Comparison of a Clinical Model, the Oral Glucose Tolerance Test, and Fasting Glucose for Prediction of Type 2 Diabetes Risk in Japanese Americans

Study subjects included second-generation (Nisei) and third-generation (Sansei) Japanese-American participants in the Japanese American Community Diabetes Study. Recruitment methods and comparison of Nisei participants with nonparticipants residing in King County, WA, have been previously described (6). Subjects with diabetes at baseline were excluded (7). This study was approved by the University of Washington Institutional Review Board, and all participants provided written informed consent.

Study subjects were examined in the General Clinical Research Center at the University of Washington at baseline and at 5–6 years and 10 years after baseline. Subjects who reported a parent or sibling with adult-onset diabetes at baseline were considered to have a positive family history of type 2 diabetes. Medication use was verified by direct observation of medication containers provided by the subject at each examination. Standing height (cm) and weight (kg) were measured without shoes in light clothing. Supine blood pressure was measured in the right arm three times by auscultation using a mercury manometer, and the average of the last two measurements was used for analysis.

Blood samples for measurement of glucose, lipids, and lipoprotein were collected after a 10-h fast. A blood sample collected 2 h after a 75-g oral glucose load was used to measure 2-h plasma glucose. Plasma glucose was assayed by an automated glucose oxidase method. Lipids and lipoproteins were measured at the Northwest Lipid Research Laboratory using described methods (8–10).

IFG was defined as fasting plasma glucose (FPG) ≥6.1 mmol/l (110 mg/dl) and <7.0 mmol/l (126 mg/dl) (7). IGT was defined as 2-h plasma glucose ≥7.8 mmol/l (140 mg/dl) and <11.1 mmol/l (200 mg/dl). IFG and IGT classifications were assigned independently; therefore, subjects with IFG at baseline may or may not have had IGT at baseline. Subjects were classified as having diabetes if any of these criteria were met: FPG ≥7.0 mmol/l (126 mg/dl), 2-h glucose ≥11.1 mmol/l (200 mg/dl), or reported use of insulin or an oral hypoglycemic medication prescribed for management of diabetes by a physician (7). Diabetes status was determined at the 10-year follow-up, independent of diabetes status at the 5- to 6-year follow-up visit.

Of 518 eligible subjects studied at baseline, 465 (89.8%) completed the 5- to 6-year examination, and 412 (79.5%) completed the 10-year examination. Reasons for missing the 5- to 6-year examination included refusal (n = 29), illness (n = 10), death (n = 6), relocation (n = 4), and inability to locate (n = 4). By 10 years after baseline, 23 subjects had died (13 of cancer, 8 of cardiovascular disease, 1 of pneumonia, and 1 of trauma). Compared with participants, subjects who missed a follow-up examination had lower baseline systolic blood pressure (P = 0.047 at 5–6 years, no significant difference at 10 years) and diastolic blood pressure (P = 0.011 at 5–6 years, P = 0.028 at 10 years) and higher HDL cholesterol (P = 0.050 at 5 years, no significant difference at 10 years). There were no significant differences in age, sex, family history of diabetes, BMI, fasting or 2-h glucose, IGT, IFG, total cholesterol, LDL cholesterol, or triglycerides.

All statistics were calculated using Stata software, version 7 for Windows (Stata Corporation, College Station, TX). All P values were two-sided. Baseline characteristics of subjects in whom diabetes developed and in those who remained nondiabetic were compared using the Wilcoxon’s rank-sum test or the χ² test.

The probability of developing diabetes at baseline was calculated for each study subject using a published clinical model (5): probability of incident diabetes = 1/(1 + e^–x), where x = −13.415 + 0.028 (age in years) + 0.661 (0 if male, 1 if female) + 0.412 (ethnicity) + 0.079 (FPG in mg/dl) + 0.018 (systolic blood pressure in mmHg) –0.039 (HDL cholesterol in mg/dl) + 0.070 (BMI in kg/m²) + 0.481 (0 if no family history, 1 if family history present). Family history is defined as the presence of type 2 diabetes in a parent or sibling. The published model coded ethnicity as non-Hispanic white = 0 and Mexican American = 1. We coded ethnicity for Japanese Americans = 1, which affects the absolute value of the calculated probability of diabetes but does not affect statistical comparisons because all subjects have the same value for ethnicity. To distinguish between the validity of the published coefficients for this model and the validity of these clinical variables for determining diabetes risk in Japanese Americans, we also calculated the probability of developing diabetes at 5–6 years and at 10 years using logistic regression models that included the same independent variables as those in the published clinical model, except for ethnicity. This model is referred to as the “clinical model using study data.” We also tested to determine whether adding one additional independent variable or first-order multiplicative interaction term improved the clinical model using study data. Nested models were compared using the likelihood ratio test.

Accuracy of predicting incident diabetes at 5–6 years or 10 years was analyzed using receiver-operating characteristic (ROC) curves (11). An ROC curve is a graph of sensitivity versus 1-specificity (or false-positive rate) for various cutoff definitions of a positive diagnostic test result. Statistical differences in the area under the ROC curves were compared using the method of DeLong et al. (12). Sensitivity, specificity, and likelihood ratio (LR) for a positive (LR+) or negative (LR–) test result were calculated for various cutoffs. The LR is the ratio of the frequency of a test result in patients with disease to the frequency of the same test result in patients without disease (13); therefore, an LR of 1.0 reflects no diagnostic value. For 2-h glucose and FPG, the cutoffs for IGT and IFG were used. We also used a cutoff for FPG of ≥5.6 mmol/l (100 mg/dl), as previously suggested (4,14). For the published clinical model, two cutoffs were randomly selected that correspond approximately to the points on the ROC curve for all ages at which the cost-benefit ratios were 1:2 and 1:4 (13). A cost-benefit ratio of 1:4 means a false-negative result is four times worse than a false-positive result, because the opportunity to prevent a serious disease (diabetes) using a low-risk intervention (lifestyle) was missed. We also used the clinical model cutoff where the number of subjects above this cutoff equaled the number of subjects with IGT so the performance characteristics of these two tests could be directly compared.

Subjects ranged in age from 34 to 75 years. BMI ranged from 16.6 to 36.9 kg/m². The incidence of diabetes was 11% at 5 years and 18% at 10 years (Table 1). Comparison of baseline characteristics between subjects in whom diabetes did and did not develop is shown in Table 1. The 5- to 6-year and 10-year incidences of diabetes were 23 and 64%, respectively, for subjects with IFG; 25 and 38%, respectively, for subjects with IGT; and 1.8 and 4.5%, respectively, for subjects without IFG or IGT. At baseline, 20 subjects had both IFG and IGT and 277 subjects had normal FPG and normal glucose tolerance.

Adding the following variables individually to the clinical model did not improve the fit of the model for predicting 5- to 6-year incidence of diabetes: ln triglycerides (P = 0.32), diastolic blood pressure (P = 0.53), waist circumference (P = 0.11), or fasting insulin (P = 0.13). Each of the following interaction terms significantly improved the fit of the clinical model using study data for predicting diabetes at 5–6 years: age × BMI (P = 0.011), age × sex (P = 0.004), and age × HDL cholesterol (P < 0.0001). No significant interaction was found for age × family history of diabetes (P = 0.67), age × systolic blood pressure (P = 0.12), age × FPG (P = 0.07), or sex × HDL cholesterol (P = 0.63).

The areas under the ROC curves for the various models tested are shown in Table 2. Because the clinical model does not account for differences in the association between incident diabetes and several independent variables with age, the results are also stratified by age. The age stratification demonstrates that the clinical model was significantly better than FPG, and comparable to the clinical model using study data and 2-h glucose, for predicting 5- to 6-year incidence of diabetes in subjects aged ≤55 years. In older subjects, 2-h glucose was significantly better than the clinical model or FPG for predicting 5-year incidence of diabetes. The clinical model using study data was significantly better than the published clinical model and fasting glucose in older subjects.

Sensitivity, specificity, and likelihood ratios are shown in Table 3. It should be noted that only 30 subjects had IFG, 20 of whom also had IGT. An especially striking result was the finding that the LR–was low for 2-h glucose in both age groups, indicating that a negative result for IGT was useful for identifying subjects of all ages in whom diabetes did not develop.

We have demonstrated that a clinical model using risk factor information and FPG was significantly better than FPG alone and was not significantly different than 2-h glucose for predicting incident diabetes at 5–6 years in Japanese Americans aged ≤55 years. However, the clinical model was significantly less accurate than 2-h glucose and was not significantly better than FPG alone for predicting 5- to 6-year incidence of diabetes in older subjects. Our findings differ from those of Stern et al. (5), who did not find evidence of significant interactions between the diabetes risk associated with independent variables in the clinical model and age. In the San Antonio Heart Study, subjects ranged in age from 25–64 years (mean 42.6–44.8), whereas our study included older subjects (range 34–75 years, mean 52.1). The inclusion of elderly subjects likely improved our ability to discern the effect of age on the accuracy of the clinical model for determining diabetes risk.

We previously reported a significant age-BMI interaction with diabetes risk in Japanese Americans, such that BMI is a strong risk factor for diabetes in subjects aged ≤55 years, but BMI was not associated with incident diabetes in older subjects (15). In the Third National Health and Nutrition Examination Survey (NHANES III) the association between BMI and diabetes was also found to be greater for subjects younger than 55 years than for older subjects (16). Although models typically perform better in the datasets used to develop them than in independent datasets, we found the clinical model using study data was significantly better than the published clinical model only among older subjects. These findings suggest that the published clinical model does not improve prediction of diabetes beyond that of FPG in older Japanese Americans because it fails to take into account the interactions between age and several variables in the model with diabetes risk.

Diabetes may also be more strongly associated with 2-h glucose than FPG in older subjects compared with younger subjects (17,18). Our findings are consistent with this observation, in that 2-h glucose was more predictive of diabetes risk than the clinical model or FPG in older subjects but not in younger subjects at the 5- to 6-year follow-up period. The effect of age on diabetes risk associated with FPG or 2-h glucose may also explain conflicting results of other studies regarding the prognostic value of IFG compared with IGT. In our study, subjects with IFG had a similar incidence of diabetes compared with subjects with IGT, although few subjects had only IFG. The incidence of diabetes was also similar in subjects with IFG compared with subjects with IGT in the Hoorn Study of Dutch men and women aged 50–75 years (19). However, Pima Indians older than 15 years with IFG reportedly have a higher incidence of diabetes than those with IGT (14), possibly because this population includes a large proportion of younger adults.

Identifying individuals at low risk for diabetes is also of interest. Absence of IGT was a useful test for identifying Japanese Americans at low risk for developing diabetes in this study, although the LR–for a clinical model cutoff ≥0.1265 was lower than that for IGT among subjects aged ≤55 years for predicting 5- to 6-year incidence of diabetes. Gabir et al. (14) suggested that lowering the cutoff for IFG may improve the performance of IFG as a prognostic test for future diabetes. We found that even if a cutoff of ≥5.6 mmol/l (100 mg/dl) for FPG is used, IGT still had a lower LR–than FPG. However, absence of IGT may not be as useful for identifying individuals at low risk for diabetes in other populations. The LR–for IGT was 0.55 in the San Antonio Heart Study (5) and 0.60 in the Mauritius study (20) compared with 0.15–0.30 in this study.

There are several limitations to this study. The sample size was small, particularly after stratification by age, which can result in a type II statistical error. This may have resulted in the failure to detect a significant advantage of the clinical model for predicting 5- to 6-year incidence of diabetes compared with 2-h glucose in younger subjects or FPG in older subjects. The sample size at the 10-year examination was even smaller and may have resulted in failure to detect an advantage of the clinical model over 2-h glucose or fasting glucose in both age groups. Because of the small number of subjects from a single ethnic group, this study is not well suited for proposing new models of diabetes risk in older subjects. Therefore, the results are useful for demonstrating the limitations of the clinical model in older Japanese Americans, but a larger study population would be needed to develop more refined models. Another limitation is that only one OGTT was performed at each visit in this study, and 2-h glucose has lower reproducibility than FPG (21,22). To minimize this problem, we evaluated diabetes risk at 10 years, independent of the findings at the 5- to 6-year visit. The area under the ROC curve was larger for 2-h glucose than for the clinical model or FPG in older subjects at the 5- to 6-year follow-up period, and in both age groups at the 10-year examination. Therefore, it is unlikely that variability in diabetes outcomes based on 2-h glucose accounts for failure to detect an advantage of the published clinical model over 2-h glucose in Japanese Americans. However, variability in diabetes outcomes based on 2-h glucose might account for the discrepancy in the level of significance for the comparisons of 2-h glucose to the clinical model at the 5- to 6-year and 10-year follow-up periods.

In summary, this study demonstrates that a recently published clinical model was significantly better than FPG alone, but not 2-h glucose, for predicting 5- to 6-year incidence of diabetes in Japanese Americans aged ≤55 years. However, the model was significantly worse than 2-h glucose and was no better than FPG alone in older subjects. The sensitivity of IFG for predicting future diabetes was poor (6–29%). Furthermore, absence of IGT seems to be more useful for identifying Japanese Americans at low risk of developing diabetes than is absence of IFG. Our findings indicate that despite the limitations of the OGTT, the 2-h glucose is a useful test for predicting future diabetes in middle-aged and elderly Japanese Americans. If mathematical models of diabetes risk are to be used as an alternative to the OGTT for defining diabetes risk in clinical practice, then further refinements that take into account the differential effects of age are needed. Further research is also needed to determine whether these findings apply to other Asian and non-Asian groups in the U.S. and in other countries.

This research was supported by National Institutes of Health Grants DK-02654, DK-31170, and HL-49292. Facility support was provided by the Clinical Nutrition Research Unit (DK-35816), the Diabetes Endocrinology Research Center (DK-17047), the General Clinical Research Center (RR-00037) at the University of Washington, and the Medical Research Service of the Department of Veterans Affairs.