93-0.99.  Longitudinal changes in brain age gap across brain regions. The distribution of brain age gaps across brain regions based on the longitudinal 1.5 T MRI sample. Brain age gaps from the MS sample are compared with the cross-sectional 3 T HC sample and residualized for age, age 2 , sex and scanner. The full brain estimates showed a significant accelerated rate of brain aging compared to chronological aging (annual increase in brain age gap 0.41 (p = 0.008)). Cohen's D effect sizes for the brain age gap between MS and HC are depicted using the colour bar. Table 2. Differences in brain age gap between the multiple sclerosis patients and healthy controls (n = 235).  Brain age gap (BAG) estimations are residualized for age, age 2 , sex and scanner. Effect sizes were calculated using Cohen's D estimates.

Time point 1 (n=73) Time point 3 (n=60) Time point 3 (n=58)
Significant differences in brain age gap are highlighted with bold (p<0.05). The two groups are matched according to age and sex. Table 3 (BAG) and Table 4 (annual rate of brain aging) show summary statistics from the multiple regressions testing for associations with demographic, clinical and MRI variables in the longitudinal MS cohort. After accounting for multiple testing, significant associations were found between BAG and brain atrophy (Cohen's D = -0.07, p = 0.01) and WMLL

DISCUSSION
Here, by using advanced cross-sectional and longitudinal MRI-based neuroimaging data as basis for brain age estimation based on machine learning, we tested the hypotheses that patients with MS on average show higher brain age than HC, and that the rate of brain aging is affected by treatment. We found an accelerated brain aging in MS patients compared to controls by cross-sectional and longitudinal brain scans and suggested that increased rates of longitudinal brain aging are associated with higher rates of brain atrophy and increasing WMLL.
We For subcortical and cerebellar brain regions we found a higher BAG in MS compared with HC (BAG 6.2 years, Cohen's D = 0.72), which was already evident at time point 1 (BAG 5.7 years, Cohen's D = 0.63). The regional variation observed in the BAG estimates in our study may reflect different affinity of MS pathology in different brain regions and aligns with previously shown lesion probability maps in MS 4 37 .
In our longitudinal patient sample, the annual rate of brain aging for global BAG exceeded that of chronological aging by 0.41 years per year (p = 0.008), which may partly be explained by chronic inflammatory processes that drive neurodegeneration in MS 3 . The association between BAG and brain aging, brain volume and brain atrophy were expected since our estimation model was based on regional and global structural MRI features.
Of notice, brain age estimation takes into account subtle and regional changes that are not included in the global brain atrophy measure. The associations between BAG and brain aging, brain atrophy, brain volume, WMLL and change in WMLL showed significant and regional differences. The associations between change in WMLLs and brain aging were significant for occipital, temporal and parietal brain regions in addition to global BAG (Supplementary Table 9). This shows that regional brain age estimation can capture regional specificity of MS pathology, which is in line with the observations that MS lesions show regional affinity 14 15 17 37 .
The significant increase in BAG in the patients receiving no treatment warrants further investigations but is of clinical interest. However, the lack of randomization and group differences in the subgroups calls for caution in the interpretations of this relative small dataset.
The preliminary findings reported by Kaufmann et al. revealed a significant association between BAG and EDSS (Fisher z = 0.23) 15 , i.e. that patients with older appearing brains have a higher clinical disease burden. In the current study, multiple regression analyses revealed nominally significant (p<0.05, uncorrected) associations between some clinical, cognitive and imaging variables and BAG as well as brain aging for specific brain regions (Supplementary Table 9 and 10). However, these associations did not survive correction for multiple testing, and further studies are needed to assess the robustness of these observations. Some limitations should be considered when interpreting the results. First, although the cross-sectional case-control comparison and the within-patient longitudinal analysis jointly suggest accelerated brain aging in patients with MS, a longitudinal sample of HCs would have enabled us to directly compare the rate of brain aging between patients and controls. Next, the current brain age model was exclusively based on gross morphometric features. The MS patients were naturalistic, hence any results based on DMT status must be interpreted with care.
In conclusion, using advanced cross-sectional imaging data and machine learning methods we report that patients with MS show evidence of increased brain aging compared to healthy controls. In the longitudinal data we found that MS patients have accelerated brain aging. Higher rates of longitudinally measured brain aging were associated with higher levels of brain atrophy and longitudinal progression of changes in WMLL. These results show that brain age estimation is a promising and intuitive tool for monitoring of the individual disease course in MS and may guide a personalized treatment approach.

ACKNOWLEDGMENTS
We thank all the patients who participated in our study. We acknowledge the collaboration with members of the Multiple Sclerosis Research Group at the University of Oslo and Oslo University hospital, especially Professor Elisabeth G. Celius.

SUPPLEMENTARY MATERIAL
Supplementary material is included in separate files.