The main findings of our study are (i) that the combination of 13 biomarkers of subclinical inflammation improved the accuracy of a risk model of incident type 2 diabetes in the MONICA/KORA cohort significantly and equally well as a combination of established cardiometabolic risk factors, (ii) that a combination of both sets of risk factors led to a further significant improvement of the accuracy of predicting type 2 diabetes compared with either set of risk factors alone, and (iii) that the improvement of accuracy of prediction models for type 2 diabetes over and above age, sex and cardiometabolic risk factors by the combination of inflammation-related biomarkers was more pronounced than for coronary events employing the same methods in the same population.
The study extends previous knowledge because it is the first study to focus on the predictive value of multiple markers of subclinical inflammation for incident type 2 diabetes. The set of 13 inflammation-related biomarkers consists of an acute-phase protein (CRP), cytokines (IL-6, IL-18, TGF-β1, MIF), chemokines (MCP-1/CCL2, IL-8/CXCL8, IP-10/CXCL10, RANTES/CCL5), adipokines (adiponectin, leptin) and soluble adhesion molecules (sE-selectin, sICAM-1) and is therefore more comprehensive than the combinations of immune mediators that were used in the ARIC cohort (leukocyte count, IL-6, four acute-phase proteins) or the MONICA/KORA cohort (CRP, IL-6, three chemokines) before 
Our study design differs in two important aspects from the design of the aforementioned Inter99 and FINRISK cohorts 
. First, we provide an estimate of the accuracy of models based on biomarkers of subclinical inflammation only (over and above age, sex and survey as essential covariates) and both compared and combined them with established cardiometabolic risk factors because we were interested in the contribution of subclinical inflammation as pathophysiological mechanism to the development of type 2 diabetes. Therefore, it was not our aim to build a risk score with optimal predictive value for incident type 2 diabetes. Second, we used nonfasting rather than fasting blood samples as fasting samples were not available from the MONICA/KORA Augsburg surveys. Although it could be argued that this represents a major limitation of our study (to be discussed below), it should be noted that the question whether inflammation-related biomarkers from nonfasting samples could be useful in the prediction of type 2 diabetes has not been addressed in comparable population-based studies and is therefore of interest.
Although we observed a substantial increase in AUC (0.044 [95% CI 0.028–0.066]) as well as large values for ΔAIC (139.8), IDI (0.061) and NRI (0.202) by the addition of inflammation-related biomarkers to a model that already contained strong cardiometabolic risk factors for type 2 diabetes, we found that this increase could be attributed to just a few biomarkers. Only adiponectin and sE-selectin increased the AUC by more than 0.010 and showed ΔAIC considerably larger than 10. Importantly, adiponectin was also one out of four biomarkers (next to apolipoprotein B, CRP and ferritin) that was included in the final prediction score derived from 31 biomarkers in the FINRISK97 cohort 
. Moreover, inclusion of adiponectin in an extensive risk score based on anthropometric, metabolic and lifestyle factors led to a small, but significant increase in the AUC in the EPIC-Potsdam Study 
. We reported before that adiponectin improved risk prediction over and above cardiometabolic and selected inflammation-derived biomarkers in the MONICA/KORA Augsburg case-cohort study 
. These data are in contrast with findings from the KORA S4/F4 cohort study, which was conducted later and independently from the MONICA/KORA surveys 1–3. In KORA S4/F4, there was no significant improvement of the AUC when adiponectin was added to a model that contained HbA1c and fasting glucose 
. Data on the impact of sE-selectin on measures of discrimination are available from a small nested case-control study within the Western New York Study. The addition of sE-selectin, serum albumin and leukocyte count improved the accuracy of a risk model for type 2 diabetes compared to a basic model based on sex, BMI and familiy history of type 2 diabetes 
Our findings on type 2 diabetes are in contrast to several other studies on the improvement by inflammation-related biomarkers of risk models already containing measure of glycemia or insulin resistance. In the Insulin Resistance Atherosclerosis Study, the addition of CRP to a prediction model for type 2 diabetes that was based on the metabolic syndrome (without or with an estimate of insulin resistance) had little impact on AUCs 
. CRP (alone or in combination with other biomarkers) also failed to improve AUCs of prediction models already containing plasma glucose glucose levels as in the Framingham Offspring Study 
and the aforementioned EPIC-Potsdam Study 
. In the Sandy Lake Health and Diabetes Project, leptin, CRP, IL-6 and serum amyloid A were included in a risk model based on cardiometabolic risk factors, adiponectin and impaired glucose tolerance, but could not improve diabetes prediction 
. Recently, the Women's Health Initiative Observational Study did not find that biomarkers of subclinical inflammation (hsCRP, IL-6, soluble tumor necrosis factor-receptor 2) and of endothelial dysfunction (E-selection, ICAM-1, vascular cell adhesion molecule-1) contribute to the prediction of incident type 2 diabetes over and above clinical risk factors and fasting glucose 
. Taken together, these data suggest that our findings may be specific for analyses based on nonfasting blood samples and that the contribution of multiple inflammation-related biomarkers to prediction models with diabetes risk factors that are used for the diagnosis of type 2 diabetes (glucose, HbA1c) may be less pronounced than for prediction models without these measures of glycemia.
An important aspect of our study is the fact that our case-cohort design allowed us to compare inflammation-related and cardiometabolic risk factors for both type 2 diabetes and coronary events as outcomes using the same methods and two largely overlapping study populations. The combination of all biomarkers and risk factors yielded almost identical AUCs for both outcomes. However, the improvement of inflammation-related biomarkers over a basic model based on age, sex and survey was considerably larger for type 2 diabetes (ΔAUC 0.111 [95% 0.092–0.149]) than for coronary events (ΔAUC 0.018 [95% CI 0.013–0.038]). This difference is confirmed by larger ΔAIC, IDI and NRI values when models for both outcomes were compared. This is most likely attributable to the higher accuracy of the basic model for coronary events (AUC 0.807) compared to type 2 diabetes (AUC 0.690) so that further improvements by additional biomarkers or risk factors can be expected to be less pronounced. Although we found a significant increase in AUC, these data are in line with data from other studies that focused on risk models for incident coronary events or cardiovascular death and that assessed the incremental predictive value of inflammation-related biomarkers. AUCs for prediction models based on cardiometabolic factors were usually in the range between 0.70 and 0.82. Although multiple promising biomarker candidates were tested, the improvement of risk models by addition of novel inflammation-related biomarkers was relatively small, especially when the basic model already had a good accuracy, and AUCs of the extended models did not increase beyond 0.82 in these studies 
. A recent study indicated that in particular N-terminal pro-brain natriuretic peptide (NT-proBNP) and sensitive troponin I may improve the prediction of risk of coronary events 
Regarding the clinical relevance of our findings, the present study did not aim at providing a simple clinical risk score, but rather at studying to which extent subclinical inflammation as one of several other mechanisms contributes to the prediction of the development of type 2 diabetes. The approach of this study was chosen to extend previous work that mainly evaluated statistical associations between inflammation-related biomarkers and incident diabetes using Cox regression models.
Overall, our data demonstrate that age (but not sex or survey) contribute a substantial part to the AUC that can be achieved with a basic risk model and with more sophisticated models involving multiple risk factors and biomarkers. Interestingly, although cardiometabolic risk factors are strongly associated with inflammation-related biomarkers, we found a significant increase in accuracy when adding inflammation-related biomarkers to a model based on age, sex and cardiometabolic risk factors. Therefore, these data are in line with a role for subclinical inflammation in the development of type 2 diabetes and indicate that in particular adiponectin and sE-selectin should be further evaluated as markers for type 2 diabetes risk in combination with other risk factors and biomarkers.
Strengths of our study include the use of the MONICA/KORA Augsburg cohort with a large number of cases and non-cases, a long follow-up period, availability of data for multiple biomarkers representing different aspects of subclinical inflammation, and the inclusion of both cases with incident type 2 diabetes and coronary events in the case-cohort study so that a direct comparison of risk factors and biomarkers for both outcomes in the same cohort using the same methods was possible.
There are also several limitations that should be pointed out. First, we did not perform oral glucose tolerance tests at baseline or follow-up so that some misclassification may have occurred and the outcome of our study was physician-diagnosed type 2 diabetes. Second, we had no data on fasting glucose and fasting insulin (HbA1c data for a subgroup of study participants only) so that we could not investigate the change in AUC by inflammation-related biomarkers over a model that contained these variables. In addition, minor variations of levels of inflammation-related biomarkers due to the nonfasting state cannot be excluded. Third, we used continuous values of biomarker concentrations in order to render results comparable to other studes 
, although consideration of sex differences or non-linear associations between biomarkers and endpoints could have led to higher accuracy of our models. Fourth, biomarkers of non-alcoholic fatty liver disease (liver enzymes such as alanine aminotransferase, aspartate aminotransferase or γ-glutamyl transferase) are relevant risk factors for type 2 diabetes 
, but were not available in our study (with the exception of γ-glutamyl transferase in S1) so that we could not include them in our cardiometabolic risk models. Finally, we did not seek for external replication of our results.
Taken together, 13 inflammation-related biomarkers measured in nonfasting serum samples significantly improved the prediction of incident type 2 diabetes and coronary events over and above cardiometabolic risk factors in the MONICA/KORA study, but this improvement was much more pronounced for type 2 diabetes. Our study could not address the question whether biomarkers of subclinical inflammation can also improve the predictive value of risk models that contain various measures of glycemia. Therefore, further research is warranted to investigate whether multiple inflammation-related biomarkers can increase the accuracy of risk models that include data on (fasting or nonfasting) glucose, insulin or HbA1c levels.