In the comment from Treviño (1), three articles are cited as having “emphasized the need for real-time assessment of blood glucose levels.” Two of the cited articles (2,3) indeed relate to real-time assessment of blood glucose levels using new technology referred to as continuous glucose monitoring (CGM). The third article Treviño refers to, by Farmer et al. (4), relates to the use of traditional finger-stick blood analysis that is communicated to a nurse via telemedicine. We are engaged in the study of CGM, so further comments will be related only to continuous sensors (2,3) and not to telemedicine.
Treviño further states that “[r]eal-time analysis is motivated by the belief that the random time series being analyzed manifests average properties that change with the passage of time. Results of such analyses are presented to patients as means (μ) ± SD.” In regard to our study (2), this statement is completely false. We do not rely on averaged properties of time series data. Patients received actual glucose concentration values transmitted every 5 min. Under no circumstances were patients presented with the mean and SD of the data in real time.
Treviño further describes an equation for determining mean and SD and discusses the reliability of such determinations on time series data. He states that “The unfortunate feature of [the equation] is that the μ and SD values generated are valid only when the G(n) is independent [(5)]”. Regardless of whether this assertion is correct, the prespecified primary accuracy analysis of Garg et al. (2) was to assess sensor bias using Deming regression. Secondary assessments of efficacy were based on a Clarke error grid analysis and Pearson correlation coefficients. Although the mean absolute relative difference was also reported in a secondary analysis, no statistical inferences were made using sample means of repeated measurements. Dr. Treviño's assertion that sample means are valid estimates of central tendency only when the observations are independent is simply not true. To back up this assertion, he cites another letter of criticism (5) to a journal. We have reviewed this (5) and can find no justification for this claim. In fact, sample mean is a valid measure of central tendency regardless of whether the observations are independent or dependent. Treviño appears to miss the context when he quotes Liang and Zeger (6). The quote is referring to the analysis of data using generalized linear models, in which case the correlation among values for a given subject must be taken into account to provide valid statistical inference procedures for regression parameters. Thus, Liang and Zeger do not state that sample means and SDs are only valid when the observations are independent. In fact, one of the methods proposed by Liang and Zeger is to estimate the regression parameters under the independence assumption and then to appropriately estimate the variances of estimated regression parameters to take the correlation between repeated measurements into account. Using these methods (bootstrapping), the underlying data from his first reference were submitted to the Federal Drug Administration, and the product was approved in March 2006. There were no differences in conclusions from this analysis compared with the first reference, except that the confidence intervals were larger but still significant.
Treviño's letter concludes with some rather broad, far-reaching assertions. He says, “In short, averages of blood glucose generated by [the equation] produce unreliable values. This contradicts the conclusion of Garg et al. ‘that [traditional] real-time continuous glucose monitoring for periods up to 72 h is accurate.’ It also conflicts with findings of Deiss et al. (3) ‘that the use of [traditional] real-time CGM has considerable potential for the management of patients with diabetes.’”
It is most striking that he inserted the bracketed word “traditional” into the quotes. It is appropriate to add bracketed words or clauses within a quote only to mark an explanatory insertion within a quote (7). In this case, the inclusion of “[traditional]” completely changes the meaning of the original quotes. In both cases, the quotes were meant to refer to the new glucose monitoring technology of CGM, not to “traditional” glucose monitoring (finger-sticks). There is nothing traditional about real-time CGM at all. It is brand new technology, just having made it to the marketplace in the last year.
The titles of the two cited articles refer to “Improvements in glycemic excursions…” (2) and “Improved glycemic control…” (3). The point of these articles was to show that when patients use continuous sensing, they improve glycemic profiles (time spent in various glycemic zones) (2) and improve their A1C values (3). These conclusions are reached by analyzing the time spent by patients in various hypoglycemic zones (2) and by analysis of A1C levels at various times (3)— not through the use of averaging of blood glucose levels.
Finally, Treviño draws on two articles that are not related in any way to continuous glucose sensing: “The unreliability of traditional analyses of blood glucose data is perhaps one of the reasons that ‘overtreating hypoglycemia has resulted in a marginally significant increase in the frequency of hyperglycemic excursions’ [4] and that ‘many patients still experience episodes of unrecognized hypo- and hyperglycemia [3].’”
We agree with both of these assertions. Both assertions relate to problems using truly “traditional” methods of glucose monitoring as described in refs. 3 and 4 (finger-sticks) and not to CGM. We provided evidence (2) that the use of CGM leads to a decrease in hyperglycemia, not an increase, and that patients experienced fewer episodes of unrecognized hypo- and hyperglycemia. The fact that, using traditional finger-stick methods, patients “overtreating hypoglycemia” suffer “unrecognized hypo- and hyperglycemia” (1) is precisely because intermittent monitoring with finger-sticks does not give them enough information. In contrast, continuous sensing reduces hyper- and hypoglycemia (2) and improves A1C levels in patients (3). In fact, more recently we reported improvement in glycemic excursions across all A1C values in subjects during the display phase of the study (8). In addition, continuous sensing highlights different contributions of fasting and postprandial glucose values at different A1C levels (9) in contrast to finger-stick intermittent glucose measurements (9).
