Investigating the relationship between BMI across adulthood and late life brain pathologies

Participants were from Insight 46, a sub-study of the MRC National Survey of Health and Development (NSHD), a birth cohort which initially comprised 5362 individuals born throughout mainland Britain in 1 week in 1946 [13]. Follow-up has included > 20 contacts since birth, including home assessments by research nurses at ages 36, 43, 53, and 69 years, and assessment at a clinical research facility at age 60–64 years. Participants were eligible for inclusion in the Insight 46 sub-study if this defined set of life course data were available, and where willingness to attend a clinic visit in London had previously been expressed. Individuals with contraindications to MRI or PET (including claustrophobia and metallic implants) were excluded. Further eligibility criteria have been described elsewhere [14]. Between 2015 and 2018, 502 participants were assessed at University College London [14], when aged 69–71 years. An overview of recruitment is provided in Figure e-1 (supplementary). Comparisons between Insight 46 participants and the larger NSHD have previously been reported [15].

Imaging was performed on a single Biograph mMR 3 T PET/MRI scanner (Siemens Healthcare, Erlangen), with simultaneous acquisition of dynamic PET/MR data, including volumetric (1.1 mm isotropic) T1 and T2-weighted Fluid Attenuated Inversion Recovery (FLAIR) sequences [14]. β-amyloid burden was assessed using [16] F florbetapir (Amyvid). PET data were processed using an automated in-house processing pipeline including pseudo-CT attenuation correction [14]. Global standardised uptake value ratio (SUVR) was calculated from cortical regions of interest (ROIs) comprising the lateral and medial frontal, anterior and posterior cingulate, lateral parietal, and lateral temporal regions, normalised to eroded subcortical white matter. Positive/negative β-amyloid status was determined using a Gaussian mixture model applied to SUVR values, taking the 99th percentile of the lower (β-amyloid negative) Gaussian as the cut-point (0.61).

Volumetric T1-weighted and FLAIR images underwent visual quality control (QC), before processing using validated automated pipelines: [14] whole-brain volume (WBV) segmentation using MAPS [17], hippocampal volumes (HV) using STEPS [18], with appropriate manual editing; and total intracranial volume (TIV) using SPM12 [16]. A validated, unsupervised automated algorithm, BaMoS (Bayesian Model Selection) [19] was used to segment white matter hyperintensities (WMH) from T1/FLAIR images, followed by visual QC and manual editing where appropriate, generating a global WMH volume (WMHV) including subcortical grey matter but excluding infratentorial regions.

Participants were classified as cognitively normal, having mild cognitive impairment (according to research criteria [20]) or dementia, based on expert consensus, informed by clinical history, informant history, MMSE [21], and cognitive performance (WMS-R Logical Memory test [22] and WAIS-R Digit symbol substitution test [23]).

Height and weight measurements were collected by standard protocol as part of NSHD assessments at ages 36, 43, 53, 60–64, and 69 years, and at the Insight 46 assessment at age 69–71 years (which for clarity will subsequently be referred to as 71 years). BMI was defined as the weight in kilogrammes divided by the square of the height (in metres). Abdominal circumference (AC), was measured to the nearest mm using a standardised protocol at 36, 43, 53, 60–64 and 69 years.

Using an approach previously employed in the NSHD, BMI change for the periods 36–43, 43–53, 53–60/64, 60/64–69 and 69–71 years, conditional on earlier measurements, was calculated as the residual from the regression of each BMI measure (from 43 years) on the earlier measure(s) for each sex, using individuals with available data at all time-points. Residuals represent a change in BMI above/below that expected on average given the earlier BMI. Residuals were standardised, allowing comparison between periods [24]. AC change variables were derived using the same approach.

Vascular risk factors selected for adjustment in statistical models included systolic blood pressure (SBP) contemporaneous with BMI measurement, and smoking status, hypercholesterolaemia, and diabetes mellitus (DM) status at the time of the Insight 46 assessment. Seated BP was measured in the upper arm twice after 5 min’ rest at ages 36, 43, 53, 60–64 and 69, and a lying BP collected after 3 min rest at age 71 years. The second BP measure was used for analyses, unless missing. Smoking status was defined by questionnaire (at age 68 years, or if missing at 60–64 years) as never-smoked, ex-smoker and current smoker. Hypercholesterolaemia status was based on self-reported use of cholesterol-lowering medication at Insight 46 assessment or random total cholesterol ≥ 5 mmol/L at 69 years. Diabetes mellitus (DM) status was based on self-reported diabetic medication use at Insight 46 assessment, HbA1c > 6.5% or a self-reported diagnosis at 69 years. APOE genotyping was performed using standard techniques and individuals categorised as APOE-ε4 carriers or non-carriers. Adult socioeconomic position (SEP) was defined as non-manual or manual, based on the occupation at 53 years according to the UK Registrar General’s Classification of Occupations.

A measure of affective symptoms was available at age 69 years derived from the general health questionnaire (GHQ-28 [25]), with a score of > 5 defined as probable anxiety or depression.

Analyses were performed in Stata v.14.1 (Stata Corp). Participants were included if they were dementia-free, and had acceptable quality amyloid PET/MR imaging. For WMH and brain volume analyses, individuals with cortical infarcts inappropriately segmented as WMH (n = 5), atypical vascular pathology (n = 1) or white matter pathologies not considered to be of vascular origin e.g. demyelination (n = 3) were excluded. For brain volume analysis individuals also needed a useable amyloid scan, since the amyloid status was included as a covariate in the fully-adjusted model. Otherwise, all individuals, including those with neurological diagnoses, with available BMI/AC data at any time-point, were included for generalisability.

BMI and AC at each visit (up to age 69 years, which was the last time-point that an assessment was performed across the entire cohort) were compared between Insight 46 participants and the larger NSHD, using unadjusted linear mixed effect models for men and women separately, using all available measurements. An unstructured residual variance-covariance matrix was used to model the correlation between repeated measures in an individual.

Due to the non-normal distribution of WMHV, generalised linear models (GLM) using the gamma distribution with log link were used to investigate relationships between BMI at each age separately and WMHV at 71 years. Linear regression was used to investigate relationships between BMI at each age and WBV and mean HV at 71 years. Model 1 adjusted for sex, TIV and scanning age. Model 2 also adjusted for SBP contemporaneous with BMI measurement. Model 3 additionally adjusted for other potential cardiovascular confounders: smoking status, diabetic status, hypercholesterolaemia status at the time of Insight 46 assessment, and adult SEP. For brain volume analyses, to explore BMI influences independently of measurable brain pathologies, model 3 also adjusted for global WMHV and β-amyloid status.

BMI change variables were then treated as the main predictor within GLM/linear regression models: Model 1(c): all BMI conditional change variables included and adjusted for sex, TIV and scanning age. Model 2(c): each conditional change variable assessed individually and adjusted for contemporaneous SBP (e.g. in the model examining BMI change between 36 and 43 years, SBP at 43 years was included in the model) and the covariates described in model 3 above. These models address whether, regardless of previous weight gain, there is a period when weight change has a particularly strong association with an imaging outcome measure at age 71 years.

We used logistic regression to test associations between BMI at ages 36 through to 71 years and β-amyloid status (positive or negative). Model 1 adjusted for sex. Model 2 further adjusted for APOE-ε4 status. Model 3 also adjusted for contemporaneous SBP. Additional vascular risk factors were not included in models due to the limited number of β-amyloid-positive individuals. Associations between BMI change and β-amyloid status were investigated using two models: Model 1(c) included all conditional change variables within the same model, adjusting for sex. Model 2(c) assessed each conditional change variable individually and adjusted for contemporaneous SBP and APOE-ε4 status. A differential influence of BMI (or change) on β-amyloid status by APOE-ε4 status was tested using an interaction term in fully-adjusted models.

In an exploratory analysis, all continuous BMI and BMI change models were repeated replacing BMI with AC, to determine whether a measure of central adiposity might be more strongly associated with imaging measures.

Model assumptions were checked with regression diagnostics, including checks of linearity by examination of residuals. Possible non-linear relationships were explored through the creation of a categorical variable defining individuals as normal weight (BMI < 25 kg/m²), overweight (25 < BMI < 30 kg/m²) or obese (BMI > 30 kg/m²) which was then used in models 1–3, replacing BMI as the independent variable. Too few individuals were underweight (BMI < 18.5 kg/m²) at any given age to create a separate category and were therefore treated as normal weight. Sensitivity analyses were performed excluding these individuals. Interactions between BMI and sex at each time point were investigated with appropriate interaction terms.

To explore the potential influence of cumulative adiposity on imaging measures, individuals were categorised by obese status at each time point and then a cumulative score calculated using the sum of the number of time-points that an individual was classed as obese. This variable was used as the predictor of interest in fully-adjusted models for each imaging outcome. Individuals missing any BMI measures were excluded from this analysis.

Ethical approvals for the wider NSHD have been described [26]. Insight 46 was approved by the Queen Square Research Ethics Committee. All participants provided written informed consent.

A data-sharing policy is in place: anonymised data will be shared by request from any qualified investigator (https://​skylark.​ucl.​ac.​uk/​NSHD/​doku.​php).

Limitations include the possibility of survival bias, and loss of individuals with pre-existent significant cognitive symptoms, which may mask possible associations between obesity and cerebral pathology. There are limitations inherent to any birth cohort. Whilst participants are broadly representative of the population born in mainland Britain in 1946, Insight 46 is a cohort consisting of exclusively white British participants, which might reduce generalisability to non-white populations. Moreover, having all been born in the same week, participants went through childhood, adolescence and midlife at the same time and are likely to have been exposed to the similar environmental and societal factors, and prior to current guidance and advice regarding weight, diet, and exercise. These factors are likely to differ from those of individuals born at other times. Individuals in Insight 46 tended to have lower BMI and AC than those in the larger NSHD cohort, although absolute differences were small. We have previously demonstrated that Insight 46 participants are healthier with lower rates of overweight/obesity at age 69 years than in the larger NSHD cohort, and, consistent with these findings, obese individuals were less likely to tolerate scanning [15] reducing the ability to detect true associations. Rates of dementia at this age are very low – 3/471 individuals in this study were diagnosed with dementia and excluded (Figure e1), and we do not think these are likely to have affected our results. There are few underweight individuals, limiting the power to detect potential U-shaped relationships between adiposity and late life cerebral pathology. Obesity may have a detrimental impact on other markers, such as lacunes (noting their relatively low prevalence in our cohort ~ 7%), which we did not investigate. Imaging was only available at a single time-point, and therefore it was not possible to determine the influence of adiposity on longitudinal imaging changes: this will be addressed in future work. Furthermore, because this cohort is largely cognitively normal, we cannot directly investigate associations with dementia prevalence at the present time.

In conclusion, we did not find consistent associations to explain the reported relationship between obesity, particularly in midlife, and late-life dementia risk, using WMHV, β-amyloid status and brain volumes as indicators of brain health. Indeed, being overweight or obese in later midlife and early late life was associated with larger hippocampal volumes, and declining BMI in the 1–2 years prior to scanning was associated with increased risk of β-amyloid positivity, which may reflect the influence of neurodegeneration on body composition. Our findings do not support interventions to tackle obesity as an effective approach towards improving later-life cerebral health, although these remain important for improving other health outcomes including cardiovascular and cancer risk. Declining BMI in later life may be a marker of preclinical AD.