Introduction

Obesity is a worldwide health problem characterized by the excessive accumulation of body fat, which leads to several comorbid conditions such as type 2 diabetes, cardiovascular disease, stroke, and various cancers1,2,3. Obesity is a multi-factorial disease affected by environmental, hereditary, and behavioral factors3,4,5. Recent studies have shown that obesity is also associated with alterations in the brain that can be explored using neuroimaging3,6,7,8.

Previous obesity-related neuroimaging studies have measured the functional connectivity of the brain using functional magnetic resonance imaging (fMRI) and found dysfunctions in cognitive control-related brain regions3,8,9,10,11. Specifically, they found that the frontoparietal and executive control networks responsible for cognitive- and inhibitory-controls were strongly associated with binge eating behaviors9,10,11. Structural alterations in reward and cognition-related brain regions have also been observed in people with obesity9,12. Collectively, these results suggest that cognitive control-related brain regions may be important in explaining the behavioral traits of obese subjects.

Relatedly, another recent neuroimaging study reported that a high burden of white matter hyperintensities (WMHs) was associated with obesity13. WMHs are brain lesions that show an aberrant increase in white matter intensity on fluid-attenuated inversion recovery (FLAIR) data. They are related to an increased risk of cognitive decline, dementia, and stroke14,15,16. Some research suggests that white matter vascularization is related to obesity and comorbid metabolic dysfunction13,17,18,19,20. However, the existing neuroimaging literature has not considered the burden of WMHs to stratify the degree of obesity. The present study aimed to address this gap in research by considering the burden of WMHs and the degree of obesity simultaneously.

Connectivity analysis is one of the representative methods to measure brain function21,22. In this study, we adopted a functional connectivity analysis based on graph theory to measure the strength of intrinsic connectivity in the brain21,22. The two fundamental factors of the analysis were nodes and edges. The graph nodes represented brain regions or networks defined using structural atlases or data-driven approaches, such as clustering or independent component analysis (ICA)23,24,25,26,27. The graph edges were defined as the strength of the connection between two different nodes28.

We hypothesized that WMHs and obesity jointly affect the function of cognitive control-related brain regions. In the present study, we aimed to explore changes in functional connectivity with respect to the burden of WMHs and the degree of obesity to assess their interaction effects on the brain connectome. We performed two-way analysis of variance (ANOVA) to compare functional connectivity among four groups stratified by the degree of obesity and WMH burden. The results of the study may provide novel insight into the neurological characteristics of people with both obesity and WMHs.

Methods

Participants

The Institutional Review Board (IRB) of Sungkyunkwan University approved the present retrospective study, which was performed in full accordance with local IRB guidelines. All participants provided written informed consent. T1-weighted, FLAIR, and resting-state fMRI (rs-fMRI) data were obtained from the UK Biobank database29 through application number 34613 entitled “Neuroimaging correlates of obesity.” Among 13,718 participants, 91 did not have waist circumference, hip circumference, or body mass index data, while 29 lacked T1-weighted, FLAIR, or rs-fMRI data and 13,416 did not have WMHs. These participants were excluded. Ultimately, 182 participants were included in the present study. Detailed demographic information is reported in Table 1.

Table 1 Demographic information of the study participants.
Full size table

MRI data acquisition

All imaging data were acquired using a 3T Siemens Skyra scanner. The imaging acquisition parameters of the T1-weighted data were as follows: voxel size = 1 mm3; repetition time (TR) = 2,000 ms; inversion time (TI) = 880 ms; matrix size = 208 × 256 × 256. The FLAIR data were acquired using the following imaging parameters: voxel size = 1.05 × 1 × 1 mm3; TR = 5,000 ms; TI = 1,800 ms; matrix size = 192 × 256 × 256. The rs-fMRI data were obtained with the following imaging parameters: voxel size = 2.4 mm3; TR = 735 ms; echo time (TE) = 39 ms; flip angle = 52°; matrix size = 88 × 88 × 64; number of volumes = 490.

Data preprocessing

The UK Biobank database provided preprocessed imaging data through the FMRIB Software Library (FSL) software30,31. To process the T1-weighted data, gradient distortion corrected data were registered onto the Montreal Neurological Institute (MNI) standard space. Non-brain tissues were then removed via inverse warping of the brain mask of the MNI standard space to the native T1-weighted space. Next, these skull-removed T1-weighted data were segmented into three tissues: cerebrospinal fluid, gray matter, and white matter. Finally, the magnetic field inhomogeneity was corrected. To process the FLAIR data, gradient distortion corrected data were registered onto the T1-weighted data to extract the brain, and the magnetic field inhomogeneity was corrected. To process the rs-fMRI data, gradient distortions and head motions were corrected, intensity normalization of the entire 4D volume was applied, as was high-pass temporal filtering with a sigma of 50 s was applied. Nuisance variables were removed using FMRIB’s ICA-based X-noiseifier (ICA-FIX) approach32.

Specification of WMHs

The UK Biobank database provided WMH masks computed from the Brain Intensity Abnormality Classification Algorithm (BIANCA) software33, which is a supervised machine learning algorithm34 of k-nearest neighbor that uses voxel- and patch-based intensity values of FLAIR data, as well as spatial coordinates of MNI standard space. Leave-one-out cross-validation was used for the training and test procedures. BIANCA produced a probability map of WMHs, which was then thresholded and binarized with a value of 0.5. However, BIANCA sometimes fails to capture small, deep WMHs35; hence, we manually adjusted the WMH masks computed by BIANCA. These adjusted WMHs were further classified into deep and periventricular WMHs. Deep WMHs showed hyperintensities with variable round- or oval-shaped clusters in the white matter on FLAIR images36. Periventricular WMHs showed hyperintensities along the walls of the ventricles, appearing as small caps, thin rims, or confluent lesions on FLAIR images37,38. The WMHs were manually annotated by two investigators (M.J.L., with 9 years’ experience in clinical neurology, and B.P., with 7 years’ experience in neuroimaging analysis). Inter-observer reliability was assessed using the dice coefficient, which yielded values of 0.93 (95% confidence interval [CI]: 0.89–0.97) for total WMHs, 0.93 (95% CI: 0.90–0.96) for deep WMHs, and 0.92 (95% CI: 0.86–0.97) for periventricular WMHs.

Functional connectivity analysis

Functional connectivity analysis was performed using a multi-scale approach. First, a large-scale network-level analysis was performed. Graph nodes (i.e., brain networks) were defined by group ICA23,24, which was performed on the temporally concatenated, preprocessed rs-fMRI data across all subjects using the MELODIC function in the FSL software31. The number of independent components (ICs) was automatically determined based on probabilistic principal component analysis23,24,39. The investigators removed noise ICs by visual inspection and by comparing them with the pre-defined resting-state networks (RSNs); ICs with a cross-correlation value below 0.2 were considered noise ICs. Secondly, a node-level analysis was performed using the automated anatomical labeling (AAL) and Brainnetome atlases to assess consistency across different parcellation schemes26,27. Graph nodes were pre-defined regions of the atlas. In both the large-scale network- and node-level analyses, the mean time series of the rs-fMRI data was extracted for each graph node (i.e., brain network/region). Pearson’s correlation was then calculated for the time series between the two different nodes. The correlation coefficients were soft-thresholded to satisfy scale-free topology using the following formula: {(r + 1)/2}β, where r is the correlation coefficient and β is the scale-free index, which was set to six40,41. The soft-thresholded correlation coefficients were then transformed into z-values using Fisher’s r-to-z transformation. Degree centrality, a graph measure that estimates the strength of the functional connectivity at a given node, was calculated by summing all edge weights connected to a given node21,22. Degree centrality values were adjusted for age and sex.

Group comparison

The degree centrality values of the brain networks/regions were compared among the four groups stratified by burden of WMHs and degree of obesity: (1) low WMH burden, non-obese (lw-no); (2) low WMH burden, obese (lw-o); (3) high WMH burden, non-obese (hw-no); (4) high WMH burden, obese (hw-o) (Table 2). The cutoff value between high and low WMH burden was the median WMH volume. There is no consensus regarding how to distinguish high and low WMH burden. Thus, in the present study, we stratified the groups using a data-driven approach based on the median WMH volume from all subjects (n = 182). However, further studies are needed to validate the usage of median WMH volume as the cutoff. To explore the differences between deep and periventricular WMHs, the median deep and periventricular WMH volumes were considered in addition to the total WMH volume. The waist-hip ratio was used instead of body mass index to stratify the obese and non-obese groups because it is a well-defined measure of metabolically unhealthy obesity, which is strongly associated with obesity-related complications such as diabetes and cardiovascular diseases42,43,44,45,46. The obese groups had a waist-hip ratio larger than 0.9 in males and 0.85 in females47. Two-way ANOVA was applied to the factors of WMH burden (low vs. high) and degree of obesity (non-obesity vs. obesity) to assess both the main effects and the interaction effects between WMH burden and degree of obesity. Both F- and p-values were calculated. Post-hoc analysis was performed using the two-sample t-test; both T- and p-values were calculated. All p-values were corrected using the false discovery rate (FDR) suggested by Benjamini and Hochberg48. The p-values of the two-way ANOVA were corrected for the number of brain regions, while those of the post-hoc analysis were corrected for the number of group comparisons.

Table 2 Demographic information of the study participants in each group.
Full size table

Results

Large-scale network-level analysis

Group ICA was performed to define large-scale brain networks. Forty-two ICs were automatically generated, and seven noise ICs were excluded. The 35 functionally interpretable ICs (mean correlation with RSN: 0.31, standard deviation [SD]: 0.14) were considered as graph nodes (Fig. 1). ICs 1–6 were visual networks, 7–12 were default mode networks, 13–26 were frontoparietal networks, 27–29 were executive control networks, 30–34 were sensorimotor networks, and 35 was an auditory network. Two-way ANOVA was performed to assess the interaction effects between WMH burden and degree of obesity using the degree centrality values. No network showed significant interaction effects. However, a significant main effect of degree of obesity was found in the frontoparietal network (IC #15; Fig. 2; F(1,178) = 13.471, p < 0.001 for total WMHs; F(1,178) = 13.299, p < 0.001 for deep WMHs; F(1,178) = 12.823, p < 0.001 for periventricular WMHs).

Figure 1
figure 1

The 35 functionally interpretable independent components (ICs).

Full size image
Figure 2
figure 2

Between-group comparison at the large-scale network-level. (A) Frontoparietal network (IC #15) mapped onto the brain surface. The color map represents the effect size of the z-statistic values derived from the FSL software. (B) Degree centrality values of all independent components (ICs) in each group. The black boxes represent the frontoparietal network (IC #15). Brain images were visualized using the BrainNet Viewer49. lw, low WMH burden; hw, high WMH burden; o, obese; no, non-obese.

Full size image

Node-level analysis

A node-level analysis using AAL and the Brainnetome atlas was performed to assess the interaction effects between WMH burden and degree of obesity at a finer level. Using the AAL atlas, significant interaction effects were found in the right orbitofrontal cortex (F(1,178) = 5.646, p = 0.0190) and right dorsal medial prefrontal cortex (Fig. 3; F(1,178) = 4.344, p = 0.0390) if periventricular WMHs were considered (Fig. 3). The post-hoc analysis revealed higher degree centrality values in the hw-o group than in the other groups for both the orbitofrontal and dorsal medial prefrontal cortices (Table 3). No interaction effects were identified when total or deep WMHs were considered. To assess the consistency of the results across different parcellation schemes, we derived additional results using the Brainnetome atlas, which was defined using multimodal (i.e., structural and functional) connectivity information27. When total WMHs were considered, a significant interaction effect was found in the right orbitofrontal cortex (A12/47o; Fig. 4; F(1,178) = 4.381, p = 0.0380). The post-hoc analysis revealed higher degree centrality values in the hw-o group than in the lw-no and lw-o groups (Table 4). No significant interaction effects were observed if deep WMHs were considered. For periventricular WMHs, the right orbitofrontal cortex (A12/47o; F(1,178) = 5.659, p = 0.0180) and left orbitofrontal cortex (A11; F(1,178) = 4.979, p = 0.0270) showed significant interaction effects. In the right orbitofrontal cortex (A12/47o), post-hoc analysis exhibited higher degree centrality values in the hw-o group than in the other groups, while in the left orbitofrontal cortex (A11), the hw-o group showed higher degree centrality values than the hw-no group (Table 4). The results derived from both the AAL and Brainnetome atlases consistently showed significant interaction effects and post-hoc results in the orbitofrontal cortex, indicating that this is a key region linking WMH and obesity.

Figure 3
figure 3

Between-group comparison at the node-level using the AAL atlas when periventricular white matter hyperintensities (WMHs) were considered. (A) Degree centrality values of all regions in each group. The inferior orbitofrontal cortex (region #16) and the dorsal medial prefrontal cortex (region #24) are represented with black boxes. (B) The three identified regions, with corresponding degree centrality values in all groups. Significant differences are shown with asterisks. Brain images were visualized using the BrainNet Viewer49. lw, low WMH burden; hw, high WMH burden; o, obese; no, non-obese.

Full size image
Table 3 Post-hoc analysis at the node-level using the AAL atlas when the periventricular WMHs were considered. Significant results are reported in bold italics.
Full size table
Figure 4
figure 4

Between-group comparison at the node-level using the Brainnetome atlas. (A) Degree centrality values of all regions for each group when total (left) and periventricular WMHs (right) were considered. The black boxes represent the orbitofrontal cortex (A12/47o and A11). (B) The identified brain regions when total (left) and periventricular WMHs (middle and right) were considered. (C) Degree centrality values of the identified regions in all groups. Significant differences are shown with asterisks. Brain images were visualized using the BrainNet Viewer49. lw, low WMH burden; hw, high WMH burden; o, obese; no, non-obese.

Full size image
Table 4 Post-hoc analysis at the node-level using the Brainnetome atlas.
Full size table

Discussion

In the present study, we used a multi-scale approach to explore differences in functional connectivity associated with WMH burden and degree of obesity. We found that in the frontoparietal network, the orbitofrontal cortex was jointly associated with WMH burden and degree of obesity, while the parietal networks were only related to the degree of obesity. These results indicate that the orbitofrontal cortex is a key region linking WMH and obesity, and that the frontoparietal network is primarily related to the degree of obesity, but not WMH burden.

Frontoparietal network is involved in the cognitive control system, controlling inhibitory behaviors50,51,52. It sends inhibitory signals to the limbic area to suppress the feeling of hunger53. Previous studies have demonstrated that perturbed connections between the prefrontal cortex, striatum, and limbic regions disrupted the balance between cognition and reward systems, leading to binge eating disorders54,55,56,57. In our previous studies, we reported that dysfunction in the frontoparietal network was associated with obesity via a mechanism involving disinhibited eating behaviors, and that participants with such dysfunction had concerns about their eating habits, shape, and weight9,10,11. These studies collectively suggest that the frontoparietal network is crucial in explaining the behavioral traits of individuals with obesity, and our current findings largely corroborate these results, linking obesity with altered functional connectivity in the frontoparietal network.

At the finer node-level, we observed that both WMH burden and degree of obesity affected functional connectivity in the orbitofrontal cortex, which controls the reward system by encoding food-related reward responses and inducing the feeling of hunger53,58,59,60,61,62. In addition, the orbitofrontal cortex is involved in the cognitive control system of inhibitory processing53,58. One previous study showed that the dysfunctional inhibitory control that leads to overeating is related to an increased demand for reward processing, suggesting links between the reward and cognitive control systems61. These studies collectively indicate that the identified regions are related to cognitive function, which is highly associated with WMHs14,15,16. Our results suggest that increased WMH burden in obesity affects altered functional connectivity in the orbitofrontal cortex that controls response inhibition and reward processing. We believe that such changes may contribute to further aberrant eating behavior, though further studies need to confirm this hypothesis.

The most important risk factors for WMHs are age and hypertension63,64,65,66. Almost half of the population has WMHs in midlife, and the incidence increases with age63,64,65,66. Previous studies have reported that the prevalence of WMHs in elderly subjects is associated with cognitive and functional impairment65, as well as with cortical thinning in the frontal areas of individuals with mild cognitive impairment and dementia66. Such changes lead to altered executive functions. These studies collectively indicate that the WMH burden is linked to cognitive decline with aging. Previous studies have demonstrated that obesity is related to a decline in cognitive function67,68. Using diffusion tensor imaging, Zhang et al. found that visceral obesity was associated with executive functions68, while Fitzpatrick et al. observed that mid-life obesity was strongly related to an increased risk of dementia67. Another previous study found that visceral obesity, one of the major risk factors of cognitive decline69,70, was associated with the presence of WMHs13. Taken together, these studies suggest that cognitive control-related function may be affected by both WMH burden and degree of obesity. To our knowledge, the current study was the first to link WMH burden with degree of obesity in terms of functional connectivity. The results may provide insight into cognitive function in individuals with both obesity and WMHs. To validate our findings, future studies should explore the associations among WMH burden, degree of obesity, eating behaviors, and cognition-related clinical scores such as Mini-Mental State Examination (MMSE) and Clinical Dementia Rating (CDR). These parameters could not be compared in our current study because the UK Biobank database does not provide these clinical scores. Further longitudinal studies are also required to fully validate how changes in cognitive function are related to WMH burden and degree of obesity.

Our present study had several limitations. First, obesity is affected by many factors, including hormones and toxins other than those listed in Table 1. Hormones such as leptin and ghrelin convey appetite-related information to the hypothalamus, regulating eating behavior71,72. These factors should be controlled for in the compared groups. However, we could not do so in the present study because we downloaded the data retrospectively from the UK Biobank database. Second, although there are many centrality measures, we only used degree centrality to quantify complex brain networks because it is a convenient graph measure to associate brain imaging with obesity9,11. Different graph centrality measures quantify different aspects of the brain network, and future works should explore these73. Third, the number of participants was relatively small compared to that in previous studies13. We had difficulty obtaining large-scale T1-weighted, FLAIR, and rs-fMRI data with sufficient WMHs. Future studies with larger samples are necessary to fully validate the results of our current study. In a similar vein, future studies should consider a multi-center/multi-database approach to verify whether the results can be generalized to many different cohorts. Lastly, total, deep, and periventricular WMHs are not independent of one another. Studies are indeterminate regarding the distinct pathophysiological backgrounds of deep and periventricular WMHs65,66,74,75,76,77,78. Periventricular WMHs show a strong association with age, hypertension, and cognitive decline, and are primarily observed in middle-aged and elderly individuals65,66,74,75, while deep WMHs are prevalently observed in young adults with migraine65,76,77. The present study was exploratory and did not aim to confirm any hypothesis regarding which WMH types are more related to obesity. Thus, we considered two subtypes of WMH (i.e., deep and periventricular), as well as total WMHs. Further studies should confirm the clear relationship between WMH subtypes and obesity.

The present study explored differences in brain functional connectivity with respect to WMH burden and degree of obesity. Among the frontoparietal network, which is largely associated with cognitive control function in individuals with obesity, the orbitofrontal cortex was identified as the key region involved in the link between WMH and obesity. The results of our study may provide a rationale for exploring the link between WMH burden and cognitive control functions in people with obesity.