The effect of BMI on hippocampal morphology and memory performance in late childhood and adolescence

Abstract

Childhood obesity is associated with negative physiological and cognitive health outcomes. The hippocampus is a diverse subcortical structure involved in learned feeding behaviors and energy regulation, and research has shown that the hippocampus is vulnerable to the effects of excess adiposity. Previous studies have demonstrated reduced hippocampal volume in overweight and obese children, however it is unclear if certain sub-regions are selectively affected. The purpose of this study was to determine how excess body weight influences regional hippocampal surface morphology and memory performance in a large cross-sectional cohort of 588 children and adolescents between 8.33 and 19.92 years of age using body mass index expressed as a percentage of the 95^th percentile cutoff (%BMIp95). We demonstrate %BMIp95 is associated with reduced radial thickness in the superior anterior region of the left hippocampus, and this relationship is predominantly driven by children younger than 14 years. We also found %BMIp95 was associated with worse performance on a spatial episodic memory task and this relationship was partially mediated by the radial thickness of the significant shape cluster. These results demonstrate the differential influence of excess body weight on regional hippocampal structure and hippocampal-dependent behavior in children and adolescents.

1. Introduction

The prevalence of childhood obesity has risen dramatically over the past 3 decades, with recent estimates that more than 30% of children are overweight or obese (Ogden, Carroll, Kit, & Flegal, 2014). While a gradual increase in adiposity is expected from childhood through adolescence (T. J. Cole & Lobstein, 2012), persistent childhood obesity leads to adverse neuropsychological and health issues (Carnell, Gibson, Benson, Ochner, & Geliebter, 2012). Obesity is an important risk factor for the development of cardiovascular disease, Type 2 diabetes, stroke, and cancer (Haines, Wan, Lynn, Barrett, & Shield, 2007; Haslam & James, 2005; Pinhas-Hamel & Zeitler, 2005). In addition to the increased risk of physiological complications later in life, childhood obesity is also associated with a higher prevalence of depression, psychiatric disorders, and cognitive issues (Allen, Byrne, Blair, & Davis, 2006; Braet, Mervielde, & Vandereycken, 1997; Erermis et al., 2004). Specifically, a growing body of research shows that obesity can increase the risk of developing memory problems in children (Cheke, Simons, & Clayton, 2016) and, over time, central nervous system pathologies such as dementia and Alzheimer’s disease (Fitzpatrick et al., 2009; Mayeux & Stern, 2012). Due to the high prevalence of and the myriad health issues associated with childhood obesity, further research is needed to establish how these changes affect the developing brain.

The hippocampal formation is a subcortical structure within the medial temporal lobe that is critical for memory and learning processes (Sheldon & Levine, 2016). Additionally, the hippocampus is implicated in a number of metabolic functions, including appetitive, ingestive and learned eating behaviors (Benoit, Davis, & Davidson, 2010; Terry L. Davidson, Kanoski, Schier, Clegg, & Benoit, 2007; Kanoski, 2012; Kanoski & Grill, 2017). The hippocampus is innervated by the hypothalamus, amygdala, and thalamus, which form networks critical for energy metabolism and weight management (T. L. Davidson & Jarrard, 1993; T. L. Davidson, Kanoski, Walls, & Jarrard, 2005), and has receptor expression for the important metabolic hormones leptin (Van Doorn, Macht, Grillo, & Reagan, 2017) and insulin (Soto, Cai, Konishi, & Kahn, 2019). Previous studies in children show obesity is associated with deficits in hippocampal-dependent processes (Cheke et al., 2016; Khan et al., 2015; A. A. Miller & Spencer, 2014), and evidence in rodents suggests this impairment is unique to juveniles (Boitard et al., 2012). Because hippocampal maturation is a protracted process (Giedd et al., 1996; Insausti, Cebada-Sánchez, & Marcos, 2010), it is important to study the unique vulnerabilities in this region due to childhood adiposity during critical periods of development.

Previous evidence suggests that the cognitive deficits observed in childhood obesity are accompanied by structural alterations in the hippocampus. In rodent models, obesity in juveniles leads to decreased neurogenesis (Boitard et al., 2012), synaptic stripping (Hao, Dey, Yu, & Stranahan, 2016) and neuroinflammatory processes (Guillemot-Legris & Muccioli, 2017; A. A. Miller & Spencer, 2014). Hippocampal pathology subsequently manifests as selective changes to cornu ammonis 1 (CA1) and CA3 subfields in the hippocampus (Sack et al., 2017). Human neuroimaging studies in vivo have demonstrated that childhood obesity is associated with reduced hippocampal volume (Bauer et al., 2015; Chaddock et al., 2010; Z. L. Mestre, Eichen, Wierenga, Strong, & Boutelle, 2017; Nouwen et al., 2017), however, few studies have explored the influence of excess body weight on hippocampal structure across development.

The hippocampus is structurally and functionally heterogeneous (Poppenk, Evensmoen, Moscovitch, & Nadel, 2013; Strange, Witter, Lein, & Moser, 2014) and it is likely that whole volumetric approaches obscure localized changes in the hippocampus associated with excess adiposity. Therefore, methods with enhanced regional specificity to body mass-related changes can provide insight into vulnerable subregions and processes. Shape analysis provides a mechanism to detect regional deformations in surface topology from triangulated meshes generated from hippocampal volumes (Shi et al., 2014; Shi, Morra, Thompson, & Toga, 2009). Radial thickness surface features provide vertex-wise measures of the hippocampal cross-sectional radius and represents local thickness estimates (Shi et al., 2009; Thompson et al., 2004). Previous studies have shown that global hippocampal shape features may be a more informative biomarker for memory performance compared to volumetry (Voineskos et al., 2015).

The purpose of this study is to characterize how body weight influences regional hippocampal structure and memory performance in a broad cohort of 588 children and adolescents between 8.33 and 19.92 years of age. Body mass index (BMI) is calculated from commonly acquired height and weight measures and is a commonly used indicator of excess body fat. BMI expressed as a percentage of the 95^th percentile cutoff (%BMIp95) is a measure of relative weight adjusted for child age and sex and is more highly correlated with adiposity in young children and in children with severe obesity compared to BMI z-scores (Freedman & Berenson, 2017; Freedman et al., 2017). Using shape analysis, we sought to characterize the relationship between %BMIp95 and hippocampal shape to identify if subregions within the hippocampus are selectively vulnerable to the effects of excess body weight in younger and older children and adolescents. We also aimed to determine the relationship between %BMIp95 and performance on episodic memory tasks. Lastly, we determined if hippocampal shape features mediate the relationship between %BMIp95 and memory performance. These findings could contribute to the current knowledge regarding the influence of childhood body weight on developing brain structures, which may have significant implications for lifelong cognitive function.

2. Materials and Methods

2.1. Subjects

Data used for this study was acquired through the publicly available Philadelphia Neurodevelopmental Cohort (PNC) research initiative at Children’s Hospital of Philadelphia (Satterthwaite et al., 2016). The PNC is a research initiative focused on characterizing the interaction between brain, behavior and genetics in children and adolescents. Subjects were included in the study if they were proficient in English, able to provide informed consent (or assent if under 18 years of age) and capable of participating in neurocognitive testing. Subjects were excluded from the study if they had major severe medical problems, neurological or endocrine disorders, history of head trauma, impaired vision or hearing, MRI contraindications such as ferrous body implants, or structural brain abnormalities (Satterthwaite et al., 2014). In total, 620 T1-weighted images from participants under the age of 20 years with height, weight, sex and age information for BMI calculations were originally considered. Following quality assurance procedures (described below), 32 participants were excluded and our final sample consisted of 588 subjects. For analyses on age effects, we split the sample into a young cohort (N=279; ages 8-13.9 years) and old cohort (N=309; ages 14-19.9 years) (Table 1).

Table 1.

Sample demographics

Cohort	N (Females)	Age (years)			%BMIp95
		Mean	SD	Range	Mean	SD	Range
Young	279 (141)	11.25	1.59	8.33 - 13.92	87.49	22.13	55.51 - 166.10
Old	309 (164)	16.48	1.51	14.00 - 19.92	84.58	17.30	57.27 - 148.09
Total	588 (305)	14.01	3.04	8.33 - 19.92	85.96	19.77	55.51 - 166.10

2.2. Body Mass Index Calculations

Weight and height were recorded during the neuroimaging session. Body mass index (BMI) was calculated for each subject using the Centers for Disease Control and Prevention’s (CDC) Child and Teen BMI calculator (https://www.cdc.gov/healthyweight/bmi/calculator.html), which is appropriate to use in children between the ages of 2 and 20 years. Child BMI was computed using the following formula:

$$BMI=703\ast \left(\frac{Weight}{Heigh{t}^2}\right)$$

where weight is expressed in pounds (lbs) and height is expressed in inches (in). We then utilized BMI expressed as a percentage of the CDC 95^th percentile cutoff (%BMIp95) for subsequent analyses because %BMIp95 is more highly correlated with adiposity in young children and in children with severe obesity compared to BMI z-scores (Freedman & Berenson, 2017; Freedman et al., 2017). %BMIp95 was calculated using the CDC SAS program for the Growth Charts (Centers for Disease Control and Prevention, 2019) expressed as:

$$\%BMIp95=\left(\frac{BMI}{BM{I}_{95\%}}\right)\ast 100$$

where BMI_95% is the BMI at the 95^th percentile for the child’s age and sex according to CDC guidelines. Obesity is defined as a BMI greater than or equal to the 95^th percentile of the CDC growth charts (Kuczmarski et al., 2002; Ogden & Flegal, 2010). Therefore, consistent with previous studies, we refer to moderate obesity as 100 ≤ %BMIp95 < 120 and severe obesity is %BMIp95 ≥ 120 (Freedman & Berenson, 2017; Freedman et al., 2017; Kelly et al., 2013).

2.3. MRI Acquisition and Image Processing

Participants were scanned on a 3T Siemens Tim Trio whole-body MRI in the Hospital at University of Pennsylvania using the Siemens software revision VB17 (Satterthwaite et al., 2014). Brain structural neuroimaging data were acquired with three-dimensional magnetization prepared, rapid-acquisition gradient-echo (MPRAGE) T1-weighted sequences with a 32-channel head coil. Scans were acquired with the following sequence parameters: Repetition time (TR) =1810 ms, echo time (TE) =3.5 ms, inversion time (TI)=1100 ms, flip angle=9°, field of view (FOV)=180x240, slice thickness=1 mm, matrix size=192x256x160, GRAPPA factor=2, voxel size=.9375x.9375x1 mm, acquisition time: 3:28 min using a right-left phase encoding direction.

FSL’s anatomical processing pipeline was used for image processing and includes bias-field correction, registration to standard MNI space, brain extraction and tissue-based segmentation (Smith et al., 2004). Bilateral hippocampal volumes were segmented from the T1-weighted image using FSL’s model-based segmentation tool, FIRST (Patenaude, Smith, Kennedy, & Jenkinson, 2011), and were visually inspected for anatomical accuracy. Intra-cranial volume (ICV) was computed per subject in cm³ using the number of voxels contained within the skull.

2.4. Quality Assurance Procedures

Quality assurance (QA) was initially carried out by the PNC prior to release. This included visual identification of excessive subject motion and removal prior to release (Satterthwaite et al., 2014). Prior to analyses for the present study, we carried out our own secondary QA assessment. Scans considered for this study were assessed for subject motion, ringing, tissue contrast and ease of identification of anatomical structures. An experienced neuroimaging researcher scored the scans on a scale from 1 through 3, where 1 = ringing artifacts in up to 1 region, good gray matter/white matter (GM/WM) contrast and no motion artifacts; 2= ringing artifacts in up to 2 regions, less sharp GM/WM boundary, and/or some blurring; and 3=ringing artifacts throughout the entire image, GM/WM contrast not differentiable and/or major motion artifacts (scan quality for all subjects considered: M ± SD = 1.06 ± .27). Scans scored as a “3” on the QC scale were excluded from analyses. Scans scored as a “2” were included in the study if subcortical structures had sharp edges and could be well-identified. After hippocampal segmentation, scans were then assessed to ensure anatomically accurate delineation of hippocampal boundaries by an individual trained in hippocampal neuroanatomy. Each hippocampal volume was overlaid on the T1w image for each subject and were visually inspected for anatomical correspondence using the anatomical landmarks outlined by (Duvernoy, Cattin, & Risold, 2013). Scans were rated either as a “Pass” or “Fail” and scans were discarded if either the left or right hippocampal volume was delineated outside the bounds of the structure on the T1w image. Of the 620 subjects originally considered for the study, 11 subjects were excluded due to biologically implausible BMI calculations due to recording errors during acquisition, 13 subjects were excluded due to subject motion and artifacts, 7 scans were excluded due to poor hippocampal segmentations and 1 scan was excluded due to pipeline processing errors, resulting in 588 subjects in the present analysis (305 female, range: 8.33-19.92 years, M=14.01, SD=3.04 years) (Table 1).

2.5. Hippocampal Shape Analysis

Shape analysis was performed using the fully-automated Metric Optimization for Computational Anatomy (MOCA) software developed by Shi and colleagues (Shi et al., 2014) and described previously (Lynch, Shi, Toga, & Clark, 2018). MOCA presents several advantages over other traditional shape analysis methods. Previous methods, such as FSL FIRST, utilizes a deformable mesh from a manually segmented training set, which may constrain the topology of the hippocampal surface and may not be representative of the developmental dataset used in this study (Patenaude et al., 2011). MOCA enables a more general mesh reconstruction through the isometry invariant generation of intrinsic surface meshes from the individual hippocampal mask boundaries (Shi et al., 2010). Additionally, previous studies have used vertex analysis, which performs statistical analyses based on the mean group difference in surface vertex positions (Patenaude et al., 2011). In contrast, MOCA describes intrinsic surface features that reflect interpretable shape properties, such as cross-sectional thickness (Shi et al., 2008). While MOCA and FSL have mechanisms that enable robust one-to-one correspondence between meshes, MOCA enables the construction of population-specific group-wise atlases and uses metric optimization to directly map between the surfaces and templates (Shi et al., 2014). Because hippocampal shape undergoes dynamic changes across childhood and adolescence (Lynch et al., 2018), we constructed an age-appropriate population averaged template for the present study.

Each segmented hippocampal volume was first converted to a triangulated mesh in native space. Laplace-Beltrami (LB) eigen-projections were then used to generate smoothed surfaces by iteratively updating vertices using surface deformation and outlier detection, resulting in robust preservation of topology while avoiding shrinkage (Shi et al., 2010). Meshes were then re-sampled to 2000 vertices. In order to probe for regional changes in hippocampal structure related to adiposity, vertex-wise radial thickness features were computed using the Reeb graph of the first LB eigen-function (Shi et al., 2008, 2009). This measure reflects local hippocampal thickness and is defined as the shortest distance from each vertex to the longitudinal core of the hippocampus. From all the surfaces, we generated the average template meshes for the left and right side with 2000 vertices per mesh by minimizing the spectral l²-distance between all subject meshes using SurfStat implemented in Matlab (www.math.mcgill.ca/keith/surfstat). Individual surfaces were mapped to the template mesh using metric optimization in a high-dimensional LB embedding space (Shi et al., 2014). The resulting conformal maps were then used to pull back the vertices of the subject surfaces onto the average template surface to establish one-to-one correspondence for statistical analyses.

2.6. Behavioral measures

Participants in the PNC were administered an abbreviated version of the Penn computerized neurocognitive battery (Penn CNB) (Satterthwaite et al., 2016) that consists of 14 tests that evaluate multiple cognitive domains, including executive function, complex cognitive processes, social cognition, motor speed and episodic memory (Gur et al., 2010). Accuracy and speed measures were available for each test. For the present study, we utilized the episodic memory tasks that assess immediate and delayed recall: (1) the Penn Word Memory Task (PWMT) tests verbal memory. Participants are provided 20 target words and 20 distractor words that are matched according to length, prevalence and imageability. The participants score reflects the number of correctly identified target words and rejected foil words; (2) the Penn Face Memory Task (PFMT) tests face memory. Participants are presented with a mixture of 20 target and 20 distractor faces matched for ethnicity, race and gender. Scores are calculated based on the number of correctly identified target faces and rejected foil faces; and (3) the Visual Object Learning Task (VOLT) of the episodic memory batter of the Penn CNB that tests spatial memory. Participants are provided 20 target and 20 distractor Euclidean shapes. The participants score reflects the number of correctly identified target shapes and correctly rejected foil shapes and this procedure is repeated following a 20 minute delay. We utilize the total correct responses for all test trials assessed with the VOLT task as a measure of declarative memory performance.

2.7. Statistical Analysis

In order to determine the effect of adiposity on hippocampal structure using shape analysis, general linear models (GLMs) were applied to each vertex in the group-wise atlas to test if radial distance is significantly associated with %BMIp95. Because we have previously shown in an expanded developmental cohort than hippocampal shape is associated with age and sex (Lynch et al., 2018), we included age, sex, and ICV as covariates of no interest. Random field theory (RFT) was used to derive significant clusters. RFT considers the spatial correlation of surface maps using both the peaks and spatial extent of smoothed statistical parameter maps using Gaussian random fields (Cao & Worsley, 1999; Worsley, Andermann, Koulis, MacDonald, & Evans, 1999). The Euler characteristic estimates the corrected p-value at a given cluster level (Friston, 1997; Woo, Krishnan, & Wager, 2014). This property is derived from the vertex-wise significance and the number of resolution elements (resels) in the image, which describes the search volume as a function of the image FWHM smoothness (Worsley, Evans, Marrett, & Neelin, 1992). In order to reduce false positives observed in surface-based anatomical analyses (Greve & Fischl, 2018), we used a stringent supra-threshold cluster level of p<.0005 and a set level threshold of p<.05 was used for the height and spatial extent thresholds, respectively. Shape analyses were also performed using Bonferroni correction at each vertex with an adjusted p < 2.5 x 10⁻⁵ (.05/2000).

Region of interest (ROI) post hoc analyses were then performed on significant clusters to better identify the relationship between %BMIp95, performance on memory tasks and regional hippocampal thickness using general linear models. To assess the relationship between the 3 memory tasks, %BMIp95 and radial thickness clusters, we controlled for multiple comparisons using Bonferroni corrected significance threshold of p < .008 (.05/6). For memory tasks where episodic memory performance is significantly associated with both %BMIp95 and radial thickness, we performed a statistical mediation analysis using nonparametric bootstrapping with 500 Monte Carlo draws (Tingley, Yamamoto, Hirose, Keele, & Imai, 2014). Statistical analyses were performed in R (version 3.6.3) using packages for linear model regression (stats version 3.6.2), statistical mediation analysis (mediation version 4.5.0) and data visualization (ggplot2 version 3.3.2).

3. Results

The children and adolescents included in this study had %BMIp95 that ranged from 55.5 to 166.1 (86.0 ± 19.8) (Figure 1). According to our criteria, 88 participants had moderate obesity (100 ≤ %BMIp95 < 120) and 37 subjects had severe obesity (%BMIp95 ≥ 120).

Figure 1. Distribution of %BMIp95.

The distribution of %BMIp95 is shown for the participants in the study. Colors denote the young cohort (8.33 – 13.9 years) and old cohort (14 – 19.92 years) used for age-stratified analyses.

3.1. Demographic associations with hippocampal structure and %BMIp95

Bilateral hippocampal volumes were significantly associated with age, left: B=28.17, t(586)=4.87, p<.001; right: B=29.05, t(586)=5.11, p<.001, and ICV, left: B=2.03, t(586)=15.00, p<.001; right: B=2.06, t(586)=15.71, p<.001. There was also a significant effect of sex on hippocampal volume bilaterally, left: t(586)=6.31, p<.001; right: t(586)=7.17, p<.001, where hippocampal volumes were larger in males (left: 3593 ± 422 mm³; right: 3723 ± 446 mm³) compared to females (left: 3374 ± 419 mm³; right: 3478 ± 373 mm³). Age and sex were each significantly associated with radial thickness clusters bilaterally (Supplementary Figure 1). These results did not change considerably after controlling for %BMIp95 (Supplementary Table 1). %BMIp95 was not significantly associated with sex, t(586)=.68, p=.50 (males: 86.54 ± 19.58; females: 85.43 ± 19.97), or ICV, t(586)=1.68, p=.09. A significant inverse relationship between age and %BMIp95, however, was observed, B=−.62, t(586)=2.33, p=.02.

3.2. BMI is associated with hippocampal volume and morphology

After controlling for age, sex and ICV, reduced %BMIp95 was significantly associated with increased left hippocampal volume (Figure 2), B=−1.57, t(586)=−2.06, p=.04. We did not observe a significant relationship between %BMIp95 and right hippocampal volume, t(586)=1.26, p=.21. Using shape analysis, %BMIp95 was significantly associated with radial thickness in a cluster that survived RFT correction after controlling for age, sex and ICV on the left superior anterior hippocampus, 39 vertices, 1.06 resels, RFT-adjusted p<.001 (Figure 3a-b). Vertex-wise significance thresholding with Bonferroni correction shows several vertices reached statistical significance in the left superior anterior cluster, peak vertex: t(586)=3.79, adjusted p=.028 (Supplementary Figure 2a). Post-hoc analyses reveal %BMIp95 is negatively associated with radial thickness in this cluster, B = −.004, t(586) = −4.88, p<.001, adjusted R² = .16 (Figure 3c). %BMIp95 was not significantly associated with radial distance in the right hippocampus.

Figure 2. The association between %BMIp95 and bilateral hippocampal volume.

(A) %BMIp95 is negatively associated with left hippocampal volume. (B) A significant relationship between %BMIp95 and right hippocampal volume was not observed. Each point represents a single observation.

Figure 3. Association between regional hippocampal radial thickness and %BMIp95.

(A) T-value maps of the linear relationship between %BMIp95 and vertex-wise radial thickness while controlling for age, sex, and ICV on bilateral hippocampal surfaces. Warm colors show hippocampal surface expansion with increasing %BMIp95 and cool colors show hippocampal surface contraction with increasing %BMIp95. (B) RFT-corrected p-value maps show clusters of contiguous vertices on the left superior anterior and left inferior lateral hippocampus where radial thickness is significantly associated with %BMIp95. (C) Post-hoc analyses of the left superior anterior cluster show %BMIp95 is associated with reduced mean radial thickness.

3.3. The effect of BMI on hippocampal structure is influenced by age

The sample was subset into younger (8-13.9 years) and older (14-19.9 years) cohorts to determine if our findings were influenced by age. In the left hemisphere, hippocampal volume was associated with %BMIp95 after controlling for age, sex and ICV in the young cohort, B=−2.16, t(277)=−2.31, p<.022, but not the old cohort, t(307)=−.90, p=.37. In the younger cohort, we found %BMIp95 was significantly associated with radial thickness after adjusting for age, sex and ICV following RFT correction in a cluster on the left hippocampus that corresponds to the lateral anterior superior surface, 29 vertices, .77 resolution elements (resels), RFT-adjusted p=.002 (Figure 4a-b). Vertex-wise significance thresholding with Bonferroni correction show the majority of vertices in the cluster reach statistical significance, peak vertex: t(277)=4.19, adjusted p=.007 (Supplementary Figure 2b). In the older cohort, %BMIp95 was not significantly associated with radial thickness in either the left or right hippocampus following RFT correction, however a region in the left anterior superior mesial hippocampus trended toward significance (Figure 4a). Post-hoc analyses confirm this age-associated discrepancy, where a significant interaction between age group and %BMIp95 on radial thickness of the significant cluster in the young cohort was observed, F(5,582)=8.15, p=.004. In this region, %BMIp95 was associated with reduced radial thickness in the young cohort, B=−.004, t(277)=−4.26, p<.001, adjusted R² = .09, but was not associated with radial thickness in the old cohort, t(307)=.23, p=.83 (Figure 4c).

Figure 4. The influence of %BMIp95 on radial thickness stratified by age.

(A) T-value maps of the linear relationship between %BMIp95 and vertex-wise radial thickness while controlling for age, sex, and ICV on bilateral hippocampal surfaces are shown for the young (left) and old (right) cohorts. (B) Significant associations between %BMIp95 and radial thickness using RFT thresholded p-values were only observed in the young cohort in the left superior anterior and left inferior anterior hippocampus. (C) In the young cohort, radial thickness of the left superior anterior cluster was negatively associated with %BMIp95, while no significant relationship was observed in this region for the old cohort.

3.4. Memory performance is associated with BMI and hippocampal shape

Tests of associations between %BMIp95, radial thickness of the superior anterior cluster and performance on memory tasks were assessed using six independent general linear models with age and sex as covariates of no interest with a Bonferroni adjusted alpha level of p<.0083 per test (.05/6). After controlling for age and sex, increasing %BMIp95 was significantly associated with fewer total correct responses on the VOLT task, B=−.014, t(586)=−2.80, p=.005, adjusted R² = .02. %BMIp95 was not significantly associated with total correct responses on the PFMT, B=−.021, t(586)=−2.61, p=.009, adjusted R² = .12, or PWMT, B=−.013, t(586)=−2.04, p=.04, adjusted R² = .03, following multiple comparisons correction. Increased radial thickness of the left anterior superior cluster extracted with shape analysis was significantly associated with more total correct responses on the VOLT task, B=−.023, t(586)=3.52, p<.001, adjusted R² = .07. Radial thickness was not significantly associated with total correct responses on the PFMT, t(586)=1.46, p=.14, or PWMT tasks, t(586)=1.80, p=.07.

We then tested whether individual differences in radial thickness of the anterior superior hippocampal cluster could statistically mediate the relationship between %BMIp95 and performance on the VOLT task. Statistical mediation analysis using nonparametric bootstrapping (1000 Monte Carlo draws) were carried out with age and sex as covariates of no interest. Including the indirect effect of average radial thickness of the ROI partially attenuated the direct association of %BMIp95 with performance on the VOLT task, B=−.010, t(586)=−2.11, p=.04, resulting in an average causal mediation effect (ACME) of −.0033, confidence interval (CI) = [−.0059, −.0001], p=.0014 (Figure 5). Replication of the statistical mediation analysis in the young cohort reveals the indirect effect of the radial thickness of the ROI reduced the direct association of %BMIp95 with VOLT task performance to non-significance, ACME = −.0040, CI = [−.0078, −.001], p=.002. In the old cohort, inclusion of the radial thickness cluster did not considerably change the direct association of %BMIp95 with VOLT task performance, ACME = −.0003, CI = [−.0011, .2250] p=.82, (Supplementary Figure 3).

Figure 5. Radial thickness partially mediates the relationship between %BMIp95 and the total correct response on the VOLT task.

Statistical mediation analysis demonstrates %BMIp95 correlates with fewer total correct responses on the VOLT spatial memory task in part through decreased radial thickness of the left anterior superior hippocampus. This model includes age, sex and ICV as covariates of no interest. Model coefficients for paths a, b, c and c’ are provided. Path c reflects the total effect and c’ reflects the direct effect. *p<.05, **p<.01, ***p<.001

4. Discussion

The hippocampus is well known for its involvement in learning and memory behaviors, however, recent studies have demonstrated its importance is metabolic functions and feeding behaviors (Kanoski & Grill, 2017). The current study sought to characterize the progressive influence of body mass on hippocampal structure in a large cross-sectional cohort of children and adolescents using whole hippocampal volumetry and shape analysis. We show %BMIp95, a relative measure of BMI, is associated with reductions to left hippocampal volume and morphological shape alterations. The relationship between %BMIp95 and hippocampal shape was particularly strong in children between the ages of 8 and 13.9 years of age. Additionally, %BMIp95 was associated with worse performance on a memory task, and we showed morphology of the left anterior superior hippocampal region partially mediates this relationship. Together, our findings lend support to previous studies in animals and humans that suggest excess weight gain is associated with alterations to hippocampal structure and behavior.

Our findings show increased %BMIp95 is associated with reduced left hippocampal volume in children and adolescents. Functional laterality of the hippocampus has been demonstrated in humans, where the left hippocampus is selectively involved in contextual memory, including associative learning (J. Miller et al., 2018) and the temporal organization of egocentric navigation (Iglói, Doeller, Berthoz, Rondi-reig, & Burgess, 2010). Because feeding behaviors rely on the integration of spatial contextual cues, the left hippocampus may be more vulnerable to the effects of increased body mass. Our findings are in agreement with previous studies that show children and adolescents with obesity have reduced left hippocampal volumes compared to normal weight children (Bauer et al., 2015; Z. L. Mestre et al., 2017). However, several cross-sectional studies that considered BMI across the weight spectrum, as opposed to testing for group differences using clinical obesity cut-offs, have not found a significant association (Masterson, Bobak, Rapuano, Shearrer, & Gilbert-Diamond, 2019; Z. Mestre et al., 2020; Perlaki et al., 2018; Yau, Kang, Javier, & Convit, 2014). It is possible that these effects are subtle in nature and by employing shape analysis, we are able to probe the dose-dependent effect of body mass on regional hippocampal structure that is typically obscured with whole-volumetric approaches.

Our shape analysis findings suggest the volumetric reductions in the left hippocampus related to %BMIp95 are localized to the superior aspect of the left hippocampal head, where increasing %BMIp95 was associated with inward surface deformations reflected by reduced radial thickness. This cluster corresponds to the putative CA1 subfield and subiculum (Supplementary Figure 4) (Insausti et al., 2010), and there is evidence in animal models to suggest these regions are particularly vulnerable to alterations associated with a western diet. Asymmetric hippocampal atrophy of the left ventral hippocampus (analogous to the anterior hippocampus in humans) has been observed in adolescent rodents fed a calorically-dense and high-fat diet (Kalyan-Masih et al., 2016), which may be associated with decreased spinal density and pyramidal neuron damage within the CA1 (Wang et al., 2016).

The relationship between %BMIp95 and reduced radial thickness in the left anterior hippocampus is driven by younger adolescents. Our results suggest childhood obesity may leave the hippocampus particularly vulnerable to structural alterations, as childhood represents a time of significant plasticity and change where adverse metabolic environments may disproportionately influence structural development. Hippocampal development is a protracted process that continues through late adolescence, however the most rapid and dynamic structural changes occur in childhood (Herting et al., 2018; Lynch et al., 2018). Damage to the developing hippocampus could therefore derail normative developmental trajectories, resulting in long lasting cognitive deficits and predisposition to metabolic and cardiovascular disease. It may also be possible that excess body weight may differentially influence hippocampal structure in children and adolescents. In a recent study in older adolescents, BMI z-score was significantly associated with reduced hippocampal T2-weighted signal intensity, but not volume (Z. Mestre et al., 2020). Together with the present results, these findings suggest BMI may influence hippocampal volume in childhood and tissue composition in late adolescence.

We demonstrate %BMIp95 is associated with worse performance on a spatial memory task. There is considerable evidence in rodent models of obesity that hippocampal abnormalities are associated with spatial episodic memory impairments (Jurdak, Lichtenstein, & Kanarek, 2008; Kanoski, Meisel, Mullins, & Davidson, 2007; Molteni, Barnard, Ying, Roberts, & Gómez-Pinilla, 2002; Valladolid-Acebes et al., 2011; Winocur et al., 2005). These findings are supported by previous studies exploring memory processes in pediatric obesity. Abdominal adiposity is associated with worse performance on relational memory tasks in overweight and obese children (Khan et al., 2015) and reduced hippocampal functional activation is observed during encoding in adolescents with severe pediatric obesity (Pearce et al., 2019). There is also evidence in adults that show excess body mass is associated with reduced memory performance (Cheke et al., 2016; Stingl et al., 2012) that is accompanied by decreased functional activation during episodic memory tasks (Cheke, Bonnici, Clayton, & Simons, 2017). The observed relationship between %BMIp95 and performance on the spatial memory task was partially mediated by radial thickness of the anterior superior hippocampal region. Spatial memory is often associated with the posterior hippocampus (Poppenk et al., 2013); however, recent evidence suggests the anterior hippocampus is preferentially involved in the encoding of spatial memories in humans (Fritch et al., 2020) and rodents (Broadbent, Squire, & Clark, 2004; Keinath et al., 2014; Wilkerson & Levin, 1999). The neural control of appetitive behaviors relies on the hippocampal-dependent integration of external visuospatial contextual stimuli with interoceptive cues related to endocrine and gastrointestinal information (Kanoski & Grill, 2017). Therefore, together with evidence of the unique involvement of the left hippocampus in contextual learning, structural alterations to the left anterior hippocampus and related visuospatial memory deficits may further contribute to obesity due to deficits in the consolidation of food memories using spatial contextual cues. The “vicious cycle” model of obesity has been proposed that links progressive hippocampal dysfunction with the development of obesity (T. L. Davidson et al., 2005; Hargrave, Jones, & Davidson, 2016; Kanoski & Davidson, 2011). The model suggests excess consumption of an unhealthy diet contributes to hippocampal dysfunction, which manifests as impaired memory inhibition of appetitive responses to food cues. Without inhibitory control of satiety signals, food cues and contexts will have increased power to evoke consummatory responses and will further promote excess consumption of an unhealthy diet, hippocampal dysfunction, and eventual cognitive decline.

While the precise mechanisms by which increased weight gain influences regional hippocampal structure and function over time are unknown, there is converging evidence that diet-induced impairments are a result of neuroinflammatory processes (Guillemot-Legris & Muccioli, 2017; A. A. Miller & Spencer, 2014; Shefer, Marcus, & Stern, 2013). These processes are associated with increased permeability of the blood-brain barrier (Banks, Farr, & Morley, 2006; Kanoski et al., 2010), reduced neurogenesis (Boitard et al., 2012), and synaptic stripping (Hao et al., 2016). Chronic central inflammation associated with high body fat can also contribute to cerebral leptin and insulin resistance (Iikuni, Kwan Lam, Lu, Matarese, & Cava, 2008; Wu & Ballantyne, 2020), which may be potential mechanisms through which excess body mass influences hippocampal structure. Loss of leptin responsiveness in animals fed a high fat diet results in reduced hippocampal synaptic transduction (Mainardi et al., 2017), while insulin resistance is associated with reduced synaptic plasticity and impaired neurogenesis in the hippocampus (Spinelli, Fusco, & Grassi, 2019). Both leptin and insulin signaling also play a critical role in the execution of hippocampal-mediated behaviors (Grillo et al., 2015; Van Doorn et al., 2017). It is unclear whether the apparent leftward structural asymmetry of the observed shape changes associated with excess body mass is due to differential expression of metabolic receptor expression in the hippocampus, and future studies should aim to elucidate the mechanisms that give rise to these differences

A strength of this study is the use of a broad cohort of 588 children and adolescents across a range of body mass in order to capture the neuroanatomical variability observed in a structure as diverse as the hippocampus. Additionally, our use of shape analysis enables the detection of localized differences in hippocampal structure associated with excess body mass. However, this study has some limitations. The cross-sectional nature of this study precludes us from exploring intra-individual hippocampal alterations due to changes in BMI. In order to fully characterize the influence of the development of excess body mass on hippocampal structure, longitudinal study designs should be employed. Additionally, information regarding lifestyle behaviors that contribute to obesity, such as physical inactivity and diet (Loprinzi & Frith, 2018), were not available. Therefore it is uncertain whether excess body mass itself or co-morbid lifestyle behaviors are driving the relationship between %BMIp95 and hippocampal-dependent processes. Previous studies have also shown that surface-based analyses are prone to false positive inflation (Greve & Fischl, 2018). In order to overcome this limitation, we used a conservative cluster-level p-value threshold (p<.0005) and replicated our analyses using Bonferroni-corrected significance levels across the 2000 surface vertices. Lastly, while BMI is an acceptable and readily available anthropometric measure for the clinical diagnosis of obesity (T. J. Cole & Lobstein, 2012; Tim J Cole, Bellizzi, Flegal, & Dietz, 2000; Jean-Philippe Bastard et al., 2006; Krebs et al., 2007), it is a non-specific measure that may not accurately quantify body fat and lean tissue mass. We mitigated this limitation by using an unbounded relative measure of BMI, %BMIp95, which is more highly correlated with adiposity in young children and those with severe obesity compared to BMI z-scores (Freedman & Berenson, 2017; Freedman et al., 2017). Nonetheless, future studies should include more precise methods that provide a more accurate profile of excess body fat relative to body weight, such as dual energy x-ray absorptiometry (DEXA), which provides an accurate measure of whole body percent fat as well as measures of regional body fat, including visceral adiposity, which is more strongly associated with cardiometabolic disease risk (Borga et al., 2018). However the ease of BMI acquisition in large studies enables the power to detect regional structural changes across a range of body weights.

In conclusion, we explored the relationship between hippocampal morphology and relative measures of body mass in a large cohort of children and adolescents between 8.33 and 19.92 years. We found %BMIp95 was significantly associated with reduced left hippocampal volume, which was localized to the anterior superior region. This difference was predominantly driven by participants younger than 14 years of age. Additionally, radial thickness in this region partially mediated the relationship between %BMIp95 and worse performance on a spatial memory task. Together, these findings provide important information regarding the influence of excess body mass on hippocampal structure and memory performance during dynamic periods of development.

Data Availability Statement

Publicly available data used in this study is available at the Philadelphia Neurodevelopmental Cohort (PNC) research initiative (https://www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000607.v3.p2). The metric optimization for computational anatomy (MOCA) software used to perform shape analysis can be downloaded at https://www.nitrc.org/projects/moca_2015/.