Predicting the Development of Diabetes in Older Adults: The derivation and validation of a prediction rule

The derivation sample consisted of subjects enrolled in the RBS who participated in a clinical research evaluation during 1984–1987. The RBS is a population-based study of white, middle or upper-middle class, community-dwelling adults in a southern California suburb. Participants with a diagnosis of diabetes or those with FPG ≥126 mg/dl were excluded. The final derivation sample population consisted of 1,549 men and women.

Demographic information (age, sex, and education), health-related behaviors (smoking and alcohol use), family history, and medication use were assessed. Weight and height were measured using standard protocols, and BMI was calculated (kg/m²). Waist circumference was measured with a flexible tape measure at the level of minimum abdominal circumference. Two resting systolic and diastolic blood pressure measurements were taken on each subject using a mercury sphygmomanometer; the mean blood pressure was used for analyses.

Morning venous blood after a 12-h fast and 2 h after a 75-g oral glucose load were performed. Plasma glucose was measured using a glucose oxidase method. HbA_1c was measured using high-performance liquid chromotography. Fasting cholesterol and triglycerides were measured by enzymatic methods with an ABA-200 biochromatic analyzer (Abbott, North Chicago, IL); HDL was assayed by precipitation, and LDL was calculated using the Friedewald formula (9).

Data collected from Health ABC study participants were used to test the prediction rule developed in the RBS. Health ABC is a prospective cohort study of 3,075 community-dwelling men and women aged 70–79 years recruited at two clinical centers in Memphis, Tennessee, and Pittsburgh, Pennsylvania. To be eligible, participants had to report no difficulty walking a quarter mile or up a flight of stairs. For our analyses, we used data gathered from 2,503 (1,550 white and 953 black) participants who did not have diabetes at baseline (1997–1998), based on self-report of diabetes, use of a hypoglycemic medication, or FPG ≥126 mg/dl.

Questionnaire variables gathered from the first clinical visit included age, self-identified race, sex, and level of education completed. Participants reported smoking history and alcohol use. Participant’s weight, height, and waist circumference was measured, and BMI was calculated.

Fasting serum specimens were collected from each participant. Lipid lipoproteins were measured by colorimetric technique (Vitros chemical methodology; Johnson & Johnson, New Brunswick, NJ), and fasting glucose and 2-h PCPG test were measured by glucose oxidase reaction (YSI 2300 Glucose Analyzer; YSI, Yellow Springs, OH).

Incident diabetes was defined by an algorithm using a physician diagnosis of diabetes in the past year, use of a hypoglycemic drug, or a FPG ≥126 mg/dl at the year 2, 4, or 6 examination.

Multivariable logistic regression models were constructed to examine which variables were most associated with an abnormal 2-h PCPG test result (≥140 mg/dl) in the RBS. Model entry criteria were P < 0.15. Accepted risk levels were used to categorize most variables (e.g., BMI <25.0, 25.0–29.9, ≥30.0 kg/m²; LDL <100 vs. ≥100 mg/dl; systolic blood pressure <140 vs. ≥140 mmHg). For other continuous variables such as age, waist circumference, and fasting glucose, recursive partitioning methods were used to select categories that optimized the data selection. Backwards elimination (P < 0.05 to retain) was used to select the final set of risk factors and checked for interactions between sex and age with other risk factors and found none significant at P < 0.05.

A clinical scoring system was created by assigning points to each risk factor by dividing the β-coefficient of each variable by 0.6 and rounding to the nearest integer. A risk score was assigned to each subject by adding up the points for each risk factor present. Sensitivity, specificity, and likelihood ratio for each risk score was calculated to predict those with abnormal glucose tolerance. We also calculated the area under the receiver operating characteristic (ROC) curves for the final multivariable model built with continuous or categorical variables and the simple scoring system.

The prediction model and the scoring system were validated using data from the Health ABC study by examining the ability of the rule to predict abnormal glucose tolerance at the baseline Health ABC visit and to predict incident diabetes over 5 years of follow-up. By using a separate cohort for validation, we tested the accuracy of the model and its geographic, racial, and methodologic transportability (10). Separate ROC curves for black and white participants in Health ABC for incident diabetes were calculated and the C-statistics were compared with that of the 2-h PCPG test alone (11).

We used SAS version 8.2 (SAS Institute, Cary, NC) and S-Plus version 6.1 (Insightful, Seattle, WA) for our analyses.

Of 1,549 participants in the RBS derivation sample, 514 (33.2%) subjects had elevated PCPG results (≥140 mg/dl). Those with abnormal PCPG were older and had higher systolic blood pressure and triglyceride levels compared with those with normal glucose tolerance (Table 1). There was no difference in mean BMI for those with normal and abnormal glucose tolerance. Mean waist circumference was larger in women with abnormal glucose tolerance compared with the women with normal glucose tolerance, but there was no difference in waist circumference for men in the two glucose subgroups. As expected, mean FPG was lower among participants with normal glucose tolerance (96.0 ± 10.4 mg/dl vs. 100.8 ± 10.4 mg/dl; P < 0.001). The mean HbA_1c levels were statistically but not clinically different between the two groups.

Of the nine variables that initially entered our multivariable logistic regression analysis, only two demographic and two laboratory variables remained significantly associated with an abnormal PCPG result (Table 2). Individuals aged ≥70 years had the highest odds of an abnormal PCPG test (OR 3.06; 95% CI 2.43–3.87). Women had higher odds of abnormal PCPG (1.60; 1.26–2.02) compared with men. Participants with triglycerides ≥150 mg/dl and FPG ≥95 mg/dl were at greater risk for abnormal glucose tolerance as well. The discrimination of the final model was modest in the derivation cohort (c = 0.73 with continuous variables, 0.71 with categorical variables). Results did not differ with sex-specific models.

Using β-coefficients from the final categorical multivariate model, we assigned points to each of the four risk factors for abnormal PCPG (Table 2). A risk score was calculated for each participant by summing the points of each risk factor that was present. The derivation cohort risk score ranged from 0 to 7 points (2.6 ± 1.4 [means ± SD]). Estimates of sensitivity, specificity, and positive likelihood ratios for each score for abnormal glucose tolerance are shown in Table 3. A score of 4 or higher could double the likelihood of an abnormal PCPG test.

To validate this rule, we applied the prediction models and simple scoring rule to the baseline data from Health ABC. The overall discrimination of the scoring system for abnormal PCPG was similar in both samples (c = 0.70 in the derivation sample, 0.66 in the validation sample). Next, we prospectively validated the prediction rule for incident diabetes in Health ABC. There were 143 incident cases of diabetes over 5 years (14.1 cases/1,000 person-years). The area under the ROC curve for the multivariable model using continuous or categorical variables was not different than the scoring system (c = 0.75 for both multivariable models, 0.71 for the scoring system). Thus, the scoring system was able to discriminate incident diabetes cases equally well as the PCPG test. When we stratified the Health ABC cohort by race, similar discrimination in both the black and white participants using the continuous model, categorical model, or scoring system was observed (Figure 1). There was no significant difference among the three models and the PCPG test in discrimination of diabetes (P > 0.10 for each comparison).

We have developed a clinical prediction rule to identify persons with abnormal glucose tolerance that predicted diabetes equally well as the 2-h PCPG test. Two demographic variables, age and sex, and two laboratory variables, FPG and triglyceride levels, were most strongly associated with abnormal glucose tolerance in a cohort of older white adults. Using a simple scoring system, a score of ≥4 points at least doubled the likelihood of an abnormal 2-h PCPG test result. Moreover, this scoring system, validated in an independent cohort of black and white older adults, was able to predict future diabetes with similar accuracy as the PCPG test. Thus, patients with a score of ≥4 points with this simple scoring rule should receive appropriate lifestyle or pharmacologic therapies in order to prevent the onset of type 2 diabetes.

Recent reports that the incidence of diabetes can be prevented or delayed in individuals with IGT by lifestyle intervention (5,6) or medication therapy (5,12,13) has highlighted the need for identification of these high-risk individuals. The American Diabetes Association (ADA) does not recommend the 2-h PCPG test for screening or diagnosis of diabetes because the test is costly, inconvenient, and has poor reproducibility (7). The ADA adopted the term of “pre-diabetes” to identify persons with IGT or impaired fasting glucose (IFG; FPG levels between 110–125 mg/dl) collectively as those at highest risk for developing diabetes. In 2003, the ADA revised the definition for IFG to include persons with fasting glucose between 100 and 125 mg/dl (14,15). The use of IFG alone to identify those with incident diabetes in our validation cohort had a sensitivity of 54% and a specificity of 82% (ROC area of 0.68). We found that the addition of age, sex, and triglyceride level to elevated FPG results increased the predictive accuracy for diabetes.

Several recent studies have evaluated the utility of laboratory and demographic variables for predicting incident diabetes. Stern et al. (8) reported the effectiveness of eight clinical variables at predicting those at risk of incident diabetes among the San Antonio Heart Study participants. The prediction equation they created was based mostly on continuous variables and requires the use of a computer or personal digital assistant to implement. Additionally, being Mexican American explained most of the predictive value of their equation. This clinical model was tested in a cohort of Japanese Americans and was useful in predicting diabetes risk in younger Japanese Americans after 5–6 years but was not useful in older participants or with longer follow-up (16). Another study that calculated a diabetes risk score utilized data from a Finnish population sample (17). The authors found seven demographic and questionnaire variables that predicted future drug-treated diabetes with ROC curve area of 0.85 and 0.87 in two consecutive cohorts. As with the San Antonio study clinical model, the Finnish model requires the use of a computer to calculate the risk score and has not been validated in different ethnic groups. The present study differs from these prior studies in that it was first a cross-sectional evaluation of predictors of abnormal glucose tolerance in one sample and then was prospectively validated for diabetes prediction in a separate cohort. Moreover, this prediction rule was derived and validated in separate samples of different racial makeup and was based on only four variables that could be calculated easily without the use of a computerized program.

Clinical decision rules attempting to refine the accuracy of clinicians’ diagnostic and prognostic ability are widely reported. Using criteria developed to grade the evidence for clinical decision rules (10), our study is limited in its applicability because we validated it only in the Health ABC study with well-functioning adults older than age 70. The age criterion for our prediction rule was met by every Health ABC participant, so age was uninformative. This may have weakened both the sensitivity and specificity of the rule in the validation cohort. This prediction rule needs to be validated prospectively in broader populations. Clinicians may consider using this rule if the patients in their clinical setting are older, well-functioning, black or white adults. However, the risk of a lifestyle intervention for those identified with abnormal glucose tolerance is low. Lifestyle modification has been recommended as the first-line intervention to prevent or delay diabetes, with the recommended goals of modest weight loss (5–10% of body weight) and modest physical activity (30 min daily) (18). Because this intervention is likely to have a variety of other benefits, the ADA has urged health care providers to counsel all overweight or sedentary individuals to adopt these changes.

We found that measures of adiposity such as BMI or waist circumference were not independently associated with abnormal glucose tolerance. We did not expect to find an association between BMI and glucose intolerance because BMI is a poor marker for total adiposity in older adults (19). But abdominal adiposity has been more closely linked with adverse metabolic consequences and has been suggested to precede insulin resistance (20). Waist circumference is an accepted criterion in the diagnosis of the metabolic syndrome, but waist circumference may reflect total body fat better than visceral fat in older adults (21). It is possible that in our derivation cohort waist circumference did not adequately represent visceral obesity, explaining the lack of association of waist circumference with glucose intolerance.

In conclusion, advanced age, female sex, FPG, and triglycerides had good ability to predict those with abnormal glucose tolerance. A score of 4 or higher doubled the likelihood of an abnormal postchallenge glucose test result. This rule could help clinicians to better identify individuals with abnormal glucose tolerance, who should be targeted for interventions to prevent diabetes.

A.M.K. was funded by the BIRCWH (5 K12 AR47659) and by grant P30-AG15272 under the RCMAR program. E.B.-C., PI for the RBS, was funded by NIA grant 5R01 AG07181 and NIDDK grant 5R01 DK31801. The Health ABC study was funded by NIA contracts N01-AG-6-2101, N01-AG-6-2103, and N01-AG-6-2106.

The authors thank Dr. Eric Vittinghoff for his statistical help with recursive partitioning and the development of models.