Area-under-the-A1c-Curve above the Normal Range and the Prediction of Microvascular Outcomes: An analysis of data from the Diabetes Control and Complications Trial

Abstract

Aims

In the Diabetes Control and Complications Trial (DCCT), mean updated A1c accounted for most of the differential risk of microvascular complications between intensive and conventional insulin therapy. We hypothesized, however, that a more precise measure of chronic hyperglycemic exposure may be the incremental area-under-the-A1c-curve above the DCCT-standardized normal range for A1c (iAUC_A1c>norm).

Methods

Using the Principal DCCT Dataset, we compared the following 3 measures of chronic glycemic exposure for their capacity to predict retinopathy, nephropathy, and neuropathy during the DCCT: mean updated A1c, iAUC_A1c>norm, and total area-under-the-A1c-curve (tAUC_A1c). For each outcome, models using each of these 3 glycemic measures were compared in the following 3 ways: hazard or odds ratio, chi-square statistic, and Akaike information criterion.

Results

The 3 glycemic measures did not differ in their prediction of neuropathy. iAUC_A1c>norm was modestly superior to mean updated A1c for predicting nephropathy (chi-square p=0.017, Akaike p=0.032). In contrast, for predicting retinopathy, both iAUC_A1c>norm (chi-square p=0.0005, Akaike p=0.0005) and tAUC_A1c (chi-square p=0.004, Akaike p=0.0004) were significantly better than mean updated A1c. Varying its A1c threshold incrementally between 37 and 53 mmol/mol (5.5% - 7.0%) inclusive did not improve the prediction of retinopathy by iAUC_{A1c>threshold} beyond that of tAUC_A1c, consistent with the concept of a continuous relationship between glycemia and retinopathy with no glycemic threshold.

Conclusions

Both iAUC_A1c>norm and tAUC_A1c were superior to mean updated A1c for predicting retinopathy. Optimal assessment of chronic glycemic exposure as a determinant of retinopathic risk may require consideration of both the degree of hyperglycemia and its duration.

Introduction

The Diabetes Control and Complications Trial (DCCT) study group reported that intensive insulin therapy can reduce the development and progression of retinopathy, nephropathy and neuropathy in type 1 diabetes (T1DM), with glycemic exposure (as determined by mean updated A1c) explaining virtually all of the difference in risk of these complications between the intensive and conventional therapy groups [1-3]. These data highlight the central importance of chronic exposure to hyperglycemia as a determinant of microvascular complications in patients with T1DM. Though assessed by the mean updated A1c in the DCCT, we hypothesized that a more precise measure of chronic hyperglycemic exposure would be the incremental area-under-the-A1c-curve above the DCCT-standardized normal range for A1c (iAUC_A1c>norm). Indeed, this measure provides a way of integrating both the degree of glycemia and the time spent above the normoglycemic range, and hence could capture the cumulative exposure to hyperglycemia. As such, by reflecting the presumed key biologic risk determinant relevant to the development of microvascular complications, it may be a stronger predictor of these outcomes than mean updated A1c. Thus, we sought to evaluate the predictive capacity of iAUC_A1c>norm as a determinant of microvascular complications in the DCCT, as compared to mean updated A1c.

Methods

The DCCT has been previously described in detail [4,5]. In brief, 1441 participants with T1DM were randomized to intensive or conventional insulin therapy and followed for mean 6.5 years, with quarterly measurement of A1c and regular assessment for microvascular complications. Using the publicly-available Principal DCCT dataset (National Technical Information Service, Alexandria,VA), we evaluated the following 3 measures of chronic glycemic exposure: mean updated A1c, mean iAUC_A1c>norm, and mean total area-under-the-A1c-curve (tAUC_A1c).

These measures were assessed because, while mean updated A1c will reflect the actual mean value for A1c over follow-up, it does not account for duration of time spent above a particular glycemic threshold (such as the upper-limit-of-normal for the DCCT-standardized A1c assay (IFCC 42 mmol/mol or DCCT 6.0%)). In contrast, iAUC_A1c>norm provides a measure of glycemic exposure that incorporates not only the degree of glycemia but also the time spent above the normal range. To illustrate this difference, consider the hypothetical example of 2 patients with the following quarterly A1c measurements: Patient A has measurements of 42 mmol/mol (6.0%), 42 mmol/mol, 42 mmol/mol, and 42 mmol/mol, while Patient B has measurements of 42 mmol/mol, 46 mmol/mol (6.4%), 38 mmol/mol (5.6%) and 42 mmol/mol. Mean updated A1c will be the same for both patients, but only iAUC_A1c>norm will capture the fact that Patient B spent time with A1c over the normal range. Finally, we also assessed tAUC_A1c since it is similar to iAUC_A1c>norm but does not assume that the upper-limit-of-normal for the DCCT-standardized assay (IFCC 42 mmol/mol or DCCT 6.0%) is an absolute threshold below which glycemic exposure does not contribute to microvascular risk.

Mean updated A1c was calculated as the mean of all quarterly A1c measures up to the visit under study, computed at 6-monthly intervals (3). Mean AUC_A1c was calculated by trapezoidal rule as a mean of the number of visits under study. For example, AUC_A1c at year 3 (12 visits) was calculated as: 0.5*[H₀+(2*H₁)+(2*H₂)+(2*H₃)+(2*H₄)+(2*H₅)+(2*H₆)+(2*H₇)+(2*H₈)+(2*H₉)+(2*H₁₀)+(2*H₁₁)+H₁₂]/12, where H₀ is baseline A1c, H₁ is first-quarterly visit A1c, H₂ is second-quarterly visit A1c, etc. Calculations for each 6-monthly or annual interval included all preceding quarterly A1c measurements since the start of the DCCT.

For each measure of glycemic exposure, we determined its predictive capacity for incidence of the following DCCT-defined [3] microvascular outcomes: persistent 3-step change in retinopathy, microalbuminuria, and neuropathy. Retinopathy and nephropathy were analyzed by time-dependent proportional hazards models and neuropathy by logistic regression model. For each outcome, models using each of the 3 A1c measures were compared in 3 ways: hazard or odds ratio, chi-square statistic, and Akaike information criterion (Table 1). For both the chi-square and Akaike information criterion statistics, pairwise comparisons were conducted using bootstrap methods (2,000 bootstrap samples were drawn from the original data set with replacement). The pairwise differences of the estimated chi-square and Akaike information criterion statistics were obtained for each bootstrap sample. The bootstrap method made it possible to estimate the sampling distribution of the pairwise difference of the chi-square or Akaike information criterion statistic. Based on this nonparametric method and the observed pairwise differences, we calculated bootstrap p-values to control type I error and determine the significance of pairwise differences. The bootstrap p-values were estimated by the proportion of those bootstrapped chi-square or Akaike values at greater than or equal to the observed pairwise difference from the original data.

Table 1.

Comparative performance of each measure of glycemic exposure for the prediction of microvascular outcomes.

	Mean Updated A1c	tAUCA1c	iAUCA1c>norm	p values for pairwise comparisonsd
				Mean Updated A1c versus tAUCA1c	Mean Updated A1c versus iAUCA1c>norm	tAUCA1c versus iAUCA1c>norm
Retinopathya (n=1440)
HR (95% CI)	1.64 (1.52, 1.77)	1.69 (1.57, 1.83)	1.68 (1.56, 1.82)	n/a	n/a	n/a
Chi-squareb	161.65	174.54	171.71	0.004	0.0005	0.170
Akaike Information Criterionc	2847.83	2837.39	2839.22	0.004	0.0005	0.264
Nephropathya (n=1365)
HR (95% CI)	1.31 (1.22, 1.41)	1.33 (1.23, 1.43)	1.33 (1.23, 1.44)	n/a	n/a	n/a
Chi-squareb	50.18	52.18	53.21	0.068	0.017	0.039
Akaike Information Criterionc	3517.83	3516.15	3515.45	0.076	0.032	0.077
Neuropathya (n=1161)
OR (95% CI)	1.53 (1.34, 1.74)	1.53 (1.35, 1.75)	1.54 (1.35, 1.75)	n/a	n/a	n/a
Chi-squareb	41.91	42.05	42.17	0.368	0.328	0.256
Akaike Information Criterionc	662.35	662.26	662.31	0.412	0.474	0.424

Abbreviations: HR, Hazard Ratio; OR, Odds Ratio; CI, confidence interval; n/a, not applicable

^aRetinopathy was defined as persistent 3-step change from baseline, graded according to the Early Treatment Diabetic Retinopathy Study scale. Nephropathy was defined as urine albumin excretion rate (AER) ≥40mg/24hr. Neuropathy was defined on clinical examination at 5 years with concurrent/prior abnormalities in nerve conduction or autonomic nervous system testing. Retinopathy and nephropathy were analyzed by time-dependent proportional hazards models since they were assessed regularly during the DCCT. Neuropathy was analyzed by logistic regression model because it was assessed only at 5 years. Patients with urine AER >40mg/24h at baseline were excluded from nephropathy models and those with neuropathy at baseline were excluded from neuropathy models.

^bA larger chi-square statistic indicates stronger evidence of an increasing trend over time in the effect of a measure of glycemic exposure on the hazard or risk of an outcome.

^cThe Akaike information criterion offers a measure of the information lost when a statistical model is used to describe reality. Thus, in this study, given a set of models using three measures of glycemic exposure for each outcome, the preferred model is the one with the minimum Akaike information criterion value, as least information is lost.

^dThe pairwise comparisons were conducted using bootstrap methods (2,000 bootstrap samples were drawn from the original data set with replacement). The pairwise differences of the estimated chi-square and Akaike information criterion statistics were obtained for each bootstrap sample. The bootstrap p-values were estimated by the proportion of those bootstrapped chi-square or Akaike values at greater than or equal to the observed pairwise difference from the original data.

Models for the prediction of retinopathy were repeated at different A1c thresholds between 37–53 mmol/mol (5.5-7.0%) for comparison of iAUC_{A1c>threshold} versus tAUC_A1c (Table 2). The purpose of this analysis was to see if there was a particular A1c threshold where iAUC_{A1c>threshold} was superior to tAUC_A1c for predicting retinopathy by both Chi-square and Akaike information criterion. All statistical analyses were performed using Stata10.0 (College Station,TX) and SAS9.2 (Cary,NC).

Table 2.

Comparison of iAUC_{A1c>threshold} and tAUC_A1c for the prediction of the risk of retinopathy, for sequential incremental A1c thresholds (T) from 37 to 53 mmol/mol (5.5 to 7.0%) inclusive.

A1c Threshold (T)		HR	Chi-square test		Akaike Information Criterion
IFCC(mmol/mol)	DCCT(%)		Value	p value for iAUCA1c>T vs tAUCA1c	Value	p value for iAUCA1c>T vs tAUCA1c
37	5.5	1.682	171.50	0.166	2838.74	0.316
38	5.6	1.682	171.55	0.169	2838.77	0.313
39	5.7	1.682	171.58	0.171	2838.85	0.293
40	5.8	1.683	171.64	0.175	2838.93	0.294
41	5.9	1.683	171.62	0.163	2839.05	0.278
43	6.1	1.684	171.81	0.180	2839.44	0.240
44	6.2	1.686	171.90	0.190	2839.74	0.205
45	6.3	1.687	171.95	0.200	2840.15	0.158
46	6.4	1.689	171.98	0.204	2840.70	0.109
48	6.5	1.690	171.94	0.202	2841.36	0.078
49	6.6	1.693	171.96	0.211	2842.08	0.047
50	6.7	1.696	172.05	0.225	2842.86	0.024
51	6.8	1.700	172.06	0.237	2843.76	0.014
52	6.9	1.705	172.29	0.264	2844.59	0.008
53	7.0	1.710	172.40	0.281	2845.56	0.005

Results

All 3 A1c measures (mean updated A1c, tAUC_A1c, iAUC_A1c>norm) were significantly associated with each microvascular outcome (Table 1). However, their comparative predictive capacities differed for each complication. While the 3 measures did not differ in their prediction of neuropathy, iAUC_A1c>norm (but not tAUC_A1c) was superior to mean updated A1c for prediction of nephropathy (chi-square bootstrap p=0.017, Akaike bootstrap p=0.032). In contrast, both iAUC_A1c>norm (chi-square bootstrap p=0.0005, Akaike bootstrap p=0.0005) and tAUC_A1c (chi-square bootstrap p=0.004, Akaike bootstrap p=0.0004) were superior to mean updated A1c for the prediction of retinopathy, with no significant difference between iAUC_A1c>norm and tAUC_A1c. Varying its A1c threshold above or below 42 mmol/mol (6.0%) from 37–53 mmol/mol (5.5-7.0%) did not improve the prediction of retinopathy by iAUC_{A1c>threshold} beyond that of tAUC_A1c (Table 2). Specifically, the chi-square statistic did not reveal a significant difference between these two measures at any threshold, although the Akaike information criterion statistic found tAUC_A1c to be superior to iAUC_{A1c>threshold} at A1c thresholds >48 mmol/mol (6.5%). Lastly, after adjustment for covariates (baseline A1c, age, gender, BMI, duration of diabetes, treatment group, and baseline value of the outcome variable), both iAUC_A1c>norm and tAUC_A1c remained superior to mean updated A1c for the prediction of retinopathy but not nephropathy or neuropathy (Online Table).

Discussion

In this analysis, all 3 A1c indices yielded comparable hazard or odds ratios for the prediction of each microvascular outcome and the overall benefits of the AUC_A1c measures over mean updated A1c were modest. Furthermore, it is recognized that neither the AUC_A1c indices nor mean updated A1c are practical measures for calculation in clinical practice. Instead, however, these indices are more relevant as research tools in clinical studies. Indeed, as shown in this analysis, AUC_A1c measures can provide insight on the impact of glycemic exposure on risk of microvascular complications, and potentially the biology underlying these relationships. In this context, three key findings emerge from the current data.

First, the demonstration that both iAUC_A1c>norm and tAUC_A1c were superior to mean updated A1c for the prediction of retinopathy supports the concept that optimal assessment of chronic glycemic exposure as a determinant of microvascular risk may require consideration of both the degree of hyperglycemia and its duration. Interestingly, Orchard and colleagues previously explored this general concept with their construct of A₁months, defined as the product of A1c above the normal range and months exposure as determined from biennial assessment in the Pittsburgh Epidemiology of Diabetes Complications study [6]. While they found that this measure was not a better predictor of microvascular complications than its components (duration and mean A1c), it should be noted that its calculation from biennial A1c assessments would have rendered A₁months a less precise measure of chronic glycemic exposure than the AUC_A1c indices obtained from the more frequent A1c measurements in the DCCT. Thus, the AUC_A1c measures in the current study may have been better able to demonstrate the predictive capacity of a cumulative glycemic exposure variable that incorporates both duration and degree of glycemia.

Secondly, it is noted that iAUC_A1c>norm and tAUC_A1c were both consistently superior to mean updated A1c for the prediction of retinopathy, but not nephropathy and neuropathy. This difference may relate to methodologic features (eg. neuropathy was assessed only at 5 years while retinopathy was assessed every 6 months) or limitations in the measurement of nephropathy and neuropathy outcomes, but also may reflect a comparatively greater influence of glycemic exposure on risk of retinopathy, as compared to the other outcomes. Thirdly, we did not detect an A1c level above which iAUC_A1c surpasses tAUC_A1c for predicting retinopathy, consistent with the concept of a continuous relationship between glycemia and retinopathy, with no glycemic threshold [3].

Finally, it should be noted that, although non-glycemic factors may also be relevant [7], the unmeasured effect of glycemic exposure in the years prior to the DCCT may have limited the achievable predictive capacity of all three A1c-based measures obtained during the trial. Thus, iAUC_A1c>norm and tAUC_A1c from the time of diagnosis ultimately warrant study for the evaluation of total cumulative glycemic exposure as a determinant of microvascular risk in T1DM.

List of Abbreviations

AER

albumin excretion rate

DCCT

Diabetes Control and Complications Trial

iAUC_A1c

incremental area-under-the-A1c-curve

iAUC_A1c>norm

incremental area-under-the-A1c-curve above the normal range for the DCCT-standardized A1c assay

tAUC_A1c

total area-under-the-A1c-curve