Introduction

Obesity is a risk factor for dementia and brain volume loss1,2,3,4. This association has become particularly concerning given the persistently increasing global prevalence of obesity, which has nearly tripled since 1975 and reached about 14% in 20195. The United States especially exhibits notable trends of obesity, with age-adjusted prevalence of obesity increasing from 30.5% to 42.4% from 1999 to 2018, and severe obesity, as defined by a body mass index (BMI) of over 40, nearly doubling from 4.7% to 9.2% in the same period6. Thus, due to the increase in the prevalence of obesity, clinicians and patients alike should pay closer attention to the pathological processes of obesity in relation to the nervous system, particularly in the context of dementia.

Obesity creates a chronic inflammatory state resulting from an altered neurohormonal axis perpetuated by adipocyte hypertrophy and altered secretion of adipokines, free fatty acids, and lipopolysaccharide-binding proteins7,8,9,10. In the central nervous system, this manifests as changes in white matter (WM) integrity, white and gray matter volume, and cognitive symptoms due to an obesity-induced breakdown of the blood-brain barrier11,12. Importantly, a distinction should be made between subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT), given their physiologically distinct metabolic profiles13. VAT adipocytes are metabolically more active than their SAT counterparts, generating a higher amount of free fatty acids that drive systemic inflammation14. Increased VAT is positively associated with insulin resistance, increased adipokine expression, and an exaggerated inflammatory response, which are known risk factors of brain atrophy13,15,16,17. Furthermore, the VAT/SAT ratio is another measure of obesity that has been demonstrated to be a correlate of cardiometabolic risk and insulin resistance, independent from BMI and VAT18. This ratio uniquely measures the propensity of the body to store visceral fat versus subcutaneous fat, and thus may provide a more individualized measurement of obesity with regards to hypertension, dyslipidemia, and diabetes among other pathologies18. These inflammatory changes in obesity are also characterized in Alzheimer’s disease (AD), and thus a pathological correlation between obesity and dementia is well-corroborated19.

Considering the pathophysiological distinction between SAT and VAT, the various health effects of obesity should also be understood in the context of metabolically healthy obesity (MHO), a subset of individuals in the obese group that have a relatively reduced risk of comorbidities20,21. Though currently there are no established criteria for MHO, close associations include lower visceral fat relative to subcutaneous fat, absence of hypertension, and normal blood pressure and lipid parameters21,22. Some proposed biological mechanisms that differentiate MHO from metabolically unhealthy obesity (MUO) are: comparatively greater insulin sensitivity in those with MHO, association with metabolic syndrome, and preserved adipose tissue function, particularly with regards to adipokine secretion and energy balance21,23. Thus, when delineating the effects of obesity on dementia, classifying obesity based on more nuanced measures such as medical history and laboratory markers is an important consideration.

WM injuries with decreased axonal density, altered diffusivity, and increased vasogenic edema were evident in chronic inflammation17,24,25,26. AD exhibits similar WM changes that are consistent with demyelination, tract deterioration, and axonal damage27,28. These WM injuries associated with neuroinflammation may be characterized by novel neuroimaging paradigms such as edge density imaging (EDI)29. EDI maps the density of network edges, which consist of white matter pathways, on a voxel-by-voxel basis from a whole-brain structural connectome that may be derived from diffusion-weighted magnetic resonance imaging sequences using probabilistic tractography. EDI and structural connectomes provide information on the spatial organization and density of WM fibers within a region, which conveys more information about WM integrity than conventional diffusion metrics, where diffusivity is averaged across WM tracts within a bundle29. Preliminary studies have suggested that EDI may improve diagnostic performance compared to diffusion tensor imaging (DTI) or neurocognitive assessments in pediatric traumatic brain injury and autism spectrum disorder with regard to WM connectivity30,31. Another study demonstrated that EDI was a reliable biomarker for associating the WM connectome with BMI, and may be comparable or superior to conventional metrics such as fractional anisotropy (FA)32. Though mean diffusivity and FA have previously been employed to characterize decreased WM integrity in both obesity and dementia, EDI remains a comparatively novel measure28,33. Given its potential in assessing WM integrity and fiber density, EDI provides promise in assessing the extent of WM damage associated with the inflammatory effects of obesity32.

Furthermore, diffusion basis spectrum imaging (DBSI) has been used to illustrate the extent of edema and inflammation within the brain34. DBSI mitigates potential errors, such as crossing fibers, associated with diffusion imaging methods by considering intracellular diffusion in the model35. As this technique is able to reliably differentiate between axonal damage, demyelination, and inflammation of brain microstructures in a variety of pathologies, DBSI, when employed to process EDI, may provide a more detailed snapshot into white matter integrity34,36.

Given the relationship between obesity and neuroinflammation, EDI in conjunction with DBSI may offer a unique perspective in helping elucidate the relationship between a proinflammatory state and WM changes in obesity. Moreover, biological sex has been shown to affect the relationship between WM microstructure and BMI37. Given that obesity is a prominent risk factor for dementia, assessing EDI’s utility in assessing white matter integrity in the context of obesity may potentially be helpful in recognizing pathological patterns of WM changes and neuroinflammation. As such, the present study investigated the relationship between metrics of body adiposity, including BMI, VAT, SAT, and VAT/SAT, and WM connectivity using EDI in a cohort of 50 healthy middle-aged individuals. We aimed to elucidate the following points: (1) is there a relationship between WM microstructure and BMI, visceral or subcutaneous adipose tissue, and (2) is there any modifying effect of obesity classification or sex in the relationship between white matter integrity and the mentioned tissue adiposity metrics.

Results

A summary of participant demographics is illustrated in Table 1. Scans from a total of 50 participants were processed for this study, including 20 men and 30 women with an average age of 50.4 ± 6.1 years. These participants were comprised of 33 white, 15 African American, and 2 Asian participants. 30 participants were classified as obese and 20 as non-obese. Those categorized as obese had a BMI average of 36.4 ± 4.3 kg/m2, while those categorized as non-obese had a BMI average of 25.8 ± 4.9 kg/m2, culminating in an overall BMI average of 32.2 ± 6.9 kg/m2.

Table 1 Demographic and obesity measure information
Full size table

Body adiposity metrics differences between BMI categories

The following t-test results are summarized in Table 1. VAT was significantly higher in the 30 obese participants (1376.2 ± 579.7 cm3) compared to the 20 non-obese (668.2 ± 412.6 cm3), t(48) = 4.7165, p < 0.0001. SAT was significantly greater in obese participants (3502.6 ± 1191.5 cm3) compared to their non-obese counterparts (1904.6 ± 699.2 cm3), t(48) = 5.3991, p < 0.0001. VAT/SAT ratios were not statistically different between obese (0.46 ± 0.32) and non-obese individuals (0.38 ± 0.24), t(48) = 0.9726, p = 0.3356. All associations held true after correcting for multiple comparisons.

MHO and MUO body adiposity metrics

Out of the 30 participants classified as obese, 14 were classified with MUO, and 8 with MHO. The obesity classification of the remaining 8 could to be determined due to insufficient clinical and metabolic data. There was no statistically significant difference in VAT between individuals with MHO (1003.9 ± 231.9 cm3) and individuals with MUO (1392.2 ± 649.3 cm3), t(20) = −1.6192, p = 0.1211. There was no statistically significant difference in SAT between individuals with MHO (3864.6 ± 1445.5 cm3) and individuals with MUO (3396.3 ± 1114.0 cm3), t(20) = 0.8521, p = 0.4043. There was no statistically significant difference in VAT/SAT between individuals with MHO (0.28 ± 0.09) and individuals with MUO (0.50 ± 0.39), t = −1.517, p = 0.1449.

Comparisons between the 8 participants with MHO and 20 participants without obesity showed significantly lower VAT in participants without obesity (1003.9 ± 231.9 cm3 vs. 668.2 ± 412.6 cm3), t(26) = 2.1534, p = 0.04074; however, this association did not hold true after correction for multiple comparisons. The 8 participants with MHO had significantly lower SAT (3864.6 ± 1445.5 cm3) than those without obesity (1904.6 ± 699.2 cm3), t(26) = 4.8553, p < 0.0001; this association held true after correction for multiple comparisons. There was no statistically significant difference in VAT/SAT between individuals with MHO (0.28 ± 0.09) and individuals without obesity (0.38 ± 0.24), t(26) = −1.1745, p = 0.2508.

Biological sex differences in body adiposity metrics

The following results are summarized in Table 1. VAT was significantly greater in the 20 male participants (1353.3 ± 676.6 cm3) compared to the 30 female participants (919.5 ± 526.6 cm3), t(48) = 2.545, p = 0.0142. SAT was not significantly different between males (2530.6 ± 1304.7 cm3) and females (3085.2 ± 1246.7 cm3), t(48) = −1.5129, p = 0.1369. VAT/SAT ratios were significantly greater in males (0.62 ± 0.34) compared to females (0.30 ± 0.16), t(48) = 4.4284, p < 0.0001. All the above associations held true after correction for multiple comparisons.

Whole-cohort EDI regressions

Linear regression analyses with TBSS in the overall cohort demonstrated a significant negative correlation between VAT/SAT and EDI, predominantly in the body and genu of the corpus callosum (CC), after controlling for age, race, and sex (pmax = 0.01, Fig. 1). BMI, VAT, and SAT were not significantly associated with EDI in the overall cohort. When BMI was added as a covariate in the regression, SAT was positively associated with EDI, predominantly in the body, rostrum, and splenium of the CC, bilateral inferior fronto-occipital fasciculi (IFOF), bilateral inferior longitudinal fasciculi (ILF), and parts of bilateral corticospinal tracts (pmax = 0.005, Supplemental Fig. 1). EDI associations with other body adiposity measures held consistently with the initial analyses.

Fig. 1: Significant regions in VAT/SAT and edge density imaging regression.
figure 1

Observed significant associations between increased VAT/SAT and decreased edge density imaging. Predominantly affected regions include the body and genu of the corpus callosum. p = significance level. Regression was performed while controlling for age, sex, and race. VAT/SAT = visceral adipose tissue: subcutaneous adipose tissue ratio.

Full size image

EDI regressions stratified by obesity

EDI was not significantly different between participants categorized as obese and non-obese after controlling for age, sex, and race. Among the 30 individuals who were categorized as obese, there was a statistically significant positive association between SAT and EDI, predominantly in the body, rostrum, genu, and splenium of the CC, bilateral IFOF, bilateral ILF, and parts of bilateral corticospinal tracts (pmax = 0.003, Fig. 2a). There was also a statistically significant negative association between the VAT/SAT ratio and EDI, predominantly in the body, rostrum, genu, and splenium of the CC, bilateral IFOF, bilateral ILF, and parts of bilateral corticospinal tracts (pmax < 0.0001, Fig. 2b). BMI and VAT were not significantly associated with EDI among participants classified as obese. Among the 20 individuals who were categorized as non-obese, there were no statistically significant associations observed between any of the body fat measures and EDI.

Fig. 2: Significant regions in body fat measures and edge density imaging regression in obese participants.
figure 2

Observed a positive associations between subcutaneous adipose tissue (SAT) and edge density imaging (EDI), and b negative associations between visceral adipose tissue (VAT)/SAT ratio and EDI. Predominantly affected regions in both regressions include the body, rostrum, genu, and splenium of the corpus callosum, bilateral inferior fronto-occipital fasciculi, bilateral inferior longitudinal fasciculi, and parts of bilateral corticospinal tracts. Regressions were performed while controlling for age, sex, and race. p = significance level.

Full size image

EDI was not significantly different between participants with MHO and those with MUO, nor between participants with MHO and those without obesity. Among the 8 individuals classified with MHO, no statistically significant associations were observed between EDI and any adiposity measure. Among the 14 individuals classified with MUO, there was a statistically significant positive association between SAT and EDI, predominantly in the body, rostrum, genu, and splenium of the CC, and right IFOF (pmax = 0.01, Supplementary Fig. 2a). There was also a statistically significant negative association between VAT/SAT and EDI, predominantly in the rostrum, genu, and splenium of the CC, and right IFOF (pmax < 0.0001, Supplementary Fig. 2b).

EDI regression results stratified by sex

EDI was significantly lower in males compared to females after controlling for age, race, and BMI, particularly within the body of the CC, bilateral IFOF, bilateral ILF, and bilateral corticospinal tracts (pmax < 0.0001, Fig. 3). Among the 20 males, there were no statistically significant associations observed between any of the body fat measures and EDI. Among the 30 females, BMI (pmax = 0.02) and SAT (pmax < 0.0001) were negatively correlated with EDI. VAT/SAT and VAT were not significantly associated with EDI in the female group.

Fig. 3: Edge density imaging differences between sexes.
figure 3

Observed areas where males had significantly lower edge density imaging compared to females. Predominantly affected regions include the body of the corpus callosum, bilateral inferior fronto-occipital and inferior longitudinal fasciculi, and bilateral corticospinal tracts. Regression was performed while controlling for age, race, and body mass index. p = significance level.

Full size image

EDI regression results stratified by race

EDI was significantly higher in white participants compared to African American participants after controlling for age, race, and BMI, specifically within the body of the CC, right IFOF, and right ILF (pmax < 0.007, Fig. 4). Among the 15 participants identifying as African American, there were no statistically significant associations observed between any of the body fat measures and EDI. Among the 33 participants identifying as white, BMI (pmax = 0.03), SAT (pmax < 0.0001), and VAT/SAT (pmax = 0.04) were negatively associated with EDI. VAT was not associated with EDI among participants identifying as white.

Fig. 4: Edge density imaging differences between white participants and African American participants.
figure 4

Observed areas where white participants had significantly higher edge density imaging compared to African American participants. Predominantly affected regions include the body of the corpus callosum and right inferior fronto-occipital and inferior longitudinal fasciculi. Regression was performed while controlling for age, sex, and body mass index. p = significance level.

Full size image

Discussion

Within the overall cohort, a negative association between VAT/SAT and EDI was appreciated, and a positive association between SAT and EDI once BMI was controlled for, suggesting that obesity plays a role in modifying the relationship between specific adiposity measures and WM integrity measured as EDI. As such, among participants categorized as obese, EDI was positively associated with SAT and negatively associated with the VAT/SAT ratio, which may indicate a neuroprotective component of SAT in contrast to VAT-related neuroinflammation and WM damage. These associations were not seen among participants without obesity. VAT tends to have higher susceptibility to cellular stress and lipid overload compared to SAT38,39. In contrast, SAT exerts less influence on glycemic control and glucose metabolism compared to VAT, and thus, preferential energy storage in SAT may confer a neuroprotective effect40. Based on the study’s results, EDI displays promise as a measure for WM integrity, as the observed regressions consistently correlate with previously-documented patterns of WM damage as a result of obesity-associated neuroinflammation24,25,29. Furthermore, these trends support the notion that obesity may play a role in the pathological manifestation of dementia41.

The associations of VAT and SAT with WM integrity are also appreciated more specifically in participants classified with MUO. Such patterns were not seen in participants classified with MHO, thus suggesting that MUO likely increases susceptibility to neuroinflammation and WM damage, which may increase risk of dementia18,42. Of note, there was an observed, though not statistically significantly lower VAT and VAT/SAT in participants with MHO than those with MUO, likely due to the small sample size of the MHO cohort (n = 8). In comparison with normal-weight individuals, SAT was decreased among participants with MHO compared to those without obesity, the mechanism of which is unclear and should be clarified with larger sample sizes; if expanded upon, this finding could potentially contribute to the literature suggesting cardiovascular risk is still increased among people with MHO compared to healthy normal-weight individuals43,44.

Anatomically, the tracts that were affected in this study are spatially consistent with those affected in dementia, including various parts of the CC, IFOF, and ILF, which have been shown to exhibit increased diffusivity in AD33,45. Functionally, these tracts contribute not only to overall structural and functional interhemispheric connectivity, but also visual processing and language comprehension, which are known areas of deficit in dementia patients45,46,47. Therefore, the EDI changes in relation to obesity, as seen in this study, are comparable to established negative associations of white matter integrity and dementia.

Of note, EDI was significantly lower in men compared to women, which may be explained by the sexual dimorphism of obesity, a phenomenon in which males tend to accrue more visceral fat, whereas females accrue more subcutaneous fat, especially prior to menopause40,48. This trend is supported by this study’s comparison of VAT, SAT, and the VAT/SAT ratio across sex differences. Interestingly, obesity measures, including SAT, seem to have a more significant negative association with EDI among women compared to men. Mechanisms for this phenomenon are unclear; however, previous studies suggest that a negative association between obesity and WM integrity, specifically FA and mean diffusivity, is more pronounced among women than men4,49,50. However, the areas of the significant associations were relatively small in these regressions compared to the whole-group regressions, putting into question the true extent of sex-dependent patterns of neurodegeneration that should be clarified with larger sample sizes. Despite these differences, given that sex and age were controlled for in the EDI regression analyses, the study’s results suggest that the neuroinflammatory and neuroprotective roles of VAT and SAT, respectively, occur consistently across males and females in an otherwise healthy middle-aged group. EDI was also significantly lower in African Americans compared to whites, which may be explained by increased WM pathology burden in African Americans than whites51,52. However, contrary to previous findings, in this sample, the effects of obesity specifically on WM integrity seem to be pronounced in whites and not in African Americans. Of note, there is a major skew in the ratio of obese to non-obese participants between whites and African Americans, and thus, these findings may not accurately reflect already-known racial trends of WM pathology. Further studies should explore the differential effects of obesity on white matter integrity, accounting for a variety of demographic factors. A limitation of this study is the omission of proinflammatory cytokines as covariates. Though some proinflammatory markers, including C-reactive protein, IL-6, IL-10, and TNF-α, were collected, data was incomplete for about half of the participants, and thus these covariates were excluded to preserve stratified sample sizes. Secondly, separate preprocessing procedures to more precisely identify MRI artifacts outside of eddy current correction, such as Gibbs artifacts, were not run due to computing restraints. Present downstream analysis results may have shown reduced variability if more comprehensive preprocessing was available, though the study does utilize newer scanners that add an improved signal-to-noise ratio compared to the scanners utilized in previous literature53. Another limitation is the relatively small sample size, especially after stratification. This was particularly problematic as measures such as VAT and SAT could vary widely between individuals, and though measures were taken to control for covariates, the wide distribution and large standard deviation of the data limited the power of the TBSS-driven regressions. Some groups, such as Asians, were too few to include in stratified analyses, while other racial groups were not represented in the cohort, thus limiting the applicability of the study’s findings to the general population. The MHO group was particularly small, numbering only 8 participants, hindering the statistical significance of tests and the interpretability of results. As a result, many of the statistical regressions, especially the associations between VAT and EDI, as well as direct EDI comparisons, were statistically insignificant, though the direction of the relationships was consistent with previous literature. Furthermore, the cross-sectional nature of this study made it difficult to assess clinical and functional outcomes of participants over time, though at this study’s time point, all were deemed cognitively healthy.

Despite these limitations, this study suggests that obesity, as quantified by VAT and SAT, manifests in differential neuroinflammatory effects at the imaging level as evidenced by EDI. A prevailing mechanism of this observation is the dichotomy of MHO and MUO, which may be differentiated by the relative amount of VAT, SAT, and the VAT/SAT ratio, with increased VAT/SAT being associated with WM damage and neuroinflammation, and increased SAT being associated with neuroprotection. Furthermore, the areas of WM damage seen in this study are spatially consistent with WM changes seen in dementia, thus corroborating the pathological association between obesity and dementia. These patterns may be applied to help identify those who may metabolically or clinically be at risk for dementia.

EDI continues to be a useful tool in assessing the extent of white matter pathology in inflammatory conditions such as obesity, which may predispose individuals to dementia. Our study demonstrates adverse neurological outcomes of MUO, as evidenced by decreased WM integrity in those with increased VAT, and the importance of distinguishing MUO from MHO. Further studies should explore detailed effects of demographic factors and proinflammatory markers, incorporate functional and clinical outcomes longitudinally, and apply similar methodology to larger cohorts and cognitively impaired patients.

Methods

Participant selection

A total of 80 participants were recruited under the Research Participant Registry/Volunteers for Health (VFH) Program at Washington University in Saint Louis (https://vfh.wustl.edu), of whom 50 had completed structural and diffusion magnetic resonance imaging (MRI) data acquisition that was successfully processed and passed quality control. All participants were cognitively normal middle-aged adults who achieved a Mini Mental Status Exam (MMSE) score of 25 or higher and/or Clinical Dementia Rating Scale (CDR®) of 054. Additional recruitment was performed through referrals from the Washington University Knight Alzheimer’s Disease Research Center (ADRC) and the Center for Human Nutrition. All participants consented to undergo an approved internal review board protocol. Clinical and laboratory tests, which included vital signs, a cholesterol panel, hemoglobin A1c, and a fasting blood glucose level, were performed for each participant. Exclusion criteria for imaging included (1) claustrophobia, (2) implanted medical devices deemed potentially unsafe for imaging, (3) women currently pregnant or breast-feeding, (4) involvement in an active bariatric surgery or obesity drug study or trial, and (5) prior bariatric surgery. All study protocols were approved by Washington University’s Institutional Review Board (IRB #202102186) and conducted in accordance with the Declaration of Helsinki.

VAT & SAT quantification

A Siemens Prisma Fit 3T scanner was utilized to acquire a transverse T1-weighted two-dimensional fast low-angle shot (FL2D) of the torso from the S1-L5 vertebral bone extending proximally for 22 slices (TR = 286 ms, TE = 3.34 ms, thickness = 10 mm). FL2D images are non-fat saturated, in which fat tissue is bright and all other tissues are dark to easily distinguish fat tissue from other tissues in the abdomen. Automatic identification of abdominal adipose tissue was performed at slices 4–11 of these images by VOXel Analysis Suite (VOXA), a MATLAB script developed institutionally55. The VAT and SAT tissue intensities within the FL2D images were normalized by removing the external background, then using a Euclidean distance transformation with MATLAB’s black-and-white distribution methods. The torso itself was identified via a radial region growing method from the centroid of the bounded box encasing the torso. Within the torso, the dark intensities inside the region defined as SAT and outside of the region defined as VAT were noted as the non-VAT subregion. Each 10 mm slice was then segmented by VOXA into VAT and SAT using erosion, fill, and dilation methods to separate VAT intensity values from SAT intensity values. Volumes were derived from the pixel dimensions and slice thickness and calculated by VOXA. For quality control, each segmentation was visually inspected and corrected if there were any regions of fat not accurately identified. VAT and SAT volumes from each of the eight segmented slices were summed to calculate overall VAT and SAT volumes.

Brain MRI acquisition & fractional anisotropy processing

Diffusion-weighted echo planar imaging (EPI) with a maximum b-value of 2000s/mm2 was utilized to acquire 98 diffusion sampling directions with a resolution of 2 × 2 × 2 mm3 voxels oriented in the axial plane (TR = 7500 ms, TE = 74 ms, flip angle = 90°). Preprocessing of the sequences, including eddy current and motion correction (eddy), was performed utilizing the FMRIB Software Library (FSL; FMRIB Analysis Group, Oxford, UK)56,57. FA was derived from DBSI and was calculated using in-house software scripted in MATLAB and Statistics Toolbox Release (MathWorks, Natick, Massachusetts)55. Images from these FA maps were extracted using brain masks created from non-diffusion-weighted (b0) images with the –mul function within the fslmaths tool (Analysis Group, FMRIB). Each slice was also visually inspected for other artifacts, such as Gibbs artifacts.

EDI map synthesis

EDI processing was performed using the Massively Parallel, Portable, and Reproducible Tractography Pipeline (MaPPeRTrac), a novel probabilistic tractography pipeline designed for high-performance computing systems58,59. MaPPeRTrac uses T1-weighted structural MR image and a diffusion MR image to generate a structural connectome through a series of three overarching stages: (1) brain segmentation and parcellation using FreeSurfer recon-all60, (2) Bayesian modeling of crossing fibers with each voxel using Markov Chain Monte Carlo sampling with FSL (bedpostx)61,62, and (3) probabilistic tractography with FSL (probtrackx2) involving computation of white matter fiber streamlines to generate a structural connectome and an EDI map61,63.

The MaPPeRTrac pipeline starts with diffusion MRI preprocessing to correct image distortions due to susceptibility and eddy currents during the image acquisition. The following steps summarize the pipeline processing: Probabilistic tractography was performed between each pair of regions out of the 82 brain regions based on the Desikan-Killiany atlas that included 68 cortical and 14 subcortical regions64. The 82 regional masks are reconstructed in the first step with recon-all and are used as the start and end points for the tractography algorithm. The affine transformation from the diffusion space to the anatomical space was computed using the FA map and the T1-weighted image. A total of 82 × 81 = 6642 tracts were computed for each connectome. For each of these tracts, the combination of regional masks other than the start- and end-point regions is used as an exclusion mask. The fiber orientation at each voxel from bedpost is used to guide tractography. The resulting image is binarized as a tract mask that informs the WM voxel pathway between the start and end regions. After processing each of the 6642 tracts, the pipeline evaluates the overlapping, or the consensus, between each pair of the A-to-B tract and the B-to-A tract, resulting in a consensus connectome29. By computing the voxel-wise summation of all binarized tract masks that passed the consensus thresholding, the number of tracts or edges passing each voxel is achieved, forming the EDI.

All processing was performed using a high-performance computing server at Washington University in Saint Louis, utilizing both central processing unit and graphical processing unit nodes in parallel.

Obesity classification and statistical analysis of parameter maps

Shapiro–Wilk’s test was used to evaluate the normality of residuals of the distribution of continuous variables. As all of the variables met the normality assumption, unpaired t-tests were performed to compare adiposity metrics between individuals stratified by obese (BMI > 30 kg/m2) and non-obese (BMI < 30 kg/m2), as well as by sex and obesity classification (MHO versus MUO). The threshold utilized to distinguish MHO from MUO was having ≤1 of the following based on the National Cholesterol Education Program Adult Treatment Panel III guidelines: hypertension (defined as systolic blood pressure ≥135, diastolic blood pressure ≥85, taking antihypertensive medications, or self-reported diagnosis of hypertension), hyperglycemia (defined as fasting blood glucose level ≥110 mg/dL, self-reported use of type 2 diabetes medications, or self-reported diagnosis of type 2 diabetes), hypertriglyceridemia (defined as fasting triglyceride levels ≥150 mg/dL or self-reported use of lipid-lowering medication), and low levels of high-density lipoprotein (defined as ≤40 mg/dL in males and ≤50 mg/dL in females, or self-reported use of lipid-lowering medications)23,65. FA and EDI maps were analyzed with the Tract-Based Spatial Statistics (TBSS) tool as part of the freely available FMRIB Software Library (FSL) v5.09 (FMRIB, Oxford, UK.)56,66,67,68,69,70. In TBSS preprocessing, individual FA maps were registered non-linearly into a 1 × 1 × 1 mm3 standard space, followed by an affine alignment to the 1 × 1 × 1 mm3 MNI space. The mean FA image was computed by averaging across registered individual FA maps and was used to produce a mean skeletonized FA image that was thresholded to remove nuisance and extra-cranial signals. Finally, the EDI maps were registered by applying the FA-based registration profiles and projected onto the mean FA skeleton, resulting in 4D-projected EDI data appropriate for voxel-wise analysis.

The Randomize tool from FSL was utilized for voxel-wise analysis of DBSI-derived FA and EDI, generating 1000 permutations of data to threshold statistical maps71. General linear model designs and contrasts were synthesized using the FEAT tool72. With this setup, multiple regressions were run with FA and EDI maps against BMI, VAT, SAT, VAT/SAT, with age, sex, and race as initial covariates, with BMI added as a covariate subsequently to assess confounding effects of BMI. The same regressions were run among males, females, whites, and African Americans, excluding the stratified variable as a covariate. Randomize was also utilized to directly compare EDI between the following groups: individuals with and without obesity, those with MHO and with MUO, males and females, and whites and African Americans. All unpaired t-tests were corrected for multiple comparisons using the Holm–Bonferroni method, using an alpha level of 0.0573. All regression analyses were corrected with threshold-free cluster enhancement (TFCE), which identifies local cluster-like spatial patterns from a statistical image74. These voxel-wise results were corrected for multiple comparisons using a TFCE threshold of 0.05.