Increasing the Accuracy of Oral Glucose Tolerance Testing and Extending Its Application to Individuals With Normal Glucose Tolerance for the Prediction of Type 1 Diabetes

The Diabetes Prevention Trial-Type 1

  1. Jay M. Sosenko, MD1,
  2. Jerry P. Palmer, MD2,
  3. Carla J. Greenbaum, MD3,
  4. Jeffrey Mahon, MD4,
  5. Catherine Cowie, PHD5,
  6. Jeffrey P. Krischer, PHD6,
  7. H. Peter Chase, MD7,
  8. Neil H. White, MD8,
  9. Bruce Buckingham, MD9,
  10. Kevan C. Herold, MD10,
  11. David Cuthbertson, MS11,
  12. Jay S. Skyler, MD1 and
  13. the Diabetes Prevention Trial-Type 1 Study Group
  1. 1Division of Endocrinology, University of Miami, Miami, Florida
  2. 2Division of Endocrinology/Metabolism, University of Washington, Seattle, Washington
  3. 3Benaroya Research Institute at Virginia Mason, Seattle, Washington
  4. 4Department of Epidemiology and Biostatistics, University of Western Ontario, Ontario, Canada
  5. 5National Institute of Diabetes and Digestive and Kidney Diseases/National Institutes of Health, Bethesda, Maryland
  6. 6Division of Informatics and Biostatistics, University of South Florida, Tampa, Florida
  7. 7Barbara Davis Center for Childhood Diabetes, Denver, Colorado
  8. 8Department of Pediatric Endocrinology and Metabolism, Washington University in St. Louis, School of Medicine, St. Louis, Missouri
  9. 9Department of Pediatric Endocrinology, Stanford University, Stanford, California
  10. 10Division of Endocrinology, Columbia University, New York, New York
  11. 11Pediatrics Epidemiology Center, University of South Florida, Tampa, Florida
  1. Address correspondence and reprint requests to Jay M. Sosenko, MD, Division of Endocrinology, University of Miami, P.O. Box 016960 (D110), Miami, FL 33101. E-mail: jsosenko{at}med.miami.edu

Abstract

OBJECTIVE—We assessed the extent to which both standard and alternative indexes from 2-h oral glucose tolerance testing predict type 1 diabetes and whether oral glucose tolerance tests (OGTTs) predict type 1 diabetes in individuals with normal glucose tolerance.

RESEARCH DESIGN AND METHODS—The prediction of type 1 diabetes from baseline OGTTs was studied in 704 Diabetes Prevention Trial-Type 1 participants (islet-cell autoantibody [ICA]-positive relatives of type 1 diabetic patients). The maximum follow-up was 7.4 years. Analyses utilized receiver-operator curves (ROCs), proportional hazards models, and survival curves.

RESULTS—ROC areas under the curve (ROCAUCs) for both the AUC glucose (0.73 ± 0.02) and an OGTT prediction index (0.78 ± 0.02) were higher (P < 0.001) than those for the fasting (0.53 ± 0.02) and 2-h glucose (0.66 ± 0.02). ROCAUCs for the 60- and 90-min glucose (0.71 ± 0.02 and 0.72 ± 0.02, respectively) were also higher (P < 0.01) than those for the fasting and 2-h glucose. Among individuals with normal glucose tolerance, OGTTs were highly predictive, with 4th versus 1st quartile hazard ratios for the 2-h glucose, AUC glucose, and OGTT prediction index ranging from 3.77 to 5.30 (P < 0.001 for all).

CONCLUSIONS—Certain alternative OGTT indexes appear to better predict type 1 diabetes than standard OGTT indexes in ICA-positive relatives of type 1 diabetic patients. Moreover, even among those with normal glucose tolerance, OGTTs are strongly predictive. This suggests that subtle metabolic abnormalities are present several years before the diagnosis of type 1 diabetes.

The accurate prediction of type 1 diabetes is essential to the success of prevention trials for this disorder. Oral glucose tolerance tests (OGTTs) have been used for the prediction of type 1 diabetes in such (1,2), but it is possible that additional information can be obtained from OGTTs that would lead to more prediction accuracy. Prevention trials have used standard criteria for glucose intolerance in their prediction algorithms, and those criteria are based on studies of type 2 diabetes (3). Given the pathogenetic differences between type 2 and type 1 diabetes, alternative OGTT indexes might more accurately predict type 1 diabetes. Moreover, the use of OGTTs for the prediction of type 1 diabetes could extend to the normal range of glycemia. With greater predictive accuracy and range of application, prevention trials could target more specific populations of individuals at risk for type 1 diabetes. We have utilized data from the Diabetes Prevention Trial-Type 1 (DPT-1) to determine whether information from OGTTs might be better utilized to predict type 1 diabetes. In particular, we have compared the prediction accuracies of alternative OGTT indexes with those of standard OGTT indexes.

RESEARCH DESIGN AND METHODS—

Individuals (n = 704) included in the analyses participated in the DPT-1 parenteral (n = 333) and oral (n = 371) insulin trials to delay the onset of type 1 diabetes; neither of these trials demonstrated a treatment effect (1,2). All participants were islet-cell autoantibody (ICA)-positive relatives of patients with type 1 diabetes who met certain 5-year risk criteria for type 1 diabetes. The algorithm for determining risk previously has been described (1,2). In addition to ICA, it was based on the presence of first-phase insulin responses (the sum of insulin levels at 1 and 3 min) on intravenous glucose tolerance tests, glucose tolerance abnormalities on OGTTs, and the presence of insulin autoantibodies. Participants were considered to be at >50% 5-year risk and thus eligible for entry into the parenteral insulin trial if either the first-phase insulin response was below a defined threshold and/or there were glucose abnormalities on the OGTT. If none of those metabolic abnormalities was present, but insulin autoantibodies were positive, the 5-year risk was considered to be 25–50% and participants were eligible for entry into the oral insulin trial. Informed consent was obtained from the participants after the nature of the procedures was explained.

Those in the parenteral insulin trial intervention group received low-dose recombinant human ultralente insulin (Humulin U; Eli Lilly) twice a day and a yearly intravenous insulin infusion, while those in the oral insulin trial intervention group received recombinant human insulin crystals at a dose of 7.5 mg per day. OGTTs were performed at 6-month (±3 months) intervals after randomization in both trials. The dose of oral glucose was 1.75 g/kg (maximum, 75 g of carbohydrate). Blood samples were obtained for plasma glucose and C-peptide measurements in the fasting state and at 30, 60, 90, and 120 min. Insulin measurements were not obtained. Diabetes was usually diagnosed at a routine visit. If the fasting glucose was ≥126 mg/dl and/or if the 2-h glucose was ≥200 mg/dl, participants were asked to return for confirmation at a follow-up visit unless clinically contraindicated. The diagnosis was confirmed if either the fasting glucose was ≥126 mg/dl or the 2-h glucose was ≥200 mg/dl. If glucose values were not confirmed at the follow-up visit, participants continued to be followed at 6-month intervals. The OGTT results were not made available to the researchers or the participants unless diabetes was diagnosed.

Laboratory measures

Plasma glucose levels were measured by the glucose oxidase method. C-peptide levels were measured by radioimmunoassay. The interassay coefficient of variation for the C-peptide assay was 6.9% in a reference pool with relatively high values and 7.8% in a reference pool with relatively low values.

Data analysis

Impaired glucose tolerance was defined both for the DPT-1 and for the analyses presented below as 2-h glucose values of 140–199 mg/dl. Impaired fasting glucose had been defined originally for the DPT-1 as fasting glucose values of 110–125 mg/dl but was defined as 100–125 mg/dl for these analyses. The latter range was chosen so that impaired fasting glucose would correspond to the most recent standard criteria (4). For that same reason, the analyses did not include the indeterminate glucose category (glucose values ≥200 mg/dl at 30, 60, and/or 90 min) used in the DPT-1.

χ2 tests and t tests were used for univariate comparisons. Proportional hazards models and survival curves were utilized to determine hazard ratio and 5-year probability estimates. Statistical differences between receiver-operator curves (ROCs) were determined nonparametrically. The duration of follow-up for the analyses began at randomization. C-peptide values in the undetectable range (<0.2 ng/ml) were assigned a value of 0.1 ng/ml. Fasting glucose measurements were obtained at −10 min and at time 0; the latter were utilized for the analyses with the exception of three individuals with missing values at time 0. Seven individuals were excluded from the analyses due to other missing values. When ties of values occurred at desired cutoffs for categories such as quartiles, approximations were made to the extent possible. The trapezoidal rule was used to calculate OGTT areas under the curve. The SAS 9.1.3 version was used for the analyses. All P values are two sided.

RESULTS—

The mean age ± SD of participants in the DPT-1 (n = 711) was 13.9 ± 9.6 years. Fifty-six percent were male subjects. The mean duration of follow-up was 3.6 ± 1.8 years (maximum 7.4 years) for those who did not develop type 1 diabetes (n = 453); in those developing type 1 diabetes, the mean time to diagnosis was 2.5 ± 1.5 years (n = 258).

Table 1 displays the coefficients for univariate associations of the occurrence of type 1 diabetes with indexes from the baseline OGTTs for the 704 participants included in the analyses. Type 1 diabetes was statistically significantly associated with all of the indexes, but the association with the log fasting C-peptide was relatively weak (P = 0.034; P = 0.074 before log transformation). As is evident from the χ2 values, type 1 diabetes was much more strongly associated with the postchallenge glucose indexes than with the fasting glucose, and the association with the area under the curve (AUC) glucose was substantially greater than that with the 2-h glucose. Type 1 diabetes was also more strongly associated with postchallenge C-peptide levels (P < 0.001 for AUC C-peptide) than with fasting levels. There were no significant differences between the parenteral and oral insulin trials in associations of type 1 diabetes with the above OGTT indexes. Also, within each trial there were no significant differences in associations of type 1 diabetes with those OGTT indexes between participants treated and untreated.

The prediction accuracies of the fasting, 2-h, and AUC glucose are compared in Fig. 1. The accuracy of the fasting glucose for predicting type 1 diabetes, as indicated by the AUC (ROCAUC), was 0.53 ± 0.02 (±SE), barely above the 0.50 level at which there is no predictive information. It was markedly lower than that for the 2-h glucose (ROCAUC = 0.66 ± 0.02, P < 0. 001 for the difference) and that for the AUC glucose (ROCAUC = 0.73 ± 0.02, P < 0.001 for the difference). The ROCAUC for the AUC glucose was significantly greater than that for the 2-h glucose (P < 0.001 for the difference). The ROCAUCs (not shown) for the 60- and 90-min glucose (0.71 ± 0.02 and 0.72 ± 0.02, respectively) were similar to that for the AUC glucose, and both were significantly higher than that for the 2-h glucose (P = 0.018 and P < 0.001, respectively, for the differences).

The positive predictive values for impaired fasting glucose (n = 46) and impaired glucose tolerance (n = 110) were 0.50 and 0.64, respectively. The positive predictive value for those with the 110 highest AUC glucose levels (to match the number above the 2-h glucose threshold for impaired glucose tolerance) was 0.68. The negative predictive values for those in the lowest quartiles of the 2-h glucose and the AUC glucose distributions were 0.80 and 0.84, respectively.

Stepwise proportional hazards modeling was utilized to develop a prediction index from the OGTT glucose and C-peptide measurements. Type 1 diabetes was significantly related to the fasting glucose and the 2-h glucose in the univariate models (see Table 1) but not in the stepwise modeling (although the P value for the 2-h glucose was <0.10). The relatively weak and inverse univariate association of type 1 diabetes with the log fasting C-peptide was stronger (P < 0.001) and positive in the final stepwise model. The equation was OGTT prediction index = 0.48(log fasting C-peptide) + 2.63 × 10−4(AUC glucose) − 3.42 × 10−3(AUC C-peptide). There was no significant difference between the parenteral and oral insulin trials in the association of type 1 diabetes with the OGTT prediction index.

As expected from fitting the data, type 1 diabetes was strongly associated with the OGTT prediction index (P < 0.001). The ROCAUC (0.78 ± 0.02) of the OGTT prediction index was significantly greater than that of the 2-h glucose and that of the AUC glucose (P < 0.001 for both differences). The positive predictive value for the 110 individuals with the highest OGTT prediction index levels was 0.77, while the negative predictive value for the lowest quartile was 0.87.

Table 2 shows the 4th quartile hazard ratios (with the 1st quartile as the reference group) for the 2-h glucose, AUC glucose, and the OGTT prediction index. The hazard ratios were highly significant for all (P < 0.001); however, they were higher for the AUC glucose and the OGTT prediction index than for the 2-h glucose. The 5-year risk ± SE estimates for the 4th quartiles were also greater for the AUC glucose and the OGTT prediction index than for the 2-h glucose.

Among the oral insulin trial participants (all with normal glucose tolerance), there were still strong associations of type 1 diabetes with the 2-h glucose (P = 0.002), AUC glucose (P < 0.001), and the OGTT prediction index (P < 0.001). The ROCAUCs of these indexes were 0.61 ± 0.03, 0.65 ± 0.03, and 0.68 ± 0.03, respectively. There were no significant differences among them, but all were significantly greater (P = 0.006 for the 2-h glucose; P < 0.001 for both the AUC glucose and the OGTT prediction index) than the ROCAUC of the fasting glucose (0.49 ± 0.03).

Table 3 shows the 4th quartile hazard ratios (with the 1st quartile as the reference group) for the 2-h glucose, AUC glucose, and the OGTT prediction index of the oral insulin trial participants. All hazard ratios were highly significant (P < 0.001 for all), although the hazard ratio for the 2-h glucose was the lowest. This pattern was also apparent for the 5-year risk estimates.

CONCLUSIONS—

As new prevention trials for type 1 diabetes are being developed, it is particularly important to maximize the accuracy of prediction of type 1 diabetes. Improved prediction accuracy would reduce concern over administering experimental treatments to individuals at lower risk and ultimately lead to a more efficient selection of those who are truly at risk. Given these considerations and the data presented above, it appears that there should be a reassessment of OGTT prediction criteria for type 1 diabetes.

The analyses compared prediction accuracies of alternative OGTT indexes with the conventional indexes used in DPT-1. They suggest that the 2-h glucose, and especially the fasting glucose, which are standard OGTT indexes, are inferior to certain alternative OGTT indexes for the prediction of type 1 diabetes in ICA-positive relatives of type 1 diabetic patients. The AUC glucose, 60-min glucose, and 90-min glucose were all significantly more predictive than the 2-h glucose. Moreover, C-peptide levels from the OGTTs were predictive of type 1 diabetes and contributed to an index that surpassed all of the predictors that only included glucose levels. It is to be noted that the predictive values shown should be considered in a relative context, since they are a function of time. However, the 5-year risk estimates show a similar pattern.

The data also indicate that OGTTs are useful for the prediction of type 1 diabetes even among ICA-positive relatives with normal glucose tolerance. The 2-h glucose was still highly predictive of type 1 diabetes in these individuals, and the contribution of OGTTs for predicting type 1 diabetes was even greater when either the AUC glucose or the OGTT prediction index was utilized. The occurrence of type 1 diabetes was several-fold higher for the 4th quartile than for the 1st quartile for all of those indexes.

It is well known that autoantibodies commonly occur years before the diagnosis of type 1 diabetes (58) and that glucose intolerance can also be present before diagnosis (910). Our observation that glycemia within the normal range is strongly predictive of type 1 diabetes suggests that there are also a number of individuals with subtle, but consequential, metabolic abnormalities for at least a few years before diagnosis. Data from another study of DPT-1 participants (11) also suggest that subtle metabolic abnormalities occur well before the onset of type 1 diabetes; glucose levels were gradually increasing at least 2 years before diagnosis.

There are no previous reports of the use of OGTT data for predicting type 1 diabetes in individuals with normal glucose tolerance. However, some studies have shown that glucose levels within the high normal range are predictive of the development of type 2 diabetes in certain populations (12,13).

The high accuracy of the OGTT prediction index is not surprising, since it was based on coefficients derived from regression modeling. Still, the OGTT prediction index serves to show that C-peptide measurements can improve prediction accuracy to some extent. The ultimate choice of which OGTT indexes to use for prediction is a function of purpose, cost, and convenience. Whereas multiple glucose and C-peptide measurements appear to yield better predictive accuracy than a single postchallenge glucose determination, the latter would be less costly and perhaps better tolerated.

The usefulness of OGTT indexes is dependent on whether the magnitude of their prediction accuracies is sufficient for particular study objectives. Moreover, since our findings were derived from a population of ICA-positive relatives of patients with type 1 diabetes, the better prediction accuracies of the alternative OGTT indexes might not extend to other populations at risk for type 1 diabetes. However, the substantially greater prediction accuracies of those indexes indicate that the findings might be generalizable. Also, the findings could be relevant to sporadic cases of type 1 diabetes, since some studies show common features between those individuals and diagnosed relatives of type 1 diabetic patients (14,15).

It should be noted that the fasting glucose and the 2-h glucose distributions are utilized both as predictors and as American Diabetes Association diagnostic criteria. Although this could conceivably have impacted the analyses, it appears that if anything, this would have provided an inherent predictive advantage for those standard indexes over the alternative indexes (e.g., AUC glucose). Since the alternative indexes were more accurate predictors even with this potential disadvantage, that consideration should not alter our conclusion.

Because the 5-year risk estimates were so high for certain groups, it is possible that criteria for the diagnosis of type 1 diabetes (essentially based on data pertaining to type 2 diabetes) need to be reassessed. This could be of importance, since it has been shown that even with current diagnostic criteria, interventions to achieve better control soon after the diagnosis of type 1 diabetes can prolong insulin secretion (16). Such interventions could be even more effective at an earlier pathogenetic stage of the disease and ultimately reduce the development of complications (1719).

DPT-1 utilized both glucose tolerance abnormalities and a decreased first-phase insulin response to effectively predict the numbers of participants required for its trials. However, the unique DPT-1 database itself has now provided evidence that the metabolic component of type 1 diabetes prediction can be enhanced and possibly simplified.

Figure 1—

Shown are the ROC curves of the fasting glucose, 2-h glucose, and AUC glucose for predicting type 1 diabetes in DPT-1 participants (n = 704). The ROCAUCs of both the 2-h glucose and the AUC glucose are much greater than that of the fasting glucose, while the ROCAUC of the AUC glucose is much greater than that of the 2-h glucose.

Table 1—

Associations of occurrence of type 1 diabetes with baseline OGTT indices from proportional hazards models (n = 704)

Table 2—

Hazard ratios and 5-year risk estimates for type 1 diabetes for 4th quartiles of baseline OGTT indices in DPT-1 participants (n = 704)

Table 3—

Hazard ratios and 5-year risk estimates for type 1 diabetes for 4th quartiles of baseline OGTT indices in DPT-1 participants with normal glucose tolerance (n = 360)

Acknowledgments

This study was sponsored by cooperative agreements with the Division of Diabetes, Endocrinology and Metabolic Diseases of the National Institute of Diabetes and Digestive and Kidney Disease; the National Institute of Allergy and Infectious Diseases; the National Institute of Child Health and Human Development; the National Center for Research Resources; the American Diabetes Association; and the Juvenile Diabetes Research Foundation. Supplies were provided by Eli Lilly, Bayer, Becton Dickinson, International Technidyne, LifeScan, the Mead Johnson Nutritionals Division of Bristol-Myers Squibb, the Medisense Division of Abbott Laboratories, MiniMed, and Roche Diagnostics.

Footnotes

  • K.C.H. is currently affiliated with the Division of Endocrinology and Metabolism, Yale University School of Medicine, New Haven, Connecticut.

    A table elsewhere in this issue shows conventional and Système International (SI) units and conversion factors for many substances.

    The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked “advertisement” in accordance with 18 U.S.C Section 1734 solely to indicate this fact.

    • Accepted September 30, 2006.
    • Received July 31, 2006.

References

| Table of Contents