Hemoglobin A_1c as a Predictor of Incident Diabetes

This historical cohort study included 12,589 patients at two Department of Veterans Affairs (VA) Medical Centers who had at least one HbA_1c test <6.5% between 1 January 2000 through 31 December 2001 (baseline HbA_1c). Institutional review board approval was obtained at both study sites before initiating the study.

Patient records were evaluated for 12 months before the baseline HbA_1c to assure that they had at least one ambulatory care visit, no HbA_1c ≥6.5%, and no diagnosis of diabetes. A diagnosis of diabetes was defined as: at least one inpatient diagnosis of diabetes mellitus (ICD-9 code of 250.*), two or more outpatient diagnoses of diabetes mellitus (ICD-9 code of 250.*), or a prescription for any medication used in diabetic treatment. This method to ascertain diabetic status has been previously validated within the VA (10). Participants with more than one HbA_1c test during the baseline period had their first test result used as the baseline. Patients were required to have at least one ambulatory care visit at any time during the follow-up period (from date of baseline HbA_1c test to 31 December 2008) or until a diabetic diagnosis was made. After the above exclusion criteria were applied, 12,375 patients were classified without diabetes and were entered into the follow-up period. The period of follow-up for a given patient varied based on when they had clinic visits during the follow-up period.

Data were abstracted electronically from the patients’ medical records. Information was collected on outpatient clinic visits, admissions, vital signs, outpatient prescriptions, comorbid diagnoses, patient demographics, and laboratory tests. Because our clinical laboratories do not label glucose measurements as fasting or nonfasting, we recorded glucose levels obtained between 0600 and 1100 h as a proxy for the fasting state. The ICD-9 codes used for the comorbid conditions were as follows: cardiovascular disease (CVD) (coronary heart disease 410–414; stroke 430–438; heart failure 428; cardiac arrest 427.1, 472.4, 427.5; inflammatory heart disease 429.01, 429.1, 420–425; and hypertension 401–405). Race and ethnicity were self-reported by patients. The study end point was whether a diagnosis of diabetes mellitus occurred during the follow-up period.

During the study, both facilities used the same methodology for measuring HbA_1c levels, which used a nonporous ion-exchange high-performance liquid chromatography to separate HbA_1c from other hemoglobin fractions and is certified by the National Glycohemoglobin Standardization Program. This method is fairly immune to the presence of hemoglobinopathies or carbamylated hemoglobin as a result of high urea concentrations. Abnormal concentrations of hemoglobins that may falsely elevate HbA_1c are clearly recognized.

The distribution of baseline HbA_1c levels was analyzed using univariate procedures and stratified into five groups (1): <4.5% (2), 4.5–4.9% (3), 5.0–5.4% (4), 5.5–5.9% (5), and 6.0–6.4%. The lowest group, with HbA_1c <4.5%, was treated as the reference group and all other groups were compared with the reference for risk calculation.

Data were analyzed using SAS 9.1 (SAS Institute, Cary, NC). Descriptive statistics were conducted on the study sample at baseline: median and range were calculated for the continuous variables; frequency and proportion were calculated for the categorical variables. The study also compared the baseline characteristics for patients who developed diabetes with patients who did not develop diabetes in the follow-up period. The nonparametric Wilcoxon rank-sum test was used to compare medians for continuous variables, and the χ² test was used to compare proportions for categorical variables. Given the large sample size, a P value of <0.0001 was used to determine statistical significance. Both logistic regression models and Cox proportional hazards models were used to compare the risk of developing diabetes with baseline HbA_1c level as the main effect (HbA_1c <4.5% as reference group). Based on the univariate results, a stepwise selection method (using an α of 0.05) was used for further evaluation of the confounders in the multiple logistic regression. Both unadjusted odds ratio/hazard ratio and multivariable adjusted odds ratio/hazard ratio were calculated. The Kaplan–Meier method was used to calculate survival probability with time, and a diabetic event probability versus time plot was developed and stratified by baseline HbA_1c groups. All of the 12,375 patients were included in the survival analysis, including those patients who died before developing diabetes. Date of entry was the date on which the initial HbA_1c measurement was obtained. For those patients who died, we used their date of death as the censor time. For patients who survived without developing diabetes, we used the last visit date as their censor time in the survival analysis calculation.

We also developed a risk model for predicting the 5-year incidence of diabetes using logistic regression. The 5-year incidence of diabetes was defined as any patient who developed diabetes within the 5-year period after the baseline HbA_1c test, and patients who did not develop diabetes within the 5-year period were used as the control group. A simple risk model was developed using baseline HbA_1c as the only predictor, and a multivariable model was developed using baseline HbA_1c, age, BMI, and systolic blood pressure (SBP) as predictors. The areas under the two receiver operating characteristic (ROC) curves (indicated by C statistic) were compared using nonparametric approaches (11). Finally, risk-calculating equations were developed from the above findings.

There were 12,589 individuals eligible for the study by having an HbA_1c test during the baseline period; 214 (1.7%) did not visit the clinic during the follow-up period and were lost to follow-up, leaving 12,375 in the study population. Baseline characteristics of the 12,375 individuals, including those who developed diabetes during follow-up, are shown in Table 1. Individuals were predominantly white men (95.4% men and 67.5% whites) with a median age of 65.9 (range 18.5– 101.5. Comparison of the demographics and clinical characteristics between these individuals and the 214 who had no follow-up visits showed no significant differences with regard to HbA_1c, age, glucose, SBP, creatinine, estimated glomerular filtration rate (eGFR), sex, and presence of cardiovascular disease. Those lost to follow-up were significantly lower with regard to diastolic blood pressure, albumin, BMI, and presence of hypertension.

During an average follow-up of 4.4 years and with an average of 140 (SD 194) ambulatory care visits, 26.9% developed diabetes. The criteria by which diabetes was diagnosed were: outpatient codes (59.9%), inpatient codes (5.2%), and new diabetes medication (34.9%). Blood pressure, BMI, glucose, serum creatinine, prevalent cardiovascular disease, and hypertension were significantly higher (P < 0.0001) in patients who developed diabetes. During the study period, there was a progressive decline in the number and percentage of individuals with clinic encounters in a given year. In the last year of the study period, 6,997 (56.5%) individuals had one or more clinic visits, 2,671 (21.6%) patients had no clinic visits, and 2,707 (21.9%) had died during the 8 years of follow-up.

Logistic regression was used to compare the groups for risk of developing diabetes in the follow-up period (Table 2). When compared with the reference HbA_1c group (<4.5%), the group with HbA_1c 4.5–4.9% was not significantly different, whereas risk of developing diabetes increased steadily for the higher HbA_1c groups (≥5.0%). The point estimates for unadjusted odds ratios were 1.57 for HbA_1c 5.0–5.4%, 4.54 for HbA_1c 5.5–5.9%, and 14.93 for HbA_1c 6.0–6.4% compared with HbA_1c <4.5% (P < 0.0001). The adjusted odds ratios by multivariable logistic regression showed similar trends with slight differences in point estimates and 95% confidence intervals. We assessed whether the number of outpatient visits affected a diagnosis of diabetes but found that it did not contribute to model fit and was not included in the logistic regression model.

Table 3 shows the results from the Cox proportional hazards models. Estimates of hazard ratios showed similar patterns as the odds ratio estimates with logistic regression. No significant difference was detected between the reference group and those with HbA_1c 4.5–4.9%, but the hazard ratio increased significantly beginning with HbA_1c ≥5.0% (P < 0.0001), with higher baseline HbA_1c associated with higher risk.

We analyzed event probability (i.e., developing diabetes) during the follow-up period (Fig. 1) differentiated by baseline HbA_1c. Similar to our other analyses, the curves for incident diabetes in those with baseline HbA_1c <5.0% intertwined, whereas those with higher baseline HbA_1c had significantly higher probability of developing diabetes during follow-up (log-rank P value <0.0001).

Risk models for a diabetic diagnosis over 5 years were developed using HbA_1c alone (model 1) and a multivariable model using HbA_1c, age, BMI, and SBP as predictors (model 2). For each predictor, a quadratic term was added into the model because of nonlinear association with risk (indicated by significant P values associated with quadratic terms). For model 1, the area under the ROC curve = 0.7543 (95% confidence interval 0.7429–0.7657), and for model 2, the area under the ROC curve = 0.7791 (95% confidence interval: 0.7687–0.7896). Comparison of the two areas showed significant improvement of the predictability of model 2 (P value <0.0001). The Hosmer and Lemeshow Goodness-of-Fit P value was 0.2827 for model 2, indicating good model fit.

Based on the above findings, we developed risk calculating equations as follows:

where HbA_1c is hemoglobin HbA_1c in percentage, age is indicated in years, BMI is measured in kg/m², and SBP is indicated in mmHg. The first equation calculates the odds of developing diabetes mellitus in 5 years, and the second equation calculates the probability of developing diabetes mellitus in 5 years using the result from Eq. 1.

We found that baseline HbA_1c was significantly predictive of the subsequent development of a diagnosis of diabetes over an 8-year period. The risk of developing diabetes increased progressively at HbA_1c levels ≥5.0%, with an odds ratio exceeding 16 in those with HbA_1c 6.0–6.4%. This latter group had a cumulative incidence of diabetes approaching 80%. Not surprisingly, significant predictors for diabetes incidence included clinical parameters, such as blood pressure, BMI, serum creatinine, prevalent cardiovascular disease, and hypertension. From these data, we also developed risk–calculating equations for determining the probability of developing a diabetic diagnosis within 5 years. We believe that these data will inform clinicians on how to risk–stratify individuals who are screened for diabetes using HbA_1c but whose levels do not reach the recommended diagnostic threshold of ≥6.5%.

Several studies have evaluated HbA_1c as a predictor of subsequent diabetes or as a tool to diagnose treatment-requiring diabetes (3–7,12–19). A number of threshold values have been previously proposed for diagnosing diabetes, such as ≥7.0% (4,7), ≥6.5% (13), >2 SD above the normal mean (i.e., >6.1%) (6). In addition, a number of HbA_1c levels have also been proposed to identify individuals at risk for diabetes (i.e., prediabetes), such as 6.1–6.9% [7], 6.0–6.4% [8], or 5.7–6.4% [1]. The implementation of new guidelines for diagnosing diabetes using HbA_1c will help standardize the way in which clinicians apply results from this test. However, there remains uncertainty on how to classify and whether to intervene in individuals whose levels fall below this threshold.

There is growing evidence that HbA_1c may not only predict diabetes but also cardiovascular disease and death (12–19). Among women without diabetes, HbA_1c levels were significantly associated with both, although the presence of other cardiovascular risk factors may contribute additionally to this risk (15,16). Selvin et al. (18) showed that, in a community-based population, HbA_1c was significantly associated with risk of developing both diabetes and cardiovascular disease independent of fasting glucose levels. As with our results, they showed that levels ≥6.0% carried the greatest risk. Therefore, these results strongly suggest that individuals with HbA_1c levels ≥6.0% should be targeted for prevention strategies to reduce not only incident diabetes but possibly also cardiovascular disease.

Many clinicians have been attracted to using HbA_1c as a screening test for diabetes since the test reflects longer-term glucose control, does not require fasting, has less day-to-day biologic variability, and is a well-accepted marker of risk of long-term microvascular complications (20). Such usage is evidenced by our large cohort of patients in whom HbA_1c levels were obtained in patients without a diagnosis of diabetes. However, prior guidelines discouraged use of HbA_1c for diagnosing diabetes, largely as a result of standardization and reproducibility issues that precluded its use in such broad settings. Current instrumentation and standardization methods (21) aligned with the Diabetes Control and Complications Trial have abrogated most of these issues. Such evidence was cited by the International Expert Committee (8) and affirmed by the American Diabetes Association (9) in their acceptance of HbA_1c for screening and diagnosis.

Significant strengths of this study are its large population drawn from two VA medical centers in different geographic regions, the ability to query a robust electronic medical record for clinical and demographic factors, and patient follow-up for up to 8 years. Important limitations include the largely white men and older population, the reliability of the administrative data set, and the selection and ascertainment bias related to patients for whom HbA_1c testing was performed. It is possible that HbA_1c tests were performed in patients who were preselected for the presence of other known risk factors for diabetes. Such selective screening might bias our results toward showing a higher risk of developing diabetes for a given HbA_1c level. To address this, our analyses controlled for many known risk factors and helped identify those factors that contribute significantly with HbA_1c in predicting risk of diabetes. To account for the possible confounding effects from sex and age, we incorporated these two variables with several other known risk factors in our 8-year follow-up models to calculate odds ratios (hazard ratios) based on baseline HbA_1c. In our 5-year risk prediction model, the sex variable did not provide significant contribution to the predictability of the model.

Other factors may limit the generalizability of our findings. We did not have information on the prevalence of smoking in the study population, which is a known risk factor for diabetes (22) and is higher among veterans than nonveterans (23). This might tend to inflate the risks of diabetes that we observed. Another potential bias is that a greater amount of medical care (e.g., increased number of clinic visits) might associate with a higher number of diabetes cases ascertained during the study. We adjusted for the number of medical visits but did not find that this significantly influenced the models. In addition, there were 214 individuals initially eligible to participate but who were lost to follow-up. Demographic and clinical characteristics of these individuals were quite similar to the remaining cohort. Those areas in which they differed were in a direction that paralleled those who did not develop diabetes during follow-up. Although their absence from the cohort can introduce bias in the ascertainment of both exposure and outcome, we believe that both the small number as well as the characteristics of these individuals makes it unlikely that this substantially affected the reported incidence of diabetes. Finally, we identified diagnosed diabetes based on medical record evidence that the patient was actually diagnosed by a clinician and/or treated with a diabetes medication. These criteria would exclude patients who have an unrecorded but true diagnosis of diabetes for whom no medications were prescribed, which could lead to an underestimation of the true risk of developing diabetes. However, the number of such patients should be small because of our use of multiple variables to establish a diagnosis of diabetes and the extended follow-up period.

In summary, we have characterized the risk of developing diabetes in patients without a diagnosis of diabetes who had a baseline HbA_1c levels <6.5%. These data show a progressive risk for developing diabetes when HbA_1c is ≥5.0%, with nominal risk below that level. We generated a risk calculator using HbA_1c and other clinical data that estimate the 5-year risk of developing diabetes. Because clinicians implement HbA_1c testing to screen for diabetes, these data may be used to help identify the risk of incident diabetes among individuals with HbA_1c levels <6.5%.

Funding for this work was provided by the VA Health Services Research and Development Service (Grants TEL-02-100 and IIR-04-045), the National Institutes of Health (Grant K24-DK063214), and Tosoh Bioscience.

No potential conflicts of interest relevant to this article were reported.

P.C. researched data, contributed to discussion, and wrote the manuscript. B.N. and P.F. researched data, contributed to discussion, and reviewed and edited the manuscript. P.R.C. researched data, contributed to discussion, and wrote the manuscript.