Pre-proofs Title : Adipose tissue distribution from body MRI is associated with cross-sec ‐ tional and longitudinal brain age in adults

There is an intimate body-brain connection in ageing, and obesity is a key risk factor for poor cardiometabolic health and neurodegenerative conditions. Although research has demonstrated deleterious effects of obesity on brain structure and function, the majority of studies have used conventional measures such as waist-to-hip ratio, waist circumference, and body mass index. While sensitive to gross features of body composition, such global anthropometric features fail to describe regional differences in body fat distribution and composition. The sample consisted of baseline brain magnetic resonance imaging (MRI) acquired from 790 healthy participants aged 18-94 years (mean ± standard deviation (SD) at baseline: 46.8 ± 16.3), and follow-up brain MRI collected from 272 of those individuals (two time-points with 19.7 months interval, on average (min = 9.8, max = 35.6). Of the 790 included participants, cross-sectional body MRI data was available from a subgroup of 286 participants, with age range 19-86 (mean = 57.6, SD = 15.6). Adopting a mixed crosssectional and longitudinal design, we investigated cross-sectional body magnetic resonance imaging measures of adipose tissue distribution in relation to longitudinal brain structure using MRI-based morphometry (T1) and diffusion tensor imaging (DTI). We estimated tissue-specific brain age at two time points and performed Bayesian multilevel modelling to investigate the associations between adipose measures at follow-up and brain age gap (BAG) the difference between actual age and the prediction of the brain’s biological age at baseline and follow-up. We also tested for interactions between BAG and both time and age on each adipose measure. The results showed credible associations between T1-based BAG and liver fat, muscle fat infiltration (MFI), and weight-to-muscle ratio (WMR), indicating older-appearing brains in people with higher measures of adipose tissue. Longitudinal evidence supported interaction effects between time and MFI and WMR on T1-based BAG, indicating accelerated ageing over the course of the study period in people with higher measures of adipose tissue. The results show that specific measures of fat distribution are associated with brain ageing and that different compartments of adipose tissue may be differentially linked with increased brain ageing, with potential to identify key processes involved in age-related transdiagnostic disease processes.

While there is a wealth of research focusing on conventional anthropometric measures such as BMI, not all individuals with a higher BMI have the same disease risk (Mulugeta et al., 2021). A study assessing 27,000 individuals from 52 countries identified abdominal obesity as one of nine key risk factors that accounted for most of the risk of myocardial infarction worldwide (Yusuf et al., 2004). However, while some obese individuals develop health problems such as lipid abnormalities and type 2 diabetes (Lacobini et al., 2019), others are metabolically healthy. Partly motivated by this heterogeneity and complexity of fat distribution and cardiometabolic health, body MRI has recently emerged as a novel opportunity to investigate adipose tissue distribution beyond anthropometric measures (Linge et al., 2018(Linge et al., , 2019(Linge et al., , 2021. Research utilising body MRI has found associations between VAT and muscle fat infiltration (MFI) and coronary heart disease (CHD) and type 2 diabetes (T2D) (Linge et al., 2018). Moreover, higher liver fat has been associated with T2D and lower liver fat with CHD (Linge et al., 2018). Cross-sectional analyses investigating body and brain MRI associations have reported negative associations between liver fat, MFI and cerebral cortical thickness while thigh muscle volume (TMV) was positively associated with brain stem and accumbens volumes (Gurholt et al., 2020).
Advanced multimodal brain MRI provides a wealth of information reflecting structural and functional characteristics of the brain. Brain age prediction using machine learning and a combination of MRI features provides a reliable approach for reducing the complexity and dimensionality of imaging data. The difference between the brain-predicted age and an individual's chronological age, also referred to as the brain age gap (BAG), can be used to assess deviations from expected age trajectories, with potential utility in studies of brain disorders and ageing Kaufmann et al., 2019). This has clear clinical implications for patient groups, where studies have reported larger brain age gaps in patients with various neurological and psychiatric disorders (Han et al., 2020;Høgestøl et al., 2019;Kaufmann et al., 2019;Pardoe et al., 2017;Sone et al., 2019;Tønnesen et al., 2020). While brain age research has frequently merged individual and age-related differences using cross-sectional data, recent evidence has demonstrated that the rate of brain ageing may be dependent on a range of life events and lifestyle factors (Cole, 2020;Sanders et al., 2021), and characteristics related to cardiovascular health and obesity, including WHR and BMI (Beck et al., 2021a;de Lange et al., 2020;Franke et al., 2014;Kolbeinsson et al., 2020;Kolenic et al., 2018;Ronan et al., 2016).
The motivation of the current study was to utilise more effective and precise descriptions of fat distribution in order to localise the subtypes of body composition that impact brain health and ageing. Our primary aim was to identify interactions between adipose tissue measures based on body MRI and tissue specific (DTI and T1-weighted) measures of brain age. We investigated cross-sectional associations of tissue specific BAG and detailed adipose tissue measures (body composition) and, for comparison, conventional anthropometric measures (BMI and WHR) used in a recent study (Beck et al., 2021a). Next, we tested for associations between longitudinal brain age and body composition at follow-up and investigated associations between each adiposity measure and longitudinal BAG.
Adopting a Bayesian statistical framework, we hypothesised that higher fat ratio, total abdominal adipose volume index (total adipose), weight-to-muscle ratio (WMR), VAT index, MFI, and liver fat percentage (liver fat) would be associated with older appearing brains, with stronger associations in the body MRI measures than for traditional anthropometric features.
Further, we hypothesised that higher fat ratio, total adipose, WMR, VAT index, MFI, and liver fat are associated with accelerated brain ageing as reflected by larger changes in BAG between baseline and follow-up.

Sample description
Two integrated studies -the Thematically Organised Psychosis (TOP) (Tønnesen et al., 2018) and StrokeMRI (Richard et al., 2018) -formed the initial sample, which included 1130 brain MRI datasets from 832 healthy participants. All procedures were conducted in line with the Declaration of Helsinki and the study has been approved by the Norwegian Regional Committee for Medical and Health Research Ethics (REC). All participants provided written informed consent, and exclusion criteria included neurological and mental disorders, and previous head trauma.
Following the removal of 68 datasets after quality checking (QC) of the MRI data (see section 2.5), the final sample included 1062 brain MRI datasets collected from 790 healthy participants aged 18-94 years (mean ± standard deviation (SD) at baseline: 46.8 ± 16.3). This included longitudinal data (two time-points with 19.7 months interval, on average (min = 9.8, max = 35.6) from 272 participants. Of the 790 included participants, body MRI data was available from a subgroup of 286 participants, with age range 19-86 (mean = 57.6, SD = 15.6). Demographic information is summarised in

DTI processing and TBSS analysis
Processing steps for single-shell diffusion MRI data in the test set followed a previously described pipeline , including noise correction (Veraart et al., 2016), Gibbs ringing correction (Kellner et al., 2016), corrections for susceptibility induced distortions, head movements and eddy current induced distortions using topup (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/topup) and eddy (http://fsl.fmrib.ox.ac.uk/fsl/fslwiki/eddy) (Andersson & Sotiropoulos, 2016). Isotropic smoothing was carried out with a Gaussian kernel of 1 mm 3 implemented in the FSL function fslmaths. DTI metrics were estimated using dtifit in FSL and a weighted least squares algorithm. Processing steps for the training set followed a similar pipeline with the exception of the noise correction procedure.
Voxelwise analysis of the fractional anisotropy (FA) data was carried out using Tract-Based Spatial Statistics (TBSS) (Smith et al., 2006), as part of FSL (Smith et al., 2004). First, FA images were brain-extracted using BET (Smith, 2002) and aligned into a common space (FMRI58_FA template) using the nonlinear registration tool FNIRT (Andersson, Jenkinson, & Smith., 2007;Jenkinson et al., 2012), which uses a b-spline representation of the registration warp field (Rueckert et al., 1999). Next, the mean FA volume of all subjects was created and thinned to create a mean FA skeleton that represents the centres of all tracts common to the group. Each subject's aligned FA data was then projected onto this skeleton.
The mean FA skeleton was thresholded at FA > 0.2. This procedure was repeated in order to extract axial diffusivity (AD), mean diffusivity (MD), and radial diffusivity (RD). fslmeants was used to extract the mean skeleton and 20 regions of interest (ROI) based on a probabilistic white matter atlas (JHU) (Hua et al., 2008)

FreeSurfer processing
T1-weighted MRI data were processed using FreeSurfer (Fischl, 2012) (Reuter et al., 2012) in FreeSurfer. Specifically, an unbiased withinsubject template space and image (Reuter & Fischl, 2011) is created using robust, inverse consistent registration (Reuter et al., 2010). Several processing steps, such as skull stripping, Talairach transforms, atlas registration as well as spherical surface maps and parcellations are then initialised with common information from the within-subject template, significantly increasing reliability and statistical power (Reuter et al., 2012). Due to the longitudinal stream in FreeSurfer influencing the thickness estimates, and subsequently having an impact on brain age prediction (Høgestøl et al., 2019), both cross-sectional and longitudinal data in the test set were processed with the longitudinal stream. All reconstructions were visually assessed and edited by trained research personnel. Cortical parcellation was performed using the Desikan-Killiany atlas (Desikan et al., 2006), and subcortical segmentation was performed using a probabilistic atlas (Fischl et al., 2002). In total, 269 FreeSurfer (SI Table   1) based features were extracted, including global features for intracranial volume, total surface area, and whole cortex mean thickness, as well as the volume of subcortical structures.

Quality control (QC) procedure
A detailed description of the complete QC procedure for the final sample is available in (Beck et al., 2021a). Briefly, for DTI we derived various QC metrics (SI Table 2), including temporal signal-to-noise-ratio (tSNR) (Roalf et al., 2016)  anomalies in the data are detected as observations that do not conform to an expected pattern to other items. Using the R package mvoutlier (Filzmoser et al., 2005), potential outliers were flagged using the Mahalanobis distance (SI Figures 1 and 2). Informed by an interactive plot using the chisq.plot function, manual outlier observations of each of these flagged values deemed none of them as true outliers, leading to no further removal from the initial 286 body MRI dataset.

Brain age prediction
We performed brain age prediction using T1-weighted and DTI data using XGBoost regression (https://xgboost.readthedocs.io/en/latest/python), which is based on a decision-tree ensemble algorithm used in several recent brain age prediction studies (Beck et al., 2021a;de Lange et al., 2019de Lange et al., , 2020Kaufmann et al., 2019;Richard et al., 2020). BAG was calculated using (predicted age -chronological age) for each of the models, providing T1 and DTI-based BAG values for each individual. The BAG estimates were residualised for age to account for age-bias (de Lange & Cole, 2020;Liang et al., 2019), where brain age is overestimated in younger subjects while they are underestimated in older subjects.

Adipose tissue measures
The raw data from 286 body MRI scans included the following measures: tissue index (VAT index), which is VAT normalised by height 2 , describing the intraabdominal fat surrounding the organs; total adipose, which is the total abdominal fat (VAT and ASAT) normalised by height 2 ; weight-to-muscle ratio (WMR), which is body weight divided by thigh muscle volume; fat ratio, which is the total abdominal fat divided by total abdominal fat and thigh muscle volume; and muscle fat infiltration (MFI), representing intramuscular fat. Figure 2 shows the body MRI for two participants. Figure

Statistical analysis
All statistical analyses were performed using R, version 3.6.0 (www.r-project.org/) (R Core Team, 2012). Bayesian multilevel models were carried out in Stan (Stan Development Team, 2019) using the brms (Bürkner, 2017(Bürkner, , 2018 package, where multivariate models are fitted in familiar syntax to comparable frequentist approaches such as a linear mixed effects model using the lme4 (Bates et al., 2015). A general overview of the brms package can be found in Bürkner (2017), while information pertaining to multilevel formula syntax implemented in the analysis can be found in Bürkner (2018).
To address the primary aim of the study, we tested for associations between tissue specific BAG and each adiposity measure. BAG (for T1 and DTI separately) was entered as the dependent variable with each adiposity measure separately entered as the independent fixed effects variable along with age, sex, and time, and subject ID as random effect. This was employed as following: To test our hypothesis that adiposity influences brain ageing we tested for associations between longitudinal changes in BAG and cross-sectional body MRI measures collected at follow-up using Bayesian multilevel models. Similarly, we tested for interactions between age and adiposity measures on BAG. For each of the models, BAG was entered as the dependent variable, with the first model running through interaction effect of each adiposity measure and time, and the second model running through interaction effect of each adiposity measure and age. Timepoint and age were controlled for in the models where appropriate while sex was controlled for in both models, and subject ID was included as the random effect term. These were employed as following: In order to prevent false positives and to regularise the estimated associations, we defined a standard prior around zero with a standard deviation of 1 for all coefficients. All variables bar sex were scaled prior to running the analyses. For each coefficient of interest, we report the mean estimated value and its uncertainty measured by the 95% credible interval interpretation, see SI Table 3.

Brain age prediction
Age prediction accuracy in the training and test sets are summarised in SI Table 4. Briefly, the models revealed high accuracy, as previously reported (Beck et al., 2021a), with R 2 = 0.72 and 0.73 for the T1 and DTI models, respectively. Figure 4 shows predicted age for each model plotted as a function of age for both timepoints.

Descriptive statistics
Descriptive statistics are shown in Table 1, while Figure 3 shows the distributions within females and males for each adiposity measure, and Figure 5 shows the correlations between adiposity measures.

Associations between BAG and adiposity measure
Posterior distributions of the estimates of the coefficients reflecting the associations between each adiposity measure and BAGs are shown in Figure 6, where parameterisation of the model can be interpreted as values increasing from 0 to 1 indicating evidence in favour of an association and values decreasing 0 to -1 indicating evidence in favour of the null hypothesis. Figure 7 shows credible intervals and evidence ratios. SI Figure 7 shows network correlation graph of correlations between adiposity measures and each BAG, and SI Table 5

Interaction effects of time and adiposity measure on brain age gap
Posterior distributions of the estimates of the coefficient for the interaction between time and each adiposity measure and DTI and T1 BAGs are shown in Figure 8, and Figure 9 shows credible intervals and evidence ratios. For full table of results see SI Table 5.

Interaction effects of age and adiposity measure on brain age gap
Posterior distributions of the estimates of the coefficient for the interaction between age and each adiposity measure and DTI and T1 BAGs are shown in Figure 10, and Figure 11 shows credible intervals and evidence ratios. For full table of reported results see SI Table 5.

Discussion
Despite increasing evidence of shared mechanisms across several metabolic conditions and cardiovascular and neurodegenerative diseases, we are yet to fully understand the complex associations of adipose tissue and brain age. The current cross-sectional and longitudinal findings support that higher measures of adipose tissueparticularly higher liver fat and MFI are associated with an older-appearing brain and faster brain ageing. Both overall brain age gap and the rates of change in brain age over time were associated with specific adipose tissue measures at follow-up, including thigh muscle fat infiltration, weight-to-muscle ratio, and liver fat.

Adiposity and brain age: cross-sectional analysis
Our findings demonstrated associations between liver fat, MFI, WMR and T1 BAG, indicating older-appearing brains in individuals with higher adipose tissue measures and supporting our hypotheses for these specific measurements. Conversely, our hypotheses for fat ratio, total adipose, and VAT index were not supported. DTI BAG associations were less prevalent, with anecdotal evidence in favour of an association for fat ratio, while moderate evidence in favour of no association was revealed for VAT index and liver fat, meaning only our hypothesis for fat ratio was supported by the results. Our hypothesis that stronger associations would be found for body MRI measures than for traditional anthropometric features was supported by results pertaining to T1 BAG, however, for DTI BAG, anthropometric features performed comparably well to body MRI measures.
The findings are in line with the hypothesis that body MRI adipose tissue measures are associated with ageing of the brain as indicated by brain MRI morphology measures.
Moreover, the findings extend previous work linking adiposity measures, including links to grey matter volume (Debette & Markus, 2010;Gunstad et al., 2005;Jørgensen et al., 2017;Ward et al., 2005) and white matter microstructural properties based on diffusion MRI (Marks et al., 2011;Mueller et al., 2011;Stanek et al., 2011;Xu et al., 2013), with conflicting evidence for white matter volume (Friedman et al., 2014;Willette & Kapogiannis, 2015). The discrepancy may be due to methodological differences, e.g., between previous regional associations and our global brain age approach. Future research estimating regional brain age models trained with more advanced diffusion MRI features may offer improved sensitivity and specificity (Beck et al., 2021b).
MFI has previously been linked to metabolic risk factors (Therkelsen et al., 2013) and insulin resistance in obesity (Goodpaster et al., 2000). Higher liver fat has been found among diabetics (Bamberg et al., 2017;Linge et al., 2018) and prediabetics without previous cardiovascular conditions (Bamberg et al., 2017). Studies have also reported no significant association between CHD or cardiovascular events and elevated liver fat (Neeland et al., 2015). Moreover, while comparing patients with non-alcoholic fatty liver disease (NAFLD) to controls, Hagström et al., (2017) reported no significant difference in cardiovascular related death. Conflicting results however, have linked NAFLD with cardiovascular disease (Brea et al., 2017), suggesting a complex interplay between regional adipose tissue and metabolic health.
This complexity has recently been corroborated by findings of 62 genetic loci associated with both higher adiposity and lower cardiometabolic risk (Huang et al., 2021), possibly reflecting protective mechanisms. Although speculative, this heterogeneity is likely also reflected in connection to brain health, which for some individual genetic architectures may offer protection against brain pathology while simultaneously increasing the risk for obesity. Further studies exploiting larger sampler are needed to pursue this hypothesis and attempt to dissect the heterogeneity using genetic data (Huang et al., 2021) in combination with brain imaging.

Adiposity and brain ageing: longitudinal evidence
Our longitudinal analyses revealed that the rate of brain ageing across the study period was associated with adiposity measured at follow-up, with evidence for WHR from our anthropometric measures, and MFI and WMR from adipose tissue distribution measures.
While evidence for WHR was present for both T1-weighted and DTI BAGs -which largely replicates a recent report from the same cohort (Beck et al., 2021a) -evidence for MFI and WMR were only observed for T1-weighted BAG, where greater adipose tissue was associated with increased rate of brain ageing. Conversely, evidence in favour of no interaction effect with time was found for BMI. This supports that BMI alone may not represent a clinically relevant proxy, as it fails to distinguish from lean mass while ignoring regional fat distribution, in particular the visceral components (Evans et al., 2012;Roberson et al., 2014).
While experimental evidence is needed to establish causality, our findings suggest that adipose tissue distribution affects brain ageing and may lead to older appearing brains in generally healthy individuals. These observations are in line with previous studies, both from clinical groups and population-based cohorts. NADLD, for example, has been previously linked to smaller total brain volume (Weinstein et al., 2018), and liver fat has been linked to smaller cortical and cerebellar structures (Gurholt et al., 2020). Moreover, negative associations have been reported between MFI and cortical structures (Gurholt et al., 2020).
While more research into adipose tissue distribution is warranted, the current results suggest that increased liver fat, muscle fat infiltration, and weight-to-muscle ratio may contribute to accelerate brain ageing.

Brain age and adiposity: interactions with age
For DTI BAG, our findings demonstrated an interaction effect with age and WHR, while T1 BAG interaction effects were present for age and WMR, fat ratio, total adipose, VAT index, liver fat, MFI, BMI, and WHR. The findings indicate that adiposity measures may be increasingly important predictors of BAG with increasing age. In contrast, we observed no evidence of interaction effects between age and BMI on DTI or T1 BAGs. Previous research has produced mixed results, with studies reporting an interaction of BMI and age on white matter volume but no interaction on cortical surface area or thickness (Ronan et al., 2016).
Further research is warranted to elucidate the degree to which associations between adiposity and brain structure change over the course of the lifespan.

Strengths and limitations
The current study benefitted from a mixed cross-sectional and longitudinal design enabling brain changes to be tracked across time. The prediction models for brain age had high accuracy, and separate T1-weighted and DTI brain age gaps provided tissue-specific measures of brain age with potential to reveal specific associations with the included adiposity measures.
Several limitations should be considered when evaluating the findings. First, while the longitudinal brain MRI data represent a major strength, the follow-up sample size is relatively small, and the body MRI data was collected at the follow-up examination only. The subsequent loss of power is reflected in the width of body MRI posterior distributions, indicating a higher level of uncertainty compared to BMI and WHR, which had available longitudinal measures and a larger sample size. Next, although body composition measures based on MRI offer high accuracy in terms of fat and muscle distributions and are therefore a potentially valuable supplement to conventional anthropometric features, future studies including biological markers such as immune and inflammation assays and lipid measurements might provide even higher specificity and opportunities for further subtyping.
Indeed, inflammation has demonstrated effects on brain function and structure (Rosano et al., 2011) and has been dubbed to have a central role in the obesity-brain connection (O'Brien et al., 2017).
Further, there is the possibility of transfer learning happening and scanner effects inducing an offset shift in the estimates, however, the current study utilises the relative estimates within the test set rather than the absolute age estimates and thus this should not introduce a problem for the conclusions drawn. Moreover, including detailed assessments of dietary routines, alcohol intake, and physical exercise is vital in order to better understand the complex processes at play. For example, physical activity has been associated with lower brain age (Dunås et al., 2021;Sanders et al., 2021) and higher grey and white matter measures (Sexton et al., 2016), while excess alcohol intake is well documented in influencing liver and brain health (Agartz et al., 1999;Rehm et al., 2010). Lastly, the current sample is predominantly ethnic Scandinavian and Northern European, restricting our ability to generalise to wider populations, and future studies should aim to increase the diversity in the study population.

Conclusion
Combining body MRI and brain age prediction based on brain MRI allows for probing individual body composition profiles and brain patterns and trajectories which may confer risk for cardiometabolic disease and neurodegenerative disorders. More knowledge and further development of automated tools for individual phenotyping in this domain may inform public health priorities and interventions. With evidence of different adiposity subtypes being differentially linked with different brain phenotypes and cardiometabolic diseases, precision methods that look at fat distribution can potentially be more informative than conventional anthropometric measures. This in turn will provide a more effective tool for development of treatment strategies that focus on individual risk of metabolic disease, as well as disentangling the associations between body and brain health.

Acknowledgements
The study is supported by the Research Council of Norway (223273, 249795, 248238,