This study used the public dataset from the nationally representative National Health and Nutrition Examination Survey (NHANES) 2005-2006 (21), which oversampled minority populations. Households were randomly assigned to morning or evening examination. Morning examination included a fasting plasma glucose test (FPG) and an oral glucose tolerance test (OGTT); 1,887 participants aged 18 years or older had valid measures for both FPG and OGTT. More detailed methods of the NHANES 2005-2006 (21) and the laboratory tests can be found elsewhere (22).

We developed 6 models to estimate all clinically relevant measures of abnormal glucose levels. Model 1 estimates IFG. Model 2 estimates IGT. Model 3 estimates prediabetes (either IFG or IGT). Model 4 estimates what we term "high-risk prediabetes" (both IFG and IGT). Model 5 estimates undiagnosed diabetes. Model 6 estimates total abnormal glucose, which includes prediabetes, undiagnosed diabetes, and diagnosed diabetes.

For the first 4 models — all estimates of prediabetes risk — we excluded the 308 people who met the criteria for frank diabetes, whether they were aware (201) or unaware (107) of this diagnosis. This left 1,579 people (of the 1,887) for estimates for models 1 through 4. For model 5, estimating undiagnosed diabetes, we excluded people who were aware they had diabetes (201), yielding a sample of 1,686 people. For the model estimating total abnormal glucose prevalence (model 6), we included all 1,887 adults with FPG and OGTT scores.

Operational definitions of IFG, IGT, prediabetes, and diabetes were developed on the basis of current diagnostic criteria of the American Diabetes Association (23). IFG was defined by an elevated FPG concentration (≥100 and <126 mg/dL). IGT was defined by an elevated 2-hour plasma glucose concentration (≥140 and <200 mg/dL) after a 75-g glucose load on the OGTT. Prediabetes was defined as having either IFG or IGT. High-risk prediabetes was defined as having both IFG and IGT; the definition was based on previous work about the increased risk of this situation (4,12,13). Diabetes was defined as having a fasting plasma glucose of 126 mg/dL or more or a 2-hour plasma glucose above 200 mg/dL. Total abnormal glucose was defined as having prediabetes of any form, undiagnosed diabetes, or diagnosed diabetes. Diagnosed diabetes was determined by individual self-report and did not include gestational diabetes, which was not assessed in the 2005-2006 NHANES.

On the basis of a literature review, we identified 11 self-reported predictor variables that were available in the NHANES and BRFSS and were known to be associated with diabetes risk for possible inclusion in each Abnormal Glucose Risk Assessment-6 (AGRA-6) model. Demographic variables included age (continuous 18-85 y), sex, self-reported race/ethnicity, and educational attainment. Behavioral variables included smoking status and participating in any leisure-time physical activities. Health condition variables included body mass index (BMI) (continuous and truncated from ≤10 to ≥100 to avoid outliers), history of hypertension, use of hypertension medication, high cholesterol, and family history of diabetes.

From these possible variables, we derived optimal logistic prediction models by using the Akaike information criterion (AIC), which selects a model that maximizes predictive power while minimizing the number of predictive variables (24). For each of the 6 outcome variables, a unique, optimal predictive model was built from the set of potential predictive variables. The main statistical analyses were performed with SAS version 9.1 (SAS Institute, Inc, Cary, North Carolina). Multiple imputations were performed with SRCware version 1.0 (University of Michigan, Ann Arbor, Michigan). All of the models took into account the complex survey design.

We validated the final models with the leave-1-out cross-validation (LOOCV) method. The LOOCV uses a single observation from the whole sample as the validation data, and the remaining observations as the training data. This process is repeated until each observation in the entire sample is used once as the validation data. The sensitivity, specificity, and positive and negative predictive values were obtained for all 6 models under LOOCV testing conditions. The area under receiver-operating characteristic curves (AUC) provides a single value that indicates the discrimination of the model (ie, its ability to identify true risk) at all possible values that could be chosen as cut points to distinguish risk from nonrisk.

In practical applications, however, specific cut points must be chosen to distinguish risk from nonrisk. Whether that cut point should prioritize identifying true positives, true negatives, or some balance between the 2 depends on the objective of the analysis and the budget of the program doing the analysis. For instance, when algorithms are used for screening purposes, it would generally be more desirable to find all cases of prediabetes, at the cost of some false positives. In this situation, the cut point delineating risk from nonrisk should be set to maximize sensitivity (finding all true positives) over specificity (identifying only true negatives). A positive finding of risk would then be followed by a laboratory test. For surveillance, on the other hand, the goal would typically be to strike a balance between types of error (false positives and false negatives). A higher specificity cut point is generally more cost-effective unless clinical priorities dominate (such as use of higher sensitivity cut points to find gestational diabetes).

Therefore, to maximize the usefulness of the AGRA-6, we present the predictive characteristics of each AGRA-6 model for 2 thresholds: 1) the high-sensitivity threshold, where a cut point is selected so that sensitivity is reached to about 0.9 (approximately 90% of positive cases will be correctly identified), and 2) the balanced-sensitivity/specificity threshold, where a cut point is selected so that sensitivity and specificity are equal. This will enable users of the tool to determine optimal cut points on the basis of local or programmatic needs and resources.

The AGRA-6 has some limitations. Computing devices — either personal computers or personal digital assistants — will generally be required to calculate risk models. This should not present a barrier for most AGRA-6 applications, but when these devices are unavailable or impractical, other tools (14,17,18) may be preferable even if they do not allow for the measure of as many subtypes of abnormal glucose risk.

Second, the AGRA-6 models have been validated by using data from the sample on which they were created, which may result in more overestimations of model performance than would be observed if the algorithms were tested in other data sets. We did this because there are no other comparable nationally representative data sets that contain laboratory tests for both FPG and OGTT. To minimize the impact of this approach, we used the LOOCV method, which is a method often used for creating testing data sets from training data sets (25).

Third, although these models were generated from a nationally representative US sample, they may not be appropriate for all US subpopulations and geographic areas or for many international populations (26). The performance of the AGRA-6 models may also vary by demographic subgroups (younger vs older, heavier vs lighter, different racial/ethnic groups), and a consideration of this variation for the AGRA-6 models and for other commonly used predictive models is an area for further study. Some of the predictive variables rely on prior access to care, including having a diagnosis of hypertension or high cholesterol and taking hypertension medication. Actual prevalence in people who lack access to care may thus be underestimated. Also, the available sample was not large enough to allow us to include Asians/Pacific Islanders as distinct subpopulations, leaving open a question regarding its usefulness for classifying risk in these groups.

Fourth, the AGRA-6 shares the limitations of any predictive model in that some people will be misclassified. Whether people are misclassified as false positives or false negatives can be manipulated to some degree by the chosen threshold levels used to delineate risk. In all public health surveillance and screening, there will be tradeoffs between precision and cost, and no option is infallible. The problems associated with misclassification must be weighed against the specific goals and budget of the program.

The health risks of type 2 diabetes can be mitigated through individual, community-based, and even structural and policy interventions (3). Lifestyle interventions can also prevent or delay the onset of type 2 diabetes among high-risk people, such as those with prediabetes (6). One major task for public health agencies and programs is to identify groups and individuals who would benefit from these interventions. The AGRA-6 allows public health organizations to identify populations and individuals who would probably benefit from these interventions and to facilitate cost-effective screening of these populations. This could further facilitate the allocation of public health resources for focused interventions to reduce the illness and death risks of prediabetes and diabetes in the United States. The AGRA-6 models should also prove useful for county, state, and national surveillance efforts to assess the progression of this epidemic.