Overweight or obesity increases the risk of cardiovascular disease among older Australian adults, even in the absence of cardiometabolic risk factors: a Bayesian survival analysis from the Hunter Community Study

To estimate the risk of cardiovascular disease (CVD) in older adults with overweight or obesity without metabolic risk factors using a Bayesian survival analysis. Prospective cohort study with median follow-up of 9.7 years. Newcastle, New South Wales, Australia. A total of 2313 community-dwelling older men and women. Participants without known CVD and with a body mass index (BMI) ≥ 18.5 kg m2 were stratified by BMI and metabolic risk to create six BMI-metabolic health categories. Metabolic risk was defined according to the International Diabetes Federation criteria for metabolic syndrome. ‘Metabolically healthy’ was defined as absence of metabolic risk factors. Bayesian survival analysis, incorporating prior information from a previously published meta-analysis was used to assess the effect of BMI-metabolic health categories on time from recruitment to CVD. Incident physician-diagnosed CVD, defined as fatal or nonfatal myocardial infarction, fatal or nonfatal stroke, angina, or coronary revascularisation procedure, was determined by linkage to hospital admissions records and Medicare Australia data. Secondary outcomes were cardiovascular mortality and all-cause mortality. From 2313 adults with complete metabolic health data over a median follow-up of 9.7 years, 283 incident CVD events, 58 CVD related deaths and 277 deaths from any cause occurred. In an adjusted Bayesian survival model of complete cases with informative prior and metabolically healthy normal weight as the reference group, the risk of CVD was increased in metabolically healthy overweight (HR = 1.52, 95% credible interval 0.96–2.36), and in metabolically healthy obesity (HR = 1.86, 95% credible interval 1.14–3.08). Imputation of missing metabolic health and confounding data did not change the results. There was increased risk of CVD in older adults with overweight or obesity, even in the absence of any metabolic abnormality. This argues against the notion of ‘metabolically healthy’ overweight or obesity.


INTRODUCTION
Obesity is a global epidemic [1] and may be an independent risk factor for cardiovascular disease (CVD) [2]. High body mass index (BMI) is also associated with increased mortality [3,4]. Studies show that overweight and obesity lead to a decrease in life expectancy [5] and greater Years of Life Lost (YLL) [6,7]. Obesity also increases the risk of developing metabolic syndrome [2].
Metabolic syndrome is characterised by the presence of any three of the following: central obesity, elevated triglyceride (≥1.7 mmol/L), low high density lipoprotein cholesterol (HDL) (<1.03 mmol/L in men, <1.29 mmol/L in women), elevated blood pressure ≥130/85 mmHg or fasting glucose ≥5.6 mmol/L. Cardiometabolic abnormalities are known risk factors for CVD. However, those with overweight and obesity without metabolic abnormalities may still be at risk for CVD. Individuals with overweight or obesity without metabolic risk factors are paradoxically referred to as 'metabolically healthy overweight' (MHOW) and 'metabolically healthy obesity' (MHO) [8][9][10].
Four systematic reviews with meta-analyses by Kramer et al. [23], Fan et al. [24], Zheng et al. [25], and Eckel et al. [26] all reported that individuals classified as MHO are at increased risk of developing CVD, however the definition of metabolic health differed across studies. Some studies defined metabolically healthy as having no metabolic risk factors, others considered metabolically healthy as having one or less risk factors, while other studies considered those with two or less risk factors as being metabolically healthy. Most previous studies of MHOW or MHO and risk of CVD used different definitions of metabolically healthy and many did not compare with a valid metabolically healthy control group. Research shows that the presence of even one metabolic risk factor confers an increased risk of CVD [15,27].
An updated systematic review with meta-analysis by our team examined the risk of CVD in MHOW and MHO, using a reference group with no metabolic risk factors, and found an increased risk of CVD in MHOW and MHO [28]. The study found that the criteria for defining metabolic health was a major cause of heterogeneity between studies, which underscores the need to use a standard reference group with no metabolic risk factors.
The only published Australian study examining the association between MHO and CVD in adults used presence of one or less metabolic risk factors to define metabolically healthy and found no increased risk of CVD [12]. Given that previous studies have produced inconsistent findings and some studies compared MHOW or MHO with a control group with metabolic risk factors, more knowledge about the health effects of the MHOW/MHO phenotype is needed. If further research confirms an increased risk of CVD this may be important for targeted preventative strategies in individuals with overweight or obesity but no metabolic abnormalities. Therefore, the aim of this study was to investigate individuals classified as having overweight or obesity with or without cardiometabolic abnormalities and the risk of CVD in a cohort of older Australian men and women.

METHODS Sample
This study utilised data from the Hunter Community Study (HCS), a population-based cohort study of community-dwelling men and women aged 55-85 years at baseline living in Newcastle, New South Wales (NSW), Australia. A detailed description of the cohort has been published already [29]. Briefly, random selection was made from the NSW State electoral roll for participants who were contacted from December 2004 to December 2007. The study used a modified Dillman recruiting strategy [30]. People who were non-English speaking and those residing in residential aged-care facilities were excluded.
Participants were asked to complete self-administered questionnaires and attend in person for a series of clinical assessments. They were also asked for consent to link their questionnaire data to Medicare Australia data so that future health events and health services interactions could be determined. Medicare is the publicly funded universal health care insurance scheme in Australia. A total of 3318 people responded to the questionnaires, a response rate of 44.5%. HCS participants were followed up in 2010 and 2013 with linked data to Medicare Australia and Pharmaceutical Benefit Scheme (PBS) and Hunter New England Area Health Service Medical Records available up to 2017. The response rate for 2010 was 74%. Participants who completed follow-up were younger and had similar gender profile compared to those who did not complete.
The current analysis excluded participants who were underweight (<18.5 kg) and those with CVD at baseline as defined below. The study was conducted according to the guidelines laid down in the Declaration of Helsinki and all procedures involving research study participants were approved by the University of Newcastle Human Research Ethics committee (code: H-820-0504).

Exposure measurement and classification of BMI and metabolic status
Metabolic syndrome was defined as presence of ≥3 of the following components according to International Diabetes Federation (IDF) criteria for metabolic syndrome: serum triglyceride ≥1.7 mmol/L; serum HDL cholesterol <1.0 mmol/L in men or <1.3 mmol/L in women or lipidlowering medication use; blood pressure ≥130/85 mmHg or antihypertensive medication use; serum fasting glucose ≥5.6 mmol/L or self-reported diabetes; and waist circumference of >102 cm in men and >88 cm in women.
"Metabolically unhealthy" was defined as having at least one of the risk factors consistent with the IDF metabolic syndrome criteria. Participants with no risk factors were classified as metabolically healthy. BMI was categorised according to World Health Organization criteria [31]. Participants were classified according to their BMI as normal weight (18.5-24.9 kg/m 2 ), overweight (25-29.9 kg/m 2 ), or obese (≥30.0 kg/m 2 ). Metabolic status was combined with BMI category to create six phenotypes as summarised below:

Outcome measurements
Primary outcome. The primary outcome in this study is incident CVD, defined as physician-diagnosed fatal or non-fatal myocardial infarction, stroke, angina, or coronary revascularisation procedure according to International Classification of Diseases 10th revision (ICD-10), CVD codes 101-199. Information was obtained by data linkage to Hunter New England Area Health Service Medical Records, Medicare Australia, or Pharmaceutical Benefits Scheme (PBS), available up to December 2017.
Secondary outcomes. Secondary outcomes were CVD mortality and allcause mortality. Death was based on data from the NSW Registry of Births, Deaths and Marriages. Cause of death was determined using ICD-10 codes from Australian Bureau of Statistics (ABS) cause of death data. Where data were not available, cause of death was imputed based on ICD-10 codes for the cause of the most recent inpatient admission before death.

Measurement of potential confounders
Based on previous research on the association between MHOW or MHO and risk of CVD, directed acyclic graphs (DAG) [32] were used to determine the confounding variables controlled for in this analysis, using content expert knowledge as well as the literature to inform the causal structure of the DAG. Confounding variables included age, gender, household income, education, smoking, alcohol intake, and physical activity level. Age, gender, and education were based on information provided in self-administered questionnaires. Education was classified into two groups of highest level of education attained: primary and secondary level schooling or University and other tertiary level study. Household income was the annual household income before tax (<$40,000 and ≥$40,000). Physical activity was measured by recorded step counts using pedometer worn by participants for seven consecutive days during waking hours and was reported as mean number of steps per day. Smoking frequency was measured by number of cigarettes per day and reported as packs per year and alcohol consumption was measured by the number of standard drinks per day defined according to National Health and Medical Research Council (NHMRC) criteria, and reported as mean alcoholic standard drinks per day [33].

STUDY DESIGN AND STATISTICAL ANALYSIS
Bayesian survival analysis, incorporating prior information from a previously published meta-analysis [28], was used to assess the effect of metabolic health and weight status groups on time from recruitment to CVD, CVD mortality, and all-cause mortality. Cox proportional hazard regression was also used in the analysis. In all models, prevalent CVD cases and those with a BMI < 18.5 kgm 2 (underweight) were excluded from all analyses.
Results from a complete case analysis are presented together with results from a multiple imputation, where missing data are completed using a Classification and Regression Tree (CART) and pooled using Rubin's rules [34,35]. Crude and multivariable adjusted models are presented. Bayesian models were fit to the data, initially with non-informative priors for all the regression model parameters, but then using the data from a previous meta-analysis to inform the prior distribution for the metabolic health/BMI status regression parameter. Flat (uninformative) priors are used for all other parameters, while normally distributed priors for the metabolic health/BMI regression parameters are used (on the log-hazard scale). Samples from the posterior distribution were obtained using the No-Uturn Sampler implemented in the R package BRMS [36], with 6000 draws from the posterior distribution and an initial burn-in period of 1000 samples. Four chains were used with random initial starting values to assess stability of convergence, and the Rhat statistic, as well as the effective sample size are reported, together with the median hazard ratio (HR), the 95% highest posterior credible interval, and the probability that the effect is positive (pb). The Rhat diagnostic statistic measures between chain variance of sampled parameter values to within chain variance, where values greater than 1.05 indicate mixing/ convergence problems. Competing risk Cox proportional hazards regression models were used to assess the relationship between metabolic health and weight status groups on time from recruitment to CVD, with death prior to CVD as a competing risk. The models included gender, baseline age, alcohol consumption, physical activity (average daily step-counts), household income, education status, and smoking status as confounding variables. For CVD-specific deaths, deaths due to other causes were treated as competing risk, allowing estimation of cause-specific hazard ratios (HR). No auxiliary variables were used, and the regression equations included each of the confounding variables. Twenty imputed data sets were created as there was <20% missing data. Subgroup analysis to examine gender effects of MHOW and MHO with CVD was undertaken. Kaplan-Meier survival curves were also constructed. Statistical analysis was completed using STATA software version 16 [37].

RESULTS
Of the 3318 participants enroled in the HCS, 584 participants with existing CVD at baseline were excluded. Twenty-one participants with no CVD were also excluded (10 underweight with missing metabolic health data, and 11 underweight with non-missing metabolic health data). Of the remaining 2713 participants, 400 had missing metabolic health data, leaving 2313 participants with complete metabolic health and BMI data allowing classification into one of the six phenotypes. A further 448 participants with incomplete confounding variable data were excluded, leaving 1865 people with complete phenotype and confounding variable data included in the complete case analysis, and a sample size of 2713 for the imputation analysis data set (Fig. 1).

Baseline characteristics of study population
Sociodemographic and baseline characteristics of the total study population according to metabolic health and BMI phenotype are reported in Table 1. Approximately 23.4% and 12.9% of the overall participants were classified as having MHOW and MHO respectively. Of the participants with obesity, 38% were classified as metabolically healthy.
Compared to the healthy normal weight group, the MHOW or MHO groups were younger, had higher a proportion of males, higher household income and level of education, fewer never smokers, and lower daily step counts. However, participants categorised as metabolically healthy and having overweight or obesity were younger, more educated with higher household income, were less likely to smoke, and more physically active than the unhealthy group ( Table 1).  Fig. 1 Flow diagram for selection of participants from the Hunter Community Study (HCS) for the study of the association between obesity-metabolic health phenotypes and risk of cardiovascular diseaese. HCS total number of participants (N = 3318). Exclusion of participants: with prior cardiovascular disease (N = 584); underweight (N = 21); and missing metabolic health measurements (N = 400). Participants free of cardiovascular disease at baseline (2734). Participants with BMI ≥ 18.5 kgm 2 (N = 2713). Participants with metabolic health measurements for six obesity/metabolic health phenotypes for follow-up (N = 2313). Participants with complete data for six obesity/metabolic health phenotypes and confounders (N = 1865).
Outcome events for CVD, CVD-related death and all-cause mortality After a median follow-up of 9.7 years for participants across all obesity phenotypes, 283 developed CVD. There were 58 CVDrelated deaths and 277 deaths from all causes.  (Table 2). In all models the Rhat was <1.05 and pd was >95%.

MHOW AND MHO AND RISK OF CVD
Adjusted complete case informative prior. In a multivariable Bayesian model of complete cases and informative prior, the adjusted HR for CVD in MHOW was 1.52 (95% credible interval: 0.96-2.36) and in MHO was 1.86 (95% credible interval: 1.14-3.08). The HRs for CVD were similarly increased in the remaining metabolically unhealthy phenotypes with even higher HRs   Table 2). The informed prior adjusted effects remained similar when missing data were imputed. In all models the Rhat was <1.05 and pd was >95%. The risks of CVD in different obesity phenotypes were not affected by gender. Kaplan-Meier survival curves showed the highest survival in the MHNW group, followed by MHOW and MHO groups. Survival in the MUNW was lower than in the MHOW and MHO groups but better than in the MUOW and MUO groups, which had the lowest survival probability (Fig. 2).

MHOW AND MHO AND RISK OF CVD MORTALITY Bayesian survival analysis
Adjusted multiple imputed data with uninformative prior. In a multivariable Bayesian model of imputed cases and uninformative prior, the adjusted HR for CVD mortality in MHOW was 5 (Table 3a). In all models the Rhat was <1.05 and pd was >95%.
Adjusted multiple imputed data with informative prior. In a multivariable Bayesian model of imputed cases and informative prior, the adjusted HR for CVD mortality in MHOW was 1.63 (95% credible interval: 0.62-4.23) and in MHO was 4.68 (95% credible interval: 1.81-12.02). The remaining metabolically unhealthy phenotypes had increased HR for CVD mortality with even higher risks observed for those classified as overweight/obese and metabolically unhealthy; MUNW (HR  (Table 3a). The uninformative prior adjusted effects on the imputed data were attenuated when the informative prior was applied to the adjusted imputed data, with increased HR for CVD mortality observed for MHO and all metabolically unhealthy groups. In all models the Rhat was <1.05 and pd was >95% except for MHOW.
The uninformed prior adjusted effects were slightly strengthened when an informative prior was applied to the adjusted imputed data, with an increased HR for all-cause mortality observed for MHO and all metabolically unhealthy groups. In all models the Rhat was <1.05 and pd was <95% in all groups except MUO.
Cox proportional hazard regression. The frequentist approach conducted for all three outcome categories showed similar results to the Bayesian analysis (Supplementary Tables 1 and 3).
Sensitivity analysis using competing risk analysis with death as a competing risk did not change the results for risk of CVD, CVD mortality or all-cause mortality (Supplementary Table 2).

DISCUSSION
In this study of 2313 older Australian men and women with median follow-up of 9.7 years, there were 283 incident CVD events. The results indicate that older adults with overweight or obesity without metabolic abnormality have increased risk of CVD.
Of the total participants studied, 23.4% were classified as MHOW and 12.9% were classified as MHO. Based on the findings from the Bayesian survival analysis there was a 52% increased risk of CVD in the MHOW group and an 86% increased risk of CVD for the MHO group, compared with the MHNW reference group. This study defined metabolically healthy participants as having no metabolic abnormality, similar to only a few other studies, which also demonstrated increased risk of CVD in MHOW and MHO [15][16][17][18]38]. Most studies, including the only Australian published study on risk of CVD in MHO, allowed one [12,39,40] or two [22,[41][42][43][44][45][46][47] metabolic abnormalities to define metabolically healthy. Hence, most studies did not use a valid metabolically healthy reference group, which may result in misclassification bias and effect estimates that are biased towards the null. Other studies did not demonstrate increased risk of CVD in MHOW and  MHO [12,[19][20][21][22] likely due to small sample sizes, or the statistical method used in the analyses. The increased risk of CVD in MHOW and MHO is consistent with other studies which found similar increased risk of CVD [13][14][15][16][17][18]43]. Two of these studies [15,16] had large sample sizes and therefore increased statistical power and more precise effect estimates. Our novel study used a Bayesian survival analysis which, by incorporating prior information, provides more precise effect estimates than frequentist analyses in populations with a relatively small number of outcomes. However, the magnitude of the effect estimates were similar even when a frequentist Cox proportional hazards regression approach was used.
Seven previous systematic reviews with meta-analyses found an increased risk of CVD in MHOW and MHO [23-26, 28, 48, 49]. However, two of these systematic reviews found no significant increase in risk of CVD when analysis was restricted to participants with no metabolic risk factors [25,26]. The findings from this study are consistent with the findings of our updated meta-analysis. This may have clinical implications where individuals with overweight or obesity and no metabolic abnormality and no established cardiovascular disease could receive targeted interventions aimed at achieving normal weight to lower their risk of developing CVD.
The findings in this study and of that in the most recent metaanalysis [28] are important in light of the US Preventive Service Task Force (USPSTF) recommendation [50] that adults aged 18 years or older categorised as having overweight or obesity are only referred for weight-loss intervention when there are additional CVD risk factors. In our study, 33% of participants were overweight or obese with no metabolic risk factors, which would render them ineligible for weight-loss intervention referral according to the USPSTF recommendation. The findings of this study indicate that being overweight or obese should be considered as a criterion for referral for such pre-emptive behavioural intervention because even without metabolic abnormality, these individuals are at significantly increased risk of CVD.
The risk of CVD mortality was increased in MHOW and MHO in the current study, similar to findings from earlier studies [41,42,51], however the effect estimates are imprecise. This contrasts with a recent systematic review and earlier studies in which individuals classified as MHO were not at increased risk of CVD mortality [13,28]. The risk of all-cause mortality was increased in those classified as MHO but not in MHOW. This is similar to earlier studies [42]. Some earlier studies however found increased risk in all-cause mortality in MHOW and MHO [13,41] but other studies did not [25,51]. The results of this study need to be interpreted with caution because of the small number of deaths from CVD or from all causes.

Strength and limitations
This study has several strengths. First, the sample consisted of a representative group of community-dwelling men and women randomly selected from the general population. Second, the use of Bayesian analysis provides more precise effect estimates given the relatively small number of outcome events. Third, the use of a metabolically healthy reference group with no cardiometabolic risk factors provides a more valid reference group with which to estimate CVD risk. Finally, although the participants in our study are predominantly older adults, the results of the study demonstrated increased risk of CVD in MHOW or MHO similar to studies with younger populations.
Limitations of this study includes the relatively small number of CVD-related events which may be due to the length of follow-up period as cohort studies generally need a longer follow-up period in diseases with long latency. The relatively small sample size also led to imprecise effect estimates and this should be considered in the interpretation of the study findings. Finally, the study did not have quantitative measurements of visceral fat thereby limiting our ability to further characterise obesity.

CONCLUSION
This prospective cohort study found an increased risk of cardiovascular disease in an older Australian population of men and women with overweight or obesity without cardiometabolic abnormalities compared to those with normal weight and no cardiometabolic abnormalities. Individuals with overweight or obesity are at higher risk for CVD events even in the absence of any additional cardiometabolic risk factors. Further research is needed to determine if weight loss interventions targeted to these groups, reduce the risk of developing cardiovascular disease.

DATA AVAILABILITY
The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.