Apolipoprotein ε4 modifies obesity-related atrophy in the hippocampal formation of cognitively healthy adults

Highlights • Age-related inverted U-shaped curve of hippocampal myelin/neurite packing.• Reduced hippocampal myelin/neurite packing and size/complexity in obesity.• APOE modifies the effects of obesity on hippocampal size/complexity.• Age-related slowing of spatial navigation but no risk effects on cognition.• CA/DG predict episodic memory and subiculum predicts spatial navigation performance.


Introduction
The world's population is aging, creating an increase in agerelated health issues, including cognitive decline ( Beard et al., 2016 ). With respect to age-related memory impairments, the hippocampal formation, that is, dentate gyrus (DG), cornu Ammonis (CA) fields, and subiculum, warrants particular attention as it is critically involved in memory processing and is affected early in the progression of Alzheimer's disease (AD). Magnetic resonance imaging (MRI) studies have, for example, shown reductions in total hippocampal volume associated with aging in cognitively healthy individuals ( Raz, 2001 ), while hippocampal atrophy remains one of the supporting diagnostic features of amnestic Mild Cognitive Impairment (aMCI) and AD ( de Flores et al., 2015a ). The next challenge is to distinguish normal age-related hippocampal The effects of three established risk factors, that is, carriage of the Apolipoprotein E ( APOE) ε4 genotype ( Angelopoulou et al., 2021 ;Chai et al., 2021 ;Feringa and van der Kant, 2021 ;Koutsodendris et al., 2021 ;Liu et al., 2013 ), a positive family history (FH) of dementia in a first-grade relative ( Alosco et al., 2014 ;Donix et al., 2010 ;Johnson et al., 2014 ;Tanzi, 2012 ;Wolf, 2012 ), and central obesity ( Arnoldussen et al., 2014 ;Beydoun et al., 2008 ;Chuang et al., 2016 ;Pedditizi et al., 2016 ;Xu et al., 2011 ) and their potential interactions were studied ( Metzler-Baddeley et al., 2019a ;Metzler-Baddeley et al., 2019b ;Mole et al., 2020 ).
The CA1, CA2/3, DG and subiculum subfields were segmented with two publicly available, automated protocols that make use of T 1 -and T 2 -weighted hippocampal images: The Bayesian inference labeling methods implemented in FreeSurfer 6.0 ( Iglesias et al., 2015 ) and the Automatic Segmentation of Hippocampal Subfields (ASHS) that utilizes multi-atlas segmentation and machine learning techniques ( Yushkevich et al., 2015b ). Both protocols have been validated with histopathological evidence in patients with epilepsy ( Menzler et al., 2021 ;Mizutani et al., 2021 ). Employing both ASHS and FreeSurfer allowed us to assess whether the choice of protocol for the labelling of hippocampal subfields had an impact on the analysis of risk effects ( Yushkevich et al., 2015a ).
Secondly, to investigate the pattern of age and risk effects not only with volumetric but also with multi-parametric microstructural MRI from diffusion neurite density and dispersion imaging (NODDI) ( Zhang et al., 2012 ) and quantitative magnetization transfer (qMT) ( Eng et al., 1991 ;Henkelman et al., 1993 ;Henkelman et al., 2001 ;Sled, 2017 ). NODDI fits a threecompartment tissue model to multishell diffusion-weighted imaging data, which allows the separation of intra-and extracellular tissue properties ( Zhang et al., 2012 ). This is useful as it enables the differentiation between the packing density of neurites, estimated with the intracellular signal fraction (ICSF), and their spatial organization, measured with the orientation dispersion index (ODI). In addition, the amount of free water in the tissue that may accumulate due to age and neurodegeneration-related atrophy can also be estimated with the isotropic signal fraction (ISOSF). Further, we applied a two-pool model to the qMT data that captured the exchange of magnetization between protons in free water and those bound to macromolecules such as lipids and proteins ( Henkelman et al. 1993 ). This model allows the quantification of the macromolecular proton fraction (MPF) that may capture apparent myelin of neurites ( Ceckler et al., 1992 ;Schmierer et al., 2007 ) and the forward exchange rate k f that may reflect tissue metabolism ( Giulietti et al., 2012 ;Harrison et al., 2015 ). Finally, we quantified the longitudinal relaxation rate R 1 for the estimation of apparent water, lipid/protein, and, to a lesser extent, iron content ( Callaghan et al., 2015 ). Support for studying neurite properties within the context of aging and neurodegeneration comes from neuropathological evidence suggesting that human aging is associated with a reduction of neocortical dendritic spine density ( Dickstein et al., 2013 ) with accompanying compensatory increases in the dendritic extent of DG granular cells ( Flood et al., 1985 ;Flood et al., 1987 ). Likewise, recent in vivo NODDI studies found age-related reductions in neocortical neurite dispersion ( Nazeri et al., 2015 ) and increases in neurite dispersion of the whole hippocampus ( Metzler-Baddeley et al., 2019b ;Nazeri et al., 2015 ). With regards to AD pathology, ODI and ICSF were found sensitive to amyloid and tau pathology in the hippocampus ( Colon-Perez et al., 2019 ) and in white matter in AD animal models ( Colgan et al., 2016 ;Colon-Perez et al., 2019 ). In addition, k f reductions were reported in the hippocampus, temporal lobes, parietal cortex and posterior cingulate in people with AD ( Giulietti et al., 2012 ) and T 1 and T 2 relaxometry measurements have been found sensitive to AD pathology in animal and human imaging studies ( Knight et al., 2016 ;Knight et al., 2019 ;Tang et al., 2018 ). These observations demonstrate that microstructural indices may provide complementary information to volumetric measurements by capturing neurite density and orientation, macromolecular, and free water-related tissue properties ( Callaghan et al., 2015 ;Giulietti et al., 2012 ;Harrison et al., 2015 ). Applying these microstructural measurements may therefore aid our understanding of the biophysical properties underpinning aging and risk ( Wolf et al., 2015 ).
Thirdly, we also explored whether individual differences in hippocampal subfield macro-and microstructure correlated with cognitive functions including episodic memory and spatial navigation, abilities known to rely on hippocampal processes ( Brown et al., 2014 ;Hartley et al., 2005 ;Hoang et al., 2018 ;Kyle et al., 2015 ).
In addition, lifestyle-related factors, such as obesity and sedentary lifestyle, may have adverse effects on the hippocampus. For instance, abdominal visceral fat has been found to be associated with volume reductions ( Anan et al., 2010 ) and increases in free water signal of the whole hippocampal formation ( Metzler-Baddeley et al., 2019a ). Obesity is related to pro-inflammatory states ( Cox et al., 2015 ) and in rodent models was found to induce microglia activation and reduce long-term potentiation in the hippocampus ( Hao et al., 2016 ). It is increasingly recognized that APOE ε4 may interact with obesity to augment disruptions in lipid, glucose, insulin, and immune response metabolism and that such adverse interactions may increase the vulnerability of medial temporal lobe structure to neurodegeneration ( Jones and Rebeck, 2018 ;Mole et al., 2020 ;Zade et al., 2013 ). However, the nature of these interactions and their impact on the subfields of the hippocampal formation remain poorly understood and require further elucidation.
It is important to note that three source of inconsistency when subdividing the hippocampal formation stems from different subfield and border segmentation protocols ( Yushkevich et al., 2015a ). Yushkevich et al. (2015a) compared 21 protocols for labelling hippocampal subfields, including those adopted here. They found considerable differences with regards to the region within each segmentation was performed, the set of the employed anatomical Displays the hippocampal subfield regions that were automatically segmented from T 1 and T 2 -weighted images using the Automated Segmentation of Hippocampal Subfields (ASHS) ( Yushkevich et al., 2015b ) and the FreeSurfer Segmentation of Hippocampal Subfields functionality (version 6) ( Iglesias et al., 2015 ). Segmentations are shown for 1 participant with coronal images along the anterior-posterior hippocampal axis on a T 2 -weighted image for ASHS and a T 1 -weighted image for FreeSurfer. Abbreviations: CA1, cornu Ammonis 1; CA2/3; SUB, subiculum; DG, dentate gyrus. "(For interpretation of the references to color in this figure legend, the reader is referred to the Web version of this article.)" labels, and the extent of specific anatomical labels. The largest discrepancies between the protocols were at the CA1/subiculum boundary and in the anterior portion relative to body and tail portions of the hippocampal formation.
For this reason, we employed the two most widely used automated hippocampal subfield labelling protocols, that is, FreeSurfer version 6.0 and ASHS. We then assessed whether the type of protocol affected the pattern of risk-related differences in macroand microstructure of the hippocampal formation. We focused our comparison on CA1, CA2/3, DG and subiculum as these regions have previously been implicated in aging and disease and were sufficiently large to extract meaningful microstructural information from diffusion and qMT images with a resolution of approximately 2mm 3 Fig. 1 . displays examples of ASHS and FreeSurfer segmentations of these subfields for one representative data set. While the microstructural measurements described above were chosen to capture complementary tissue properties ( Wolf et al., 2015 ), they also share some overlapping information that can cause redundancies in the data analysis and reduce the statistical power of the analyses ( Chamberland et al., 2019 ). We and others have previously demonstrated that Principal Component Analysis (PCA) can be successfully employed to reduce the dimensionality of multi-modal brain measurements and extract meaningful components of the underlying data structure ( Bourbon-Teles et al., 2017 ;Chamberland et al., 2019 ;Geeraert et al., 2020 ;Metzler-Baddeley et al., 2017 ;Penke et al., 2010 ).
Here we adopted this approach to study potential dissociations between age-and age-independent risk effects on principal components of hippocampal macro-and microstructure. More specifically, we modelled main and interaction effects of APOE , FH, and WHR on three principal components that reflected apparent myelin/neurite packaging, size/complexity, and metabolism of hippocampal gray matter. Differences between hippocampal subfields and between protocols in the pattern of risk effects were modelled by including subfield and protocol as independent factors, which allowed for the testing of interaction effects between risk and these factors. These analyses controlled for the effects of age, sex ( Veldsman et al., 2021 ) and verbal intelligence ( Boyle et al., 2021 ), as we aimed to gain a better understanding of age and sexindependent effects of APOE , FH and obesity and their potential in-teractions. Differences between subfields in the pattern of age and risk effects were further assessed with post-hoc tests.
Finally, we assessed age and risk effects on cognitive functions including episodic memory and spatial navigation and explored brain-function correlations between the hippocampal subfield macro-and microstructure and cognition.

Materials and methods
CARDS received ethical approval from the School of Psychology Research Ethics Committee at Cardiff University (EC.14.09.09.3843R2). Participants provided written informed consent in accordance with the Declaration of Helsinki.

Participants
Participants between the age of 38 and 71 years were recruited from the local community via Cardiff University volunteer panels, notice boards and local poster advertisements. A detailed description of the CARDS sample can be found in Mole et al. (2020) . All participants had a good command of the English language and were without a history of neurological and/or psychiatric disease, head injury with loss of consciousness, or drug or alcohol dependency. A total of 166 CARDS volunteers underwent MRI scanning at the Cardiff University Brain Research Imaging Centre (CUBRIC). Seven participants did not complete the MRI protocol, including the high resolution T 2 images of the hippocampus, due to claustrophobia and/or feeling uncomfortable. Participants' intellectual function was assessed with the National Adult Reading Test-Revised (NART-R) ( Nelson, 1991 ) and cognitive impairment was screened for with the Mini Mental State Exam (MMSE) ( Folstein et al., 1975 ). One person with a MMSE score of 26 was excluded from the analysis. Thus, the current analysis was based on 158 datasets (see Table 1 ).

Assessment of dementia risk factors
Central obesity was assessed with the Waist to Hip Ratio (WHR) following the World Health Organisation's ( World Health Organisation, 2008 ) recommended protocol for measuring waist and hip circumference. Central obesity was defined as a WHR ≥0.9 for  ( Folstein et al., 1975 ); NART, national adult reading test ( Nelson, 1991 ); SD, standard deviation. a Based on waist-hip ratio ≥0.9 for males and ≥0.85 for females ( Organisation, 2008 ). males and ≥0.85 for females. Individuals were categorized as centrally obese (WHR + ) or normal WHR (WHR-).
Saliva samples were collected with the Genotek Oragene-DNA kit (OG-500) for DNA extraction and APOE genotyping. APOE genotypes ε 2, ε 3, and ε 4 were determined by TaqMan genotyping of single nucleotide polymorphism rs7412 and KASP genotyping of single nucleotide polymorphism rs429358. Genotyping was unsuccessful for one individual. Participants were categorized into those who carried at least one ε4 allele ( APOE ε4) and those who did not ( APOE ε4-).
Participants also self-reported their family history of dementia, that is, whether a first-grade relative was affected by Alzheimer's disease, vascular dementia or any other type of dementia. Two participants could not provide information about their family history (FH). The remaining participants were categorized into those with a positive FH (FH + ) and those without (FH-).

Cognitive assessment
Immediate and 30 minutes delayed verbal and visual recall were assessed with the Rey Auditory Verbal Learning Test (RAVLT) ( Rey, 1941 ;Schmidt, 1996 ) and the complex Rey Figure Test ( Rey, 1941 ). Short term topographical memory was measured with the 4 Mountains Test ( Chan et al., 2016 ). Spatial navigation was assessed with a virtual Morris Water Maze Task ( Hamilton et al., 2002 ) where participants had to find and navigate to a hidden platform in a water pool. This task also included a motor control condition without a visible platform. Working memory capacity and executive functions were assessed with computerized tests from the Cambridge Brain Sciences battery ( Hampshire et al., 2012 ;Owen et al., 2010 ). Working memory capacity was tested with digit and spatial span, distractor suppression with an adapted version of the Stroop test (Double-Trouble), problem solving with a version of the Tower of London task (the Tree task), abstract reasoning with grammatical reasoning and the odd-one-out task, as well as the ability to manipulate and organize spatial information with a self-ordered spatial span task. In addition, participants performed a paired-associate learning (PAL) and a choice-reaction time task. Cognitive outcome measures were the number and latencies of correct responses as well as spatial navigation path length.
In brief, the FreeSurfer version 6.0 cross-sectional segmentation pipeline firstly requires the fully automated analysis of the T 1 -weighted images with reconall ( http://surfer.nmr.mgh.harvard. edu/ ) that involves skull stripping, correction for motion artefacts and field inhomogeneities, the registration of native data to and from the standard Talairach space, parcellation of cortical regions and the segmentation of subcortical structures as well as the hippocampi ( Fischl, 2012 ). These processing steps were followed by a hippocampal subfield segmentation pipeline that utilises both T 1 and T 2 -weighted images to identify anatomical landmarks for the bilateral segmentation of 12 subfields, that is, CA1, CA2/3, CA4, subiculum, presubiculum, parasubiculum, molecular layer of the subiculum and CA fields, granule cell layer of the dentate gyrus, fimbria, hippocampus-amygdala transition area, hippocampal tail, and fissure ( Iglesias et al., 2015 ). This pipeline is based on a statistical atlas of the hippocampus constructed from manual segmentation labels from both in vivo and high-resolution ex vivo data that were used to develop a Bayesian inference algorithm for the automatic segmentation of the hippocampus proper from T 1 and T 2weighted structural images ( Iglesias et al., 2015 ).
The ASHS software ( https://sites.google.com/site/hipposubfields ) implements a multi-atlas segmentation technique that involves registering the target MRI with a bank of T 2 -weighted atlas MRIs including manually labeled subregions. Here the ASHS UPenn PMC Atlas package was chosen for this purpose. Multi-atlas label fusion is then applied to select a consensus segmentation based on the shared similarity between target and atlas images. Systematic segmentation errors are minimized with a learning-based bias correction technique. Joint label fusion and corrective learning are then repeated by bootstrapping seeded by the segmentation results from the previous phase. ASHS leads to the segmentation of 10 region of interests, that is, CA1, CA2, CA3, DG, subiculum, entorhinal cortex, Brodmann area 35, Brodmann area 36, collateral sulcus and miscellaneous regions. To allow meaningful extraction of lower resolution microstructural data, ASHS CA2 and CA3 regions were combined using the fslmaths utility from the Oxford Centre for Functional MRI of the Brain (FMRIB) Software Library (version 6.0) ( Jenkinson et al., 2012 ) .
For the purpose of comparing between the two segmentation techniques and for allowing meaningful extraction of lower resolution microstructural values, the present analysis focused on CA1, CA2/3, DG and subiculum regions that were shown to be affected by aging and AD ( de Flores et al., 2015a ). For quality control, FreeSurfer and ASHS labels of these regions were visually inspected by authors BC and PAG.
Mean hippocampal subfield and intracranial volumes (ICV) were extracted for each brain. Subfield volumes were then adjusted for ICV to correct for inter-individual differences in head size using the formula (subfield volume x 10 0 0)/ ICV (from FreeSurfer or ASHS respectively).

HARDI and qMT data processing
A detailed description of the microstructural data processing has been provided previously ( Metzler-Baddeley et al., 2019a ;Metzler-Baddeley et al., 2019b ). In brief, the dual-shell HARDI data were split and b = 1200 and 2400 s/mm 2 data were corrected separately for distortions induced by the diffusion-weighted gradients and motion artifacts with appropriate reorientation of the encoding vectors ( Leemans and Jones, 2009 ) in ExploreDTI (Version 4.8.3) . EPI-induced geometrical distortions were corrected by warping the diffusion-weighted image volumes to the T 1 -weighted anatomical images ( Irfanoglu et al., 2012 ). After preprocessing, the NODDI model ( Zhang et al., 2012 ) was fitted to the HARDI data with the fast, linear model fitting algorithms of the Accelerated Microstructure Imaging via Convex Optimization (AMICO) framework ( Daducci et al., 2015 ) to gain ISOSF, ICSF, and ODI maps.
Using Elastix ( Klein et al., 2010 ), MT-weighted GRE volumes were co-registered to the MT-volume with the most contrast using a rigid body (6 degrees of freedom) registration to correct for inter-scan motion. Data from the 11 MT-weighted GRE images and T 1 -maps were fitted by a two-pool model using the Ramani pulsed-MT approximation ( Ramani et al., 2002 ). This approximation provided MPF, k f and R 1 maps. MPF maps were thresholded to an upper intensity limit of 0.3 and k f maps to an upper limit of 3.0 using the FMRIB's fslmaths imaging calculator to remove voxels with noise-only data.
All microstructural maps were spatially aligned with the hippocampal subfield masks by co-registration with the T 1 -weighted anatomical space as reference image with linear affine registration (12 degrees of freedom) using FMRIB's Linear Image Registration Tool. Spatial alignment of microstructural maps to ASHS hippocampal subfield masks involved an additional warping to the T 2 -weighted space with FMRIB's Linear Image Registration Tool.

Statistical analysis
Statistical analyses were conducted in SPSS version 26 ( IBM, 2011 ). All volumetric, microstructural, and cognitive data were examined for normal distribution and for outliers, defined as above or below 3 times the interquartile range (75th percentile value -25th percentile value).

Missing data
FreeSurfer hippocampal subfield segmentations could be performed for all 158 datasets and ASHS segmentations for a total of 153 datasets. For FreeSurfer one volume and for ASHS two volume measurements were excluded as outliers. For the microstructural data, 1% of the data were excluded as outliers for FreeSurfer and 5% for ASHS segmentations. For the cognitive data 2.3% of the data were either missing or excluded as outliers.
For the principal component analysis (PCA), each participant's volumetric and microstructural data of the 16 hippocampal subfield segmentations [2 (FreeSurfer, ASHS) x 2 (left, right) x 4 (CA1, CA2/3, DG, subiculum)] were concatenated to form n = 2528 observations. The dependent variables were represented by the 7 brain measurements (ICV-adjusted volumes, MPF, k f , R 1 , ODI, ISOSF, ICSF). Bartlett's test of sphericity and the Kaiser-Meyer-Olkin (KMO) test were used to check that the data were suitable for PCA [KMO = 0.57, Chi 2 (21) = 2842.2, p < 0.001]. PCA was then carried out using a procedure with orthogonal Varimax rotation of the component matrix. Components were extracted based on the Kaiser criterion of including all components with an eigenvalue > 1 ( IBM, 2011 ), by inspecting Cattell's scree plots ( Cattell, 1952 ), and by assessing each component with regards to their interpretability. Component loadings that exceeded a value of 0.5 were considered as significant. The dimensionality of the cognitive data for all 158 participants was also reduced with PCA using the same procedure as described above [KMO = 0.61, Chi 2 (630) = 2463.37, p < 0.001].
Each participant's PCA least squares regression component scores ( DiStefano et al., 2009 ) were subsequently entered as dependent variables in a multivariate analysis of covariance (MAN-COVA) that tested for main and interaction effects of the risk factors APOE genotype ( ε 4 + , ε 4-), FH (FH + , FH-), and WHR (WHR + , WHR-) as well as for the effects of the segmentation protocol (FreeSurfer, ASHS) and hippocampal subfield segmentations (bilateral CA1, CA2/3, DG, subiculum). Age, sex and IQscores from the NART-R ( Nelson, 1991 ) were included as covariates. Similarly, MANCOVA tests for risk effects on cognitive component scores, while controlling for age, sex, and IQ, were completed.
Significant omnibus effects were further investigated with posthoc comparisons using univariate analysis of covariance (ANCOVA) and independent t-tests. Relationships between brain structure and cognitive component scores were studied using hierarchical linear regression models that first entered age and sex as independent variables, and then in a second model volumetric and microstructural measurements o f all left and right CA1, CA2/3, DG, and subiculum regions segmented with ASHS and FreeSurfer in a stepwise fashion to predict the variance in the cognitive component data. First and post-hoc models were corrected for multiple comparisons with a False Discovery Rate (FDR) of 5% using the Benjamini-Hochberg procedure ( Benjamini and Hochberg, 1995 ). The 5% FDR was applied to all statistical tests that related to the same theoretical inference ( Lakens, 2014 ).
All reported p -values, unless stated otherwise, were Benjamini-Hochberg adjusted ( p BHadj ) and two-tailed. Information about effects sizes was provided with the partial eta squared index η p 2 for MANCOVA analyses and R 2 for regression analyses. Fig. 2 displays the correlation matrix of the microstructural variables that were entered into the PCA. The variables MPF, R 1 and ICSF were positively correlated with each other, as were ODI and volume. ISOSF was negatively correlated with ODI, MPF, R 1 , ICSF and volume. There were no correlations with other measures for k f . Consistent with this cross-correlation pattern, PCA extracted three components that explained 67% of the variance in the hippocampal macro-and microstructural data ( Table 2 ). The first component explained 28.2% of the data variance and had positive loadings > 0.5 from MPF, R 1 and ICSF sensitive to myelin and neurite packing. The second component explained an additional 24.5% of the variance and had positive loadings > 0.5 from ODI and ICVadjusted volumes and a negative loading from ISOSF and thus captured tissue size and complexity. The third component explained 14.3% of variance and had a high loading from k f only and may reflect tissue metabolism.  ( Fig. 3 A) but larger PC2 values [t(2248) = 11.2, p BHadj < 0.001] ( Fig. 3 B). With regards to the subfields, subiculum was associated with the largest PC1 myelin/neurite packing values ( Fig. 3 C), while CA2/3 had the lowest PC2 size/complexity values ( Fig. 3 D). No effects were observed for the PC3 Metabolite component.

Exploring the pattern of age and risk effects across hippocampal subfields
To explore the patterns of age and risk effects across the 4 hippocampal subfields CA1, CA2/3, DG, and subiculum, separate AN-COVAs were carried out on the PC data for each subfield concatenated across hemisphere and protocol.

PCA of the cognitive data
Four principal components were extracted that explained together 44% of the variance in the cognitive data ( Table 3 ). The first PC labelled "Verbal Memory" explained 15.2% of the variance and had high loadings > 0.5 from RAVLT measures. The second PC "Spatial Navigation First Move" explained 13.5% of the variance with loadings from first move latencies in the virtual spatial navigation task. The third PC "Spatial Navigation Path Length" explained an additional 9.5% of the variance with loadings from spatial navigation path length. The fourth PC "Visual Memory" explained 5.8% of data variance and had loadings from immediate and delayed recall of the Rey figure and from the 4 Mountains Test.
There were no effects of risk. Table 4 summarizes the results of the linear hierarchical regression analyses. As age and sex had significant effects on the cognitive components, they were first entered as independent variables prior to testing for the effects of hippocampal subfield macro-and microstructure in a stepwise fashion. A final model that included contributions from left CA1 ODI and right DG ISOSF explained 14% of the Verbal Recall data, while 24% of the Visual Recall data were explained by right CA2/3 ICSF and right CA1 volume and ISOSF. For spatial navigation first move latencies, 31% of the data were explained by a model including age, sex, and left CA2/3 R 1 , while for path length 19% of the data were explained by right subiculum ODI and left subiculum ISOSF and k f .

Discussion
Dissociating the effects of healthy aging on hippocampal subfields from those related to genetic and lifestyle risk of dementia could be key to effectively targeting interventions for older-age memory impairments. Thus, the primary objective of the present study was to investigate age and age-independent effects of three major risk factors, that is, carriage of the APOE ε4 genotype, a positive family history (FH) of dementia, and central obesity, on the macro-and microstructure of the hippocampal formation in a sample of 158 cognitively healthy adults.
We characterized properties of the hippocampal formation with subfield volumes based on T 1 and high-resolution T 2 -weighted images as well as with microstructural measurements from NODDI and qMT imaging to gain complementary information about apparent myelin, neurite packing, free water, and metabolism. Accounting for overlapping information between the various MRI measurements, PCA was employed to reduce the data dimensionality to three principal components that reflected myelin/neurite packing, size/complexity, and metabolic tissue properties. These components were then investigated across the main subfields of the hippocampal formation, that is, CA1, CA2/3, DG, and subiculum, which were previously shown to be particularly vulnerable to the impact of aging and disease ( de Flores et al., 2015b ). Subfields were segmented using two widely employed and freely available, automated segmentation protocols, that is, ASHS and FreeSurfer (version 6). This allowed us to assess any potential interaction effects between the type of protocol and risk factors on the analysis of hippocampal subfield properties.
While type of labelling protocol and hippocampal subfields were associated with absolute differences in component measures, they did not interact with the risk factors, suggesting that risk effects did not significantly differ between ASHS and FreeSurfer segmentations. Overall ASHS relative to FreeSurfer segmentations resulted in larger size/complexity but lower myelin/neurite packing estimates ( Fig. 3 ). ASHS segmentations of CA1, CA2/3 and DG were larger, but subiculum labels were smaller than those of FreeSurfer ( Fig. 4 ).
These differences were caused by disagreements in the anatomical labels between the two segmentation protocols that are known to be most pronounced at the CA1/subiculum boundary and within the anterior portion of the hippocampal formation ( Yushkevich et al., 2015a ). In addition, both protocols differed in the region of the hippocampal formation that was labelled and the number of subfield segmentations. While FreeSurfer labels 12 regions (CA1, CA2/3, CA4, subiculum, presubiculum, parasubiculum, molecular layer of the subiculum and CA fields, granule cell layer of the dentate gyrus, fimbria, hippocampus-amygdala transition area, hippocampal tail, and fissure) ( Iglesias et al., 2015 ), ASHS segments ten structures (CA1, CA2, CA3, DG, subiculum, entorhinal cortex, Brodmann area 35, Brodmann area 36, collateral sulcus and miscellaneous regions) ( Yushkevich et al., 2015b ). As many of these subfields were too small to extract meaningful lower resolution microstructural information, we focused on the analysis of the 4 main subfields of the hippocampal formation that were labelled by both protocols (CA1, CA2/3, DG and subiculum).
However, these subfield labels did not correspond to exactly the same regions across the two protocols. Notably, FreeSurfer provides a separate label for the "molecular layer" and voxels classified as such were missing from its CA field segmentations while they were part of the ASHS labels. Similarly, the DG label from FreeSurfer only contains the unmyelinated granule cell layer while ASHS labels the whole DG structure including molecular, granule, and inner polymorphic layers. These differences are evident in the example of the ASHS and FreeSurfer segmentations for one dataset displayed in Fig. 1 . Together with disagreements about the CA1/subiculum border, they account for the observed size/complexity differences between the two protocols.
However, there were also similarities between the protocols, that were consistent with the known cellular composition and microstructural properties of different hippocampal subfields ( Benes et al., 1994 ;Geurts et al., 2007 ;Seress et al., 1993 ;Seress et al., 1992 ;Seress et al., 1994 ). Notably, myelin/neurite packing signals were largest in the subiculum for both ASHS and FreeSurfer ( Fig. 4 ). This is consistent with evidence from histological studies using myelin-staining of a higher proportion of myelinated axons in the subiculum compared to other hippocampal regions such as CA1 or the molecular and granule cell layers of the DG ( Benes et al., 1994 ;Geurts et al., 2007 ;Seress et al., 1994 ). In addition, both ASHS and FreeSurfer labelled CA2/3 as the smallest structure, while CA1 and DG were comparable in size/complexity consistent with the known anatomy of the hippocampus.
It should also be noted that the final ASHS segmentation outputs were in high resolution T 2 -weighted space while those for FreeSurfer were in T 1 -weighted space ( Fig. 1 ) suggesting differences in the use of multi-spectral information between the pipelines. A recent study found significant variations in FreeSurfer volume estimations of hippocampal subfields depending on the input images (T 1 and/or T 2 ) ( Seiger et al., 2021 ). It is therefore possible that differences in the use of standard T 1 and high-resolution T 2 -based information may have contributed to the discrepancies observed between the two protocols. Importantly though for our primary research question, despite these significant differences between the ASHS and FreeSurfer labels, there were no interaction  effects between the type of protocol and any of the three risk factors, suggesting that a comparable risk pattern was observed across both protocols. With regards to risk factors, we observed that central obesity as measured with the WHR was associated with reductions in the myelin/neurite packing and the size/complexity but not the metabolic component ( Fig. 6 ). These findings are consistent with accumulating evidence that obesity is associated with adverse effects on the hippocampus and memory functions ( Anan et al., 2010 ;Dekkers et al., 2019 ;Mueller et al., 2012 ;Spencer et al., 2017 ;Stranahan, 2015 ;Willette and Kapogiannis, 2015 ). For instance, a recent analysis of data from 12,0 0 0 participants (45-76 years of age) of the UK Biobank study, reported that total body fat was related to smaller volumes in subcortical structures and the hippocampi in men ( Dekkers et al., 2019 ). This study also reported obesity-related differences in whole brain white matter microstructure measured with fractional anisotropy and mean diffusivity. Similarly, previous analyses of CARDS data found WHR to be positively correlated with hippocampal atrophy [as estimated with the free water signal (ISOSF)] and negatively with fornix MPF and k f ( Metzler-Baddeley et al., 2019a ). In this study, WHR was an estimate of abdominal visceral (r = 0.3, p = 0.001) but not subcutaneous fat fractions (p = 0.7) while Body Mass Index estimated subcutaneous (r = 0.4, p < 0.001) but not visceral fat ( p = 0.3) ( Metzler-Baddeley et al., 2019a ). As Body Mass Index had no effects on brain microstructure, these findings suggest that the correlations between WHR and hippocampal and fornix microstructure were driven by excessive visceral rather than body fat per se , consistent with accumulating evidence for visceral fat-related adverse effects on the hippocampus, brain white matter, and mortality ( Anan et al., 2010 ;Koster and Schaap, 2015 ;Koster et al., 2010 ;Lampe et al., 2019 ). Comparable to Dekkers et al. (2019) , men were more centrally obese and had higher visceral fat fractions and higher hippocampal ISOSF ( Metzler-Baddeley et al., 2019a ).
Central obesity, notably excessive visceral fat, is associated with multiple metabolic alterations affecting blood cholesterol, glucose, and insulin levels, that can lead to cardio-and cerebrovascular disease and Type 2 diabetes ( Dommermuth and Ewing, 2018 ;Whitmer et al., 2007 ). Central obesity at midlife may also be accompanied with systemic, low-grade inflammation ( Cox et al., 2015 ;Guillemot-Legris and Muccioli, 2017 ). Diet-induced obesity in animal studies has been shown to trigger inflammation in the hippocampus, which in turn impaired synaptic functioning and spatial memory ( Hao et al., 2016 ). In humans, individuals with a genetic polymorphism associated with pro-inflammatory state and AD risk had smaller CA1-2, CA3-DG and subiculum than healthy controls ( Raz et al., 2015 ). All of those obesity-related metabolic changes in lipid, glucose and immune responses are thought to contribute to the risk of developing dementia in older age ( Businaro et al., 2018 ;Profenno et al., 2010 ;Ricci et al., 2017 ).
Furthermore, it is also increasingly recognized that obesity may interact with genetic risk factors, notably APOE ε4 ( Jones and Rebeck, 2018 ;Mole et al., 2020 ;Zade et al., 2013 ). Indeed, here we observed interaction effects between APOE and WHR on the size/complexity of the hippocampal formation. While all obese individuals showed reduced size/complexity, this effect was only significant for APOE ε4 non-carriers ( p < 0.001) but not for APOE ε4 carriers ( p = 0.1) ( Fig. 7 ). As can be seen in Fig. 7 , APOE ε4 carriers did not show a significant obesity effect because size/complexity was attenuated in normal-weighted APOE ε4 carriers. Thus, central obesity appeared to be related to atrophy in the hippocampal formation regardless of an individual's APOE or FH status (as no effects of FH were present). However, APOE ε4 carriage alone appeared also to be associated with adverse effects on the hippocampus and these may have masked any metabolic and vascular benefits of a healthy WHR. Or expressed differently, keeping a healthy weight may not be sufficient to compensate for APOE ε4-driven hippocampal atrophy. These risk effects were present for data collapsed across all hippocampal subfields. There was no evidence for any subfield specific vulnerability to the impact of obesity and APOE although trends were observed for CA1, CA2/3 and DG but not the subiculum. Future larger studies are required to clarify whether these regions are particularly susceptible to obesity and APOE ε4 related tissue changes and why this may be.
A previous analysis of data from the Framingham Offspring cohort also highlighted complex synergistic effects between APOE and obesity ( Zade et al., 2013 ). They found several APOE-related modifications of correlations between individual differences in WHR on one hand and brain structure and cognition on the other.
For instance, APOE ε4 carriers showed stronger negative relationships between WHR and executive and memory functions as well as larger correlations between WHR and white matter hyperintensities. Interestingly, they also reported a stronger correlation between WHR and frontal brain volume for APOE ε4 non-carriers similar to our observations here.
In addition, it is likely that obesity effects will be modulated by other polygenic risk factors than APOE ε4 ( Woo and Reifman, 2018 ). While the CARDS sample was too small to quantify AD and/or obesity-related polygenic risk hazards ( Escott-Price et al., 2014 ;Escott-Price et al., 2015 ), we included family history of dementia as a variable that captures environmental and genetic risk factors beyond APOE ε4. In the present analysis of hippocampal microstructure, we did not find any effect of FH, but previously we observed widespread interaction effects between APOE , FH, and WHR on white matter microstructure, that were most pronounced in the right parahippocampal cingulum ( Mole et al., 2020 ). More specifically, APOE ε4 carriers with a positive FH, showed obesityrelated reductions in apparent myelin MPF while no effects were observed for individuals without a FH. These risk-effects on apparent myelin were moderated by hypertension and inflammationrelated blood markers ( Mole et al., 2020 ). The precise nature of these complex synergistic effects between obesity and APOE ε4 remain elusive and require further investigation. A recent animal study ( Jones et al., 2021 ) points to inflammation and neuronal plasticity mechanisms underpinning interaction effects between obesity and APOE genotype. In this study, a highfat diet increased gliosis and immediate-early gene expression only in APOE ε3 but not APOE ε4 knock-in mice. This suggested early dysregulation of adaptive inflammatory mechanisms in APOE ε4 mice that may make the brain more vulnerable to insults and damage in the long run. In addition, APOE ε4 is also known to lead to changes in glucose, insulin, and lipid metabolism and altered betaamyloid production ( Jones and Rebeck, 2018 ;Jones et al., 2019 ).
Together, these findings suggest that APOE ε4 and obesity lead to metabolic alterations, including inflammatory processes, and may adversely interact with each other and with other genetic factors on brain structure and function. We propose that interaction effects between APOE ε4 and obesity on the hippocampal formation may increase that region's vulnerability to subsequent neurodegeneration. The effects of risk on the hippocampal formation were observed while accounting for age and sex and interaction effects between WHR and APOE were only present for the size/complexity component. In contrast, aging was associated with a non-linear inverted U-shaped curve of the myelin/neurite packing component akin to the trajectory of white matter microstructure across the lifespan in humans and rhesus monkeys ( Bartzokis et al., 2010 ;Kubicki et al., 2019 ;Slater et al., 2019 ;Yeatman et al., 2014 ). Myelin/neurite packing increased between the thirties and fourties seconds, when it peaked, remained relatively stable in the fifties and sixties and declined in the seventies ( Fig. 5 A). Age effects were present in all four subfields of the hippocampal formation, that is, CA1, CA2/3, DG, and subiculum. We propose that this observed age trajectory in the myelin/neurite packing component most likely reflects the maturation of white matter pathways within the hippocampal formation, such as the perforant path, mossy fiber, and Schaffer collateral pathways ( Zeineh et al., 2017 ). Indeed, age-related differences in the myelin basic protein (MBP) expression were found in the CA1, CA2/3 and DG regions of gerbils such that white matter fibers of the perforant pathway, the mossy fibers and Schaffer collaterals were reduced in older relative to younger gerbils ( Ahn et al., 2017 ) Our findings accord with a study ( Douaud et al., 2014 ) that employed a data-driven analysis of gray matter structural variation and identified a brain network comprising prefrontal, intraparietal, posterior cingulate, and medial temporal lobe regions whose lifespan pattern mirrored brain development and age-related degeneration. The hippocampus forms part of this network which matures during adolescence and young adulthood into midlife and shows heightened vulnerability to accelerated neurodegeneration in older age ( Douaud et al. 2014 ). As we did not observe any age-related differences in the size/complexity and metabolic components, we propose that hippocampal changes across midlife and early older age may be primarily driven by changes in white matter myelin and neurite packing rather than a loss of neurons and synapses.
Finally, we tested for the effects of risk factors on cognitive performance and explored the relationship between cognitive performance and hippocampal subfield macro-and microstructure. The dimensionality of the cognitive data was reduced to four components reflecting verbal and visual episodic memory as well as spatial navigation first move latencies and path length. We did not observe any risk effects on cognitive components. Additional analysis of risk factors on individual cognitive measurements did not reveal any risk effects either ( APOE: p = 0.5, FH: p = 0.6, WHR: p = 0.3). This may suggest that risk-related macro-and microstructural tissue changes in the hippocampal formation precede any cognitive impairment and/or were too subtle to induce any functional impairments in this sample of cognitively healthy adults. In contrast, age was associated with larger latencies in the spatial navigation task due to older people taking longer to plan and initiate their first move in the virtual water maze. In addition, females performed better in the verbal recall test than males ( Metzler-Baddeley et al., 2019a ).
With regards to brain-cognition relationships we observed dissociations between left and right hemisphere contributions to verbal and visual recall. While verbal recall was predicted by microstructural differences in left CA1 and right DG, visual recall was only predicted by macro-and microstructural differences in rightlateralized CA1 and CA2/3 regions. This pattern of dissociations accords with previous evidence from epilepsy patients that verbal and visual episodic memory rely more on the left and right medial temporal lobes, respectively ( Kaplan et al., 1994 ;Wagner et al., 2009 ). Furthermore, we also observed functional dissociations between the subiculum and CA fields and DG. Episodic memory performance relies on the ability to complete as well as separate patterns, with CA regions, notably CA3, playing an important role in pattern completion and DG in pattern separation ( Baker et al., 2016 ;Keinath et al., 2020 ;Pilly et al., 2018 ;Poli et al., 2018 ). Thus, we observed contributions from CA fields and DG to both visual and verbal recall performance. In contrast, microstructural differences in the subiculum were predictive of path length performance in the spatial navigation task. This finding accords with evidence that the subiculum is crucially involved in the representation of space and displays boundary-dependent neural activity ( Lee et al., 2018 ;O'Mara, 2005 ). Thus, the here observed pattern of specific associations between hippocampal subfields and cognitive performance provides support for the dissociation between subicular representation of space involved in environmental navigation and episodic memory function mediated by CA and DG regions in the human hippocampus. These findings also highlight the possibility that, although risk factors had no apparent impact on cognition in the current study, they may predispose an individual for future episodic memory and navigational deficits.
In summary, we provide novel evidence for dissociations between age and age-independent risk effects on hippocampal subfield macro-and microstructure. Non-linear age effects in myelin/neurite packing were observed in all subfields of the hippocampal formation. Central obesity was associated with reduc-tions in myelin/neurite packing and size/complexity across all subfields, with APOE genotype modifying the effects of obesity on size/complexity. Notably, APOE ε4 carriers did not seem to benefit of a healthy Waist-Hip-Ratio as much as non-carriers. Age and sex were significantly related to performance differences in spatial navigation and recall but no effects of risk factors on cognition were present. We also provide evidence for dissociations between contributions of the subiculum to spatial navigation and of CA and DG regions to episodic memory performance as well as of left-right lateralization effects on verbal and visual recall. It remains to be determined if the observed risk-related hippocampal macro-and microstructural differences may precede any future cognitive decline.

Disclosure statement
The authors declare no competing financial and/or non-financial interests.