HbA1c shows its mettle in predicting diabetes risk

Create PDF

Dr. Selvin

The commonly held belief that one-quarter to one-third of diabetes cases are undiagnosed is a misconception in the literature, according to another new study published Oct. 24 in the Annals of Internal Medicine, led by Dr. Selvin (“Identifying trends in undiagnosed diabetes in U.S. adults by using a confirmatory definition: a cross-sectional study”; doi:10.7326/M17-1272).

“Almost exclusively, prior studies of the prevalence of undiagnosed diabetes have relied on looking at single elevated fasting glucose, elevated HbA_1c, or elevated two-hour glucose, and if you have an elevated fasting glucose and no other indication, they classify it as undiagnosed diabetes. This is not consistent with guidelines from our major diabetes organizations, which state that an elevated test result should always be confirmed with a second test.” The implication, Dr. Selvin points out, is that prior studies have overestimated the prevalence of undiagnosed diabetes.

The problem with prior epidemiologic studies is that they do not use definitions of undiagnosed diabetes using confirmatory testing. “A person with elevated glucose of 127 but normal HbA_1c would not be classified as having diabetes in clinical practice, yet our epidemiological studies say that person has undiagnosed diabetes.”

In reality, using definitions of undiagnosed diabetes that are more consistent with clinical practice, “we see that undiagnosed diabetes is fairly uncommon. Our study suggests that health care providers are doing a good job with screening and diagnosis.” While it’s true that diabetes is at epidemic proportions, Dr. Selvin adds, people who have diabetes are largely being identified in clinical practice.

Many of us have published a lot of papers on the question of whether HbA_1c is as good a predictor as glucose, and have examined the question of where does HbA_1c fit most usefully in clinical care,” says Dr. Meigs, co-director of the clinical effectiveness research group at Massachusetts General Hospital. “Quest was interested in collaborating on this because they have an HbA_1c assay and were interested in a paper that would get specifically at the information utility of HbA_1c in various health settings.”

In fact, Quest service centers are one place where people might go to get a blood test and have only an HbA_1c performed; that’s the first of the four scenarios the authors considered (HbA_1c only). “Another scenario is ‘come see me in my office today,’ where you would take a medical history, draw a spontaneous, nonfasting blood test, including an HbA_1c level. The clinic visit is a more information-rich environment than a clinical lab service center,” Dr. Meigs says.

The authors decided that since the question was big enough, “we would use two different, representative population studies where there are lots of cases of diabetes and lots of measurements of HbA_1c and careful follow-up.” Framingham has the advantage of including middle-aged participants, right around the age where people develop diabetes, “but everyone in it is white,” Dr. Meigs says. “ARIC, although the visits are a little less frequent, has a black population, younger people, and lots of long follow-up. And they’re generally comparable studies, so we were able to use them together to combine some of the models to get a robust estimate of the marginal information value of HbA_1c using all of the test results.”

Dr. Meigs

One of the most important findings is that in all four scenarios studied, HbA_1c improved prediction or improved discrimination, he says. “Discrimination is expressed in the so-called c-statistic, or concordance statistic, which is a probability that if you have two people in front of you, you can guess which one has higher risk. If the probability is 70 or 80 percent, which is the kind of value we were getting in our models and our papers, then that is clinically useful. Then the question we asked—if you add HbA_1c, are you even better at assessing risk?”

For comparison, Dr. Meigs notes, the criterion standard in the field is the Framingham heart attack prediction model, which takes age, sex, smoking, diabetes, cholesterol levels, and blood pressure and returns the probability of having a heart attack. The discrimination of that model is about 75 percent.

It’s hard to improve a c-statistic, he says. “You need a marginal information that is high—higher than most tests we have. HbA_1c does improve the c-statistic just a little bit, but it is significant because it is a strong diabetes risk factor.” The meaning of adding a test like HbA_1c to the mix has to be contextualized in terms of how bad for health the condition is that you’re looking for, he points out. “Type 2 diabetes is increasing in frequency and leads to very poor health outcomes, so we think that anything that demonstrably improves our ability to find people at risk is a good thing.” This is one reason the study’s conclusion is that HbA_1c has definite clinical uses in a range of scenarios.

The Diabetes Care study includes additional useful information for those exploring different clinical scenarios. “We published online a huge supplement to this paper giving the parameters of every single model we ran” for providers who want to know what model to use to detect diabetes in their settings. “You’d need a computer to do this. But you can take the parameters and use the regression equations we’ve published to estimate for a person what their risk would be.” Judging from his past published papers, Dr. Meigs explains, “People are interested in programming the values in the models into their own system, for prediction or to compare their model outputs with ours.”

The two cohorts of the Framingham and ARIC studies, with their extended follow-up, were important to the study. “Diabetes is a disease that takes a while to present itself, and if you want to really understand its prediction,” Dr. Meigs says, “having 20 years of follow-up will show the long-term predictiveness of a test. But we also looked at the short term. If you’re an 18- or 20-year-old person, you might have one encounter with the health care system, then not show up again for five or 10 years. We wanted to know if you measured HbA_1c once, even a long time ago, does it still predict? HbA_1c does predict; it provides a very good picture of diabetes risk over the short term and the long-term future.”

One reason is that HbA_1c is an excellent biomarker. “The test is very biologically stable.”

For that reason it can serve different purposes. “HbA_1c can be used for screening people to diagnose diabetes and get them into treatment. In treatment, it can reveal your average blood sugar so that today we can decide whether we want to change treatment or change prevention strategy. We can also use it to ask, if we measure this now, what are the chances you’ll get the disease in the future on the basis of the test, and can that information be used to reduce the chances of getting diabetes?” LDL cholesterol testing for heart attack is used in similar ways, he notes.

Dr. Meigs is cautious in predicting the impact of a study like this one on clinical care. “We’re modeling what people are actually doing, where such data have been hard to come by. Certainly a common practice is that people use the test as part of routine screening annually for people they are worried about as being at risk of developing diabetes. We studied particular scenarios reflecting what people are already doing. If we’d found that in certain areas the test really wasn’t useful, then we would have been arguing for a change in practice. I think what we wind up saying with the Diabetes Care paper is, however you end up using it, HbA_1c is a useful test, and here’s the range of its utility in terms of prediction and discrimination.”

But the Diabetes Care study could help in establishing thresholds for prediabetes, Dr. Meigs says. “The thresholds for establishing the different states of diabetes and prediabetes are arbitrary and decided by expert committees. They draw a line across what are often normal distributions, and say above and below some value is good or bad for health, and people still argue about those values.”

“In all fairness, they are reasonable values. They’re based on evidence of complications of diabetes increasing at certain values. There’s no hard threshold in any of this risk, and for some complications there’s an inflection point where rates start to be higher than is acceptable. What we’re showing here in this paper is that, even within the ranges people are saying are safe, there is still some risk. Within the range of prediabetes, there is gradation of risk as well, so if your HbA_1c is starting to get closer to the diagnostic threshold, the risk is very high. We’re saying that if you find a person with a value that is close to the threshold, that person is very likely to progress over the line, compared to a person with slightly lower values.”

The study will be useful, he predicts, when people inevitably argue about what the right thresholds are for screening and treating. “We’ve not just published the equations, but we’ve also split up the distributions into some finer gradations and are able to show in some of the paper figures this greater increased risk.” In addition, the study identifies how many people could be expected to fall into these risk groups, “so if a health plan is deciding which people it should focus on when doing screening, we give some information to guide that decision.”

As possible follow-up research to the study, Dr. Meigs says, a next step “is getting bigger cohorts and studies that are not white. ARIC does offer a good population of African Americans in the U.S. But the chronic problem in health epidemiology and population science is the absence of minority subjects in studies, so a more representative sample of African Americans is needed, and then further, Latinos, Asian Americans, East Asians, South Asians, and so on.”

A paper by Aaron Leong, MD (of Massachusetts General Hospital and first author of the Diabetes Care study), and Drs. Meigs and Selvin and others, published Sept. 12, reveals the danger of ignoring possible genetic differences in interpreting HbA_1c results (Wheeler E, et al. PLoS Med. 2017;14[9]:e1002383).

“People with a certain mutation, carried by about 11 percent of people of African ancestry in America, have HbA_1c values on average that are almost one unit lower than people without it. If you screen them using the approaches we talk about in our community-based cohort papers, you’re going to miss them.” In the Diabetes Care study, Dr. Meigs says,“we don’t get into this threshold issue. We just say the models work equally well. And that’s fine for a population, but when you’re looking at an individual patient, it might matter if they have a genetic mutation.” How to bring this complication into practice is a problem with which the field is still struggling.
[hr]

Anne Paxton is a writer and attorney in Seattle.