• obesity;
  • brain structure;
  • aging;
  • BMI;
  • MRI


  1. Top of page
  2. Abstract

Although a link between body mass index (BMI) and brain volume has been established in several cross-sectional studies, evidence of the association between change in BMI over time and changes in brain structure is limited. Using data from a cohort of 347 former lead workers and community controls with two magnetic resonance imaging scans over a period of ∼5 years, we estimated cross-sectional and longitudinal associations of BMI and brain volume using both region of interest (ROI) and voxel-based morphometric (VBM) methods. We found that associations of BMI and brain volume were not significantly different in former lead workers when compared with community controls. In the cross-sectional analysis, higher BMIs were associated with smaller brain volumes in gray matter (GM) using both ROI and VBM approaches. No associations with white matter (WM) were observed. In the longitudinal analysis, higher baseline BMI was associated with greater decline in temporal and occipital GM ROI volumes. Change in BMI over the 5-year period was only associated with change in hippocampal volume and was not associated with change in any of the GM ROIs. Overall, higher BMI was associated with lower GM volume in several ROIs and with declines in volume in temporal and occipital GM over time. These results suggest that sustained high body mass may contribute to progressive temporal and occipital atrophy. Hum Brain Mapp 35:75–88, 2014. © 2012 Wiley Periodicals, Inc.


  1. Top of page
  2. Abstract

Obesity has rapidly increased in prevalence across the United States in the last three decades [Baskin et al., 2005; Flegal et al., 2010]. The health consequences of overweight and obesity, which include type 2 diabetes, coronary heart disease, osteoarthritis, and certain cancers, among others, have been well established [Must et al., 1999]. Additionally, there is a growing body of evidence linking obesity with greater cognitive decline and higher risk of neurodegenerative diseases, such as Alzheimer's disease [Beydoun et al., 2008; Cournot et al., 2006; Whitmer et al., 2005]. However, the mechanisms by which obesity may impact cognitive function are not clearly understood. A better understanding of the associations between obesity and brain volume may elucidate the role of obesity as a risk factor for neurodegenerative diseases.

One measure of obesity, body mass index (BMI), has been correlated with atrophy in several areas of the brain. Although prior studies of BMI and brain volumes have been published, these generally had small sample sizes, were cross-sectional in design, used either region of interest (ROI) or voxel-wise approaches, but not both, and most did not control for intracranial volume (ICV) in the analyses [Driscoll et al., 2011; Gustafson et al., 2004; Ho et al., 2010a, 2010b; Pannacciulli et al., 2006; Raji et al., 2010; Taki et al., 2008; Walther et al., 2010]. In addition, we found just two studies looking at the longitudinal association of BMI and brain volume, although neither of these had both BMI and brain volume measured at multiple time points, and neither found any longitudinal association [Driscoll et al., 2011; Gustafson et al., 2004].

In this study, we leverage rich longitudinal data on BMI and brain volume while avoiding limitations of prior work. First, we consider analysis of brain volume on two scales: an ROI approach that targets large-scale diffuse, nonlocalized effects in the brain as well as highly localized effects in smaller substructures, and voxel-based morphometry (VBM) that investigates localized brain volumetrics without a priori identification of regions. Second, in our multivariate regression models, we adjust for a measure of ICV. Although clearly ICV is an important predictor of regional and voxel-level brain volume, it has not consistently been controlled for in studies of the BMI-brain volume association. Third, we consider separate gray matter (GM) and white matter (WM) ROIs and conduct voxel-wise analyses separately for GM and WM partitions, which enables the detection of differences in the BMI–brain volume association in the two types of tissue. Fourth, by studying longitudinally BMI measures and magnetic resonance imaging (MRI) scans from two time points, we can assess several features of the relation between BMI and brain volume, including temporality, persistence, and reversibility. Finally, we conduct our study on a cohort of 347 middle-aged men having MRI scans at two visits, the largest study to thoroughly investigate the BMI–brain volume association in healthy American subjects.


  1. Top of page
  2. Abstract

The study population, study design, data collection, image acquisition, and other research methods have been reported previously [Schwartz et al., 2000; Schwartz et al., 2007, 2010; Stewart et al., 1999; Stewart et al., 2006]. The most relevant information will be briefly summarized in the following sections.

Study Population and Design

Male subjects from a population of former lead workers from a chemical manufacturing plant in the eastern United States as well as population-based controls with no prior occupational lead exposure were recruited in three study phases. In phase I (1994–1997), the initial cohort was enrolled, and during phase II (2001–2003) additional study participants were recruited and the first MRI was obtained [Stewart et al., 2006]. Subjects who completed the first MRI were invited for a second MRI in phase III (2005–2008) [Schwartz et al., 2007]. There were 352 individuals from phases I–III having two acceptable MRI scans that satisfied automated processing procedures. Of these, five (1.4%) were missing data for one of the covariates needed for the present study (see “Statistical Analysis” section) and so our final cohort consisted of 347 participants. The Johns Hopkins Bloomberg School of Public Health Committee on Human Research reviewed and approved each phase of the study, and all participants provided written informed consent.

Data Collection

Data were collected at seven clinic visits over time. Information on age, race, smoking history, health outcomes and other study variables were collected for each subject with an interviewer-administered questionnaire. At each visit height and weight were recorded, from which we calculated BMI (in kg/m2). The BMI measures obtained closest to the date of MRI acquisition for the two scans were used in subsequent analyses. Apolipoprotein E, a gene whose ε4 allele has been associated with Alzheimer's disease and impaired cognitive function [Small et al., 2004], was genotyped in phases I and II using methods described elsewhere [Stewart et al., 2006]. Tibia lead, a measure quantifying cumulative lead exposure, was collected from former lead workers at the first visit. Because bone lead levels naturally degrade over time, current tibia did not adequately capture the highest levels of lead exposure, as time since exposure varied across subjects and also influences this measure. Thus, current tibia lead was used to derive peak tibia lead (PTL), an estimate of lead exposure at the end of employment using previously published methods [Stewart et al., 1999]. Higher PTL has been associated cross-sectionally with lower cognitive test scores [Stewart et al., 1999], with longitudinal decline in performance on cognitive tests [Schwartz et al., 2000], and with lower brain volumes using both ROI and voxel-wise approaches [Schwartz et al., 2010; Stewart et al., 2006].

MRI Acquisition and Preprocessing

Images were obtained using a General Electric 1.5-T Signa model at the first time point. A 3-T General Electric scanner was used for the second MRI. T1-weighted images, axial proton density/T2 images, and fluid-attenuated inversion recovery images were acquired as reported previously [Schwartz et al., 2010]. For eighteen MRIs at the first scan and nine at the second, image quality was poor, and so these were excluded from subsequent volumetric analysis. The two scans were acquired, on average, ∼5 years apart.

Images were preprocessed and then segmented into GM, WM, and cerebrospinal fluid partitions, as previously documented [Goldszal et al., 1998; Stewart et al., 2006]. To transform brain images into a standard template space while preserving individuals' volumetric information, regional analysis of volumes in normalized space (RAVENS) was applied [Davatzikos, 1996]. The resulting RAVENS images consist of a 3D array of voxels (three-dimensional pixels), where the value at each voxel for each subject represents the volume for that location in the standardized space. To account for changes in scanner technologies and pulse sequences between MRI acquisitions, we used the previously validated Consistent Longitudinal Alignment and Segmentation for Serial Image Computing (CLASSIC) algorithm [Xue et al., 2006]. CLASSIC uses 4D image segmentation to jointly segment multiple 3D images collected over time, a process done before calculating RAVENs images. Using previously published methods [Shen and Davatzikos, 2002; Stewart et al., 2006], regional image analysis was conducted to obtain volumes of 20 nonmutually exclusive, prespecified ROIs.

Statistical Analysis

We first conducted a cross-sectional analysis of the association between BMI and brain volume using both ROI and VBM approaches. We then examined the association of longitudinal change in ROI volumes with (1) BMI at baseline, and (2) change in BMI from the first to second MRI.

Cross-sectional analysis
ROI approach

We considered four cross-sectional models. In the first model, the expected volume of the kth ROI (k = 1, …, 20) for the ith individual at the first MRI time point was modeled as

  • display math(1)

where race is an indicator comparing non-whites to whites and APOE is a vector of indicator variables comparing the genotypes 24, 34, 44, and 22 or 23 to the reference genotype 33, age1 is age at the time of the first scan, inline image is the average age of the cohort (to center age), smoke1 is a vector of smoking status indicators at the first MRI (comparing current and prior to never smokers), ICV1 is ICV obtained at the first scan, and BMI1 is BMI measured at the visit corresponding to the first MRI (baseline BMI). ICV was calculated as the sum of the volume of the GM, WM, and cerebrospinal fluid. The second cross-sectional model we considered was the same as (1) except it also included a quadratic term for the centered age variable; the third model contained the second but also added an indicator variable for control versus former lead worker status; and the fourth was the same as the second but also added PTL. Note that since PTL was only measured for former lead workers, the fourth model was fit only for that subset of the cohort.

VBM approach

Separately for the GM and WM partitions, we conducted VBM to identify clusters of contiguous voxels for which there was an association with BMI. For each of the V voxels, we used least squares estimation to fit the third cross-sectional model (Equation (1) with the control status indicator and quadratic age term), where yikrepresented the volume at the kth voxel for subject i. From the fitted models, we calculated the observed t-statistics inline image where inline image is the estimated coefficient of BMI1 for the kth voxel and inline image is its estimated standard error (k = 1,…,V). We then identified the set of voxels for which the test statistics Tk were below a particular threshold, since it was hypothesized (as evidence from previous studies has suggested) that brain volume was negatively associated with BMI. Specifically, we identified all voxels k such that Tk < −3.1, where the threshold corresponds approximately to an unadjusted P-value of 0.001. We then located all sets of conjoining voxels (clusters) among the full set of voxels below the threshold.

To assess the statistical significance of the largest cluster, we performed a permutation test as follows. Let yi = (yi1,…, yiV) be the vector of voxel-wise volumes for the ith subject. We repeated the process just described for 250 permutations of the image data yiacross subjects, identifying all clusters for each permutation. By permutation of the imaging data, we mean that the entire vector yiwas permuted relative to covariates, thus breaking any associations. By virtue of not having permuted within a vector, spatial covariances were retained. A P-value for the largest cluster of the observed data was then calculated as the proportion of times that the size of the maximal permuted cluster exceeded the size of the maximal observed cluster [Nichols and Holmes, 2002].

Longitudinal analysis

We considered two models of longitudinal change,

  • display math(2)
  • display math(3)

where y1ik, y2ikare the brain volumes of the kth ROI for the ith subject measured at visits one and two, respectively, and E(y2ik- y1ik) denotes the expected change. These two models quantified two types of relation between BMI and longitudinal change in volume. Model (2) evaluated persistent versus progressive associations, whereas (3) evaluated whether the association was reversible. The term γ′kzi, the adjustment for potential confounders is given by:

  • display math

where control is the indicator of control status (versus former lead worker).


  1. Top of page
  2. Abstract

Descriptive Statistics

Table 1 shows summary statistics of cross-sectional and longitudinal covariates of interest for the present analysis and Figure 1 shows the distribution of baseline BMI and change in BMI between the two MRI time points. Additional descriptive statistics for the full cohort stratified both by former lead worker versus control status and by the number of MRI scans (0, 1, or 2) have been previously reported [Schwartz et al., 2010].

Table 1. Select summary statistics over two visits, by baseline BMI category
  Baseline BMI 
 AllNormalOverweightObeseP-value by BMI Category
  1. a

    Visit 1 corresponds to the first MRI time point and Visit 2 to the second MRI acquisition. Obese: BMI ≥ 30, Overweight: 25 ≤ BMI < 30, Normal: BMI < 25.

N (%)347 (100)47 (13.5)158 (45.5)142 (40.9) 
Age, (yrs), mean (SD)     
Visit 1a60.1 (7.8)61.5 (8)60 (8.2)59.8 (7.3)0.416
Visit 265.1 (7.8)66.5 (8)65 (8.2)64.7 (7.4)0.381
BMI, (kg/m2)     
Visit 1, mean (SD)29.3 (4)23.5 (1.2)27.7 (1.4)33.1 (2.9) 
Visit 2, mean (SD)29.9 (4.2)24.8 (2.8)28.4 (2.5)33.3 (3.1) 
APOE genotype, N (%)     
ε2/22 (0.6)0 (0)2 (1.3)0 (0)0.233
ε2/343 (12.4)10 (21.3)13 (8.2)20 (14.1) 
ε2/411 (3.2)2 (4.3)6 (3.8)3 (2.1) 
ε3/3211 (60.8)23 (48.9)98 (62)90 (63.4) 
ε3/475 (21.6)11 (23.4)38 (24.1)26 (18.3) 
ε4/45 (1.4)1 (2.1)1 (0.6)3 (2.1) 
Control status
N (%)42 (12.1)7 (2)18 (5.2)17 (4.9)0.810
Peak tibia lead, μg/g
N (% of controls)294 (96.4)39 (97.5)134 (95.7)121 (96.8)0.841
Mmean (SD)21.5 (15.6)21.1 (14.9)20.9 (14.3)22.3 (17.2)0.759
Smoking status, N (%)     
Never115 (33.1)18 (38.3)47 (29.7)50 (35.2)0.489
Current52 (15)6 (12.8)29 (18.4)17 (12) 
Prior180 (51.9)23 (48.9)82 (51.9)75 (52.8) 
Non-white, N (%)26 (7.5)1 (2.1)13 (8.2)12 (8.5)0.342

Figure 1. Distribution of baseline BMI and change in BMI. Vertical lines correspond to BMI category cutoffs, Obese: BMI ≥ 30; Overweight: 25 ≤ BMI < 30; Normal: BMI < 25. [Color figure can be viewed in the online issue, which is available at]

Download figure to PowerPoint

An average (SD) of 5.0 (0.4) years passed between the two MRI acquisitions (range: 3.5–6.1). The majority of subjects (86%) were either overweight or obese, with BMIs ranging from 21.0 to 43.3 (Fig. 1). BMI increased an average (SD) of 0.5 (2.4) kg/m2 between the two MRI time points. The majority of individuals maintained a BMI within 3 kg/m2 of the baseline measurement, which corresponds to a change in weight of 6.6 pounds (assuming no change in height), while 31 individuals gained weight and 17 lost weight above this threshold. We found no statistically significant difference in the distribution of each of the covariates across BMI categories (Table 1).

Figure 2 shows boxplots of the change in the regional volumes between the two MRI acquisitions. In general, across subjects and ROIs, there was a decrease in volume at the second visit as compared to the first. The one exception was for the occipital WM, which increased by an average of 0.56 cm3 (95% CI = 0.38–0.73).


Figure 2. Boxplots of the change in ROI volume* across the 347 study participants. *Volume at the second MRI subtracted from volume at the first MRI, divided by the standard deviation of the change in volume across subjects. Mean (SD) of the unscaled ROIs (cm3) shown on right side of plot. GM = gray matter, WM = white matter. [Color figure can be viewed in the online issue, which is available at]

Download figure to PowerPoint

Cross-Sectional Analysis

ROI approach

We first assessed whether the association of BMI with brain volume differed across former lead workers as compared to community controls. For each model and each ROI, we added an interaction term between the indicator for former lead worker/control status and BMI, finding that in general this did not significantly improve the model fit. The one exception suggested that in the insula, higher BMI may be associated with lower brain volume among former workers (P = 0.001), but the association was not statistically significant among controls (P = 0.33). Thus we concluded that our population of mostly former lead workers was, for the most part, representative of the general population in regards to the cross-sectional association of BMI and brain volume. We therefore combined the former lead workers and controls into a single group for estimating the association between BMI and brain volume.

Table 2 displays the estimated coefficients θk for the ROI analysis, across each of the four models described in “Cross-sectional analysis” section. We found that higher BMI was associated with lower total brain volume, as well as broadly in GM, in addition to other more localized regions. In particular, a negative association was identified in the frontal, parietal, temporal, and occipital regions of the GM, as well as in the medial structures, insula, amygdala, and cingulate gyrus. Overall, higher BMI was associated with lower volume for 10 of 20 examined ROIs. These were either all GM or smaller structures consisting primarily of GM. We did not find an association in any of the WM regions.

Table 2. Cross-sectional associations of BMI with region-of-interest volumes from the first MRI, controlling for different sets of covariates
 Model (1)Model (2)Model (3)Model (4)
  1. a

    Bold values denote significance at the 0.05 level.

  2. Model (1) adjusts for race, APOE genotype, age, smoking status, and ICV; Model (2) adds quadratic term for age to Model (1); Model (3) adds former lead worker status to Model (2); and Model (4) adds PTL to Model (2). Note that in Model (4), the association is among those subjects who had PTL measured (none of the controls and all but 11 of the former lead workers).

Total brain0.989a0.3351.0420.3311.0350.3310.8930.358
Total GM1.3920.2991.4250.2981.4230.2981.5600.331
Total WM0.4030.3350.3830.3350.3880.3360.6660.370
Frontal GM0.4360.1130.4400.1130.4390.1130.4950.126
Parietal GM0.1600.0670.1650.0670.1650.0670.1900.075
Temporal GM0.3340.0970.3490.0960.3500.0960.3970.105
Occipital GM0.1340.0460.1360.0460.1350.0460.1190.050
Frontal WM−0.0500.160−0.0510.161−0.0510.161−0.0170.170
Parietal WM−0.0340.108−0.0440.108−0.0430.108−0.0450.120
Temporal WM0.0630.1010.0530.1010.0520.1010.1260.111
Occipital WM0.0660.0710.0660.0710.0680.0710.0870.077
Medial structures0.2270.0610.2290.0610.2270.0610.2430.068
Entorhinal cortex0.0020.0060.0020.0060.0020.0060.0030.007
Cingulate gyrus0.0920.0350.0930.0360.0930.0360.1120.039
Ventricular volume0.0450.1470.0620.1460.0660.1460.1360.156
Internal capsule−0.0140.013−0.0140.013−0.0140.013−0.0070.014

As a sensitivity analysis, we checked for a nonlinear association of BMI by adding a quadratic baseline BMI term to each of the first two cross-sectional models (Models (1) and (2) in Table 2), and found that it did not significantly improve the model fit. Finally, we compared our ROI results from the third cross-sectional model (shown in Table 2) to results from (i) an analysis that did not include ICV, and (ii) one that used an individual's baseline height as a surrogate for ICV. Table 3 shows that the associations from the analysis including ICV were attenuated if ICV was removed from the model, and that three highly significant regions became only marginally significant (occipital and parietal GM) or nonsignificant (total brain volume). When baseline height was used as a surrogate for ICV, the associations of BMI with ROI volumes became further attenuated as compared to the model that adjusted for ICV directly.

Table 3. Comparison of cross-sectional associations of BMI with region-of-interest volumes when model does not control for intracranial volume
 Without ICVWith height
  1. a

    Bold values denote significance at the 0.05 level.

  2. Models adjust for the same covariates as cross-sectional Model (3) (results shown in Table 2) except for ICV, that is, race, APOE genotype, smoking status, quadratic age association, and former lead worker status. Associations shown from model that excludes ICV (left column) or that replaces ICV with baseline height (right column).

  3. ICV = intracranial volume

Total brain−0.8321.300−0.5611.288
Total GM1.337a0.6131.2230.609
Total WM0.5040.7960.6620.789
Frontal GM0.4170.1770.3900.176
Parietal GM−0.1550.091−0.1340.090
Temporal GM0.3310.1520.3160.152
Occipital GM−0.1270.066−0.1200.066
Frontal WM−0.0090.3050.0450.303
Parietal WM−0.0220.1690.0140.167
Temporal WM0.0760.1810.1040.181
Occipital WM0.0780.0950.0880.095
Medial structures0.2130.106−0.2020.106
Entorhinal cortex0.0030.0070.0030.007
Cingulate gyrus0.0880.044−0.0820.044
Ventricular volume0.0740.1540.0720.155
Internal capsule−0.0120.017−0.0100.016
VBM approach

Figure 3 shows histograms of the number of contiguous voxels of the maximal cluster across permutations for gray and WM VBM analyses. For the GM partition, the largest observed cluster consisted of 7,445 contiguous voxels, and there were four permutations with larger maximal clusters. The corresponding P-value of the maximal observed cluster was 0.02. The second largest cluster, consisting of 7,001 contiguous voxels, had a P-value (with respect to the distribution of maximal permuted clusters) also of 0.02. Figure 4 shows images of the largest and second largest observed cluster of the GM. For the WM partition, the largest observed cluster consisted of 1,619 contiguous voxels. Fifty-two of the permutations had larger maximal clusters, and the corresponding P-value of the maximal observed cluster was 0.21.


Figure 3. Distribution of sizes of the maximal clusters of the permuted data sets for gray and white matter VBM. Vertical line denotes the size of the largest observed cluster. [Color figure can be viewed in the online issue, which is available at]

Download figure to PowerPoint


Figure 4. Images of the largest (left) and second largest (right) observed gray matter cluster. Red, yellow, and blue regions correspond to unadjusted voxel-wise p-values of <10−5, <10−4, and <10−3, respectively.

Download figure to PowerPoint

Longitudinal Analysis

Table 4 shows estimates from the longitudinal analysis, and Figure 5 shows residual plots for the temporal GM ROI under model (2) and for the hippocampus ROI under model (3). Under the model of baseline BMI and longitudinal change in regional volumes (2), seven regions had P-values less than 0.05. In comparison with the cross-sectional models which suggested associations primarily in GM, both gray and WM regions were identified, and only two of these (temporal and occipital GM) were consistent with ROIs associated in the cross-sectional analysis. Higher baseline BMI tended to be associated with a smaller decline in WM ROI volumes over time, but with larger declines in the two GM ROI volumes over time. Under the model of change in BMI and change in volumes (3), experiencing a larger weight gain (or less weight loss) over the 5-year period was associated with experiencing a smaller decline only in hippocampal volume by the second visit.

Table 4. Longitudinal associations of BMI with region-of-interest volumes
 Model (2)Model (3)
  1. a

    Bold values denote significance at the 0.05 level.

  2. Model of the change in ROI volume versus BMI at the first visit (Model 2) and versus the change in BMI from the first to second visit (Model 3). Both models adjust for ROI volume at the first visit, age at the first visit, change in age between the visits, race, APOE genotype, smoking status, former lead worker status, and ICV. Model (3) additionally controls for baseline BMI.

Total brain−0.0410.2700.6640.469
Total GM−0.3480.2110.1640.361
Total WM0.402a0.1570.4730.275
Frontal GM−0.0940.0610.1030.105
Parietal GM−0.0480.0320.0520.055
Temporal GM0.1210.050−0.0180.088
Occipital GM0.0630.022−0.0100.039
Frontal WM0.2360.0670.1650.117
Parietal WM0.1440.0380.0720.068
Temporal WM0.0540.0400.1000.071
Occipital WM0.0370.0210.0630.037
Medial structures−0.0520.0350.0040.060
Entorhinal cortex0.0010.0020.0030.004
Cingulate gyrus−0.0050.0130.0380.023
Ventricular volume−0.0150.0350.0500.062
Internal capsule0.0120.0050.0090.009

Figure 5. Residual plots from the longitudinal models. For the temporal GM ROI, left panel shows residuals from linear regression of the longitudinal change in volume y2y1on the set of covariates excluding baseline BMI in model (2) plotted against residuals from linear regression of baseline BMI on the same set of covariates. For the hippocampus ROI, right panel shows residuals from regression of the longitudinal change y2y1 on the set of covariates excluding change in BMI in model (3) plotted against residuals from regression of change in BMI on the same set of covariates. Solid lines correspond to a loess smooth and dashed lines corresponds to a linear trend. [Color figure can be viewed in the online issue, which is available at]

Download figure to PowerPoint


  1. Top of page
  2. Abstract

In this manuscript we have pursued a large-scale investigation of the potential relationship between obesity, as measured by BMI, and brain structure, as measured by processed MRI. The study is notable in its scale, with a large number of subjects measured longitudinally, and scope, having processed separately for GM and WM both regional and voxel-based volumetric summaries. To our knowledge this is the first study to investigate the longitudinal association of BMI with brain volume using repeated measures of both BMI and imaging data. Moreover, we applied computationally difficult, yet robust statistical methodology for voxel-level analyses by using permutation testing.

We found intriguing relationships between brain structure and obesity. Primarily in GM, higher baseline BMI was associated cross-sectionally with lower brain volume at baseline and longitudinally with further decline in volume subsequent to the first MRI. Cross-sectional ROI results were robust to model specification and suggested a diffuse impact across the cortex. Cross-sectional VBM demonstrated the possibility for some localized association, identifying two large clusters in the frontal GM in both hemispheres. WM voxel-based analyses did not yield any significant associations. The two longitudinal models we considered quantified different features of the potential temporal association. For the model regressing change in ROI volume against baseline BMI adjusted for baseline volume, associations with two of the ROIs associated in the cross-sectional model were found, suggesting that BMI may lead to progressive atrophy in the temporal and occipital GM regions. These regions were not significantly associated in the model regressing change in ROI volume versus change in BMI, adjusted for baseline BMI, providing no evidence that the association of BMI with change in volumes over time was reversible.

An interesting secondary finding in the manuscript was the importance of a good measure of ICV. In adults, ICV clearly represents a form of intrinsic volumetric capacity for calibration. Using height as a surrogate was insufficient and drastically changed results.

Table 5 summarizes previous studies of the association between BMI and brain volume [Driscoll et al., 2011; Gustafson et al., 2004; Ho et al., 2010a, 2010b; Pannacciulli et al., 2006; Raji et al., 2010; Taki et al., 2008; Walther et al., 2010]. Two of these studies had sample sizes larger than the present analysis, but they did not include longitudinal measurements of both BMI and imaging data. Additionally, only two of the studies adjusted for ICV in their ROI or voxel-wise regression models [Driscoll et al., 2011; Walther et al., 2010]. A third study used as outcome measure the percentage of ICV that was GM in a global volumetric analysis, though their voxel-wise analysis did not similarly adjust for ICV [Taki et al., 2008]. Our cross-sectional findings are consistent with results from previous studies identifying negative associations between BMI and total GM, as well as GM more specifically in the frontal, parietal, temporal, and occipital regions, and in smaller structures (Table 5). Across studies, higher BMI was generally associated with lower brain volumes, though there were some exceptions, including increased brain volumes in specific regions in both the gray [Pannacciulli et al., 2006; Taki et al., 2008] and WM [Walther et al., 2010]. In addition, for some studies only the potential for atrophy was assessed and so any possible positive associations could not be identified [Gustafson et al., 2004; Raji et al., 2010].

Table 5. Overview of previous voxel-based morphometric and longitudinal studies of the BMI–brain volume association
ReferencesDesignCohortBMI VariableMethodsMain findingsCovariate Adjustment
  1. VBM, voxel-based morphometry; TBM, tensor-based morphometry; GM, gray matter; WM, white matter; ROI, region of interest; MCI, mild cognitive impairment; AD, Alzheimer's disease

Pannacciulli et al. (2006)Cross-sectional60 nondiabetic Caucasians (24 obese and 36 lean)Binary: obese (BMI ≥ 30) versus lean (BMI < 25)VBM (separately for GM and WM segmentations)GM density reduced in obese versus lean in R. cerebellum, L. post-central gyrus, R. frontal operculum, R. and L. putamina, R. and L. middle frontal gyri; GM density elevated in obese versus lean in L. calcarine cortex, L. middle occipital gyrus, L. inferior frontal gyrus, and R. cuneus; negative association between GM density of L. post-cenral gyrus in obese but not lean; WM density elevated in obese versus lean in striatal regionSex, age, handedness, global tissue density
Taki et al. (2008)Cross-sectional1,428 Japanese including 27 obese and 273 overweightCategorical: 0 (BMI < 20), 1 (BMI 20–24.9), 2 (BMI 25–29.9), 3 (BMI ≥ 30)Assessed association of BMI with “GM ratio” (whole brain); VBM (GM segmentation)Negative association between BMI and GM ratio (% GM volume in intracranial volume) in men but not women; for men, negative correlation between GM volume and BMI category in bilateral medial temporal lobes, anterior lobe of the cerebellum, occipital lobe, frontal lobe, precuneus, and midbrain; positive correlations in bilateral inferior frontal gyri, posterior lobe of the cerebellum, frontal lobes, temporal lobes, thalami, and caudate headsage, lifetime alcohol intake, history of hypertension, and diabetes mellitus
Raji et al. (2010)Cross-sectional94 healthy elderly subjects (14 obese, 51 overweight, and 29 normal BMI)Categorical: normal, overweight, and obeseTBMBMI associated with atrophy in frontal, temporal, and subcortical brain regions in unadjusted model. In adjusted models, those with BMI > 30 showed atrophy in frontal lobes, anterior cingulate gyrus, hippocampus, and thalamus compared to those with BMI 18.5–25; Atrophy in basal ganglia and corona radiata of WM among overweight (BMI 25–30). Overall brain volume did not differ between overweight and obese subjectsage, gender, race, type 2 diabetes mellitus
Ho et al. (2010, 2010a)Cross-sectional226 healthy elderly subjectsContinuousTBMBMI associated with atrophy in frontal, temporal, parietal, and occipital lobe regions, as well as in brain stem and cerebellar regionage, sex, education level, and physical activity
Ho et al. (2010, 2010b)Cross-sectional700 patients with MCI or ADContinuousTBMNegative association between BMI and brain volume in frontal, temporal, parietal, and occipital lobesage, sex, and years of education
Walther et al. (2010)Cross-sectional95 community- dwelling females, ages 52–92ContinuousVBM (separately for GM and WM segmentations)Higher BMI associated with lower GM volume in L. orbitofrontal, R. inferior frontal, R. precentral gyri, posterior region containing parahippocampal, fusiform, and lingual gyri, and R. cerebellar regions; higher WM volume in frontal, temporal, and parietal lobesage and ICV; sensitivity analysis included hypertension indicator
Gustafson et al. (2004)Longitudinal BMI; single CT scan290 middle-aged Swedish womenContinuousROI (Four regions: temporal, frontal, occipital, and parietal lobes); atrophy visually ratedFound atrophy in temporal lobe but not frontal, occipital, or parietal lobes; in multivariate analysis only age and BMI significant predictors; Increased risk of temporal atrophy of 13-16% per 1.0kg/m2 of BMI; Did not find relation between change in BMI and atrophy; waist-to-hip ratio not associated with atrophyage, diastolic blood pressure, serum triglycerides, education, smoking, socioeconomic status, presence of psychiatric disorder;
Driscoll et al. (2011)Longitudinal MRI; single BMI152 community- dwelling participants, ages 56–86Continuous and Binary (BMI ≥ 30 versus BMI < 25)ROI (16 regions, including both GM and WM)No associations found between BMI at age 50 and subsequent changes in brain volume among those who did not become impaired at follow-up visits; BMI ≥ 30 associated with greater decline in GM volume (total, frontal, temporal), precuneous, cingulate, and orbitofrontal gyri, and higher continuous BMI with greater decline only in cingulate when individuals who later became impaired were includedBackward step-wise selection on ICV, sex, race/ethnicity, education, and smoking status

To our knowledge only one previous study had imaging data available from multiple visits, from which they estimated the association between BMI at age 50 and subsequent longitudinal trajectories of brain volume ROIs [Driscoll et al., 2011]. When considering the entire cohort, which included 17 individuals who later became cognitively impaired, that study found that obese individuals (BMI ≥ 30) had greater decline in temporal GM, which was one of the regions negatively associated with baseline BMI in both our cross-sectional and longitudinal analyses, as well as decreased volume in frontal and total GM, which were associated in our cross-sectional analysis, and in a few smaller structures (Table 5). However, these associations were only identified when BMI was modeled as a categorical variable; when BMI was modeled as continuous just the cingulate had P < 0.05. In addition, no positive associations with WM ROIs were identified, which differs from our results. Another study considered longitudinal BMI measures, though imaging data was only available from a single time point and atrophy was visually rated in four regions rather than using processed volumetrics on a finer regional scale [Gustafson et al., 2004]. While the methodology used was different, that study also found that obesity was associated with atrophy of the temporal lobe [Gustafson et al., 2004].

In addition to strengthening the evidence of the BMI-brain volume association, longitudinal study designs can provide complementary insights. Since Driscoll et al. (2011) considered the interaction of the longitudinal brain volume trajectory with a single BMI time point prior to the MRI scans, their study allows the investigation into whether age-related decline in brain volume is modified by mid-life BMI. Other questions of interest that are not directly addressed by this design include: at which age(s) in life is BMI most associated with changing brain structure? Are changes in BMI associated with changes in volume? On what time scale might obesity play a role in changing brain volume? Finding, for example, that high BMI in one's 30s was most predictive of future decline in brain volume or that weight loss might reverse decline would have important clinical and public health implications. Using data from two time points, we explored the association of baseline BMI with baseline brain volume, baseline BMI with the subsequent change in volume, and change in BMI with change in brain volume, so as to assess persistence, progression, and reversibility [Bandeen-Roche et al., 2009] of the BMI-brain volume association. Our findings suggest that the association of higher BMI with lower brain volume is persistent and may be progressive in some lobar GM. Further large-scale studies investigating the association between BMI measured at multiple time points and longitudinal image acquisitions would further elucidate the complex associations among obesity, aging, and brain volume.

While no WM ROIs were associated in the cross-sectional analysis, it is possible that the positive longitudinal associations seen in WM are more interesting than artifactual. A recent cross-sectional study of 95 females reported positive associations in extensive WM regions in a voxel-wise analysis [Walther et al., 2010], though most studies have either found little association in WM regions or did not conduct separate gray and WM analyses. Changes in brain structure, as well as lesions that primarily occur in WM, can dramatically impact segmentation and registration. Thus the positive longitudinal structural changes associated with BMI and WM may be real associations from structural changes that are highlighted, perhaps in the wrong direction, through image processing. Of course, one cannot discount the possibility that the associations are potentially due to chance or confounding.

Though the majority of study participants were former lead workers, sensitivity analyses demonstrated that these individuals were not appreciably different than community-based controls in the BMI-brain volume association. In particular, including a BMI-by-control status interaction term made no difference in model fit except in a single ROI. Additionally, restricting the model to just include the former lead workers and adjusting for PTL did not materially impact our findings (Table 2).

It is worth noting the complexities inherent in studying longitudinal brain volumetrics. First, it is important to avoid any causal interpretations of associations as BMI is correlated with many factors (many unaccounted for) that are presumably associated with brain structure, such as physical activity and cardiovascular health. Second, voxel-based analyses have lower power, by virtue of the number of tests performed, and presume spatial localization of associations. Nonlocalized associations, or localized associations that vary in location across subjects would not be detectable by VBM, with the latter case being difficult to detect by any method. In our ROI-based analyses, we presented all results from a range of models rather than formally adjust for multiple comparisons. A Bonferroni correction in this setting would likely be highly conservative as the ROIs are overlapping regions, with some completely contained within others. The robust significant cross-sectional ROI associations across a broad range of GM areas, including the whole GM, suggest a diffuse, nonlocalized potential impact of BMI on brain volumetrics.

The relation of obesity with cognitive decline [Whitmer et al., 2005] and dementia [Beydoun et al., 2008] may be explained in part by preceding changes in brain structure. For example, parallel to our findings on brain volume, one study found a cross-sectional association of higher BMI with lower cognitive scores, and that higher baseline BMI was associated with an additional decline in scores at a follow up visit, but that change in BMI was not significantly correlated with change in cognitive function [Cournot et al., 2006]. Other studies looked at the relation of measures of obesity with specific cognitive tasks and generally found higher BMI associated with poorer performance on memory, executive function, language, and visuomotor tasks [Smith et al., 2011].

It is worth hypothesizing the potential mechanistic pathways for an impact of BMI on brain structure. Obese individuals may be resistant to the effects of leptin [Enriori et al., 2006], a hormone involved in appetite and the regulation of food intake, despite having higher leptin levels than the normal-weight [Considine et al., 1996]. One VBM study of 32 young adults found fasting plasma leptin concentrations were associated with changes in GM regional volumes [Pannacciulli et al., 2007], and results from another study of three adults suggested that leptin replacement may lead to subsequent increases in GM volume among the leptin deficient [Matochik et al., 2005]. It has been established that enlarged adipocytes among the obese may induce a low-grade state of systemic inflammation [Greenberg and Obin, 2006; Gregor and Hotamisligil, 2011]. Higher levels of inflammatory markers have been associated with lower total brain volume [Jefferson et al., 2007], and females taking anti-inflammatory drugs were found to experience a smaller decline in brain volume associated with normal aging than those not taking the drugs [Walther et al., 2011]. Insulin resistance and diabetes, which are highly correlated with obesity, could also play a role in brain volume deficits [Korf et al., 2007; Tan et al., 2011], though one VBM study found that when BMI, fasting plasma insulin levels, and type II diabetes variables were included together in the model, only BMI remained independently associated with lower brain volumes [Raji et al., 2010].

The combination of increasing obesity prevalence and a large aging population in the United States have exacerbated the public health burden in recent years, and these trends are expected to continue in the coming decades [Wang et al., 2011]. The implications of this and previous studies that higher BMI may also be a risk factor for changes in brain structure and for neurodegenerative disorders could further strain public health resources if obesity-related cognitive decline compounds the normal effects of aging. Consequently, obesity prevention programs and interventions targeting obesity as a risk factor have the potential not just for reductions in type 2 diabetes and improved cardiovascular health but also for more robust brain health in mid- and late-life.


  1. Top of page
  2. Abstract