Glycated hemoglobin (A1c) has been referred to as the gold standard for glycemic control. At the American Diabetes Association (ADA) 2007 Scientific Sessions -- June 22-26, 2007; Chicago, Illinois -- it was reported by Dr. David Nathan (Harvard University Medical School, Boston, Massachusetts) that using hemoglobin A1c to assess glycemic control will be replaced by average blood glucose. Justification for the change is that it "will add clarity for diabetic patients looking to manage their disease." Specifically, Dr. Nathan reported that "patients have a glucose problem, not a hemoglobin problem" and that "reporting glycohemoglobin results as an A1c-derived average glucose would have the advantage of reporting chronic glycemia in the same units as the patients' self-monitoring of daily glycemia"; see a related article by Michael O'Riordan for further elaboration.
The problem is classic in the analysis of random time series. Specifically, given that there is an accepted mathematical relationship between mean plasma glucose (MPG) and A1c, the intent is to measure the A1c of a patient at some time and from that measurement assign a glucose average over, say, the previous 90-120 days. The logic is that we can replace, with impunity, one set of reference values, namely, those used to define the "standard" relationship, with a second set, namely, those from the patient's own history.
One problem with this strategy is that the accepted mathematical relationship between MPG and A1c was derived doing interindividual analysis, ie, by simultaneously measuring both 'true' levels of plasma glucose (not MPG) and A1c in randomly selected different individuals and determining from these measurements a representative regression line defining MPG = f(A1c). Data values used in interindividual analysis are therefore independent; therefore, standard methods for computing mean and variance are applicable. A glucose average for one individual, however, is determined by repeatedly measuring the level of glucose over time, and then determining the arithmetic average of the data. This type of analysis is denoted intraindividual analysis. Intraindividual data are strongly correlated, thus invalidating the applicability of "garden-variety" statistical methods. The mean for one individual is ideally determined at a given point in time from a collection (ensemble) of time histories, ie, both the single observed series and (the many) essentially equivalent unobserved series (the unobserved ones sometimes called alternative sample paths). For this reason, mean and arithmetic average are not the same. To proceed viably, intraindividual analysis requires incorporating sensible assumptions about the correlation between successive data values.
For example, suppose that 3 different patients have A1c values of 5.6, 7.8, and 9.2. From the "standard" relationship between glucose and A1c, these values -- respectively -- correspond to MPG values of 108, 174, and 216. Because these values are determined from data that are independent, there's no direct analogy between them and the arithmetic average glucose of the patients in question. "Independent" means that the magnitude of any particular glucose value is not influenced by a previous value. Such independence is untrue for successive (time series) glucose values of any particular patient.
A second problem is the labile nature of the mean of a patient's glucose history, a feature not uncommon in diabetics. "Lability" in the mean invalidates the applicability of standard statistical methods to intraindividual analysis. A simple arithmetic average over time cannot replicate lability, and the related implications for variance computation and regression analysis are unacceptable; the attendant polysemy is overwhelming. The issue here is not about mean and average glucose levels, nor is it about the difficulties associated with A1c measurement (cf paragraph 1). It's about the benefits of reporting glycemic control to patients in terms of meaningful individualized average magnitudes (not just meaningful glucose units) rather than "generic" multipatient A1c-derived magnitudes.
The bottom line is that, in quantifying standardized relationships, the data used in the assessment can be measured and recorded in any order without contaminating the sought-after equation. In the assessment of an arithmetic average the measured values can be recorded only in the order defined by the patient. This patient-defined order engenders from differences in metabolism caused by diet, exercise, drugs, and other factors. These factors are unique to each individual, and 2 distinct individuals can have the same average glucose yet have different correlation structures (and variances). Thus, the 2 subject data sets are not interchangeable. In short, averages of uncorrelated data are not transformable into averages of correlated data, with or without A1c as the go-between. Thus, A1c should remain the gold standard; instead of abandoning it, we should refine it to incorporate "sensible assumptions about the correlation between successive data values."
Reader Comments on: Should Diabetes Healthcare Abandon Its Gold Standard?
See reader comments on this article and provide your own.
Readers are encouraged to respond to the author at email@example.com or to Paul Blumenthal, MD, Deputy Editor of MedGenMed, for the editor's eyes only or for possible publication as an actual Letter in MedGenMed via email: firstname.lastname@example.org
Cite this: Should Diabetes Healthcare Abandon Its Gold Standard? - Medscape - Nov 01, 2007.