Glucose Variability: Timing, RiskAnalysis, and Relationship toHypoglycemia in DiabetesDiabetes Care 2016;39:502–510 | DOI: 10.2337/dc15-2035

Glucose control, glucose variability (GV), and risk for hypoglycemia are intimatelyrelated, and it is now evident that GV is important in both the physiology andpathophysiology of diabetes. However, its quantitative assessment is complexbecause blood glucose (BG) fluctuations are characterized by both amplitude andtiming. Additional numerical complications arise from the asymmetry of the BGscale. In this Perspective, we focus on the acute manifestations of GV, particularlyon hypoglycemia, and review measures assessing the amplitude of GV fromroutine self-monitored BG data, as well as its timing from continuous glucosemonitoring (CGM) data.With availability of CGM, the latter is not only possible butalso a requirementdwe can now assess rapid glucose fluctuations in real time andrelate their speed and magnitude to clinically relevant outcomes. Our primarymessage is that diabetes control is all about optimization and balance betweentwo key markersdfrequency of hypoglycemia and HbA1c reflecting average BGand primarily driven by the extent of hyperglycemia. GV is a primary barrier to thisoptimization, including to automated technologies such as the “artificial pan-creas.” Thus, it is time to standardize GV measurement and thereby streamlinethe assessment of its two most important componentsdamplitude and timing.

Although reducing hyperglycemia and targeting HbA1c values of 7% or less result indecreased risk of micro- and macrovascular complications (1–4), the risk for hypo-glycemia increases with tightening glycemic control (5,6). Consequently, hypogly-cemia has been implicated as the primary barrier to tight control (7,8). Thus,patients with diabetes face a lifelong optimization problem: reducing average gly-cemic levels and postprandial hyperglycemia while simultaneously avoiding hypo-glycemia. A strategy for achieving such an optimization can only be successful if itreduces glucose variability (GV). This is because bringing average glycemia down isonly possible if GV is constraineddotherwise blood glucose (BG) fluctuations wouldinevitably enter the range of hypoglycemia (9). This numerical assertion was re-cently reflected by a clinician’s Perspective stating: “The selection of a glycemic goalin a person with diabetes is a compromise between the documented upside ofglycemic controldthe partial preventionor delay ofmicrovascular complicationsdandthe documented downside of glycemic controldthe recurrent morbidity and poten-tial mortality of iatrogenic hypoglycemia” (8). Such a two-sided approach to theassessment of glycemic control is necessary because HbA1c fails to capture GV andthe attendant risks associated with the extremes of hypo- and hyperglycemia. Thiseffect is illustrated in Fig. 1, which presents 15-day glucose traces of two subjectswho had identical HbA1c of 8.0%, but very different degrees of GV. As seen inFig. 1, subject 1 and subject 2 had similar average glucose throughout the courseof observation, but subject 1 (Fig. 1A) had visibly higher amplitude of glucose

1Center for Diabetes Technology, University ofVirginia, Charlottesville, VA2Department of Information Engineering, Uni-versity of Padova, Padova, Italy

Corresponding author: Claudio Cobelli, cobelli@dei.unipd.it.

Received 17 September 2015 and accepted 21January 2016.

© 2016 by the American Diabetes Association.Readersmayuse this article as longas thework isproperly cited, the use is educational and not forprofit, and the work is not altered.

Boris Kovatchev1 and Claudio Cobelli2

502 Diabetes Care Volume 39, April 2016

PER

SPECTIVES

INCARE

fluctuations, i.e., higher GV, which re-sulted in seven episodes of moderate hy-poglycemia (#50 mg/dL) and eightepisodes of moderate hyperglycemia($350 mg/dL). In contrast, subject 2 hadno such hypo- or hyperglycemic episodes.Thus, GV is a major determinant of a per-son’s risk for hypoglycemia or hyperglyce-mia, which in turn is reflected by the riskfor complications associated with ex-treme glucose fluctuations (10).Many studies, however, encountered a

purely analytical problemdhow to mea-sure GV. A recent review argued that“there are better markers to assess therisk of diabetes than GV” (11). This argu-ment depends heavily on the metricsused to quantify GV, and it can be simi-larly stated that there are better metricsto assess GV that reflect more accuratelythe risk of diabetes. The problem is thatdifferent ways of glucose monitoring re-sult in different resolutions of observa-tion of GV. For example, HbA1c capturesonly slow fluctuations in average glyce-mia taking place over several months,self-monitoring of BG (SMBG) a fewtimes a day can capture daily variationin BG fluctuations, and contemporarycontinuous glucose monitoring (CGM)can take the GV monitoring resolutionto the scale of minutes. With this analyt-ical peculiarity in mind, we now focuson the causes, metrics, and clinical

manifestations of GV, with a special em-phasis on measuring the amplitude andthe timing of glucose fluctuations andtheir relevance to underlying physiologyand patient behavior.

CAUSES OF GV

In health, glucose metabolism is tightlycontrolled by a hormonal network in-cluding the gut, liver, pancreas, andbrain to ensure stable fasting BG levelsand transient postprandial glucose fluc-tuations. In other words, BG fluctuationsin type 1 diabetes result from the activ-ity of a complex metabolic system per-turbed by behavioral challenges. Thefrequency and extent of these chal-lenges and the ability of the person’ssystem to absorb them determine thestability of glycemic control. The degreeof system destabilization depends oneach individual’s physiological parame-ters of glucose–insulin kinetics, includ-ing glucose appearance from food,insulin secretion, insulin sensitivity,and counterregulatory response. A ma-jor source of GV is the rapid onset of hy-perglycemia due to the consumption of“high–glycemic index” foods. Moreover,sustained hyperglycemia engendered byfoods with simple carbohydrates andhigh fat (classically pizza) may challengetherapy. Such foods often have hedonicqualities (e.g., comfort foods) leading to

repeated challenges to glycemic stabil-ity. There is strong evidence that feedingbehavior is abnormal in both uncon-trolled diabetes and hypoglycemia andthat feeding signals within the brainand hormones affecting feeding, suchas leptin and ghrelin, are implicated indiabetes (12–14). Insulin secretion andaction vary with the type and durationof diabetes. In type 1 diabetes, insulinsecretion is virtually absent, which de-stroys the natural insulin–glucagon feed-back loop and thereby diminishes thedampening effect of glucagon on hypo-glycemia. In addition, insulin is typicallyadministered subcutaneously, whichadds delays to insulin action and therebyamplifies the amplitude of glucose fluc-tuations. In type 2 diabetes, increasedinsulin resistance and progressive loss ofb-cell function result in higher and pro-longed postprandial glucose excursions.Although observed in type 2 diabetes, im-paired hypoglycemia counterregulationand increased GV in the hypoglycemicrange are particularly relevant to type 1diabetes: It has been shown that glucagonresponse is impaired (15), and epineph-rine response is typically attenuated aswell (16). Antecedent hypoglycemia shiftsdown BG thresholds for autonomic andcognitive responses, thereby further im-pairing both the hormonal defenses andthe detection of hypoglycemia (17). Stud-ies have established relationships be-tween intensive therapy, hypoglycemiaunawareness, and impaired counterregu-lation (16,18–20) and concluded that re-current hypoglycemia spirals into a“vicious cycle” known as hyperglycemia-associated autonomic failure (HAAF) (21).Our studies showed that increased GVand the extent and frequency of low BGare major contributors to hypoglycemiaand that such changes are detectable byfrequent BG measurement (22–25). Inaddition, a study involving 34 subjectswith type 1 diabetes found that higher in-sulin sensitivity and lower epinephrine re-sponse during hypoglycemia measuredin a hospital setting were related to in-creasedGVand risk for hypoglycemiamea-sured in the field, irrespective of HbA1c andother patient characteristics (26).

PRINCIPAL COMPONENTS OFGLUCOSE FLUCTUATION:AMPLITUDE AND TIMING

GV in diabetes is a process that developsin time. Most traditional metrics of GV

Figure 1—Fifteen-day glucose traces of two subjects who had identical HbA1c of 8.0% butdifferent degrees of GV. High GV in subject 1 was reflected by numerous episodes of bothhypo- and hyperglycemia (A), whereas low GV in subject 2 resulted in no such episodes (B).

care.diabetesjournals.org Kovatchev and Cobelli 503

focus on the amplitude of glucose fluc-tuations and ignore their timing compo-nent. Figure 2 presents the principalcomponents of GV: amplitude andtime (Fig. 2A). Amplitude is the projec-tion of the BG data along the y-axis(Fig. 2B) and presents the extent ofBG fluctuations without placing themin time. Timing is the projection ofthe data along the x-axis (Fig. 2C), whichis representative of the duration of var-ious eventsdin our example, the dura-tion of time spent below, within, andabove a predetermined BG target rangeof 70–180 mg/dL.Traditionally, the metrics of ampli-

tude of GV have relied on episodic BGreadings, e.g., SMBG data, whereas themore contemporary metrics of the tim-ing of GV rely on CGM. These distinc-tions are clarified below.

MEASURING THE AMPLITUDEOF GV

MetricsThe traditional statistical calculation ofBG includes standard deviation (SD)(27), coefficient of variation (CV), orother metrics, such as the M-value in-troduced in 1965 (28), the mean ampli-tude of glucose excursions (MAGE)introduced in 1970 (29), the glycemiclability index (30), or the mean absoluteglucose (MAG) change (31,32). As

recently shown, the multitude of GVmetrics introduced over the years canbe reduced to fewer representative in-dices (33). Table 1 presents some of themost widely used GV metrics with theirformulas.

Risk AnalysisThe last three of these metricsdthe lowBG index (LBGI), high BG index (HBGI),and average daily risk range (ADRR)dare based on a transformation of theBG measurement scale (f(x) in the for-mulas in Table 1), which aims to correctthe substantial asymmetry of the BGmeasurement scale. Numerically, thehypoglycemic range (BG ,70 mg/dL) ismuch narrower than that in the hyper-glycemic range (BG .180 mg/dL) (34).As a result, whereas SD, CV, MAGE, andMAG are inherently biased toward hy-perglycemia and have a relatively weakassociation with hypoglycemia, the LBGIand ADRR account well for the risk ofhypoglycemic excursions. The analyticalform of the scale transformation f(x)was based on accepted clinical assump-tions, not on a particular data set, andwas fixed 17 years ago, which made theapproach extendable to any data set(34). On the basis of this transformation,we have developed our theory of riskanalysis of BG data (35), defining a com-putational risk space that proved to be

very suitable for quantifying the extentand frequency of glucose excursions.The utility of the risk analysis has beenrepeatedly confirmed (9,25,36–38). Wefirst introduced the LBGI and HBGI,which were specifically designed to besensitive only to the low and high end ofthe BG scale, respectively, accountingfor hypo- and hyperglycemia withoutoverlap (24). Then in 2006, we intro-duced the ADRR, a measure of GV thatis equally sensitive to hypo- and hy-perglycemic excursions and is predic-tive of extreme BG fluctuations (38).Most recently, corrections were intro-duced that allowed the LBGI and HBGIto be computed from CGM data withresults directly comparable to SMBG(39).

Graphs

Variability Grid Analysis

Besides glucose traces and histogramsshowing the spread of BG data, thereare very few graphs specifically aimedto visualize the amplitude of GV. Onesuch plot is the variability grid analysis(VGA), a method visually presenting GVat a population level (40,41). The VGAis a minimum/maximum plot of the BGreadings for a subject taken over a cer-tain observation period (e.g., 72 h). Theminimum BG (the lower 2.5th percen-tile) is plotted on the x-axis, which isinverse coded from 110 mg/dL to,50 mg/dL. The maximum BG (the up-per 2.5th percentile) is plotted on they-axis. Thus the difference (y 2 x) repre-sents the BG range observed for a sub-ject. The plot is split into zones. A point inthe A-zone (light green) indicates thatthe subject was never below 90 mg/dLand never above 180 mg/dL; i.e., thecontrol was optimal. B-zones (darkgreen) indicate various degrees of sub-optimal, but still acceptable, control;the C-zones (yellow) indicate overcor-rection of hypo- or hyperglycemia, result-ing in either maximum above 300 mg/dLorminimumbelow70mg/dL; the D-zones(orange) indicate even a higher degree ofGV, whereas the E-zone (red) indicatesBG readings below 70 mg/dL and above300 mg/dL within the same observationperiod (Fig. 3).

The data points in Fig. 3 come from apreviously discussed data set containingSMBG readings for 335 subjects withtype 1 (n = 254) and type 2 (n = 81) di-abetes (38). Figure 3A presents subjects

Figure 2—Principal components of GV. Glucose fluctuations are a process in time that hastwo dimensionsdamplitude and time (A). Projected along its amplitude axis, this process ismeasured by metrics such as SD or MAGE (B). Projected along its time axis, this process isassessed by temporal characteristics, such as time within target range and time spent inhypo- or hyperglycemia (C).

504 Glucose Variability and Risk Analysis of Diabetes Diabetes Care Volume 39, April 2016

in good control with most of the datapoints (70.4%) in the A- and B-zones.Figure 3B presents subjects at poor controlwith most of the data points (89.6%) inthe C-, D-, and E-zones. The subject splitwas done along the median ADRR forthe population. The subjects presentedin panel A had a mean HbA1c of 7.6% andmean ADRR of 22.9; these subjects re-corded 0.54 BG episodes #39 mg/dLand 3.27 BG episodes .400 mg/dL perperson in the subsequent 3 months. Thesubjects presented in panel B had amean HbA1c of 8.2% and mean ADRRof 48.0; these subjects recorded 3.32BG episodes #39 mg/dL and 13.6 BGepisodes .400 mg/dL per person in thesubsequent 3months. Thus, the differencein GV depicted by the VGA plots in Fig. 3and quantified by the ADRR resulted in aseveral-fold difference in subsequent ex-treme hypo- and hyperglycemic episodes

recorded over 3 months of follow-upobservation (38).

MEASURING THE TIMING OF GV

MetricsA recent review of GV (42) stated that“to avoid distortion of variability to thatof glycemic exposure, its calculationshould be devoid of a time component,”and a counterpoint presented in thesame issue of Diabetes argued that thetiming of BG fluctuations is important(32). Indeed, the same amplitude of glu-cose fluctuation can yield quite differentclinical results depending on the speedof glucose transition. For example,among the many criticisms of MAGE, itwas noted that MAGE was originally de-veloped using 1-h data spacing but wasthen used with 7-point glucose profilesand with CGM data, both of which havenot been validated (43). As noted in the

previous section, oneof themetrics used tomeasure GV amplitudedMAGdattemptsto capture this effect and includes a timingcomponent (DT in the denominator),which accounts for the duration of timeoverwhich BG fluctuations are observed.As a result, the same fluctuation ampli-tude over longer time would yield alower MAG. Although apparently useful,considering time as an additional com-ponent of GV adds an additional layerof complexity, which so far has pre-cluded widespread clinical use of time-dependent GV metrics. Attempts tofocus on specific temporal characteris-tics of the data, such as the variabilityof periodic phenomena, brought aboutthe mean of daily differences (MODD),which is designed to assess interday gly-cemic variation and therefore circadianperiodicity (44). A generalization of thisapproach to a finer-resolutionmeasure-ment of GV was presented by the con-tinuous overlapping net glycemic action(CONGA) metric that calculates the dif-ference between a current BG readingand a reading taken (n) hours earlierand then takes the SD of these differ-ences (44). If the BG readings are morefrequent than (n) hours, CONGA willtake into its calculation overlappingtime windows. The order of CONGAdepends on clinical considerations:CONGA1 (for n = 1) corresponds totime period of 1 h, CONGA2 to 2 h, andso forth (44). A straightforward version ofCONGA is the SD of the BG rate of change(45), which can be considered a CONGAof order 1 as long as the time intervalbetween the BG readings coincideswith the step of the CONGA procedure(Table 2).

Graphs: Poincaré Plot of GlucoseDynamicsWhereas in Fig. 2C we illustrated thetiming of glucose control with a tradi-tional graph depicting a person’s timewithin target range, here we will focuson a plot used in nonlinear dynamics tovisualize the dynamics of the investi-gated system. Such a representation be-comes possible with the availability ofCGM time series data, a sequence offrequent, equally spaced in time BGdeterminations that capture a high-resolution picture of the dynamics ofBG fluctuations (41,45). The plot (knownalso as recurrence plot or lag plot) isnamed after the French mathematician

Table 1—Metrics of the amplitude of GV

Metric Formula Meaning

SD

SD ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi∑ðxi 2 xÞ2

n2 1

sVariation of BG fluctuations

around the mean

CV SD

Mean

Magnitude of variabilityrelative to the mean

MAGE MAGE ¼ ∑ln, if l . SD,

where l is each BG increaseor decrease (absolute value)exceeding SD

The MAGE exceeding theSD of glucose variation

MAGchange MAG ¼ ∑jxi 2 xiþ1j

DT

Summed absolute differencesbetween sequential readingsdivided by the time betweenthe first and last BGmeasurement (MAG includesa timing component and isnot a purely amplitude metric)

LBGILBGI ¼ ∑rlðxiÞ

n, where

rlðxiÞ ¼ 22:77 z fðxiÞ2 if fðxiÞ,0;and 0 otherwise;

fðxiÞ ¼ ðlnðxiÞ1:084 2 5:381Þfor several BG readings x1,. . .,xn

Measure of frequency andextent of hypoglycemiadesigned to amplifyhypoglycemic excursionsand ignore hyperglycemia

HBGIHBGI ¼ ∑rhðxiÞ

n, where

rhðxiÞ ¼ 22:77 z fðxiÞ2 if fðxiÞ.0;and 0 otherwise;

fðxiÞ ¼ ðlnðxiÞ1:084 2 5:381Þfor several BG readings x1,. . .,xn

Measure of frequency andextent of hyperglycemiadesigned to scalehyperglycemic excursions andignore hypoglycemia

ADRR

where LRj = max(r1(x1), . . .r1(xk))and HRj = max(rh(x1),. . .rh(xk))for several BG readings x1,. . .,xktaken within day #j, j = 1,2. . .,M

Average of the maximal dailyamplitudes of glucoseexcursions across several days,designed to be equally sensitiveto hypo- and hyperglycemia

In all formulas, x = BG level.

care.diabetesjournals.org Kovatchev and Cobelli 505

Jules Henri Poincaré (1854–1912) whomade fundamental contributions tothe mathematics and physics of dynam-ical systems. The idea is straightfor-ward: Given a series of BG readings,BG(t1), BG(t2),. . ., BG(tn), taken at equallyspaced points in time, t1, t2, . . ., tn, eachpoint of the plot has coordinates BG(ti-1)on the x-axis and BG(ti) on the y-axis (Fig.4). Thus, the difference (y 2 x) of thecoordinates of each data point repre-sents the BG change occurring betweentimes ti and ti-1. Intuitively, larger differ-ences between the coordinates of

sequential BG points will signify fasterBG fluctuations, i.e., amore unstable glu-coregulatory system. Alternatively, asmaller, more concentrated plot wouldindicate system stability. It is also worthnoting that a Poincaré plot can indicatecertain characteristics of system behav-ior. For example, an elliptic plot wouldindicate an oscillating system, such asthe perfect oscillation of BG values be-tween 70 and 250mg/dL imagined in Fig.4A. The boundaries of the plot are alsoindicative of the minimum and maxi-mum BG levels achieved during the

observation period, thereby visualizingthe risk for hypo- and hyperglycemia towhich the patient was exposed duringthe observation period. Thus, a moreconcentrated plot, such as the one inFig. 4C, indicates a stable patient,whereas a more scattered Poincaréplot, such as the one in Fig. 4B, indicatessystem irregularity or poor glucose con-trol. We should also note that nonstan-dard applications of Poincaré-type plotshave been used as well; for example, thetime lag associated with CGM can be vi-sually estimated by plotting the codynam-ics of reference versus CGM readings atdifferent time lags (46).

CASE STUDIES ON GV

Example 1: Prediction of the Riskfor Extreme BG FluctuationsIn this example, we illustrate themetricsof GV with respect to their ability to pre-dict future extreme glucose fluctua-tions, e.g., assess the risk for hypo- andhyperglycemia, separately and in combi-nation. In a large study of 335 patientswith type 1 and type 2 diabetes, wecompared the LBGI, HBGI, and ADRR toestablished measures of GV, includingSD, MAGE, and others (38). All GV met-rics were computed from a month ofSMBG data and were then fixed andtested as predictors of low (BG ,40mg/dL and BG ,70 mg/dL) and high(BG .180 mg/dL and BG .400 mg/dL)events in the subsequent 3months. Cor-relations between GV metrics and sub-sequent extreme glycemic events aregiven in Table 3. Table 3 illustrates sev-eral effects of GV measurement:

1. Metrics that are based on nonsym-metrized BG readings are predictiveprimarily of hyperglycemia and ac-count poorly for future hypoglycemicepisodes (SDofBG,M-value, andMAGE).

2. The predictive power of these metricsfor future extreme hypoglycemia is sim-ilar to that of average BG and HbA1c, afinding consistent with the DiabetesControl and Complications Trial (DCCT),which concluded that only about 7% ofsevere hypoglycemic episodes can beaccounted for by HbA1c (6).

3. The ADRR, which is based on symme-trization of the BG scale, balances theprediction of hypo- andhyperglycemia.

4. Separate independent assessmentsof the risks for low and high BG ex-tremes are provided by the LBGI and

Table 2—Metrics of the timing of GV

Metric Formula Meaning

SD of the BGrate of change SD ¼

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi∑ðRi 2 RÞ2

n2 1

s,

where Ri ¼ xðiÞ2 xði2 1ÞDt

is the BG rate of changeat time (i) and R is theaverage BG rate of change

Variation of the speed of BGfluctuation; particularly applicableto BG readings that are equally spacedin time, such as CGM data

MODD∑tk

t¼t1

jxðtÞ2 xðt2 24hÞjk

,

where k is the numberof available data pairs24-h apart

Intraday GV computed from all24-h time intervals where pairedreadings are available at the beginningand at the end of 24 h

CONGAffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi∑tk

t¼t1

ðDðtÞ2DÞ2k2 1

s,

where D(t) is thedifference betweenBG at time t and BGtaken n hours earlier,and D is the average ofthese differences

MAGE exceeding the SD of GV

In all formulas, x = BG level.

Figure 3—VGA. A split of a population of 335 below the median ADRR (A) and above the medianADRR (B) illustrated by the VGA. Each data point has BG coordinates (minimum, maximum) for asubject during the observation period. Subjects in A are in good control with most of the datapoints (70.4%) in the A- and B-zones; subjects in B are in poor control with most of the datapoints (89.6%) in the C-, D-, and E-zones.

506 Glucose Variability and Risk Analysis of Diabetes Diabetes Care Volume 39, April 2016

HBGI, which is consistent with theobservation that “risks associatedwith hypoglycemia are different fromthose associated with hyperglycemiain type, timing, and predictability,and they have no interaction” (42).

Example 2: Effect of IsletTransplantation on GVTo illustrate the effect of treatment ob-served via CGM, we used previouslypublished 72-h glucose traces observedbefore and 4 weeks after islet transplan-tation (45). Analysis of the BG rate ofchange (measured in mg/dL/min) de-rived from CGM data was suggestedas a way to evaluate the dynamics ofBG fluctuations on the time scale of min-utes (41,45). A larger variation of the BG

rate of change indicates rapid and morepronounced BG fluctuations and there-fore a less stable system. To illustratethis point, Fig. 5A and B presents histo-grams of the distribution of the BG rateof change over 15 min. It is apparentthat the pretransplantation distributionis more widespread than the distribu-tion posttransplantation. Numerically,this effect is reflected by the SD of theBG rate of change, which was reducedfrom 1.58 to 0.69 mg/dL/min, and by19.3% of BG rates outside of the [-2, 2]mg/dL/min range in Fig. 5A versusonly 0.6% BG rates outside that rangein Fig. 5B (45).

The same effect was also captured bythe Poincaré plots in Fig. 4B and C. Theseplots present data for the same patient

before and after islet transplantation,confirming that the surgical treatmenthas stabilized the metabolic system ofthis person (45).

Example 3: Effects of MedicationTreatment

Pramlintide

SMBG records were analyzed from arandomized, double-blind, placebo-controlled study of the effects of pram-lintide on intensively treated patientswith type 1 diabetes. Two groupsdpramlintide (n = 119) or placebo (n =129)dwere matched by age, sex, andbaseline HbA1c. Compared with placebo,pramlintide significantly attenuated thepreprandial to postprandial BG rate ofchange (F = 83.8, P , 0.0001). Conse-quently, in pramlintide-treated pa-tients, both the average postmeal BG(151.2 vs. 174.6 mg/dL) and the post-prandial HBGI (4.8 vs. 7.1) were signif-icantly lower than in the patients givenplacebo (P, 0.0001). The reduction inpostprandial hyperglycemia in pram-lintide-treated patients occurred with-out increased risk for preprandialhypoglycemia as quantified by the LBGI(47).

Exenatide

In this study, variability analyses wereapplied to SMBG data from patientswith type 2 diabetes suboptimally con-trolled with oral therapy plus exenatideor insulin glargine as a next therapeuticstep. Exenatide-treated (n = 282) andinsulin glargine–treated (n = 267)

Figure 4—Poincaré plot of BG fluctuations depicting system dynamics. A: Plot of perfectly cyclic glucose fluctuations, an unrealistic but illustrativescenario. B: Plot of unstable BG fluctuations of a person in poor control. C: Plot of stable BG fluctuations of a person in good control.

Table 3—Correlations of variability measures with subsequent extreme glycemicevents

Metric of average glycemia andGV computed from SMBGduring month 1 of study

Correlation of metric with the number of extremeBG episodes observed during months 2–4 of study*

BG ,40 mg/dL BG .400 mg/dL

HBA1c (baseline laboratorymeasurement) 20.04 0.58

Mean BG 20.15 0.58

SD of BG 0.15 0.58

M-value 0.16 0.68

MAGE 0.17 0.56

LBGI 0.59 20.06

HBGI 20.05 0.65

ADRR 0.44 0.45

*At this sample size, correlations above 0.4 were considered significant at P = 0.001, which wasan appropriate significance level accounting for themulticollinearity of the consideredmeasures(39).

care.diabetesjournals.org Kovatchev and Cobelli 507

patients were matched by age, HbA1c,BMI, and duration of diabetes. GV wasreduced on exenatide compared withglargine as indicated by the ADRR(mean 6 SEM, 16.33 6 0.45 vs.18.54 6 0.49, P = 0.001), LBGI, andHBGI. The study concluded that al-though average glycemic control wassimilar for both treatment groups, exe-natide minimized GV and the risk forglycemic extremes to a greater degreethan glargine (48).

Lixisenatide

A recently presented analysis evaluatedthe impact on GV of lixisenatide (LIXI)versus placebo as an add-on to basal in-sulin in type 2 diabetes. Data from 1,198patients (665 LIXI, 533 placebo) indi-cated significant GV decrease on LIXI: SDdecrease on LIXI was 28.28 mg/dLversus 23.89 mg/dL on placebo, P =0.003; MAG decreased 26.80 versus21.48 mg/dL, P , 0.001; MAGE de-creased 216.02 versus 27.02 mg/dL,P = 0.003; and the HBGI decreased23.65 versus 20.88, P , 0.001. TheLBGI remained unchanged, indicatingno increase in the risk for hypoglycemiaas a result from treatment. The studyconcluded that when added to insulinLIXI significantly reduced GV and post-prandial glucose excursions without in-creased risk of hypoglycemia (49).

CONCLUSIONS

Diabetes control is all about optimiza-tion and a balance, or “trade-off,” be-tween two key markers: frequency of

hypoglycemia and HbA1c, which reflectsaverage BG and is primarily driven bythe duration and the extent of hypergly-cemia (8). As shown by several studies,GV is a primary barrier to this optimiza-tion. However, although GV has richerinformation content than just averageglucose (HbA1c), its quantitative assess-ment is not straightforward becauseglucose fluctuations carry two compo-nents: amplitude and timing.

The standard assessment of GV ismeasuring amplitude. However, whenmeasuring amplitude we should bemindful that deviations toward hypo-glycemia are not equal to deviationstoward hyperglycemiada 20 mg/dLdecline in BG levels from 70 to 50 mg/dLis clinically more important than a20 mg/dL raise of BG from 160 to180 mg/dL. We explained how to fixthat with a well-established rescaling oftheBGaxis introducedmore than15yearsago (34). For convenience and becausethe risks for hypo- and hyperglycemiaare clinically independent (42), wewouldalso suggest using two independent in-dices that have been specifically de-signed to account for the variation inthe low BG range (LBGI [22]) and highBG range (HBGI [24]) (see example 2).

In addition, we should be mindful ofthe timing of BG fluctuations. Thereare a number of measures assessingGV amplitude from routine SMBG, butthe timing of readings is frequently ig-nored even if the information is avail-able (42). Yet, contrary to widespreadbelief, BG fluctuations are a process in

time and the speed of transition fromone BG state to another is of clinicalimportance. With the availability ofCGM, the assessment of GV timing be-came not only possible but also required(32). Responding to this necessity, weshould keep in mind that the assess-ment of temporal characteristics of GVbenefits from mathematical computa-tions that go beyond basic arithmetic.Thus, some assistance from the theoryand practice of time series and dynami-cal systems analysis would be helpful.Fortunately, these fields are highly de-veloped, theoretically and computation-ally, and have been used for decades inother areas of science (e.g., engineering,physics, meteorology, and environmen-tal science, to mention a few). The com-putational methods are standardizedand available in a number of softwareproducts and should be used for theassessment of GV.

We should emphasize that diabetesresearch and clinical practice are nostrangers to, and have benefited from,advanced mathematical modeling andcomputing methods (50). For instance,the minimal model of glucose–insulindynamics has been used for over 30years to assess insulin sensitivity andb-cell function (51). In this Perspective,we argue that techniques known inmathematics for decades can be verybeneficial to the understanding of glu-cose metabolism in diabetes and inhealth. There is no doubt that the timingof glucose fluctuations is clinically im-portant, but there is a price to pay for

Figure 5—Histograms of the BG rate of change before (A) and after (B) islet transplantation, illustrating increased system stability after theprocedure.

508 Glucose Variability and Risk Analysis of Diabetes Diabetes Care Volume 39, April 2016

its accurate assessmentda bit higherlevel of mathematical complexity. This,however, should not be a deterrent. Ap-propriate techniques and expertise areincreasingly available in the diabetesfield, and these resources should beused for the common good. Examples3 and 4 illustrate the importance of tim-ing components for the assessment ofdiabetes treatments.Last but not least, we should note

that a modern field of increased re-search and industrial effortsdthe artifi-cial pancreas or the closed-loop controlof diabetesdwould not be possiblewithout real-time assessment of GVand real-time reaction to glucose fluctu-ations. Thus, we should put to rest thedispute of whether GV deserves atten-tion and should focus on standardizing,via scientific consensus, themethods forthe assessment of its most importantcomponentsdamplitude (Table 2) andtiming (Table 3).

Funding. This study was supported by NationalInstitutes of Health/National Institute of Diabe-tes and Digestive and Kidney Diseases projectR01 DK051562 “Biobehavioral triggers, mecha-nisms, and control of glucose variability inT1DM” and by the Seventh Framework Pro-gramme of the European Union project600914 “Models and simulation techniquesfor discovering diabetes influence factors”(MOSAIC).Duality of Interest. B.K. has received researchsupport from Amylin Pharmaceuticals, AnimasCorp., BectonDickinson, Dexcom, Insulet, RocheDiagnostics, Sanofi, and Tandem Diabetes Care;holds patents related to metrics of GV in di-abetes; and is a shareholder in InSpark Technol-ogies and TypeZero Technologies. C.C. hasreceived research support from Amylin Pharma-ceuticals, Dexcom, Insulet, Roche Diagnostics,and Sanofi. No other potential conflicts of in-terest relevant to this article were reported.

References1. Lachin JM, Genuth S, Nathan DM, Zinman B,Rutledge BN; DCCT/EDIC Research Group. Effectof glycemic exposure on the risk of microvascu-lar complications in the Diabetes Control andComplications Trial–revisited. Diabetes 2008;57:995–10012. Santiago JV. Lessons from the Diabetes Con-trol and Complications Trial. Diabetes 1993;42:1549–15543. The Diabetes Control and Complications TrialResearch Group. The effect of intensive treatmentof diabetes on the development and progres-sion of long-term complications in insulin-dependent diabetes mellitus. N Engl J Med1993;329:977–9864. UK Prospective Diabetes Study (UKPDS)Group. Intensive blood-glucose control with

sulphonylureas or insulin compared with con-ventional treatment and risk of complicationsin patients with type 2 diabetes (UKPDS 33). Lan-cet 1998;352:837–8535. American Diabetes Association Workgroupon Hypoglycemia. Defining and reporting hypo-glycemia in diabetes. Diabetes Care 2005;28:1245–12496. The Diabetes Control and Complications TrialResearch Group. Hypoglycemia in the DiabetesControl and Complications Trial. Diabetes 1997;46:271–2867. Cryer PE. Hypoglycaemia: the limiting fac-tor in the glycaemic management of type I andtype II diabetes. Diabetologia 2002;45:937–9488. Cryer PE. Glycemic goals in diabetes: trade-off between glycemic control and iatrogenic hy-poglycemia. Diabetes 2014;63:2188–21959. McCall AL, Kovatchev BP. The median is notthe only message: a clinician’s perspective onmathematical analysis of glycemic variabilityand modeling in diabetes mellitus. J DiabetesSci Technol 2009;3:3–1110. Hirsch IB. Glycemic variability and diabetescomplications: does it matter? Of course itdoes! Diabetes Care 2015;38:1610–161411. Bergenstal RM. Glycemic variability and di-abetes complications: does it matter? Simplyput, there are better glycemic markers! Diabe-tes Care 2015;38:1615–162112. Fernstrom JD. Branched-chain amino acidsand brain function. J Nutr 2005;135(Suppl.):1539S–1546S13. Gelling RW, Overduin J, Morrison CD, et al.Effect of uncontrolled diabetes on plasma ghre-lin concentrations and ghrelin-induced feeding.Endocrinology 2004;145:4575–458214. German JP, Wisse BE, Thaler JP, et al. Leptindeficiency causes insulin resistance induced byuncontrolled diabetes. Diabetes 2010;59:1626–163415. Gerich JE, Langlois M, Noacco C, Karam JH,Forsham PH. Lack of glucagon response to hy-poglycemia in diabetes: evidence for an intrinsicpancreatic alpha cell defect. Science 1973;182:171–17316. Amiel SA, Sherwin RS, Simonson DC,TamborlaneWV. Effect of intensive insulin ther-apy on glycemic thresholds for counterregu-latory hormone release. Diabetes 1988;37:901–90717. Dagogo-Jack SE, Craft S, Cryer PE. Hypo-glycemia-associated autonomic failure in insulin-dependent diabetes mellitus. Recent antecedenthypoglycemia reduces autonomic responses to,symptoms of, and defense against subsequent hy-poglycemia. J Clin Invest 1993;91:819–82818. Amiel SA, Tamborlane WV, Simonson DC,Sherwin RS. Defective glucose counterregulationafter strict glycemic control of insulin-dependentdiabetes mellitus. N Engl J Med 1987;316:1376–138319. Cryer PE, Gerich JE. Glucose counterregula-tion, hypoglycemia, and intensive insulin ther-apy in diabetes mellitus. N Engl J Med 1985;313:232–24120. White NH, Skor DA, Cryer PE, LevandoskiLA, Bier DM, Santiago JV. Identification oftype I diabetic patients at increased risk for hy-poglycemia during intensive therapy. N Engl JMed 1983;308:485–491

21. Cryer PE. Hypoglycemia: Pathophysiology,Diagnosis, and Treatment. Oxford UniversityPress, 199722. Kovatchev BP, Cox DJ, Gonder-Frederick LA,Young-Hyman D, Schlundt D, Clarke W. Assess-ment of risk for severe hypoglycemia amongadults with IDDM: validation of the low bloodglucose index. Diabetes Care 1998;21:1870–187523. Kovatchev BP, Cox DJ, Farhy LS, StraumeM,Gonder-Frederick L, ClarkeWL. Episodes of severehypoglycemia in type 1 diabetes are preceded andfollowed within 48 hours by measurable distur-bances in blood glucose. J Clin Endocrinol Metab2000;85:4287–429224. Kovatchev BP, Cox DJ, Kumar A, Gonder-Frederick L, Clarke WL. Algorithmic evaluationof metabolic control and risk of severe hypogly-cemia in type 1 and type 2 diabetes using self-monitoring blood glucose data. Diabetes TechnolTher 2003;5:817–82825. Cox DJ, Gonder-Frederick L, Ritterband L,Clarke W, Kovatchev BP. Prediction of severehypoglycemia. Diabetes Care 2007;30:1370–137326. Pitsillides AN, Anderson SM, Kovatchev B.Hypoglycemia risk and glucose variability indi-ces derived from routine self-monitoring ofblood glucose are related to laboratory mea-sures of insulin sensitivity and epinephrinecounterregulation. Diabetes Technol Ther 2011;13:11–1727. Rodbard D. New and improved methods tocharacterize glycemic variability using continu-ous glucose monitoring. Diabetes Technol Ther2009;11:551–56528. Schlichtkrull J, Munck O, Jersild M. The M-value, an index of blood-sugar control in dia-betics. Acta Med Scand 1965;177:95–10229. Service FJ, Molnar GD, Rosevear JW,Ackerman E, Gatewood LC, Taylor WF. Meanamplitude of glycemic excursions, a measure ofdiabetic instability. Diabetes 1970;19:644–65530. Ryan EA, Shandro T, Green K, et al. Assess-ment of the severity of hypoglycemia and gly-cemic lability in type 1 diabetic subjectsundergoing islet transplantation. Diabetes2004;53:955–96231. Hermanides J, Vriesendorp TM, Bosman RJ,Zandstra DF, Hoekstra JB, Devries JH. Glucosevariability is associated with intensive careunit mortality. Crit Care Med 2010;38:838–84232. DeVries JH. Glucose variability: where it isimportant and how to measure it. Diabetes2013;62:1405–140833. Fabris C, Facchinetti A, Sparacino G, et al.Glucose variability indices in type 1 diabetes:parsimonious set of indices revealed by sparseprincipal component analysis. Diabetes TechnolTher 2014;16:644–65234. Kovatchev BP, Cox DJ, Gonder-Frederick LA,Clarke W. Symmetrization of the blood glucosemeasurement scale and its applications. Diabe-tes Care 1997;20:1655–165835. Kovatchev BP, StraumeM, Cox DJ, Farhy LS.Risk analysis of blood glucose data: a quantita-tive approach to optimizing the control of insu-lin dependent diabetes. J Theor Med 2000;3:1–1036. McCall AL, Cox DJ, Crean J, Gloster M,Kovatchev BP. A novel analytical method forassessing glucose variability: using CGMS in

care.diabetesjournals.org Kovatchev and Cobelli 509

type 1 diabetes mellitus. Diabetes Technol Ther2006;8:644–65337. Kovatchev BP, Cox DJ, StraumeM, Farhy LS.Association of self-monitoring blood glucoseprofiles with glycosylated hemoglobin in pa-tients with insulin-dependent diabetes. Meth-ods Enzymol 2000;321:410–41738. Fabris C, Patek SD, Breton MD. Are risk in-dices derived from CGM interchangeable withSMBG-based indices? J Diabetes Sci Technol2015;10:50–5939. Kovatchev BP, Otto E, Cox D, Gonder-Frederick L, Clarke W. Evaluation of a new mea-sure of blood glucose variability in diabetes.Diabetes Care 2006;29:2433–243840. Magni L, Raimondo DM,Man CD, et al. Eval-uating the efficacy of closed-loop glucose regu-lation via control-variability grid analysis. JDiabetes Sci Technol 2008;2:630–63541. Clarke W, Kovatchev B. Statistical toolsto analyze continuous glucose monitor data.

Diabetes Technol Ther 2009;11(Suppl. 1):S45–S5442. Service FJ. Glucose variability. Diabetes2013;62:1398–140443. Rodbard D. The challenges of measuringglycemic variability. J Diabetes Sci Technol2012;6:712–71544. McDonnell CM, Donath SM, Vidmar SI,Werther GA, Cameron FJ. A novel approach tocontinuous glucose analysis utilizing glycemic var-iation. Diabetes Technol Ther 2005;7:253–26345. Kovatchev BP, Clarke WL, Breton M,Brayman K, McCall A. Quantifying temporal glu-cose variability in diabetes via continuous glu-cose monitoring: mathematical methods andclinical application. Diabetes Technol Ther2005;7:849–86246. Kovatchev BP, Shields D, Breton M. Graph-ical and numerical evaluation of continuous glu-cose sensing time lag. Diabetes Technol Ther2009;11:139–143

47. Kovatchev BP, Crean J, McCall A. Pramlin-tide reduces the risks associated with glucosevariability in type 1 diabetes. Diabetes TechnolTher 2008;10:391–39648. McCall AL, Cox DJ, Brodows R, Crean J,Johns D, Kovatchev B. Reduced daily risk of gly-cemic variability: comparison of exenatide withinsulin glargine. Diabetes Technol Ther 2009;11:339–34449. Umpierrez GE, DiGenio A, Goldenberg R,et al. Lixisenatide added to basal insulin reducesglycemic variability in T2DM patients (Abstract).Diabetes 2014;63(Suppl. 1):A26050. Cobelli C, Man CD, Pedersen MG, BertoldoA, Toffolo G. Advancing our understanding ofthe glucose system via modeling: a perspec-tive. IEEE Trans Biomed Eng 2014;61:1577–159251. Cobelli C, Dalla Man C, Toffolo G, Basu R,Vella A, Rizza R. The oral minimal model method.Diabetes 2014;63:1203–1213

510 Glucose Variability and Risk Analysis of Diabetes Diabetes Care Volume 39, April 2016