How Valid Are the New Hypoglycemia Definitions for Use in Clinical Trials?

The recent changes to hypoglycemia definitions, as proposed by the International Hypoglycaemia Study Group (IHSG) and widely accepted by some regulators and major diabetes associations, were poorly evidence-based, relying on pathophysiological studies and small clinical studies (1). In this issue of Diabetes Care, Heller et al. (2) attempt to remedy this, and they present some useful information on the effect of choice of level for hypoglycemia on statistical power and effect size when comparing insulins. The definitions are for use in clinical trials, as the IHSG emphasizes, and not for application to individual people with diabetes seen in clinical practice (1).

Broadly, using data from the SWITCH 1 (type 1 diabetes [T1D]) and SWITCH 2 (type 2 diabetes [T2D]) studies of insulin degludec versus insulin glargine 100 units/mL (3,4), the difference in event rates declines below the earlier American Diabetes Association (ADA) threshold of <3.9 mmol/L (<70 mg/dL), increasing in inverse proportion to the absolute decline in event rates, at least down to 2.8–2.9 mmol/L (50–52 mg/dL) where the number of events leads to uncertainty. This is best espoused in the authors' Fig. 2A, where for SWITCH 2 the central estimate for difference in rate of confirmed symptomatic events between insulins falls with each 0.1 mmol/L drop in threshold from 3.9 to 3.0 mmol/L (70 to 54 mg/dL). More importantly, the upper confidence limit also falls despite the overall confidence interval (CI) widening, meaning higher certainty that the difference is real. Below 3.0 mmol/L, the lower event rate leads to increasing uncertainty, with loss of statistical difference by 2.7 mmol/L (49 mg/dL).

The T1D study, where the majority of hypoglycemia events will be influenced by mealtime and not basal insulin but in which the event rate is much higher, appears confirmatory of the T2D study (authors' Fig. 2B), with a small improvement in hypoglycemia compared with control appearing when using a progressively lower threshold below 3.9 mmol/L down to 2.5 mmol/L (45 mg/dL). No similar data analysis for nocturnal hypoglycemia, which might be influenced more by a difference between two basal insulins, is given.

Notable also from Fig. 2 is that exclusion of asymptomatic hypoglycemia appears to improve discrimination at any threshold with no detriment to the upper confidence limit despite lower event rates.

Other data are available in the literature but can be difficult to integrate because symptomatic confirmed hypoglycemia is often combined with severe hypoglycemia (making sense from the patient perspective) and the threshold is sometimes 3.1 mmol/L (56 mg/dL), although that is perhaps not materially different from 3.0 mmol/L. However, the EDITION series of studies of insulin glargine 300 units/mL do present equivalent data using <3.0 and ≤3.9 mmol/L for documented symptomatic hypoglycemia. In T1D, no material effect on the central estimate versus control can be discerned, but the CIs are wider at the lower cutoff of <3.0 mmol/L (5). However, in a very small study, central estimates were lower using <3.0 mmol/L, and despite widening CIs the upper limit for nocturnal hypoglycemia then fell to 0.82, having been 1.11 at ≤3.9 mmol/L (6). In the same article, continuous glucose monitoring (CGM) numbers of values below the two cutoffs fell by 40%, but so did difference between insulins; power thus seems similar, but estimates and CIs are not given.

In a primary composite analysis of the T2D EDITION studies, the effect on overall event rate is dramatic, with an 83% drop in number of events using the lower cutoff (data for both insulins combined) and a halving of proportion of people affected (53% to 24%) (7). For documented symptomatic hypoglycemia, data are available for 1 year of exposure (8), and for incidence (proportion of people affected) as well as event rate, the former a more powerful statistic in people with T2D. Overall, however, in the two articles, whether for incidence or rates, and whether including asymptomatic or not, there are no consistent effects on central estimates, but statistical power is adversely affected by a lower hypoglycemia definition with wider CIs.

The BRIGHT study also presents data for ≤3.9 and <3.0 mmol/L thresholds in T2D (9). Underpowering for hypoglycemia is a problem for that study but particularly for the lower threshold where the CIs are about double (in logarithmic terms) those for ≤3.9 mmol/L, with the data presented being for all biochemically confirmed hypoglycemia.

Novo Nordisk's U.S. Food and Drug Administration prescribing information for insulin degludec (last revised November 2018) notes that the primary definition of hypoglycemia used in its registration trials was of episodes “with or without the presence of hypoglycemic symptoms” (10). While the IHSG article was not clear as to whether asymptomatic hypoglycemia should be included as “events” (1), and while the current article gives data both with and without such episodes, the authors here use “IHSG level 2” to mean all episodes, with and without symptoms (2). Presently, the U.S. Food and Drug Administration is not happy to accept comparative hypoglycemia data based on studies with variation in “intensity of glucose control, background therapies, and other intrinsic and extrinsic patient factors” (10). Accordingly, the SWITCH studies are not included in the current U.S. insulin degludec label. The subjective nature of self-measured plasma glucose, in particular in open-label studies, is also understood to be an issue, not least because the frequency of self-measurement will affect the frequency of detection of nonsymptomatic episodes and thus the proportions of data contributed by symptomatic events and biochemical detections.

Use of biochemically confirmed symptoms as given in the current report would appear to bring the definition closer to the patient experience, at a cost of missing possible “clinically significant” hypoglycemia, and, in some studies, losing useful statistical power. The IHSG is nevertheless further motivated to consider inclusion of nonsymptomatic detection by the need to give a definition for CGM-detected hypoglycemia (<3.0 mmol/L [<54 mg/dL] for 20 min) (1). Some other studies have addressed the issue of symptomatic and asymptomatic episodes by presenting data on both all documented hypoglycemia and documented symptomatic hypoglycemia (5,7,8), and going forward and with increasing use of CGM in randomized controlled trials (RCTs), this seems wise. Similarly, it seems sensible, as in the study by Heller et al., to present both a primary composite (clinically significant confirmed hypoglycemia or severe hypoglycemia) and its components (2).

The IHSG has not advised changing the long-established ADA definition of severe hypoglycemia (requiring assistance), though it is now referred to as “IHSG level 3” (1). The presentation of DEVOTE (Trial Comparing Cardiovascular Safety of Insulin Degludec Versus Insulin Glargine in Patients with Type 2 Diabetes at High Risk of Cardiovascular Events) data in the current article does not then add new information (11). This is also true of severe hypoglycemia/level 3 in the SWITCH studies, where in T2D the number of episodes is too small (SWITCH 2, n = 53) to do more than indicate there is not a safety issue compared with insulin glargine 100 units/mL (4).

The present article does not help with the underlying issue of what is “clinically important” hypoglycemia (and is not intended to) (2), except insofar as supporting symptomatic/severe hypoglycemia as an adequate measure compared with the addition of self-measured asymptomatic episodes. That is likely to change with increasing use of CGM. While the DEVOTE and CAROLINA studies both failed to show any reduction in cardiovascular events when hypoglycemia event rate is reduced (11,12), levels of 3.0 mmol/L (54 mg/dL) are also important with regard to reduced hypoglycemia awareness, with asymptomatic cognitive impairment, defective counterregulation, and increased risk of severe hypoglycemia (1).

The issue of study power from the change in definition, while not found to be real in the SWITCH studies, might be further addressed by applying the approach used by Heller et al. (analysis by 0.1 mmol/L intervals from 3.9 down to 2.5 mmol/L) in other studies, most obviously other published insulin RCTs including other studies on insulin degludec and insulin glargine 300 units/mL. Indeed, data might be available from insulin studies going back a couple of decades. But, as noted above, the advent of valid CGM is changing the approach used in RCTs (6), and the emphasis then becomes asymptomatic hypoglycemia, albeit at the same “clinically significant” level. Because CGM detects more hypoglycemia than intermittent self-measurement of plasma glucose, the issue of statistical power is more easily addressed, and data collection is more objective. Current CGM metrics often use time in range of 3.9–10.0 mmol/L (40–180 mg/dL), but analysis at additional levels (time below 3.0 mmol/L, events as defined above) is computationally easy. Methodological traps related to validity of measurements with CGM at lower levels, and statistical analysis of repeated measures, require careful attention with this technology.

Lastly, these discussions have little relevance to observational pharmacoepidemiological (“real-life”) studies, where the recording of symptomatic hypoglycemia is invariably too erratic to be useful, indeed even for severe hypoglycemia/IHSG level 3. Here, only hospitalization for the primary indication of management of hypoglycemia seems valid and is thus independent of cutoffs for measured glucose level.

In summary, the deeper analysis of data from the well-powered SWITCH studies does support that the new IHSG level 2 definition compared with the previous ADA definition can result in statistically more robust findings when comparing insulins. However, other similar analyses on lower-power studies may not support that conclusion. Exclusion of asymptomatic events may not be detrimental. We can look forward to future data from studies using CGM that may add to our understanding of different cutoff levels as applied to clinical trials.