Scientific Exploration With Continuous Monitoring Systems: An Early Assessment of Arrhythmias During Hypoglycemia

Hypoglycemia is a leading concern in treating type 1 diabetes (T1D). Nocturnal hypoglycemia is especially feared because of reports of death presumed to be due to cardiac arrhythmias. However, until recently, little objective information has been available on the kinds of abnormal rhythms that might be associated with hypoglycemia, their frequency, and the clinical settings in which they occur. Because both glucose levels and electrocardiographic patterns can now be continuously monitored during everyday life, these questions can be directly studied.

Results from exploratory analyses of such data were reported in the May 2017 issue of Diabetes Care by Novodvorsky et al. (1). This was a study of 37 young adults with T1D who used continuous glucose monitoring (CGM) concurrently with continuous electrocardiographic monitoring (CECM) for up to 96 h during daily activities. Besides confirming the feasibility of concurrent monitoring over extended periods of time, the study produced some novel findings. Among other observations, the authors reported that in comparison with periods when glucose was normal, periods of hypoglycemia were associated with increased risk of atrial ectopic beats during the day but not at night. In contrast, the investigators reported that the risk of slow heart rate (bradycardia) was greater during nocturnal hypoglycemia than when glucose levels were normal during the night but was less during daytime hypoglycemia than when glucose levels were normal during the day.

This difference in CECM patterns accompanying hypoglycemia at different times of day is a new and potentially important finding. However, questions raised about the statistical methods used in this study have led to further analyses. In an online Comment in this issue of Diabetes Care, Home and Lachin (2) suggest that the original analysis may have led to an erroneous conclusion about bradycardia accompanying nocturnal hypoglycemia. Their assessment of additional data provided to them by the original investigators is presented in the Comment. In an online Response, the original investigators present their own further analyses and conclusions (3).

To understand how these experienced investigators came to disagree on whether nocturnal (as opposed to daytime) hypoglycemia is, or is not, differentially associated with potentially dangerous bradycardia, the study’s design, definitions, and statistical assumptions must be reviewed. This was an observational study with the stated purpose of exploring whether electrocardiographic changes during hypoglycemia differed between day and night (1). The investigators cast a wide net, deriving various measures of outcome from the CECM data. These included occurrence of abnormal atrial beats, single or multiple abnormal ventricular beats, intervals of bradycardia, variations in intervals between normal beats, and changes in the QT interval (repolarization). There was no primary end point, and no adjustment for multiple statistical comparisons was planned. Consequently, the results must be considered hypothesis generating rather than conclusive. The main definition of hypoglycemia used for analysis of cardiac rhythms was the occurrence of an hour-long interval with mean glucose ≤3.5 mmol/L. Comparisons were made between the number of cardiac events occurring during a given hour of hypoglycemia and the number during an hour of normal levels (between 5 and 10 mmol/L) observed in the same patient at the same time of day on another occasion. Based on these (and other) predefined statistical procedures, one main conclusion was that there were differences in risk of bradycardic events during nocturnal versus daytime hypoglycemia in this group of patients, taken together. The distribution of bradycardic events between the individual patients in the study was not considered in this analysis.

The originally reported excess of nocturnal bradycardic beats was more than sixfold, with narrow CIs. Further analyses by Home and Lachin (2) noted that only two of the 37 participants had any bradycardic beats during nocturnal hypoglycemia. The original investigators extended this finding, showing that when a single patient with a large excess of bradycardia was excluded from the predefined analysis, no excess of events during nocturnal hypoglycemia was present in the other 36 patients.

What can be concluded from this report, the challenge to it, and the further analyses? To begin with the main question, whether there was an excess of bradycardia during nocturnal hypoglycemia, we suggest the original conclusion may be partly right and partly in error. Extended bradycardia during nocturnal hypoglycemia noted in the single patient might conceivably have been due to faulty measurements, an unrecognized confounding factor other than hypoglycemia, or the play of chance. But the possibility that this individual represents a relatively small group of patients who are uniquely at risk cannot be dismissed. However, the lack of any signal that hypoglycemia at night was associated with bradycardia in the other 36 study participants provides reassurance for people with T1D in general. It is consistent with the observation that hypoglycemia is common, but deaths believed to be due to arrhythmias are rare.

Would a complete reanalysis of the data from the study by Novodvorsky et al. (1) add significantly to the present findings? We think not, mainly because the data available are quite limited. A total of 58 h of observation during nocturnal hypoglycemia for 37 study participants, averaging 1.6 h per individual, is a very modest sample. And, if only a small subgroup of people with T1D is postulated to be at risk, more analyses of the present data are unlikely to provide additional insights. As an exploratory effort, the current study has done its job by showing that concurrent measurement of glycemic and cardiac changes is feasible in this population of patients, by suggesting that cardiac effects of hypoglycemia may differ by time of day, and by posing the challenge of identifying who may be at greatest risk. Future studies will build on this base.

Beyond these present clinical conclusions, this early study using new methods of data collection poses important questions to be addressed in future studies. First, how common might high-risk individuals be among all those with T1D? Calculation of CIs (4) around the ∼3% prevalence of individuals possibly at risk in this small group (1 of 37) suggests a 95% likelihood of the true prevalence being between 0.1% and 14.2% in the general population. Future studies may require larger numbers of patients and efforts to include those who are considered at greatest risk. Analyses of such studies should include both the numbers of participants affected and the risk of events overall.

Further questions concern both potential bias and statistical precision. The 37 participants in the study by Novodvorsky et al. (1) might have collected up to 96 h of observations each, a total of 3,552 h. However, concurrent CGM and CECM data were available for 2,395 h, or 67% of the potential hours. For the analysis of nocturnal hypoglycemia that is in question here, there were 58 h of evaluable data, and only 4 h of hypoglycemia-contained bradycardic beats. For comparison, consider the recent FLAT-SUGAR study focused on glycemic variability in patients with long-duration type 2 diabetes (T2D) and high cardiovascular risk (5). In that study, 92 patients were treated for 26 weeks with basal insulin together with either prandial insulin or prandial exenatide. Each patient was to collect 7–10 days of CGM and CEGM data concurrently at the end of treatment. Hypoglycemia defined as 20 min below 3.9 mmol/L occurred 345 times in 68 patients, with a total of 29,948 min of hypoglycemia. The mean time of hypoglycemia was 326 min (∼5 h) for all participants in the study and 440 min (7.3 h) for those with at least one event. Even with this longer exposure to hypoglycemia during which arrhythmias might be identified, just 9 patients had 12 episodes of tachycardia and 8 patients had 9 bradycardic events. As a result, the investigators of this study in T2D did not attempt a detailed analysis of arrhythmias. To obtain enough events for reliable analysis, future studies in both T1D and T2D should not only include high-risk individuals to increase the number of events and resulting precision but also aim for more complete data collection for longer intervals of time to reduce potential biases. Also, if the aim of a study is to assess the effects of both hypoglycemia and the time of day at which it occurs, the statistical plan should include a formal assessment of heterogeneity of the relationship between hypoglycemia and risk of arrhythmias for day versus night.

Another concern is less obvious but may be equally important. How should hypoglycemia be defined for this purpose? Should brief episodes of hypoglycemia, perhaps 10–20 min long as in the FLAT-SUGAR study (5), be included? Is it appropriate to use 3.5 mmol/L as a threshold level, or should the 3.0 mmol/L level now recommended for clinical studies (6) be used instead? Should prior levels and rapidity in decline of glucose be included in the analysis to determine whether a rapid decline from a high average level is more likely to cause arrhythmias? That approach would have the added advantage of including all glycemic data collected rather than excluding measurements between 3.5 and 5 mmol/L and also those above 10 mmol/L.

Finally, this experience reminds us of the need for caution and humility. Opening new areas of investigation resembles exploration of the seas by early mariners. Better ships and celestial navigation provided these explorers a means and some guidance, but they did not know what they would find or what the risks might be. Similarly, modern methods of measurement and analysis have taken us out of an experimental comfort zone. Gregor Mendel counted how many of his pea plants had yellow or green peas, thereby deducing some rules of genetics. More recently, Yalow and Berson (7) measured immunoreactive insulin in the blood of 14 healthy adults and 17 with T2D and, by noting that levels in T2D were higher, opened a new chapter of diabetes research. Analyzing these observations was straightforward and on a human scale (as opposed to involving molecular observations, thousands of data points, or years of follow-up). Now, we have electronic methods to collate and analyze ordinary clinical and laboratory measurements in tens or hundreds of thousands of people studied for decades. Epidemiologic analyses of high dimensional data are done not at a human scale, where the methods used can be easy to grasp and the primary data directly viewed. Instead, we rely on potentially complex algorithms based on assumptions and prior information to summarize complex series of data, and the system delivers derived numbers that look official but may not be scrutinized as easily as simpler collections of data. These powerful tools allow important insights, but they also can mislead us when faulty assumptions, missing data, or unmeasured covariates confound the analysis.

Now, with availability of ways to collect physiologic data continuously over extended periods of time, this problem also applies to clinical studies of limited numbers of participants but greatly increased numbers of individual measurements. Wearable devices can provide ongoing measurements not only of glucose and electrocardiographic patterns but also of blood pressure, oxygen saturation in blood, and body movements—and surely the list will grow. These devices generate countless data points. Measurement of tissue glucose levels every 5 min by CGM provides >2,000 reports per week. At a pulse rate of 72 per minute, CECM recordings will describe >725,000 beats in a week. These are not human-scale numbers, and they defy simple analysis. We are learning to use digital technology, but not doing it perfectly every time. In the study by Novodvorsky et al. (1), an unanticipated asymmetry in distribution of data confounded the original analytic plan. All of us—epidemiologists, clinical physiologists, statisticians, and journal editors and reviewers alike—will have to recognize the potential pitfalls of electronic data capture and analysis, rigorously scrutinize summary results, and remain modest in our conclusions.

Further studies exploring the relationships between hypoglycemia and cardiac arrhythmias are already under way and will clarify the issues discussed above. The ultimate goal is to determine which people with diabetes may be at risk for harm from hypoglycemia and how to protect them. Progress to that end will be—as suggested by Fuller Albright (8)—stepwise. We can thank Novodvorsky et al. for their exploratory study and both this group and Home and Lachin for their additional analyses that have moved us another step forward.