Study population

Data were drawn from the ongoing Whitehall II study where all men and women aged 35–55 working in the London offices of twenty civil-service departments were invited to the study with no inclusion/exclusion criteria; 10,308 were recruited in 19,585–1988, and the response rate was 73% [15]. The baseline consisted of a clinical examination and a standard self-administered questionnaire. The clinical examination is undertaken by research nurses who follow a protocol elaborated by the research team, with examination undertaken in central London premises hired for this purpose or at participants’ homes for those unable to travel to London. Each wave takes around 2 years to complete; follow-up clinical examinations have taken place approximately every 4 to 5 years since baseline (1991, 1997, 2002, 2007, 2012, 2015, and 2020) using the same protocol. Linkage to electronic health records of the UK National Health Service (NHS) was used to obtain records of health outcomes until March 31, 2019. Participants’ written informed consent and research ethics approval were renewed at each contact; the latest was from the Joint UCL/UCLH Committee on the Ethics of Human Research (reference number 85/0938).

Metabolic-obesity phenotypes

Components of the metabolic-obesity phenotypes were measured six times (1991, 1997, 2002, 2007, 2012, and 2015) for each participant and data were extracted from measurements taken at < 60 (range: 40 to 59.9 years), 60 to < 70 (range: 60 to 69.9 years), and ≥ 70 (range: 70 to 84 years) using multiple waves of the study (Additional file 1: Fig. S1). When data were available at several time points within an age category, the measure closest to age 55, 65, and 75 years was chosen for the three groups, respectively.

Weight was measured to the nearest 0.1 kg on digital Soehnle electronic scales with participants in light clothing. Height was measured to the nearest 1 mm using a stadiometer with participants standing erect in bare feet with the head in the Frankfurt plane. BMI was calculated as weight (kg) divided by height (m) squared in kg/m2. Participants were classified based on their BMI as non-obesity (BMI < 30 kg/m2) or obesity (BMI ≥ 30 kg/m2) [16]. Underweight participants (BMI < 18.5 kg/m2) were removed from the analyses. Metabolic health was measured using components of the metabolic syndrome [17]. As in previous studies [6,7,8], the waist circumference criterion was excluded due to collinearity with BMI (variance inflation factor > 100 for both BMI and waist circumference in all study waves, suggesting high collinearity).

Poor metabolic status was defined as a prevalence of ≥ 2 of the following criteria: (a) elevated serum triglycerides (≥ 150 mg/dL [1.7 mmol/L], or use of lipid-modifying drugs); (b) low HDL-C (in men: < 40 mg/dL [1.0 mmol/L] and in women: < 50 mg/dL [1.3 mmol/L], or use of lipid-modifying drugs); (c) elevated blood pressure (systolic blood pressure ≥ 130 mmHg and/or diastolic blood pressure ≥ 85 mmHg, or use of antihypertensive drugs), blood pressure was the mean of two measurements using a sphygmomanometer with the participant in a sitting position after 5 min of rest; and (d) elevated serum fasting glucose (≥ 100 mg/dL [5.6 mmol/L]) or use of glucose-lowering drugs.

Blood samples were handled according to standard protocols, with participants in a fasting state (≥ 8 h fasting or ≥ 5 h for afternoon visits). Venipuncture of the antecubital vein in the left arm used to draw blood, collected in plain and fluoride Sarstedt (Neumbrecht, Germany) monovettes. Plasma or serum was immediately moved into microtubes and stored at – 70 °C. HDL-C was measured by precipitating non-HDL-cholesterol with dextran sulfate-magnesium chloride using a centrifuge and measuring cholesterol in the supernatant fluid. Serum triglycerides were determined by the enzymatic colorimetric method (glycerol-3-phosphate oxidase/phenol and aminophenazone). Serum glucose was measured using the glucose oxidase method (YSI MODEL 2300 STAT PLUS Analyzer, YSI Corporation, Yellow Springs, OH, USA) [18]. The assays were performed by research-accredited laboratories in the London area; technical error was estimated by assaying blinded duplicate samples for 5% of subjects and coefficients of variation were 2.0 to 6.6%.

Metabolic-obesity phenotypes were defined based on obesity (yes/no) and poor metabolic status (yes/no) and included: metabolically healthy non-obesity (MHNO), metabolically unhealthy non-obesity (MUNO), metabolically healthy obesity (MHO) and metabolically unhealthy obesity (MUO).

Dementia

Dementia was ascertained by linkage to three electronic health records databases (the national Hospital Episode Statistics (HES), the Mental Health Services Data Set (MHSDS), and the National Statistics Mortality Register) until March 31, 2019. Dementia cases were identified based on ICD-10 codes F00-F03, F05.1, G30, and G31. The NHS provides most of the health care in the UK, including in- and outpatient care. Ascertainment of all-cause dementia using the HES data has a sensitivity and specificity of 78.0% and 92.0% [19]. The sensitivity in our study is likely higher as we also used data from the MHSDS and the mortality register. The date of dementia was defined as the earliest date at which dementia had been diagnosed via any register.

Cognitive test battery

The cognitive function test battery from measurements in 1997, 2002, 2007, 2012, and 2015 was used in the analyses. The battery included tests of (a) memory, assessed using a 20-word free recall test where a list of one or two-syllable words was presented to participants who then had to write as many words as they could recall within 2 min; (b) reasoning, assessed in 10 min via the Alice Heim 4-I test [20], which is composed of a series of 65 verbal and mathematical items of increasing difficulty, and inductive reasoning tests by measuring the ability to identify patterns and infer principles and rules; and (c) phonemic and semantic fluency, where participants were asked to recall in writing as many words beginning with “s” (phonemic fluency) and as many animal names (semantic fluency) as they could in one minute for each test.

Individual test scores at each wave were standardized to a z-score (mean 0, standard deviation 1) using the mean and standard deviation of the baseline 1997 measure. In addition, a global cognitive score was created by averaging all four standardized tests and then re-standardizing the resulting score, leading to a score of mean 0 and standard deviation of 1. The global z-score is useful as it minimizes measurement error inherent in each individual test [21].

Covariates

Sociodemographic factors included age, sex, ethnicity (white and non-white), education (high, intermediate, or low), and marital status (married, widowed, and single). Health-related behaviors included were smoking (never, former, and current smoker), alcohol consumption (no consumption, 1–14 units per week, and > 14 units per week), consumption of fruits and vegetables (less than daily, once a day, and twice or more a day) and time spent in moderate and vigorous physical activity (hours per week). Cardiovascular disease (CVD) was defined as a history of stroke (assessed with the MONICA-Ausburg stroke questionnaire; and from ICD-10 codes I60–64), coronary heart disease (CHD; assessed from 12-lead resting electrocardiogram recording; and from ICD-10 codes I20–25), and/or heart failure (assessed from ICD-10 code I50). Data on covariates were extracted in a similar manner as the metabolic-obesity phenotypes at < 60, 60 to < 70, and ≥ 70 years, and were concurrent to the measure of these phenotypes for each analysis.

Statistical analysis

We tested whether the associations of metabolic-obesity phenotypes with cognitive decline and dementia varied by sex and found no evidence of differences (all p for interaction > 0.05), leading us to combine men and women in the analyses.

Association of metabolic-obesity phenotypes with incident dementia

We first examined the association of obesity, poor metabolic health, and components of metabolic status (at < 60, 60 to < 70, and ≥ 70 years) with dementia using Cox proportional hazards regression with age as the timescale. The proportional hazards assumption was examined by plotting Schoenfeld residuals and found not to be violated in the analyses. The beginning of the follow-up for incident dementia was age at assessment of the exposure (at < 60, 60 to < 70, and ≥ 70 years), prevalent dementia cases at the start of the follow-up were excluded. Participants were censored at the date of record of dementia, death, or end of follow-up (March 31, 2019), whichever occurred first. Cause-specific hazard models were used to account for competing risk of death. All analyses were first adjusted for sociodemographic factors and birth-cohort effects using 5-year bands of birth-year (model 1), and then for health-related behaviors (model 2) and mutually adjusted (model 3). We then examined the association of metabolic-obesity phenotypes (4 groups with MHNO phenotype as the reference) with incidence of dementia using the approach described above, without model 3.

We performed additional analyses to examine the robustness of our findings. First, we used inverse probability weighting to repeat the main analyses to take missing data into account. Attrition over the study period led to analyses on a smaller number of participants in the analysis on exposure measured at older ages. Inverse probability weighting allowed us to check if the results are affected by missing data [22]. This involved first calculating the probability of being included in the analytical sample using logistic regression, in a model that included demographic (age, sex, and ethnicity), socioeconomic (educational level, occupation, and marital status), behavioral factors (physical activity, smoking status, alcohol consumption, fruit and vegetables consumption, 28-item General Health Questionnaire, SF-36 Physical, and Mental Health Summary Scales), as well as BMI and metabolic components at the 1991 wave, chronic diseases (CHD, stroke, diabetes, chronic obstructive pulmonary disease, cancer, and dementia) during the follow-up, and stepwise-selected interactions between covariates. The inverse of these probabilities was used as weights in the Cox regression. Weights were calculated separately for each age-stratified analysis at ages < 60, 60 to < 70, and ≥ 70 years. Second, as ethnic groups may have different thresholds of obesity [16], we repeated the analyses on only “white” participants; the non-white group was too small to allow further analyses in this group. Third, to examine the role of prevalent CVD in our analyses, we excluded participants with CVD at baseline in the analyses. Fourth, as even one metabolic abnormality may affect dementia risk [23], we repeated the main analyses using an alternative definition with unhealthy metabolic status defined as the prevalence of ≥ 1 instead of ≥ 2 metabolic components.

Association of metabolic-obesity phenotypes with cognitive decline over 18 years

In complementary analyses, we examined the association of metabolic-obesity phenotype components (separate models) and phenotypes (4 groups) at the 1997 wave of data collection with cognitive decline between 1997 and 2015 using linear mixed models [24], with time of follow-up as the timescale. These models consider the fact that repeated measures on the same individual are correlated, and use all available data during the follow-up period. Individual differences in cognitive performance in 1997 and the rate of cognitive decline were estimated by fitting both the intercept and slope as random effects. The analyses included terms for metabolic-obesity phenotype, time of follow-up, time2, and the interaction of metabolic-obesity phenotype with time terms, and were adjusted for age and sociodemographic factors (Model 1), and then for health-related behaviors (Model 2) at the baseline in these analyses (1997 wave). All models included interactions of covariates with time, and interactions of covariates with time2 when p < 0.05.

Sensitivity analyses: association of trajectories of metabolic-obesity phenotypes between 1991 and 2002 with incidence of dementia (2002 to 2019) and 12-year cognitive decline (2002 to 2015)

While the focus of the main analyses was on the role of age at measurement of metabolic-obesity phenotypes we examined the association between trajectories of these phenotypes and both incident dementia and cognitive decline in sensitivity analyses. To allow sufficient follow-up for dementia we used data on metabolic-obesity phenotypes from measures in 1991, 1997, and 2002 to construct trajectories using group-based trajectory modeling [25]. The STATA Traj package’s censored normal model was used for these analyses. To determine the optimal number of trajectories, analyses were repeated to obtain 4 to 6 trajectories. Linear and quadratic functional forms were used to choose the best-fitting models, and the optimal trajectory shape and number of groups were determined based on the following criteria: (1) the number of participants within each trajectory group (≥ 5% of the total sample size); (2) the average posterior probability of each trajectory group (≥ 0.70); and (3) the lowest BIC/AIC value [26, 27]. The trajectories identified by these analyses were used as the exposure in Cox regression (for incident dementia) and linear mixed models (for cognitive decline between 2002 and 2015); covariates in these analyses were drawn from the 2002 wave.

Analyses were undertaken using STATA version 16.1 (StataCorp). A two-sided p value < 0.05 was considered statistically significant.