The Fallacy of Average: How Using HbA_1c Alone to Assess Glycemic Control Can Be Misleading

As expounded by Todd Rose in his book The End of Average (1), the mean of a measurement made among a large number of individuals is relevant for describing a population or group but often is not applicable for a given individual and can be misleading. Hemoglobin A_1c (HbA_1c) provides a good example of this. HbA_1c, which reflects blood glucose concentrations over 3–4 months, is a valuable metric for comparing treatment groups in a randomized trial, for assessing glycemic trends in a population over time, or for cross-sectional comparisons of glycemic control in different populations, and it is the only metric of glycemic control that has been strongly associated with chronic diabetic vascular complications. However, it has been debated whether, for an individual patient, the HbA_1c level is the best marker for complication risk or whether the level of glycemia with which the HbA_1c is associated is an equal or better marker of the risk of complications. Well recognized is the fact that HbA_1c may not accurately reflect glycemic control in the presence of a hemoglobinopathy, hemolytic anemia, or other conditions that affect red blood cell life span or interfere with glucose binding to hemoglobin. However, what is not widely appreciated is that even when no such diagnosed condition is present, HbA_1c may not be a good indicator of an individual’s glycemic control because of the wide range of mean glucose concentrations and glucose profiles that can be associated with a given HbA_1c level. It has been postulated that this mean glucose–HbA_1c discordance is due to interindividual variation in red blood cell life span (2,3).

This distinction in utilizing HbA_1c to compare groups versus its use in determining glycemic control for an individual was illustrated in a recent study we and others conducted assessing racial differences in the mean glucose–HbA_1c relationship (4). The study showed that on average HbA_1c levels in blacks are about 0.4% (4.4 mmol/mol) higher than those of whites for a given mean glucose concentration determined with continuous glucose monitoring (CGM). However, importantly the data also showed that the interindividual variation in HbA_1c for a given mean glucose concentration within race substantially exceeds the average degree of variation between races.

The wide range of mean glucose concentrations associated with a given HbA_1c level is not a new observation. It has been known since at least the 1990 publication of Yudkin et al. (5) and has been consistently demonstrated in numerous studies in individuals with prediabetes, type 1 diabetes, and type 2 diabetes (6–13), including the A1c-Derived Average Glucose (ADAG) study, which produced the widely used conversion table to estimate mean glucose for an HbA_1c level (14).

The ADAG study was conducted in 2006–2007, utilizing CGM, which was not as accurate as current generation CGMs, as well as blood glucose meter measurements to determine the mean glucose concentration. The analysis was conducted on a data set with a median of 13 days of CGM measurements plus 39 days of fingerstick blood glucose measurements. To assess the mean glucose–HbA_1c relationship with current CGM technology and a greater amount of data, we pooled data collected in 387 participants (age range 20–78 years, 83% white, 315 with type 1 diabetes and 72 with type 2 diabetes) in three randomized trials using the Dexcom G4 Platinum CGM System with an enhanced algorithm, software 505 (Dexcom, Inc., San Diego, CA) (4,15,16). Mean glucose concentration was determined for each participant using up to 13 weeks of CGM data (median amount of CGM data 66 days) and plotted versus HbA_1c measured following the collection of the CGM data at the Northwest Lipid Research Laboratories, University of Washington, Seattle, WA, using nonporous ion exchange high-performance chromatography (TOSOH, Biosciences, Inc., South San Francisco, CA).

As shown in Fig. 1, in the compiled data from the three studies, there is a wide range of mean glucose concentrations for a given HbA_1c level. For an HbA_1c of 8.0% (64 mmol/mol), the 95% prediction interval for mean glucose concentration is 155 to 218 mg/dL, substantially overlapping the CI for HbA_1c of 7.0% (53 mmol/mol) of 128 to 190 mg/dL and HbA_1c of 9.0% (75 mmol/mol) of 182 to 249 mg/dL. So, an HbA_1c of 8.0% (64 mmol/mol) could be associated with good, fair, or poor glycemic control as judged by potential mean glucose levels of 128 to 249 mg/dL. These results are quite similar to the results of the ADAG study (Table 1). Thus, estimating glycemic control by HbA_1c alone may not be accurate for some patients. As a result, utilizing HbA_1c alone to judge health care provider performance in treating patients with diabetes may be problematic.

The potential impact of mean glucose–HbA_1c discordance is illustrated in the post hoc analysis of the Action to Control Cardiovascular Risk in Diabetes (ACCORD) study data by Hempe et al. (17). ACCORD, a study of type 2 diabetes, had the unexpected finding of an increased mortality rate in the intensively treated group (N = 10,251, mean age 62 years, median HbA_1c 8.1%), which had an HbA_1c target of <6.0%. Hempe et al. (17) showed that a higher mortality rate was present only in the subjects in the intensive treatment group whose HbA_1c level was higher than the level predicted from fasting glucose concentration and that such subjects were more likely to have experienced severe hypoglycemia than those with an HbA_1c lower than predicted. Regardless of whether this finding is a possible explanation for the ACCORD mortality results, this analysis illustrates the problem and in some cases potential danger of determining a patient’s treatment regimen and glycemia goal from HbA_1c alone without knowledge of the patient’s mean glucose–HbA_1c relationship and glucose profile.

The best way to determine whether a given HbA_1c might be over- or underestimating a patient’s level of glycemic control is with CGM. CGM technology has advanced to where this can be done accurately and easily. For patients not already using CGM, a blinded CGM sensor could be worn once to compute a mean glucose concentration that can be compared with the patient’s HbA_1c. Ideally CGM data should be obtained for at least 14 days immediately preceding the measurement of HbA_1c during a period when diabetes treatment and glycemic control are reasonably stable (18). Several studies have demonstrated that an individual’s mean glucose–HbA_1c relationship tends to be reasonably constant over time (7,19–22). Although interval blinded CGM is useful for identifying patterns of glycemic control, a single blinded 14-day CGM wear to measure mean glucose concentration should be sufficient to estimate HbA_1c to determine how well the actual HbA_1c measurement estimates overall glycemic control for the patient. We recognize that this may not be realistic currently for all patients with diabetes, especially those with type 2 diabetes, but as sensor technology advances, that could become part of standard practice.

With knowledge of an individual’s mean glucose concentration, a CGM-estimated HbA_1c can be determined from the plot shown in Fig. 2 or by plugging the mean glucose concentration into the following formula: 3.38 + 0.02345 × [mean glucose] (23,24). Then, to inform how well an HbA_1c measurement estimates the mean glucose concentration for a patient, the estimated HbA_1c (eA1C) can be compared with the observed HbA_1c, which has been referred to as the hemoglobin glycation index (observed HbA_1c minus predicted HbA_1c) (9).

While potentially better than HbA_1c in understanding an individual patient’s glycemic control, mean glucose itself is an average, and different degrees of glycemic variability and many different glycemic patterns could produce similar mean glucose concentrations and similar HbA_1c levels. Figure 3 shows 2 weeks of CGM data (up to 288 sensor glucose measurements/day) displayed as a modal day or an ambulatory glucose profile (AGP) for four patients with type 1 diabetes using multiple daily injections of insulin. While each patient has a central laboratory–measured HbA_1c of 8%, the AGP glucose patterns vary greatly and would each lead to different clinical advice for lifestyle changes or insulin adjustments. This is where retrospective review of CGM data has considerable benefit. CGM profiles provide far more information than just the mean glucose concentration by identifying patterns of hyperglycemia and hypoglycemia as well as potentially dangerous high or low glucose concentrations that are often missed with self-monitoring of blood glucose. The results of secondary analyses of two major studies (the Examination of Cardiovascular Outcomes with Alogliptin versus Standard of Care [EXAMINE] trial and the Atherosclerosis Risk in Communities [ARIC] Study) (25,26) that found an association between hypoglycemia and cardiovascular events emphasize the importance of understanding a patient’s glucose profile with CGM to potentially identify patients who may be at high risk for these events. Thus CGM by providing more clinical insights than HbA_1c or self-monitoring of blood glucose measurements can help optimize and personalize glucose control and diabetes management (27).

We have written this Perspective to raise awareness of the need to know a patient’s actual mean glucose concentration, ideally by using CGM, if HbA_1c is to be used to assess a patient’s glycemic control and make diabetes management decisions. As long as HbA_1c is being used to define a glycemic target, we hope that eA1C becomes a standard metric used by clinicians and patients in assessing the level of glycemic control. Beyond that, a patient’s CGM glucose profile, or AGP, has considerable value for optimizing diabetes management. Estimating glycemic control from HbA_1c alone is in essence applying a population average to an individual, which can be misleading. In this era of personalized, precision medicine, there are few better examples than this one with respect to the fallacy of applying a population average to a specific patient rather than using specific information about the patient to determine the optimal approach to treatment.

Funding and Duality of Interest. Funding from the Leona M. and Harry B. Helmsley Charitable Trust supported in part the analyses performed and the writing of the manuscript and some components of AGP development and evaluation. The National Institute of Diabetes and Digestive and Kidney Diseases (DK108611) supported work to refine the AGP for clinical and research use. R.W.B.’s nonprofit employer has received consultant payments on his behalf from Animas, Insulet, and Tandem and research grants from Novo Nordisk, Boehringer Ingelheim, Takeda, Animas, Dexcom, Bigfoot, and Tandem with no direct personal compensation to R.W.B. R.M.B.’s nonprofit employer has received consultancy payments from Abbott Diabetes Care, Amylin, Bayer, Boehringer Ingelheim, Calibra, Eli Lilly, the Leona M. and Harry B. Helmsley Charitable Trust, Hygieia, Johnson & Johnson, Medtronic, Novo Nordisk, Roche, Sanofi, and Takeda and grants from Abbott Diabetes Care, Bayer, Becton Dickinson, Boehringer Ingelheim, Calibra, Dexcom, Eli Lilly, the Leona M. and Harry B. Helmsley Charitable Trust, Hygieia, Johnson & Johnson, Medtronic, Merck, National Institutes of Health, Novo Nordisk, Roche, Sanofi, and Takeda with no direct personal compensation to R.M.B. R.M.B.’s employer receives royalites from the Betty Crocker Diabetes Cookbook, and R.M.B. holds stock in Merck. D.M.M.’s nonprofit employer has received research grants from Abbott Diabetes Care and Dexcom for which there was no personal compensation. No other potential conflicts of interest relevant to this article were reported.