Nathan et al. (1) present a regression equation that they propose be used to convert measured values of A1C into estimated average glucose (eAG) values, in accordance with recommendations (2) and subject to the condition that their study fulfill its a priori–specified criterion. The criterion declared is that for ≥90% of patients, average glucose values, calculated from continuous glucose monitoring and 7-point glucose measurements over a 3-month span, fall within 15% of the average glucose using the study-wide average glucose as opposed to the A1C regression equation. This criterion was indeed fulfilled (with surprising narrowness: 89.95% of subjects fell within the ±15% limits) (1). Using the proposed equation thus involves an error limit in eAG of >15% in approximately 10% of cases. Yet, eAG is intended as a reference for the interpretation of glucose measurements, for which the internationally accepted maximum error limit is ±10% (3) or, if biological variation is taken into account, ±6.9% (4).
The SD of the prediction error of the Nathan equation is 15.7 mg/dl, which is equivalent to a random error of ∼10%. The accepted maximum tolerable random error, taking biological variation into account, is 2.9% (4). Using the medically tolerable total error limit of ±6.9% in the Bland-Altman analysis, instead of an arbitrary ± 1.96 SD, would make the proportion of patients falling within limits much smaller than 90%.
In keeping with the heteroscedastic model of Nathan et al., the 95% CI for eAG grows linearly: 44 to 107 mg/dl between A1C values of 5 and 12%. For an A1C of 7%, eAG is 154 mg/dl and its 95% CI 123–185 mg/dl, the upper end of which corresponds with an A1C level above the threshold for patient reevaluation (8%); for an A1C of 8%, the lower 95% CI corresponds with an A1C of 6.7% and the upper limit with an A1C of 9.2%.
Nathan et al. report that there was no significant difference between the regression equation calculated using only continuous glucose monitoring data and the equation calculated using only data from 7-point fingerstick tests—a method similar to that employed in the Diabetes Control and Complications Trial (DCCT). That the scatter of the data around the regression line in the study by Nathan et al. was smaller than that in the DCCT is attributed to the higher frequency with which 7-point profiles were obtained. Yet equally plausible explanations are that the study by Nathan et al. included about three times fewer subjects and that not all of these subjects were type 1 diabetic patients—as opposed to entirely type 1 diabetic population of the DCCT. Furthermore, all of the diabetic patients were required to satisfy a condition providing a degree of assurance of the stability of their A1C values.
In view of the above, it seems doubtful whether eAG values are clinically useful: the potential error for individual patients is analytically and clinically unacceptable. The relationship between A1C and diabetes complications is well established. Equally clearly, increased average glucose leads to increased A1C (as is reflected by the study by Nathan et al., among others). The assumption underlying the proposal to express A1C as eAG—that A1C depends exclusively on average plasma glucose—is not justified (5).
Article Information
No potential conflicts of interest relevant to this article were reported.
References
Readers may use this article as long as the work is properly cited, the use is educational and not for profit, and the work is not altered. See http://creativecommons.org/licenses/by-nc-nd/3.0/ for details.
