Respective Contributions of Glycemic Variability and Mean Daily Glucose as Predictors of Hypoglycemia in Type 1 Diabetes: Are They Equivalent?

It is well known that fear of hypoglycemia is a main barrier to optimizing insulin therapy in persons with type 1 diabetes (T1DM) (1,2); they prefer adhering to less stringent glycemic targets rather than experiencing severe hypoglycemic episodes that can be associated with acute and chronic cardiovascular complications (3–5). The seminal Diabetes Control and Complications Trial (DCCT) clearly demonstrated that intensive treatment in persons with T1DM resulted in higher rates of severe hypoglycemia than occurred in those with less well-controlled diabetes when expressed in terms of glycated hemoglobin (HbA_1c) (6). The hypothesis that abnormally high daily glucose fluctuation, between peaks and nadirs, is also associated with a higher frequency of hypoglycemic episodes (7,8) has been confirmed in both observational and interventional studies (9–14), and it has been mathematically modeled by Rodbard (15). Consequently, from a “glucocentric” viewpoint, when managing persons with T1DM, therapeutic targets should include ambient hyperglycemia, time in range (16–20), short-term glycemic variability (21–23), and the threshold for hypoglycemia (24,25). Although it is highly likely that mean daily glucose (MDG) concentration and glycemic variability are the two main explanatory factors that contribute to an increased risk of hypoglycemia, their respective contributions remain to be adequately investigated. This observational and retrospective study of subjects with T1DM who were undergoing continuous glucose monitoring (CGM) aims to determine the relative priority of MDG concentration and glycemic variability (coefficient of variation for glucose [%CV]) in determining the time spent with glucose below 3.9 and 3.0 mmol/L (17–20), levels that represent hypoglycemia.

After an initial screening of 163 persons with T1DM who underwent 3 days of blinded CGM, a total of 100 persons were entered into the study, of whom 88 were receiving multiple daily insulin injections and only 12 were receiving continuous subcutaneous insulin infusion (CSII). The relatively low percentage of subjects treated with CSII is simply a result of the study being conducted between 2006 and 2012, when CSII was used less than it is today. In all patients treated with multiple insulin injections, the basal insulin was glargine U100 administered before the evening meal. Pregnant patients and those younger than 20 years of age, those who had recently experienced an illness, and those who had been treated with steroids during the 3-month period preceding the investigation were excluded. All participants (65 men and 35 women) attended the outpatient clinic of the University Hospital of Montpellier (France) for more than 3 years and had a diabetes duration of 5 years or more (mean duration of diabetes, 28 years). In addition, their respective insulin treatments were stable for a minimum of 3 months before the investigation. CGM was conducted on an ambulatory basis over 3 days. Patients were excluded from the final analysis if the glucose monitoring was unexpectedly disrupted, if an insufficient number of capillary blood tests were obtained, or if calibrations were inappropriate. As recommended by the manufacturer of the CGM technology (the second-generation MiniMed System; Medtronic, Northridge, CA), glucose was determined four times daily from capillary blood via fingersticks. An acceptable calibration was based on an accuracy criterion defined by a correlation coefficient >0.79 between paired readings. As mentioned above, all glycemic profiles were monitored on an ambulatory basis, and all CGM system sensors were inserted by trained health care professionals at the diabetes outpatient clinic. The study was conducted in accordance with the Declaration of Helsinki; each participant gave oral consent in accordance with a European directive (26). No ethics committee approval was required because of the noninterventional design of the study.

The three consecutive days of CGM avoided a weekend; the CGM sensor was inserted on day 0 (before 1200 h) and removed on day 3 at the same time of day. Chronic hyperglycemia (based on HbA_1c) was assessed at baseline (study day 0) by using a high-performance liquid chromatography assay (Menarini Diagnostics, Florence, Italy) (27).

The data recorded on the day the glucose sensor for GGM was inserted were excluded from the analysis in order to avoid any bias due to insufficient glucose stabilization between the sensor and the interstitial fluid during the first 24 h after insertion of the device. On the basis of two validated 24-h glycemic profiles (calculated from glucose readings obtained at 5-min intervals on study days 1 and 2), the total MDG, short-term glycemic variability (%CV), and the presence of hypoglycemia (both symptomatic and asymptomatic) were assessed, in addition to the cumulative time spent below three glycemic thresholds: 3.0, 3.45, and 3.9 mmol/L. Because 100 patients were included in this study, and because two 24-h glycemic profiles were calculated for each patient, 200 glycemic profiles were suitable for analysis. The %CV was derived from the following equation: ([SD of the 24-h mean glucose value]/[24-h mean glucose]) × 100. In accordance with our own previous findings (11), international consensus on the use of CGM has adopted the %CV as a reference to separate stable from labile diabetes (17). The %CV has the main advantage of being the simplest metric for assessing short-term within-day glucose variability, and it does not need to be adjusted for the MDG value. Because of its characteristics (28), the low blood glucose index was not used in this study because it is necessarily highly correlated with the incidence of hypoglycemia (29); thus, it is more a marker for use in evaluating the risk of hypoglycemia and less a potential causative factor for explaining the occurrence of hypoglycemic episodes. According to the recommendations of the International Hypoglycaemia Study Group (24), we considered glucose concentrations <3.0 mmol/L detected by CGM as unequivocally indicating the presence of clinical hypoglycemia. A glucose value <3.9 mmol/L from plasma venous samples, however, serves as an alert threshold (25). Because many uncertainties exist concerning the agreement between plasma venous glucose values and those recorded by CGM, we decided to determine the duration of time for which glucose was below 3.0 and 3.9 mmol/L, with an intermediate glucose threshold set at 3.45 mmol/L, to indicate the presence or absence of hypoglycemia. The time spent below the selected glucose thresholds was recorded only when the duration exceeded 15 min, that is, when three consecutive measurements, taken at 5-min intervals, indicated a glucose value below the threshold.

For descriptive analysis, variables were expressed as the mean ± SD or the median (interquartile range [IQR]). Groups were compared by using either the Student t test or the Wilcoxon-Mann-Whitney rank sum test, according to whether the data were normally distributed.

The statistical analysis was oriented toward the goodness of fit for identifying and investigating the contributions of the independent (explanatory or predictor) variables that can explain or predict the time spent below the three glucose thresholds (dependent or outcome variables). Because the time spent below the thresholds was computed over a 24-h interval, all parameters recorded throughout that period (the 24-h mean glucose concentration, within-day glucose variability, and the daily amount of insulin injected) were selected as independent variables. To this preliminary list were added two other variables: age and BMI; both can play a potential role in determining the amount of time spent below the glycemic thresholds.

A univariate analysis was conducted by calculating the simple linear regression between the dependent variables (Y) and each potential explanatory variable (X). This statistical procedure was used for removing the independent variables that did not enter the model. Statistical significance, given as a P value, was calculated after the time spent below glucose thresholds (normally expressed as percentage from 0 to 100%) had been converted into the arcsine of the square root of the dependent variable (30). This transformation was applied to the dependent variables to obtain an underlying distribution in agreement with the assumption of model linearity. In addition, the calculations used linear mixed-effect models because two values for the dependent variables were obtained from each patient over two consecutive days of glucose monitoring. The relationship between the independent and dependent variables was expressed as a simple linear regression function: Y = βiX + α, in which each partial regression coefficient (βi) was an estimate of the relationship between the dependent variable (Y) and one given explanatory variable (X). To reflect the initial units (Y, expressed as a percentage), the βi coefficients were transformed back to proportions and then expressed as unitless standardized coefficients that provide indications of relative importance of the explanatory variables (X) in determining the value of the dependent variable (Y).

To estimate the respective influence of each remaining explanatory variable on the dependent variable, a sensitivity analysis was performed by using an arborescent regression (classification and regression tree [CART]) (31). The method with this approach includes all potential explanatory variables at the top of the regression tree and sequentially selects those with the best splitting in order to establish a hierarchy, allowing predictions of their role as primary or secondary key players in the expression of the dependent variable. During the first step of this top-down approach, the data for the independent variable that showed the best sensitivity were separated into two homogeneous subsamples; a cutoff (node) value served, via a pruning procedure, either to reject the data exhibiting a poor predictive value or to maintain those exhibiting a high predictive value. The latter subsample with the highest predictive value underwent the same procedure, splitting into two subsamples, which ended when the independent variable showed its lowest sensitivity.

To gain further insight into the best fit between the dependent variables and the value for each explanatory variable entering the model, the dependent variables were transformed into ln. The relationships (rearranged as increasing or decreasing simple exponential curves, as appropriate, after reverse transformation of the ordinate scale of ln) between dependent and independent variables were statistically tested by using coefficients of correlation (r values). The r values were compared after they had been converted into their corresponding Fisher z values.

Finally, in the subset of 24-h glycemic profiles exhibiting MDG concentrations ≤7.8 mmol/L, which correspond to estimated HbA_1c levels <6.5% (<48 mmol/mol) (32)—that is, to near-normal glycemic control (33)—the time spent below 3.0 mmol/L was divided into tertiles by using the Wilcoxon-Mann-Whitney rank sum test for unpaired, asymmetrically distributed data; these tertile values allowed for further statistical comparisons between medians and IQRs.

Analyses were performed with the R software package, version 3.5.0 (The R Foundation [www.r-project.org]).

In the population considered as a whole (100 patients, 200 glycemic profiles), the mean ± SD and range for each of the main parameters are described as follows: HbA_1c, 8.27% ± 0.99% (67 ± 10.9 mmol/mol), 5.3–10.9% (34.3–95.9 mmol/mol) (n = 100); average MDG concentrations, 9.0 ± 2.4 mmol/L, 4.4–17.8 mmol/L (n = 200); %CV, 37.9% ± 11.5%, 82.0–94.8% (n = 200); daily insulin dose, 0.67 ± 0.20 units/kg/day, 0.37–1.21 units/kg/day (n = 100); age, 55.7 ± 13.1 years, 20.0–83.0 years (n = 100); BMI, 24.8 ± 3.1 kg/m², 19.1–34.9 kg/m² (n = 100).

The 200 24-h glycemic profiles recorded among the 100 participants were divided into two categories according to whether no time or a positive amount of time was spent below the selected glucose thresholds of 3.0, 3.45, and 3.9 mmol/L (Table 1). As expected, the number of glycemic profiles with no time spent below the selected thresholds was smaller when the glucose threshold was set at 3.0 mmol/L rather than 3.9 mmol/L. MDG concentration was significantly lower, and %CV significantly higher, when glycemic profiles with glucose values below the thresholds were compared with those profiles with no time spent below each of the three glucose thresholds. Daily insulin dose, age, and BMI did not differ between the groups, irrespective of whether time was spent below the glucose thresholds.

When the time spent below the selected thresholds (the dependent variables [Y]) was expressed as its arcsine transformation (i.e., arcsin√Y), the univariate mixed regression analysis showed strong positive relationships (P < 0.001) with the %CV and strong negative relationships (P < 0.001) with MDG concentration. The unitless standardized partial coefficients of regression (βi) for the thresholds (3.9, 3.45, and 3.0 mmol/L) were positive for the %CV (β₁ = +1.31, +1.08, and +0.75, respectively) and negative for MDG (β₂ = −1.07, −0.58, and −0.35, respectively), indicating that the time spent below the selected glucose threshold increased as %CV increased and as MDG concentrations decreased. In contrast, the remaining independent variables—daily insulin dose, age, and BMI—did not correlate with the time spent below the three glucose thresholds. When the thresholds were set at 3.9, 3.45, and 3.0 mmol/L, the respective P values were 0.92, 0.91, and 0.60 for daily insulin dose; 0.25, 0.16, and 0.24 for age; and 0.17, 0.13, and 0.07 for BMI. As a consequence, the latter three variables were no longer considered to be explanatory and were thus removed from the model used in statistical analysis.

Because the %CV and MDG are both highly statistically significant (P < 0.001), we used the CART method to decipher whether one is more significant than the other. Different results were obtained when the three selected thresholds were tested (Fig. 1). At 3.9 mmol/L, MDG was the first explanatory variable, but when the glucose threshold was set at 3.0 or 3.45 mmol/L, the %CV appeared as the primary factor associated with a higher risk of hypoglycemia (Fig. 1). When the threshold was set at 3.0 mmol/L, the %CV was the first variable that entered the top of the regression tree. The first node of %CV for partitioning the tree was found at 47.3%. The subsample with the %CV below 47.3% was subsequently pruned from the regression tree, and in that subsample the mean terminal value of the time spent below 3.0 mmol/L was 1.7%. The subsample with a %CV ≥47.3% was retained and split further. As expected, the second explanatory variable entering the regression tree was MDG, with a splitting node at 6.8 mmol/L. The subsamples with MDG ≥6.8 mmol/L or <6.8 mmol/L ended with mean times spent below 3.0 mmol/L of 6.8 and 26.0%, respectively, as indicated at the bottom of the terminal branches.

The exponential relationships are illustrated in Fig. 2A and B. The relationship between the percentage of time spent below 3.0 mmol/L (54 mg/dL) (dependent variable [Y]) and MDG (independent variable [X], expressed as milligrams per deciliter) was described by a simple decreasing exponential curve: Y = 23.66 e^−0.009X (r = −0.358, P < 0.001). When Y was plotted against the %CV (X), the relationship was represented by a simple increasing exponential curve: Y = 0.93 e^0.043X (r = 0.509, P < 0.001). A comparison between the two coefficients of correlation did not show any significant difference after testing the corresponding z values after a Fisher transformation (z = 1.87, P = 0.065), even though the r value for %CV seemed to be slightly better than that for MDG.

Among the 200 daily glycemic profiles, 65 exhibited mean glucose concentrations ≤7.8 mmol/L. After dividing that subset on the basis of %CV, three tertiles were selected: %CV <34% (n = 21) (tertile 1); %CV between 34 and 44.1% (n = 22) (tertile 2); and %CV >44.1% (n = 22) (tertile 3). As illustrated in Fig. 3, percentages of time spent below 3.0 mmol/L (median [IQR]) were statistically significantly higher in tertile 3 (15.6% [20.5%]) than in tertile 1 (0.0% [2.7%]; P < 0.001) and tertile 2 (3.1% [5.5%]; P < 0.01). No difference was found between tertiles 1 and 2.

Several key messages can be drawn from this study. First, the two main factors for predicting risk of hypoglycemia in persons with T1DM are low or near-normal daily glucose levels and abnormally high glucose fluctuations from peaks to nadirs. Second, the role of these two factors was approximately equivalent across the range of glucose thresholds from 3.9 to 3.0 mmol/L, even though excessive within-day glucose fluctuations seemed to be increasingly involved in predicting the risk of hypoglycemic episodes and in precipitating their onset (with a lower glucose threshold used to define hypoglycemia). In addition, the coefficient of correlation for the exponential curve depicting the relationship between the time spent below 3.0 mmol/L and the %CV was greater than that for the MDG value (as an explanatory variable), even though the differences were not statistically significant. Similar results derived from a less sophisticated mathematical analysis were found for the %CV; these results were reported recently at the 78th Scientific Sessions of the American Diabetes Association (34). Nevertheless, all these results reinforce the opinion that, in addition to achieving near-normal glycemia for preventing the development or progression of micro- and macrovascular complications (6,35–38), it is crucial to try to reduce as much as possible the magnitude of glucose fluctuations in order to limit the risk of hypoglycemia, a main challenge for improving the quality of life of persons with T1DM (7–14).

This position is supported by two additional findings. First, in all groups, the mean %CV was below 36% (the limit that separates stable from labile diabetes [11,17]) when the patients did not spend any time below the selected thresholds—whatever the values chosen (from 3.9 to 3.0 mmol/L). By contrast, the mean %CV was always above 36% when the patients spent a positive amount of time above the aforementioned glucose thresholds. Second, the computation by tertiles of %CV in the subset with near-normal MDG values showed that the amount of time spent below 3.0 mmol/L was rarely positive when the %CV was less than 34%. Therefore, maintaining the %CV below this threshold should be a suitable recommendation for limiting the risk of hypoglycemia when the MDG concentration is stable and maintained below 7.8 mmol/L (a value that corresponds to near-normal glucose) (32,33). Even though the two thresholds of 36 and 34% are very close, they represent quite different things: a %CV above 36% is synonymous with frequent hypoglycemic episodes (labile diabetes), regardless of MDG concentration, whereas achieving a %CV below 34% ensures a very low risk of hypoglycemia when tight glycemic control is obtained in terms of overall glucose exposure. Until recently, only a few people with T1DM concomitantly attained such targets (low %CV and near-normal MDG concentration) with conventional insulin treatments consisting of either multiple injections daily or CSII through a pump (39). However, the expanded use of novel devices for CGM (17,40,41) and ceaseless progression toward the implementation of more sophisticated systems for closed-loop insulin delivery (42) both raise promising expectations for improving glucose homeostasis in persons with T1DM. Even though it is highly likely that a reduction in short-term glycemic variability is one of the key players in reducing the incidence of hypoglycemia, the observation of a strong relationship between %CV and the time spent below various glucose thresholds does not allow us to distinguish whether glucose variability is a cause or a consequence of the time spent below the recommended target glycemic range (usually set between 3.9 and 10 mmol/L) (17,18,20). It should be noted that this study included patients who were carefully instructed to ingest moderate amounts of refined carbohydrates in order to prevent excessive glycemic rebounds after periods of hypoglycemia and thereby avoid any subsequent additional deterioration of glucose variability. Furthermore, because the patients were blinded to their CGM data throughout the entire monitoring period, all episodes of asymptomatic hypoglycemia were not corrected with therapeutic interventional measures in the absence of any warning signs. Consequently, there are many reasons to consider excessive glucose variability as more of a causative factor than a consequence of hypoglycemia.

This study has some limitations. It has an observational design and a short duration, although all CGM was conducted on an ambulatory basis. Finally, the CGM sensors may have been inaccurate when glucose values were within the lower range.

In conclusion, to address the question raised in the title of this article, it seems that excess short-term glycemic variability, as assessed on the basis of the within-day %CV, is at least as important as—and perhaps slightly more important than—the MDG level in preventing the risk of hypoglycemia, especially when the threshold of hypoglycemia is set at the lowest level (3.0 mmol/L). In addition, striving to achieve a %CV below 34% in patients who already have satisfactory glycemic control in terms of chronic glucose exposure (MDG concentration ≤7.8 mmol/L) should be one of the main objectives in the management of T1DM.

Funding. This study was conducted with the help of academic funds provided by the University of Montpellier.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. L.M. designed the study, collected and interpreted data, and wrote the manuscript. A.W., C.C., E.R., and D.O. designed the study, interpreted data, and critically revised the manuscript. N.M. carried out the statistical analysis. L.M. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.