Detection of Undiagnosed Diabetes and Other Hyperglycemia States: The Atherosclerosis Risk in Communities Study

The Atherosclerosis Risk in Communities (ARIC) study (10) initially investigated 15,792 men and women aged 45–64 years selected by probability sampling from four U.S. communities in 1987–1989. Of these, 11,656 (74%) returned to the fourth ARIC visit (1996–1998), during which an OGTT, physical measures, and interview were performed. Of the returnees, we excluded 1,507 individuals as having previously known diabetes (report of physician diagnosis or use of hypoglycemic medications); 1,779 lacking information permitting ascertainment of diabetes status at all four ARIC visits; 52 lacking information on basic covariates; 1 with an improbable anthropometric value; and, due to small numbers, 31 individuals who were neither white nor African American. After these exclusions, 8,286 individuals remained. Comparison of these subjects with those not studied and documented not to have diabetes (4,6,17) showed that those not studied were slightly older (54.5 vs. 53.8 years), more likely to be male (47 vs. 44%), and notably more likely to be African American (35.4 vs. 15.9%). Those not studied were also more frequently hypertensive (39 vs. 26%) and had slightly higher BMI (27.3 vs. 27.0 kg/m²) and fasting glucose (5.6 vs. 5.5 mmol/l) at ARIC baseline. Institutional review boards approved the study protocol.

We assessed hypertensive medication use and parental history of diabetes by interview, obtained physical measures with participants fasting and with an empty bladder, and performed a standard 75-g OGTT after an overnight fast. Overweight was defined as a BMI ≥25 and <30 kg/m², and obesity as ≥30 kg/m². Waist girth was measured at the umbilical level. Blood pressure was determined as the mean of two standardized measurements; hypertension was defined as blood pressure ≥140/90 mmHg and/or the use of antihypertension medication.

All analytes were determined at central laboratories according to standard protocols. Plasma glucose was measured by a hexokinase assay, insulin by nonspecific radioimmunoassay, and triglycerides and HDL cholesterol (HDL-C) by enzymatic methods. We defined diabetes as a fasting plasma glucose (FPG) ≥7.0 mmol/l or a 2-h postchallenge value ≥11.1 mmol/l; IFG as a fasting value from 6.1 to 6.9 mmol/l; and IGT as a 2-h glucose from 7.8 to 11.0 mmol/l.

In consonance with the National Cholesterol Education Program’s definition of the metabolic syndrome (11), we defined central obesity as a waist circumference >88 cm (35 in) for women and >102 cm (40 in) for men; high triglycerides as ≥150 mg/dl (1.70 mmol/l); low HDL-C as <40 mg/dl (1.03 mmol/l) for men and 50 mg/dl (1.29 mmol/l) for women; and raised blood pressure as ≥130/85 mmHg or the use of antihypertensive medication.

We produced rules for detecting diabetes or IGT based on logistic regression model risk functions generated on a randomly selected half of the sample (learning sample) with fasting glucose <6.1 mmol/l. Variables considered were sex, ethnicity, age, parental history of diabetes, use of antihypertensive medication, height, various measures of obesity (weight, BMI, waist-to-hip ratio, and waist circumference), systolic blood pressure, fasting glucose, HDL-C, triglycerides, and fasting insulin. Through a backward modeling strategy, we removed variables having lesser impact on the diagnostic properties of the risk functions, as reflected by changes in the area under the receiver operator characteristic (ROC) curve (12). We next examined risk functions in which continuous variables were categorized, following a similar modeling strategy as described above, and finally transformed the most promising risk function into a simplified clinical score by approximating the function’s β coefficients to integers.

Using the second random half of the sample (testing sample), we calculated each subject’s probability of having diabetes or IGT from the risk equation (13). We determined diagnostic properties of rules based on fasting glucose cut points and on cut points of these probabilities. The case detection rate was expressed as the proportion of total cases detected by the rule (sensitivity). All analyses were performed with SAS software (Cary, NC).

This sample of 8,286 participants consisted of 3,165 (38%) white men, 3,803 (46%) white women, 457 (6%) African-American men, and 861 (10%) African-American women. These individuals, aged 53–75 years and not previously known to have diabetes, had a high prevalence of overweight (41%), obesity (32%), central obesity (63%), hypertension (44%), high triglycerides (34%), and low HDL-C (41%). Based on fasting and 2-h postchallenge plasma glucose, 966 (12%) had diabetes and 2,684 (32%) IFG or IGT. By 2-h plasma glucose alone, 881 had diabetes (10.6%) and 2,369 (28.6%) had IGT. By fasting glucose alone, 419 (5,1%) had diabetes and 1,069 (12.9%) had IFG.

Table 1 compares the screening properties of fasting glucose cut points in the detection of postchallenge diabetes or IGT. The cut point for IFG (≥6.1 mmol/l) detected 32.6% of the 1,649 postchallenge abnormalities and 66.2% of 459 postchallenge diabetes cases. Lowering the cut point to 5.6 mmol/l (100 mg/dl) further increased the detection rate (sensitivity 61.5 and 84.1%, respectively) at the expense of identifying 45.7% of the sample as screen positive. Screening only those with BMI ≥25 kg/m², a serial strategy that reduces the number of glucose tests needed, lowered the overall detection rate at each glucose cut point.

We next investigated various screening steps that could follow an initial fasting glucose. First, we developed risk function models in the learning half of the sample with fasting glucose <6.1 mmol/l. Factors offering greatest predictive ability were, in descending order: fasting glucose, triglycerides, age, systolic blood pressure and the use of antihypertensive medication, height, HDL-C, waist circumference, parental history of diabetes, ethnicity, and sex. For practical purposes, risk functions were developed with and without lipid variables. Once the three most predictive factors were included in the model, incremental gains in diagnostic properties diminished rapidly with the addition of further factors. When weight was used instead of waist circumference in the equation, the area under the ROC curve decreased only slightly; no improvement occurred when fasting insulin was added (data not shown).

Second, we evaluated in the testing sample with fasting glucose <6.1 mmol/l the screening properties of fasting glucose cut points and clinical detection rules defined by the probability of having diabetes or IGT (Table 2). Risk functions used to define these rules and to estimate these probabilities are shown in the footnotes. We first evaluated a risk function based on the most readily available clinical information, thus excluding lipids. Clinical rules were generated to identify a range of clinically useful proportions of screen positivity (30–50%). To allow comparison with screening based on fasting glucose alone, we also generated rules identifying as screen-positive proportions equivalent to those of the glucose cut points 5.6 mmol/l (34.4%) and 5.3 mmol/l (60.4%). A rule with screen positivity of 34.4%, equal to that of the 5.6 mmol/l cut point, detected 48.9% (vs. 42.9%) of the 1,112 cases of diabetes or IGT and 58.7% (vs. 52.9%) of the 155 cases of diabetes. We obtained similarly superior diagnostic properties for a clinical rule with 60.4% screen positivity. Detection rates improved slightly when information on lipids was included in the risk function. For example, a rule with lipids with screen positivity of 30% detected 47.3% (vs. 44.1% without lipids) of cases of diabetes or IGT and the same proportion (56.1%) of cases of diabetes. An increase in the area under the ROC curve for the risk equation including lipids was similarly small (0.73 vs. 0.69; see footnote, Table 2).

Table 2 also shows rules based on categorical factors. The first one attributes one point each for age ≥65 years, a positive parental history of diabetes, being African American, having a large waist circumference, having hypertension, and having a fasting glucose >5.3 mmol/l. A score of ≥4 out of 6 total points identified 36.3% of the sample with fasting glucose <6.1 mmol/l as positive, 52.0% of the sample with diabetes or IGT, and 57.4% with diabetes. Of note are the large jumps in the percent of the sample deemed screen positive between cut points of this score. The second categorical rule, based on metabolic syndrome abnormalities (11), produced detection rates inferior to the above rule: the 28.5% of the sample having three or more abnormalities contained only 40.6% of all hyperglycemia cases and only 40% of the cases of diabetes.

Table 3 shows the outcomes of seven screening strategies in the entire testing sample. If the OGTT, which represents a certain gold standard for these strategies, were applied to all subjects, it would identify 44.1% of the sample as screen positive and detect and explicitly classify all hyperglycemia cases.

Screening with fasting glucose using the conventional IFG cut point identified 17.3% of the sample as screen positive (diabetes or IFG), detecting 68.8% of the 497 cases of diabetes but only 28.1% of the 1,330 cases of IFG/IGT. Lowering the fasting glucose cut point to 5.6 mmol/l identified 45.7% of subjects as screen positive, increasing the detection rate to 85.3% for diabetes and 57.7% for IFG/IGT, while leaving unaltered the sample fraction explicitly labeled as diabetes or IFG. It detects 51.7% of IGT cases. If labeling is desired among those with FPG <6.1 mmol/l, an OGTT could then be applied. This approach decreases screen positivity to 28.8% while explicitly classifying all the hyperglycemia cases detected.

Another approach is to additionally screen all those with fasting glucose <6.1 mmol/l using a detection rule composed of clinically available diabetes risk factors. Applying such a clinical detection rule, set to identify those at “high probability” (upper 30%) of having hyperglycemia, this strategy detects 86.3% of the total cases of diabetes and 58.4% of the cases of IFG/IGT. This strategy detects 52.4% of IGT cases in the sample. With this strategy, 42.1% of the sample are identified as screen positive, yet only 17.3% actually receive a diagnostic label of hyperglycemia (diabetes, IFG, or IGT). If a clinical diagnosis of hyperglycemia is considered desirable to initiate clinical preventive actions, then an additional step could apply an OGTT to those considered at high probability of having hyperglycemia by this clinical rule. This step does not increase the detection rate, it decreases screen positivity to 29.1%, and it permits explicit classification of those with glucose abnormalities, at a cost of an OGTT in 24.8% of the sample.

Finally, screening with fasting glucose alone applied only to those with BMI ≥25 kg/m² reduced those tested to 73.3% of the sample, at the expense of considerably lower overall case detection rates. All strategies assume that those with IFG will not be further classified by 2-h postload testing. If such classification were performed, then an additional 12.1% of the sample would undergo OGTT testing.

Evidence from five clinical trials that type 2 diabetes can be prevented or delayed by lifestyle or pharmacological interventions in individuals with IGT (2–4,14,15) has encouraged screening for IGT (16). Such screening will inevitably find cases of IFG and diabetes, for which the benefits of early intervention are less clear. Screening based on an OGTT clearly labels individuals with IGT. Furthermore, it detects a subset of undiagnosed diabetes at higher risk for cardiovascular disease and all-cause mortality than does fasting glucose (6,7). However, the clinical utility of the OGTT has been questioned (17), and it imposes inconveniences related to test variability, glucose load, time spent in the laboratory, and cost.

Fasting glucose is readily available and simpler. It is considered equally predictive of future diabetes as a 2-h value (5). Our results indicate that almost two-thirds of those classified as IFG have either diabetes or IGT by a 2-h value, providing further justification for the use of this fasting category as the basis for preventive action. The major limitation of the IFG cut point is its low sensitivity to detect postchallenge hyperglycemia.

Screening based on clinical detection rules costs little and uses readily available clinical information. Being composed of cardiovascular risk factors, such rules identify those at higher risk for cardiovascular disease while avoiding diagnostic labeling of predisease. Several studies have analyzed strategies based on clinical information aimed at identifying individuals at high probability to have (18–22) or develop (23) diabetes. However, little has been done to develop a practical means to identify not only those who currently have undiagnosed diabetes but also those with other states of asymptomatic hyperglycemia meriting clinical intervention (24).

To this aim, we evaluated strategies based on combinations of fasting glucose, clinical detection rules, and the OGTT. For all strategies, a fasting glucose was the first step, followed by additional screening dependent on the FPG value. The best results—detecting >85% of the cases of diabetes, 58% of the cases of IFG/IGT, and 52% of the cases of IGT—were obtained with two strategies: the first one used an FPG cut point of 6.1 mmol/l and then applied a clinical detection rule to those below this cut point; the second one used an FPG cut point of 5.6 mmol/l and then applied an OGTT to those with FPG <6.1 mmol/l.

Because the clinical detection rule used in the first of these strategies is based on cardiovascular and diabetes risk factors, it identifies higher-risk subjects who would logically derive greater benefit from preventive interventions. Applying an OGTT to those positive by the detection rule, while not changing the detection rate, reduces the fraction deemed screen positive from 42.1 to 29.5% and provides a diagnostic hyperglycemia label for all positives, which could be of motivational value for interventions.

The diagnostic properties of the rules shown in Table 2 suggest that one could use a categorical score instead of a continuous variable risk function when this is deemed more practical. Because the score includes African-American ethnicity, it is of less value in non-U.S. settings. A score based on metabolic syndrome abnormalities, though less discriminating diagnostically, could also be used when the relevant information is available. It has the advantage, as do our continuous variable clinical detection rules, of greater generalizability to populations with other ethnic admixtures.

The appropriate form of a clinical detection rule merits discussion. Options run from simple to complex scores, and they include nomograms and questionnaires. Our clinical prediction rules were designed to require either a very simple summation (the categorical rule) or the assistance of some form of calculator. With the increasing availability of personal digital assistants and web-based calculators, a rule determined by entering data into a preestablished calculator may be more practical than all but the simplest of manual calculations. Our risk equations were developed using readily available variables. In this regard, we favored the use of weight, rather than waist circumference, in risk functions for continuous variables, given its more widespread availability in clinical databases and the inability of waist circumference to provide a clinically relevant increment in test sensitivity and specificity when used in place of weight. Health care providers serving large populations could apply continuous risk function rules to administrative or patient care databases already at hand.

Study limitations merit note. Although our study sample presents fewer African Americans and a slightly lower risk profile for diabetes than the initial ARIC sample, we believe it unlikely that fasting glucose or a clinical detection rule would present importantly different diagnostic properties in those not available for study. A more important limitation of our analyses, and of almost all published studies in the literature, is that we define and validate cases of hyperglycemia based on a single measurement rather than using the repeated measurement necessary for a clinical diagnosis. This diagnostic criteria overestimates the prevalence of undiagnosed abnormalities—12% for diabetes and 32% for lesser hyperglycemic states in our sample—and thus, in absolute terms, it overestimates the yield of screen-positive cases the strategies would produce. In this regard, because variability is greatest for the OGTT and probably lowest for predictive equations, screening strategies using the OGTT would be most affected and those using clinical detection rules would be least affected by within-subject glucose variability. An additional limitation, in this regard, is that we are screening and diagnosing hyperglycemia with the same glucose test. However, this use of the screening value as part of the diagnostic test is a current diagnostic practice. A similar limitation is the fact that we used fasting and 2-h determinations that were not independent, being obtained from the same OGTT. However, with the increasing availability of precise, rapid means of fasting glucose determination, one can apply a two- or three-step strategy at a single visit, dismissing those negative at the first step and proceeding immediately to the second step. Another limitation of our study is that we lacked HbA_1c levels for comparison. Finally, the limited range of age and the high prevalence of obesity in our sample, as well as the fact that all participants were either white or African American and lived in four U.S. communities, may limit the generalizability of our findings.

Nevertheless, we provide validation in a large sample of a practical means of identifying subjects who may benefit from diabetes preventive actions. Although the two strategies showing the best diagnostic properties had similar performances, we favor that which is based on the IFG cut point followed by the clinical detection rule because it identifies those presumably at greatest risk. Additionally, if OGTT testing is desired, the IFG/clinical detection rule strategy requires somewhat fewer OGTTs to detect the same fraction of hyperglycemia cases. Whether those positive by this strategy need an OGTT before preventive lifestyle counseling is debatable.

Although we believe that these strategies developed and validated with ARIC data have direct clinical application, an important future step is a collaborative effort to pool data from several large studies internationally so as to develop and more fully validate hyperglycemia screening strategies.

In conclusion, screening strategies based on fasting glucose, clinical detection rules, and/or an OGTT are effective and practical in the detection of cases of hyperglycemia for which preventive action is likely to beneficial. The adoption of a given strategy will depend, however, not only on its diagnostic properties but also on the effectiveness of interventions and on resources available for screening, diagnostic confirmation, and intervention.

Support for this study was provided by National Heart, Lung, and Blood Institute contracts N01-HC-55015, N01-HC-55016, N01-HC-55018, N01-HC-55019, N01-HC-55020, N01-HC-55021, and N01-HC-55022 and National Institute of Diabetes, Digestive and Kidney Diseases Grant 5R01-DK56918-03. B.B.D. and M.I.S. received support from a Centers of Excellence Grant of CNPq (the Brazilian National Research Council).

The authors thank the staff and participants in the ARIC study for their important contributions.