Plasma adiponectin levels and type 2 diabetes risk: a nested case-control study in a Chinese population and an updated meta-analysis

Results from previous prospective studies assessing the relation between adiponectin and type 2 diabetes (T2D) were not entirely consistent, and evidence in Chinese population is scarce. Moreover, the last meta-analysis did not examine the impact of metabolic variables on the adiponectin-T2D association. Therefore, we prospectively evaluated the adiponectin-T2D association among 571 T2D cases and 571 age-sex-matched controls nested within the Singapore Chinese Health Study (SCHS). Furthermore, we conducted an updated meta-analysis by searching prospective studies on Pubmed till September 2016. In the SCHS, the odds ratio of T2D, comparing the highest versus lowest tertile of adiponectin levels, was 0.30 (95% confidence interval: 0.17, 0.55) in the fully-adjusted model. The relation was stronger among heavier participants (body mass index ≥23 kg/m2) compared to their leaner counterparts (P for interaction = 0.041). In a meta-analysis of 34 prospective studies, the pooled relative risk was 0.53 (95% confidence interval: 0.47, 0.61) comparing the extreme tertiles of adiponectin with moderate heterogeneity (I 2 = 48.7%, P = 0.001). The adiponectin-T2D association remained unchanged after adjusting for inflammation and dyslipidemia markers, but substantially attenuated with adjustment for insulin sensitivity and/or glycaemia markers. Overall evidence indicates that higher adiponectin levels are associated with decreased T2D risk in Chinese and other populations.

The pooled relative risk (RR) comparing the highest to lowest tertile of adiponectin concentrations was 0.53 (95% CI: 0.47, 0.61) using the random-effects model (Figure 1), and results were similar when using the fixed-effects model (RR 0.54; 95% CI: 0.50, 0.59). A moderate heterogeneity was observed (I 2 = 48.7%; P = 0.001), and the meta-regression analysis suggested evidence of effect modification by level of adjustment (P = 0.08) (data not shown). We further performed stratified analysis by level of adjustment, and the association was attenuated after further adjustment for glycaemia markers and/or insulin sensitivity markers, while additional adjustment for lipids and/or inflammatory markers did not materially change the association (Supplemental Table S5). When stratifying by baseline participant characteristics, the inverse association was consistently observed among all subgroups (Supplemental Table S5).
Twelve studies (14 data points because two studies reported results for men and women separately 10,28 ) provided data for dose-response meta-analysis (Supplemental Table S6). Cubic spline regression model suggested a linear dose-response relationship (P = 0.33 for nonlinearity) (Figure 2). The RRs of T2D with per 5 µg/mL increment in adiponectin levels were 0.75 (95% CI: 0.73, 0.77) in the random-effects model. No significant publication bias was detected (P = 0.12 in the Egger's test).

Discussion
In this nested case-control study, we found a strong dose-dependent association between higher plasma adiponectin levels and lower T2D risk, which seemed to be independent of established T2D risk factors, including lipids, inflammatory biomarker (hs-CRP), random glucose and insulin. Moreover, the inverse association was Cases (n = 571) Controls (n = 571) P-value b Age (years) at blood taken 59.6 (6. significantly stronger among overweight/obese subjects than their lean counterparts. Furthermore, adiponectin significantly improved T2D risk reclassification in this population. In an updated meta-analysis which pooled data from 34 prospective studies including the current study, participants in the highest tertile had almost half the risk of developing T2D compared to those in the lowest tertile (RR 0.53; 95% CI: 0.47, 0.61), and the adiponectin-T2D association appeared to be linear. A previous meta-analysis of 13 prospective studies (2623 incident T2D cases and 14,598 participants) reported an inverse association between adiponectin levels and T2D risk (RR 0.72 for per 1-log µg/mL increment in adiponectin levels) across different populations 5 , but no study was done in Chinese population. In the current study, we not only observed the inverse association in a Chinese population after comprehensive adjustment, but also conducted an updated meta-analysis with a much larger sample size (7504 incident T2D cases and 64,696 participants), which provided strong evidence for the role of adiponectin in the pathogenesis of T2D across diverse populations, a consistent finding despite methodologic variations such as adiponectin assays, diagnostic criteria for T2D and follow-up durations.
Several lines of evidence have hypothesized that adiponectin may act against T2D development by lowering blood glucose through improving insulin sensitivity. Mechanistic studies have shown that adiponectin improves insulin sensitivity by stimulating glucose utilization and fatty acid oxidation in the skeletal muscle and liver through improving AMP-activated protein kinase 4 . Animal studies have shown that adiponectin levels drop in parallel with reduced insulin sensitivity, before the increment in glycemic levels 3 . A recent Mendelian randomization study in humans provided further evidence between increased adiponectin levels and improved insulin sensitivity using ADIPOQ SNPs as instruments 39 . Additionally, we found in our current meta-analysis that the adiponectin-T2D association was substantially attenuated in studies that adjusted for insulin sensitivity and/or glycaemia markers compared to studies that did not adjust for either one of them, suggesting that the association was at least partially mediated through insulin and glucose pathways. Moreover, the stratified analysis from the current meta-analysis suggested that lower adiponectin levels may contribute to an increased T2D risk through pathways independent of other important T2D pathogenic mechanisms such as dyslipidemia and inflammation. Although adiponectin is an adipose-specific protein and its concentration reduces with increasing level of obesity 40 , the current cohort study together with previous studies 10, 11,22 have shown that adjustment for markers of  obesity, such as BMI or CT-assessed abdominal fat area, did not materially change the association between adiponectin and T2D risk, suggesting that low adiponectin levels could be associated with higher T2D risk through pathways that may not be related to systemic or regional fat deposition. Moreover, our study found stronger relation of adiponectin with T2D risk in overweight/obese people than their leaner counterparts. In line with our results, a 4-year follow-up study of elderly British men found stronger association among obese participants (BMI ≥30.0 kg/m 2 ) than non-obese counterparts 32 . Similarly, a Japanese cohort study reported stronger association in overweight/obese individuals (BMI ≥25.0 kg/m 2 ) than in normal weight people 22 . Some other studies also found that the association between adiponectin levels and T2D were stronger in participants with higher metabolic risk profile 11,17,41 . A recent Mendelian randomization study in a Swedish population also observed a stronger causal relation between adiponectin and insulin sensitivity in participants with higher BMI 39 . Although the underlying mechanism for the observed interaction with obesity is still unclear, the Mendelian randomization study hypothesized that the observed interaction with BMI may be due to the interaction between adiposity and the ADIPOQ SNPs, in which the transcription factors could have been regulated by the inflammatory mediators to stimulate the binding site and thus cause up-regulation of the related gene 39 . However, given the multiple statistical tests for interaction conducted and that some studies did not observe significant interactions with BMI 13,27,30 , our observed interaction with BMI should be interpreted with caution.
In the current study, adding adiponectin beyond traditional risk factors and glucose markers (random glucose or HbA1c) did not improve discrimination (assessed by AUC), but improved risk reclassification (assessed by NRI and IDI) for T2D. In consistent with our results, two cohort studies from Switzerland and Germany did not observe significant improvement in AUC when adding adiponectin alone to base models that included fasting glucose 16,34 . In contrast, results from a cohort in Hong Kong and EPIC-Potsdam study have shown that adiponectin improved AUC significantly beyond glucose markers 26,37 , and interestingly, in a previous study conducted in Finland, adiponectin improved AUC and NRI significantly in the primary cohort (FINRISK97) but not in the validation cohort (Health 2000) 6 . In addition, a T2D score comprising adiponectin and other five biomarkers has shown better predictive utility (assessed by AUC) than HbA1c and fasting glucose in a Danish population 33 , and the predictive utility of the score was further validated in two cohorts from Finland 35 and US 36 using AUC and NRI.
Our study has several strengths. First, we have adjusted for well-established T2D factors and used comprehensive statistical methods to explore the predictive utility of adiponectin. Moreover, compared to the previous meta-analysis 5 , our updated meta-analysis included more prospective studies with a tripled sample size; we have also included results from models that have adjusted for a comprehensive panel of T2D biomarkers (lipids and markers of inflammation, glycaemia and insulin sensitivity) to examine their impact on the adiponectin-T2D association, and further performed stratified analysis by these factors to explore potential heterogeneity among subgroups. Furthermore, we have performed a dose-response meta-analysis among 12 studies, and observed a linear association between adiponectin and T2D. However, the current study has some limitations as well. First, we measured adiponectin levels only once and some measurement errors are inevitable. In addition, the height and weight were self-reported and residual confounding was possible. Hence, this could lead to non-differential misclassifications and result in an underestimation of the true effect size of adiponectin on T2D risk. Furthermore, information of more precise surrogates for insulin resistance (such as fasting insulin, HOMA-IR, and insulin sensitivity index) and glycaemia markers (such as fasting glucose) were not available in the study, thus, random insulin and glucose levels were used as covariates in our models. Moreover, incident T2D was obtained from self-reported information, thus undiagnosed T2D may exist. However, we have used HbA1c as a selection criterion (<6.0%) for controls to minimize bias due to undiagnosed T2D. In addition, we also conducted sensitivity analysis restricting to cases with HbA1c <6.5% and <6.0% at the time of blood collection and their respective controls, and observed similar adiponectin-T2D association. Furthermore, the present study was conducted in a middle-aged and elderly population with a higher T2D incidence and adiponectin levels, and the findings may not be applicable to younger age group 28 . However, in the current meta-analysis, we did not observe any heterogeneity of results by age group, suggesting the finding will be most likely to be observed in younger people as well.

Materials and Methods
Study population. The design of the SCHS was described in detail previously 42 . Briefly, the SCHS was established between 1993 and 1998, and recruited 63,257 Chinese adults aged 45-74 years. At recruitment, an in-person interview was conducted using a structured questionnaire to collect information of diet, lifestyle habits and medical history. Follow-up I interviews were conducted via telephone between 1999 and 2004 to update selected lifestyle habit and medical history. A total of 52,322 participants were re-contacted successfully, and among them, a total of 32,535 participants donated their bio-specimens during follow-up I visits. Follow-up II interviews were conducted via telephone from 2006 to 2010, and 39,528 participants were re-contacted successfully. Among the 32,535 participants who donated bio-specimens, 25,477 (78.3%) were re-contacted for follow-up II interviews. The study protocol was approved by the Institutional Review Boards at the National University of Singapore and the University of Pittsburgh, and all methods were performed in accordance with the relevant guidelines and regulations. Informed consent was provided with completion of the baseline interview.

Ascertainment of diabetes and other covariates.
History of physician-diagnosed diabetes was asked at baseline and both follow-up interviews using the question: "Have you been told by a doctor that you have diabetes?" If the answer was "yes", participants were also asked for the age at which they were first diagnosed. The robustness and accuracy of the self-reported diabetes data has been confirmed in a validation study: among 1651 cohort participants who reported history of diabetes either at baseline or follow-up I interview, 98.9% were confirmed by medical records or telephone interview. Some participants (n = 619) refused or were not available for the validation study, but their main characteristics (age, sex, BMI etc.) were similar to those in the validation study 43 . Body weight and height were self-reported at baseline and both follow-ups. BMI is calculated as weight (kilograms) divided by height (meters) squared. For those with missing weight or height, BMI was calculated using imputed weight or height derived from the linear regression equation: Weight = y-intercept + gradient × height,  Table 4. Results of the 33 prospective studies included in the meta-analysis. Abbreviations: OR, odds ratio; CI, confidence interval; TNF-1 , the tumour necrosis factor 1; PAI-1, plasminogen activator inhibitor 1; TC, total cholesterol; LDL-C, LDL cholesterol; HDL-C, HDL cholesterol; TG, triglycerides; TG/HDL-C, the ratio of TG to HDL-C; CRP, C-reactive protein; hs-CRP, high-sensitivity CRP; IGT, impaired glucose tolerance; IFG, impaired fasting glucose; GGT, gamma-glutamyltransferase; ICAM-1, intercellular adhesion molecule 1; S 1 , insulin sensitivity index; eGFR, the epidermal growth factor receptor; CHD, coronary heart disease; DBP, diastolic blood pressure; MI, myocardial infarction. where values for the y-intercept and gradient were derived from gender-specific weight-height regression lines obtained from all cohort participants with known heights and weights.

Establishment of nested case-control study.
For the current analysis, we established a nested case-control study of 571 cases and 571 matched controls within this cohort. All cases and controls were free of physician-diagnosed T2D, cardiovascular disease and cancer at baseline interview as well as the time of blood collection during 1999 and 2004. Cases were a total of 571 participants who subsequently reported to be diagnosed with incident T2D during follow-up II visit (2006)(2007)(2008)(2009)(2010). Controls were chosen from the remaining participants who did not develop T2D or cardiovascular disease at follow-up II, and were matched for age (±3 years), date (±6 months) of blood collection, sex and dialect group with the cases on a 1:1 ratio. Furthermore, the selected controls were screened for the presence of undiagnosed T2D at the time of blood donation by HbA1c measurements. All matched controls with HbA1c ≥6.0% (42 mmol/mol) were considered ineligible for the study and a replacement control with the same matching criteria was randomly chosen among the remaining eligible subjects.
Laboratory procedures. Random morning peripheral blood samples were obtained, and frozen plasma aliquots from the cases and controls were analyzed simultaneously in the same batch at the National University Hospital Reference Laboratory, Singapore. Adiponectin levels were measured by ELISA/Evolis (Bio-Rad Laboratories, Hercules, CA). The within-assay and between-assay coefficients of variation were 3.9-5.9% and 6.3-7.0%, respectively. Plasma hs-CRP levels and blood lipids (total cholesterol, TG and HDL-C) were measured via colorimetric method on a chemistry analyzer (AU5800 Analyzer, Beckman Coulter, Brea, CA). HbA1c was measured by HPLC method using Bio-Rad Variant II ™ System (Bio-Rad Laboratories, Hercules, CA) in red blood cells.
Statistical analysis. Study participants were divided into tertiles and the lowest tertile served as the reference group. We used conditional logistic regression models to compute the OR and corresponding 95% CI between adiponectin and T2D with adjustment for age (continuous), education level (primary school and below, secondary or above), smoking status (never, ever smoker), alcohol consumption (never, ever drinker), weekly moderate-to-vigorous physical activity levels (<0.5, ≥0.5 hours/week), history of hypertension (yes, no), BMI (continuous), and tertiles of biomarkers, including the ratio of TG to HDL-C, hs-CRP, random glucose and random insulin levels. We also calculated the T2D risk associated with per one log mg/L increment in adiponectin levels, in order to compare with previous studies 5 . To examine the impact of potential selection bias, we examined the adiponectin-T2D association among cases with HbA1c <6.5% or <6.0% and their matched controls. We stratified the analysis by sex, BMI, physical activity, smoking habits and hs-CRP level using unconditional logistical regression models with additional adjustment for sex and dialect group, and we used restricted cubic spline regression with 3 knots at 25 th , 50 th and 75 th percentiles of adiponectin concentrations to examine the linearity of adiponectin-T2D association. The predictive utility of adiponectin was examined by establishing three logistic regression models. Base model 1 included education level, physical activity, history of hypertension, BMI, TG, HDL-C, and hs-CRP; base model 2 additionally included random glucose and insulin. Since previous studies have shown that HbA1c levels outperform glucose levels in predicting T2D 44,45 , we further established base model 3 to substitute random glucose from base model 2 with HbA1c levels. The improvement in discrimination was examined by comparing AUC between each base model and the model plus plasma adiponectin levels 46 . Moreover, due to the limitation of AUC such as its insensitivity to model improvement 47 , we also evaluated the category-free NRI and IDI 48 . The goodness-of-fit of all models were assessed by Akaike information criteria (AIC), where lower AICs indicate better model fit.

Meta-analysis.
We performed an updated systematic review and meta-analysis of studies evaluating the association between adiponectin and T2D in adult populations since the date of the previous meta-analysis 5 . We followed the guidelines in Preferred Reporting Items for Systematic Reviews and Meta-Analyses and the Meta-analysis of Observational Studies in Epidemiology 49 . We searched systematically on PubMed for prospective human studies (including cohort, case-cohort, and nested case-control studies) assessing adiponectin-T2D associations between April 10, 2009 (the last date of literature research in the previous meta-analysis 5 ) up to September 10, 2016. Our searches included both MeSH terms and key words without restrictions on language (Supplemental Material).
Studies were excluded if they were: (1) irrelevant (not using adiponectin as the exposure or T2D as the outcome); (2) review, editorial, commentary, meta-analysis, animal or experimental study; (3) not prospective study (case-control, cross-sectional, and genetic studies). Identified studies were screened for titles and abstracts first, and then potentially relevant articles were reviewed for full-text (Supplementary Figure S2). Studies included in the previous meta-analysis 5 were also reviewed for full-text and added into the current meta-analysis. Data extraction was conducted independently by 2 authors (Y.W., R.M.). Extracted information included study characteristics (title, authors, publication year, study name, study design, follow-up length), participant characteristics (location, ethnic origin, sample size, number of incident T2D cases, gender composition, mean age or age range), adiponectin assay, T2D assessment, and analysis strategy (statistical models and covariates controlled). In case of multiple studies involving the same cohort, the most updated or relevant study was selected. We extracted relative risk estimates from the most fully-adjusted multivariable models. Studies qualities were assessed by the Newcastle-Ottawa Quality Assessment Scale 50 .
For the meta-analysis, RRs were used as the common measure of association, and ORs or hazard ratios (HRs) were considered as equivalent to RRs because of the low incident rate in most studies. To enhance consistency Figure 2. The association between concentrations of adiponectin and type 2 diabetes in the updated metaanalysis. The Solid line represents point estimates of relative risk for the adiponectin-diabetes association, and the dotted lines represent the upper and lower bound of 95% CIs. Cubic spline analysis was used to examine the association between adiponectin concentrations (categorical) and risk of developing type 2 diabetes using the most fully-adjusted models from reported studies. P = 0.33 for nonlinearity in the cubic spline regression model. The summary estimates were obtained from the most fully-adjusted models of each study using a randomeffects model. The data markers indicate the adjusted relative risks (RRs) comparing extreme tertiles of adiponectin levels. The size of the data markers indicates the weight of the study, which is the inverse variance of the effect estimate. The diamond data markers indicate the pooled RRs. and improve comparability between studies, we transformed the originally reported RRs (per unit, per SD, per log increment, quartiles or quintiles) to a uniform comparison involving extreme tertiles of adiponectin levels using standard statistical methods which were described previously 51,52 . When RRs were reported in subgroups rather than the total sample in a study, a within-study RR was attained first using a fixed-effect analysis. RRs and corresponding 95% CIs were pooled by the DerSimonian-Laird random-effects models 53 , and fixed-effect models were used in sensitivity analyses. Between-study heterogeneity was evaluated by Cochran χ 2 and I 2 Statistic 54 . Meta-regression analyses were used to examine the influence of certain factors on observed associations, including location, ethnicity, age, sex, follow-up duration, number of T2D cases, adiponectin assay, ascertainment of T2D, level of adjustment, and study quality. Stratified analyses were performed to evaluate the influences of baseline participant characteristics on the results. Potential for publication bias was examined by Egger's test 55 . To detect any nonlinear association, restricted cubic spline regression model with three knots at 25%, 50% and 75% percentiles adiponectin distribution was applied. When the nonlinear hypothesis was rejected, a two-stage, log-linear, dose-response regression function was then used for both the fixed-effect and random-effect models to estimate the risk of T2D per 5 µg/mL increment in adiponectin levels 56 . Studies which reported case number, total number of participants, as well as adiponectin median value in each category were included in the dose-response meta-analysis. All P values were two-sided, and data were analyzed with STATA version 14 (Stata Corp, College Station, Texas). Data availability. The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

Conclusion
In conclusion, the strong, dose-dependent association between increased adiponectin levels and decreased T2D risk has been demonstrated in a Chinese population, and adiponectin may be a useful marker for T2D prediction among Chinese. Our meta-analysis strengthens the evidence that adiponectin is involved in T2D pathogenesis, and further clinical studies are needed to investigate the feasibility of targeting adiponectin through pharmacological, dietary and physical activity interventions to reduce the risk of T2D in high-risk population.