Sensitivity of Traditional and Risk-Based Glycemic Variability Measures to the Effect of Glucose-Lowering Treatment in Type 2 Diabetes Mellitus

Abstract

Background:

Here we assess associations between glycemic variability (GV) measures and outcomes from glucose-lowering therapy in patients with type 2 diabetes (T2DM) to identify the metrics most sensitive to treatment response.

Methods:

Data from 1699 patients in 6 previously reported studies in adults with T2DM treated with basal insulin and/or oral glucose-lowering drugs were included in a post hoc meta-analysis. Using 7-point blood glucose (BG) profiles we compared the GV metrics standard deviation (SD), mean amplitude of glycemic excursion (MAGE), mean absolute glucose (MAG), low and high BG risk indices (LBGI, HBGI), and average daily risk range (ADRR). Treatment-related changes in GV and risk status and associations between end-of-trial GV/risk metrics with treatment outcomes (end-of-trial glycated hemoglobin A1c[A1C] level ≥7.0%, hypoglycemia, and composite outcome of A1C <7.0% and no hypoglycemia), were evaluated.

Results:

Significant changes from baseline to end of treatment were observed in all measures (all P < .0001), with the largest reduction following treatment for HBGI (–65.5%) and ADRR (–43.3%). The baseline risk classification for hyperglycemia based on the risk categories of HBGI improved for 66.8%, remained unchanged for 29.8%, and deteriorated for 3.3% of patients (chi-square P < .0001), while the risk for hypoglycemia did not change. HBGI showed the strongest association with A1C ≥7.0% at the end of treatment, and LBGI showed the strongest association with symptomatic hypoglycemia.

Conclusions:

During glucose-lowering therapy in T2DM, HBGI and LBGI offer insights into hyperglycemia and trends toward hypoglycemia, respectively; ADRR may be the optimal GV measure responsive to hypo- and hyperglycemic treatment effects.

Glycemic variability (GV) in patients with type 2 diabetes mellitus (T2DM) has been associated with oxidative stress and an increased risk of cardiovascular complications.¹ As a result, there is growing interest in GV as an emerging factor linked to increased risk of long-term complications of diabetes.² Recent reviews on this subject have explored the importance of GV as a clinical outcome, but the actual metrics to be used for its assessment are still hotly debated.^3-6 The most widely used methods to quantify GV include standard deviation (SD) and mean amplitude of glycemic excursions (MAGE),⁷ both of which are exclusively dependent on the magnitude of glucose excursions. We should clarify that, while in most applications MAGE is used with continuous glucose monitoring data, here we use the discrete form of this measure introduced by Service in a 2001 reanalysis of 7-point blood glucose (BG) profile DCCT data.⁸ In addition to amplitude, the more recently introduced the mean absolute glucose (MAG) change per unit time, which includes a timing component (eg, the length of the observation period), thereby accounting for the pace of glucose fluctuations.^6,9,10 A different approach, taking into account the asymmetry of the BG measurement scale, was taken in 1997 with the introduction of a mathematical transformation of the BG scale, which converted every BG reading into a risk value for hypo- or hyperglycemia.¹¹ A detailed risk analysis base was developed,¹² and 2 indices were introduced, the “low BG Index” (LBGI) and the “high BG Index” (HBGI), specifically designed to capture the frequency and extent of hypo- and hyperglycemic glucose excursions.^13,14 The average daily risk range (ADRR) was subsequently designed as a risk-based GV metric that is equally sensitive to both hypoglycemia and hyperglycemia.¹⁵ These indices have been associated with negative patient outcomes, such as severe hypoglycemia^13,16 and high hemoglobin A1c (A1C) or frequent hyperglycemia,^14,15,17 as well as with the effects of medication (including pramlintide,¹⁸ exenatide,¹⁹ and lixisenatide²⁰) in both type 1 and type 2 diabetes. A recent review provides a summary of research studies and clinical applications of the ADRR.²¹ Two authors have recently proposed to combine data from hypo- and hyperglycemia. Rodbard suggested several options for graphical displays based on the percentage of glucose values (or percentage of time) in hypoglycemic, target, and hyperglycemic ranges;²² Vigersky proposed the use of a composite metric, the hypoglycemia-A1C score (HAS), as the basis for a combined outcome that incorporated both the rate/severity of hypoglycemia and hemoglobin A1C, plus additional parameters such as weight change, cost, and quality of life.²³ Most recently, analysis of the same data used for this manuscript showed a robust association between virtually all GV measures and clinical outcomes.²⁴ Nevertheless, challenges in measuring GV (such as the large number of indices available and their interpretation) further complicate its utilization in clinical research and may delay its adoption in the clinical care of patients.²⁵ Determining the most appropriate GV metrics may be of potential use to clinicians when designing the best treatment strategies for patients and assessing therapeutic responses. This analytical study was conducted to (1) help identify GV measures that are most sensitive to treatment response in patients with T2DM and (2) evaluate the applicability of risk-based GV metrics in measuring response to glucose-lowering therapy.

Methods

Study Design and Patients

This was a post hoc pooled analysis of data from 6 previously reported insulin glargine studies conducted by Sanofi, or predecessor companies, from 1997 to 2007.^26-31 Eligible studies were required to be at least phase 3, prospective, randomized, controlled, and conducted in adult patients (>18 years) with T2DM treated with insulin glargine, a comparator insulin, and/or oral antidiabetes drugs (OADs) for at least 24 weeks to achieve target fasting plasma glucose (FPG) ≤ 100 mg/dl. Strict, predefined insulin titration algorithms were required for the studies to be included. Complete daily 7-point BG profiles (at 08:00, 10:00, 12:30, 14:30, 18:00, 20:00, and 23:00) were required for inclusion of patient data. Insulin glargine was required to be used as basal insulin only and no other prandial or bolus insulin was permitted as part of the insulin treatment regimen (short courses of regular insulin therapy for emergency medical reasons were permitted). Clinical studies that used concomitant basal or basal-bolus insulin regimens were excluded from the analysis. Complete 7-point BG profiles were available for 1699 patients; 60.4% (1026/1699) treated with insulin glargine and 39.6% (673/1699) treated with comparators (OADs, n = 140; NPH insulin, n = 235; 30/70 NPH insulin, n = 139; insulin lispro, n = 159). As previously reported,²⁴ baseline patient demographics and characteristics were: average (SD) age 59.4 (9.4) years; 42.3% women, average (SD) duration of diabetes 9.0 (6.0) years; average (SD) A1C 8.69% (0.95%) and average (SD) FPG 193.5 (47.5) mg/dl.

Clinical Endpoints

As the recommended treatment target for most patients with T2DM is an A1C of 7.0%,² the clinical outcomes evaluated in this analysis were: A1C levels ≥ 7.0% at 24 weeks to indicate persistent hyperglycemia (ie, failure of treatment), the on-trial development of symptomatic hypoglycemia (≥1 event), and the composite outcome of both A1C < 7.0% at 24 weeks and no hypoglycemia during the trial. In all 6 studies symptomatic hypoglycemia was defined as any reported hypoglycemia with symptoms. Associations between baseline SD, MAGE, MAG, and HBGI with A1C at 24 weeks and changes in A1C from baseline to week 24 are reported in another publication.²⁴

Traditional Glucose Variability Measures^5,6

GV was assessed, using 7-point BG profiles, by SD, MAGE (mg/dl), and MAG (mg/dl) at week 24. These traditional GV metrics were calculated using their literature definitions:

$SD=\sqrt{\frac{{\displaystyle \sum {(x_i-\overline{x})}^2}}{k-1}}$, where number of concentrations is k = 7, and the average BG is $\overline{x}$

$MAGE={\displaystyle \sum \frac{\lambda }{n}}$, if λ > SD, where n is the number of observations with λ > SD, and λ is each BG increase or decrease (absolute value) exceeding SD

$MAG=\frac{{\displaystyle \sum \left|x_i-x_{i+1}\right|}}{\Delta T}$, where ΔT is the time between the first and last BG measurement

Risk-Based GV Measures^11,12

These metrics were computed using the same 7-point BG profiles at week 24 and the function$f(x_i)=(\ln {(x_i)}^{1.084}-5.381)$, a mathematical transformation of the BG scale that symmetrizes the distribution of BG values:¹¹

$$LBGI=\frac{{\displaystyle {\sum }_{i=1}^n{r}_l(x_i)}}{n},$$

$$r_l(x_i)=22.77\cdot f{(x_i)}^2\mathrm{if}f(x_i)<0,\mathrm{and}0\mathrm{otherwise}$$

$$HBGI=\frac{{\displaystyle {\sum }_{i=1}^n{r}_h(x_i)}}{n},$$

$$r_h(x_i)=22.77\cdot f{(x_i)}^2\mathrm{if}f(x_i)>0,\mathrm{and}0\mathrm{otherwise}$$

$ADRR=\frac{1}{M}{\displaystyle \sum _{j=1}^M(L{R}^j+H{R}^j)}$, where M = days of measurement,

Where $L{R}^j=\max (r_l(x_1),\mathrm{\dots }{r}_l(x_k))$ and $H{R}^j=\max (r_h(x_1),\mathrm{\dots },r_h(x_k))$are the maximum hypo- and hyperglycemia risk values for day #j, j = 1,2, . . ., M.

We should note that several of these metrics are used in a nontraditional way and are based on 7-point daily profiles instead of their customary source data: (1) MAGE is traditionally based on continuous monitoring data, but is used here as proposed by Service in 2001;⁸ (2) MAG is typically used with sparse observations, but the formula and the meaning of the metric are preserved when used with 7-point profiles; (3) ADRR was introduced as a predictor of extreme hypo- and hyperglycemia over the subsequent several months,¹⁵ which required a month of SMBG with a testing frequency of 3-5 readings/day for robust prediction; ADRR, however, can be used a metric reflecting glucose variability with any number of days of observation as long as a good daily profile can be derived (eg, by 7 BG readings in a day).

Glycemic Risk Classification

As opposed to the traditional metrics of GV that describe only the magnitude of glucose fluctuations, all risk-based measures have well-defined outcome-related categories identifying subjects at low, moderate, and high risk for hypoglycemia (LBGI), hyperglycemia (HBGI), or both hypo- and hyperglycemia (ADRR). These risk categories are

• LBGI ≤ 2.5 (low); 2.5 < LBGI ≤ 5 (moderate); LBGI > 5 (high) risk for hypoglycemia¹³

• HBGI ≤ 4.5 (low); 4.5 < HBGI ≤ 9 (moderate); HBGI > 9 (high) risk for hyperglycemia^12,19

• ADRR ≤ 20 (low); 20 < ADRR ≤ 40 (moderate); ADRR > 40 (high) risk for both¹⁵

Statistical Analyses

Student t tests were used to test the significance of the change in GV from baseline to week 24. Cross-tabulation was used to assess the transition of subjects from 1 risk category to another as a result of treatment. The diagnostic utility of each GV or glycemic risk measure was evaluated using the area under the curve (AUC) of the receiver operating characteristic (ROC) curve. The ROC curve plots the true positive rate (specificity) versus the false positive rate (1-specificity) of diagnostic tests to assess the test performance and select an optimal cut-off point.³² The AUC of the ROC curve can range from 0 to 1; the closer this value to 1, the better the diagnostic value of the test, with a practical lower limit of 0.5. The AUC of the ROC was calculated for each measure for A1C levels ≥ 7.0% at 24 weeks, hypoglycemia (≥ 1 events), and the composite of A1C < 7.0% and no hypoglycemia. Patients who did not achieve the composite of A1C < 7.0% and no hypoglycemia could fail for 3 different reasons: failure to achieve A1C < 7.0%, experiencing hypoglycemia, or both. A 4 × 1 ANOVA with the composite endpoint as a factor was conducted to clarify how each of the GV measures contributed to the different types of treatment failure. For all GV measures, the incremental odds ratio (OR) for achieving each endpoint was also calculated.

Results

Baseline to End-of-Study Changes in A1C, Hypoglycemia, and GV

As previously reported,²⁴ the mean A1C at end of week 24 was 7.0% (±0.9%), a change of -1.65% (±1.03) from baseline to week 24. This improvement was accompanied by a 68.5 (±55.7) mg/dl reduction in FPG and a 60.8 (±43.2) mg/dl reduction in mean BG. Overall, 1035 patients (60.9%) had symptomatic hypoglycemic events, with an event rate of 9.29 per patient-year. Statistically significant changes in all GV measures were observed over the 24 weeks of treatment with both insulin glargine and comparators; P < .0001 for all metrics, as presented in Table 1, which in addition to previous results²⁴ now includes ADRR as well.

Table 1.

Baseline, Week 24, and Percentage Change From Baseline to Week 24.

Measure of GV	Baseline (mean, SD)	Week 24 (mean, SD)	Percentage change	P value
A1C, %	8.69 (0.95)	7.04 (0.91)	−19	<.0001
SD, mg/dl	43.2 (17.6)	39.0 (18.7)	−9.7	<.0001
MAGE, mg/dl	63.1 (28.9)	55.6 (28.0)	−11.8	<.0001
MAG, mg/dl	20.6 (9.2)	18.2 (9.5)	−12.0	<.0001
LBGIa	0.07 (0.3)	0.5 (0.8)	—	—
HBGI	15.4 (10.5)	5.3 (5.4)	−65.5	<.0001
ADRR	28.9 (15.0)	16.4 (11.6)	−43.3	<.0001

^aFor most subjects, the baseline LBGI was 0 (ie, no BG values below 112.5 mg/dl), which makes the calculation of percentage change unreliable.

Baseline to End-of-Study Changes in Glycemic Risk Classification

The absolute changes in GV metrics reported in the previous paragraph imply that the risk status of the patients should change as well. Table 2 presents the subject transitions between low-, moderate-, and high-risk categories from baseline to the end of study for LBGI, HBGI, and ADRR. Cross-tabulation and chi-square analysis indicate that, overall, the subjects:

Table 2.

Baseline Versus End-of-Study Categories of Risk for Hypoglycemia (LBGI), Hyperglycemia (HBGI), and Overall Risk-Based Glucose-Variability (ADRR).

A. Hypoglycemia risk categories based on LBGI		End-of-study number of subjects (% of all)
		Low	Moderate	High
Baseline number of subjects (% of all)	Low (LBGI ≤ 2.5)	1636 (96.3)	52 (3.1)	6 (0.4)
	Moderate (2.5-5.0)	4 (0.2)	0 (0.0)	0 (0.0)
	High (LBGI > 5.0)	1 (0.1)	0 (0.0)	0 (0.0)
B. Hyperglycemia risk categories based on HBGI		End-of-study number of subjects (% of all)
		Low	Moderate	High
Baseline number of subjects (% of all)	Low (HBGI ≤ 4.5)	160 (9.4)	28 (1.6)	8 (0.5)
	Moderate (4.5-9.0)	251 (14.8)	80 (4.7)	21 (1.2)
	High (HBGI > 9.0)	556 (32.7)	328 (19.3)	267 (15.7)
C. Hypo- and hyperglycemia risk categories based on ADRR		End-of-study number of subjects (% of all)
		Low	Moderate	High
Baseline number of subjects (% of all)	Low (ADRR ≤ 20)	455 (26.8)	64 (3.8)	7 (0.4)
	Moderate (20-40)	545 (32.1)	236 (13.9)	32 (1.9)
	High (ADRR > 40)	176 (10.4)	149 (8.8)	35 (2.1)

Subjects on the diagonal retained their baseline risk status; shaded area below the diagonal indicates risk reduction; shaded area above the diagonal indicates risk increase.

• Retained their low baseline level of risk for hypoglycemia, with only 3.5% of all subjects increasing their risk category (chi-square P = .94) (Table 2A);

• Significantly reduced their risk for hyperglycemia, with 66.8% of subjects improving their HBGI risk category and 3.3% increasing their risk (chi-square P < .0001; Table 2B);

• Reduced their GV, with 51.3% reducing their ADRR category and 6.1% increasing their ADRR category (chi-square P < .0001; Table 2C).

Associations Between GV Measures at 24 Weeks and Clinical Endpoints

All end-of-trial GV measures were significantly associated with week 24 A1C ≥ 7.0% and ≥1 symptomatic hypoglycemic event during the trial, with each GV measure showing AUC values > 0.5 for the ROCs (Table 3). LBGI showed the strongest association with symptomatic hypoglycemia alone; HBGI showed the strongest association with week 24 A1C ≥ 7.0% alone, and both ADRR and HBGI had the best associations with week 24 composite outcome A1C < 7.0% and no hypoglycemic events. Given that LBGI and HBGI were designed to look at the low and the high end of the BG scale, respectively,^11,12 these results are consistent with the definition of these indices.

Table 3.

Area Under the ROC Curves for Outcomes: A1C ≥ 7.0%; ≥1 Symptomatic Hypoglycemic Event, or Composite Endpoint of A1C < 7.0%; and No Hypoglycemic Event by Week 24 GV Measure.

AUC (SD)	SD	MAGE	MAG	LBGI	HBGI	ADRR
A1C ≥ 7.0%	0.55 (0.014)	0.55 (0.014)	0.55 (0.014)	N/Aa	0.68 (0.013)	0.59 (0.014)
≥1 symptomatic hypoglycemic event	0.56 (0.014)	0.56 (0.014)	0.55 (0.014)	0.60 (0.014)	N/Aa	0.54 (0.014)
A1C < 7.0% and no hypoglycemia	0.57 (0.018)	0.57 (0.018)	0.57 (0.018)	N/A	0.62 (0.017)	0.59 (0.018)

Shaded cells mark the strongest associations.

^aLBGI and HBGI were specifically designed to look at the low and high ends of the BG scale, respectively; thus, each measure’s use in predicting the reverse is inconsistent with its definition.

The ROC curves and the optimum thresholds for LBGI, HBGI, and ADRR are presented in Figure 1. For the A1C ≥ 7.0% outcome, the optimum thresholds for the endpoints ADRR and HBGI were 14.8 (true positive rate 55%, false positive rate 40%) and 4.1 (true positive rate 62%, false positive rate 32%). The ORs for these thresholds were 1.84 and 3.37, indicating that patients with week 24 ADRR ≥ 14.8 and HBGI ≥ 4.1 were approximately 2 and 3 times more likely to have end-of-trial A1C ≥ 7.0%. For ≥1 symptomatic hypoglycemic event during the trial, the best threshold for endpoints ADRR and LBGI were 13.0 (true positive rate 56%, false positive rate 51%) and 0.1 (true positive rate 59%, false positive rate 44%), with ORs of 1.17 and 1.73, indicating that patients with endpoint LBGI ≥ 0.1 were about twice as likely to have a hypoglycemic event. With the composite endpoint of both week 24 A1C < 7.0% and no on-trial symptomatic hypoglycemic events, the HBGI had the largest AUC, which is consistent with the primary treatment goal—reduction in hyperglycemia. The best threshold for endpoint HBGI was 3.3 (true positive rate 57%, false positive rate 40%) with an OR of 2.03 (95% CI 1.57-2.62), indicating that patients with endpoint HBGI < 3.3 were twice as likely to achieve the A1C < 7.0% goal without hypoglycemic events. Similarly, ADRR < 11.7 was also relevant to this endpoint with an OR 1.50 (95% CI 1.23-1.82).

Figure 1.

ROC curves of ADRR, HBGI, and LBGI at week 24 as the predictor and week 24 A1C ≥ 7.0%, ≥1 symptomatic hypoglycemic event and the composite of week 24 A1C < 7.0%, and no symptomatic hypoglycemic event as the outcomes. Values in the graphs are the optimal thresholds for each classification.

Finally, Table 4 assesses the ability of the GV measures to differentiate the various components of treatment failure (ie, failure to achieve A1C < 7.0%, experiencing hypoglycemia, or both). As seen in column 4 of Table 4, all GV metrics differentiated extreme cases well (those patients who had both A1C > 7.0% and hypoglycemia) from all other subgroups. However, in terms of intermediate outcomes, HBGI was the best predictor of when failure to achieve target A1C was due to inadequate control of the high end of the postprandial glucose excursions (Table 4, column 3). In addition, only the LBGI revealed when occurrence of hypoglycemia was due to overtreatment (Table 4, column 2). No other variability measures, except for ADRR, were able to differentiate any intermittent subgroups.

Table 4.

ANOVA of Week 24 GV Measures by Types of Treatment Failure.

Outcome at week 24 mean values[P value]	Type of treatment failure at week 24
	A1C < 7.0% with ≥1 hypoglycemic event	A1C ≥ 7.0% with no hypoglycemic events	A1C ≥ 7.0% and ≥1 hypoglycemic event
SD
A1C < 7.0% with no hypoglycemic events	35.1 vs 37.9 (ns)a	35.1 vs 38.0 (ns)	35.1 vs 43.9 (P < .0001)
A1C < 7.0% with ≥1 hypoglycemic event		37.9 vs 38.0 (ns)	37.9 vs 43.9 (P < .0001)
A1C ≥ 7.0% with no hypoglycemic events			38.0 vs 43.9 (P < .0001)
MAGE
A1C < 7.0% with no hypoglycemic events	49.8 vs 54.4 (ns)	49.8 vs 53.4 (ns)	49.8 vs 63.0 (P < .0001)
A1C < 7.0% with ≥1 hypoglycemic event		54.4 vs 53.4 (ns)	54.4. vs 63.0 (P < .0001)
A1C ≥ 7.0% with no hypoglycemic events			53.4 vs 63.0 (P < .0001)
MAG
A1C < 7.0% with no hypoglycemic events	16.7 vs 17.9 (ns)	16.7 vs 17.7 (ns)	16.7 vs 19.9 (P < .0001)
24 A1C < 7.0% with ≥1 hypoglycemic event		17.9 vs 17.7 (ns)	17.9 vs 19.9 (P = .0004)
A1C ≥ 7.0% with no hypoglycemic events			17.7 vs 19.9 (P = .0005)
ADRR
A1C < 7.0% with no hypoglycemic events	13.4 vs 15.1 (ns)	13.4 vs 17.2 (P < .0001)	13.4 vs 19.7 (P < .0001)
A1C < 7.0% with ≥1 hypoglycemic event		15.1 vs 17.2 (P < .0055)	15.1 vs 19.7 (P < .0001)
A1C ≥ 7.0% with no hypoglycemic events			17.2 vs 19.7 (P = .0015)
LBGI
A1C < 7.0% with no hypoglycemic events	0.5 vs 0.8 (<0.0001)	0.5 vs 0.2 (P < .0001)	0.5 vs 0.4 (P = .0088)
A1C < 7.0% with ≥1 hypoglycemic event		0.8 vs 0.2 (P < .0001)	0.8 vs 0.4 (P < .0001)
A1C ≥ 7.0% with no hypoglycemic events			0.2 vs 0.4 (P = .0119)
HBGI
A1C < 7.0% with no hypoglycemic events	3.6 vs 3.7 (ns)	3.6 vs 7.1 (P < .0001)	3.6 vs 7.1 (P < .0001)
A1C < 7.0% with ≥1 hypoglycemic event		3.7 vs 7.1 (P < .0001)	3.7 vs 7.1 (P < .0001)
A1C ≥ 7.0% with no hypoglycemic events			7.1 vs 7.1 (ns)

Shaded cells include comparisons that are clinically relevant to distinct types of treatment failure.

^aTo account for multiple comparisons, only results with P < .01 are considered significant.

Conclusions

This analysis of nearly 1700 patients with T2DM involved in 24-week clinical trials of glucose-lowering therapies demonstrates that ADRR is the overall GV measure most responsive to treatment. ADRR demonstrated a percentage change 4-fold greater when compared with the other commonly used 2-sided GV measures (SD, MAGE, MAG) and was sensitive to both reduction of hyperglycemia (the primary treatment goal) and concurrent occurrence of hypoglycemia. In addition, the LBGI and the HBGI (which are 1-sided metrics selectively assessing the hypo- or hyperglycemic end of the BG scale) were better associated with the frequency of hypoglycemia and with failure to achieve target A1C, respectively, than any 2-sided GV measure. Week 24 LBGI was the best predictor of symptomatic hypoglycemic events during treatment intensification, and week 24 HBGI was the best predictor of week 24 A1C levels. Here we should note that in Table 3 and Figure 1, HBGI was slightly better than ADRR in its association with the composite outcome A1C < 7.0% without symptomatic hypoglycemic events. This is due to the fact that HBGI is not sensitive to hypoglycemia by design and therefore ignores all low BG readings, while these readings are accounted by ADRR and deter ADRR from showing improvement in some patients who reduced their A1C but registered some (albeit nonsymptomatic) hypoglycemia in their daily profiles.

From a different perspective, LBGI and HBGI could be used selectively to predict the reasons for failing to achieve specific endpoints, for example, to distinguish those who would or would not have hypoglycemia while keeping A1C < 7.0% (LBGI), or those who would or would not achieve A1C < 7.0% regardless of their frequency of hypoglycemia (HBGI). As we have established, achieving the composite endpoint can fail due to more than 1 reason, but differentiation of the specific reasons for failure cannot be achieved by any overall GV metric (Table 4). In this case, selective GV measures may be useful in deciphering the various reasons for treatment failure. Specifically, failure to bring A1C under control may be due to inadequate control of the high peaks of postprandial glucose excursions, as measured by the HBGI. In contrast, the occurrence of hypoglycemia may be due to overtreatment as indicated by LBGI. Failure to control both ends of the scale simultaneously would be due to inadequate control of the overall GV as indicated by virtually all 2-sided variability metrics.

The analysis and evaluation of GV in diabetes is a developing field with a multitude of proposed GV metrics, many of which overlap, and for which there is no gold standard.⁴ This is because, as opposed to the mean BG, which is easily quantified by A1C measurement, the magnitude and timing of BG fluctuations are mathematically difficult to measure, compare, and contrast. To the best of our knowledge, the only nonempirical approach to diabetes-specific quantifying of GV is our risk-analysis theory¹² proposed 15 years ago that takes into account the peculiarities of the BG measurement scale. Over the years, the indices derived from this risk analysis (LBGI, HBGI, ADRR) have consistently proven to be sensitive to severe hypoglycemia (LBGI^13,16), hyperglycemia, and A1C (HBGI^14,17), extreme bidirectional glucose fluctuations (ADRR¹⁵), and the effects of medications.^18-20 As a result, the LBGI was included in the 2005 report of the American Diabetes Association Workgroup on Hypoglycemia as a metric of the frequency and extent of hypoglycemic episodes.³³ Most recently, in the same database used for the analyses presented in this manuscript, the LBGI and the HBGI were shown to be predictive of the clinical outcomes during treatment intensification in T2DM.²⁴

Now we have taken advantage of a large available sample size of about 1700 patients with T2DM and of the extensive patient-level data derived from randomized controlled studies to present new comparisons of several widely used GV metrics and to introduce the use or risk analysis and risk classification for quantifying treatment outcomes.

The new analyses augmenting our previous report²⁴ include

• Classification of subjects according to their risk status for hypoglycemia, hyperglycemia, or both, and the change of this risk status as a function of treatment intensification (Table 2). We should note that such a classification is only possible with metrics that have established risk cut-off points—a feature not available to any of the traditional measures of GV

• Establishing the predictive value of several GV metrics in terms of their ORs for predicting a particular outcome (Table 3 and Figure 1).

• Deciphering the various reasons for treatment failure (Table 4).

The major limitation of this study is that with these data GV metrics cannot be directly related to the physiology of diabetes complications as speculated by other studies. For example, mechanisms potentially linking GV to macrovascular complications through excessive protein glycation and oxidative stress³⁴ cannot be derived from these data. Similarly, we cannot address the possible correlation between swings in postprandial hyperglycemia and cardiovascular disorders.³⁵ The relationships established here are strictly observational and can be extrapolated to physiology and complications only through mediating variables, such as A1C.

Furthermore, these analyses are post hoc and were not prespecified in the design of the studies collecting the data; thus, these retrospective analyses may be subject to selection bias and confounding. However, any selection bias would be equally applicable to all GV metrics used in this work and, therefore, their relative sensitivity to treatment outcomes should be preserved. Nevertheless, the utility of GV measures in identifying patients at greatest risk of not meeting glycemic targets or of experiencing hypoglycemic events should be evaluated in future prospective studies.

In conclusion, this study shows that HBGI and LBGI may, in concert, provide the most clinically useful information about GV that is specific and selective to the outcomes of treatment intensification in T2DM: week 24 HBGI was the optimal predictive variable for week 24 A1C levels and week 24 LBGI was the optimal predictive variable for the development of symptomatic hypoglycemic events. The concurrent use of these 2 indices may therefore assist the clinical treatment decisions to optimize management of patients with T2DM. If a single combined metric is preferred, the ADRR was the optimal bi-directional GV measure to account for both reduction of hyperglycemia and trends toward increase in hypoglycemic events.