Safety of a Novel Continuous Glucose Monitoring–Informed Insulin Bolus Calculator Mobile Application for People With Type 1 or Type 2 Diabetes Free

This clinical trial was an open-label, industry-sponsored study conducted at two sites (Chicago, IL, and Baltimore, MD). The single-arm study design was selected to collect safety data that were used to support the submission of the investigational software to the U.S. Food and Drug Administration (FDA) for clearance as a medical device. Each site received institutional review board (IRB) approval (the Northwestern University IRB for the Chicago site and the Advarra IRB for the Baltimore site).

Fifty-four participants were enrolled in this study. Major inclusion and exclusion criteria are shown in Table 1. Informed consent to participate in the trial was obtained before enrollment. There were three study visits: baseline (visit 1), a virtual or in-person visit 7 days after the baseline visit (visit 2), and an in-person visit 30 days after the baseline visit (visit 3).

At visit 1, the mobile app was installed on each participant’s mobile device (Figure 1). The mobile app received CGM data via an application programming interface (API) supplied by the CGM manufacturer (Dexcom, San Diego, CA). The native CGM app remained installed on each participant’s mobile device. The principal investigator (PI) configured the calculator for each participant per usual clinical practice. The settings included the insulin-to-carbohydrate ratio (ICR), correction factor (CF), target glucose, and duration of insulin action (DIA). Participants were asked to use the insulin calculator in the mobile app whenever it was time to take an insulin bolus. The DDS17 questionnaire was administered for the first time during this visit.

Visit 2, scheduled 7 days after visit 1, could be in person or virtual. The purpose of this visit was to make sure participants were not having difficulty using the mobile app and, in the judgment of the PI, were not having excessive hyperglycemia or hypoglycemia. Calculator parameter adjustments were permitted as deemed necessary by the PI.

Visit 3, 30 days after visit 1, was an in-person visit at which the mobile app was removed from the participant’s mobile device. Participants returned to their usual method of insulin dosing. The DDS17 questionnaire was administered a second time at this visit.

This study was initially designed and powered with a safety end point to demonstrate that the mean TIR in the 30 days of using the mobile app was noninferior to baseline TIR. Because TIR was found to be improved at the end of the study, both noninferiority and superiority end points were examined in the final analysis. Three population segments were defined a priori. The intention-to-treat (ITT) group contained all enrolled participants. The complete cases (CC) group consisted of all ITT participants who completed visit 3 with ≥90% CGM sensor wear time. The per protocol (PP) group consisted of all CC participants who used the calculator ≥30 times during the 30-day study period. These participants would have complied with the protocol sufficiently to ensure that their data would be likely to represent the effects of using the calculator.

Demographic and survey data and A1C values were recorded in study documents at each study site. CGM data were transmitted securely to a remote server where they were de-identified and sent to the statistics team for analysis. The CGM API acquired the 30 days of CGM data prior to the use of the app. The day when the app was installed was excluded from the analysis. The 30 days of pre-study data were considered baseline and compared with data from the 30 days of the study period. The dataset thus contained 60 days of data for each participant.

Fifty-four individuals with diabetes using CGM enrolled in the study. Their demographic and clinical characteristics and insulin calculator parameters are shown in Table 2. There was heterogeneity in sex, race, ethnicity, and type of diabetes. Thirty percent of the cohort were ≥65 years of age, which provided perspective on the usability of the app in an age-group generally considered less technologically savvy than younger individuals.

Mean TIR values at baseline and during the study period for the three groups are shown in Table 3. A poolability analysis found that the two research sites were not statistically different with respect to the primary end point values at baseline or post-treatment or in terms of change from baseline (t test P values 0.43, 0.23, and 0.43). Mean TIR improved by 3.8% (95% CI 0.7–6.9%) from 69.2 to 73.0% (P = 0.017) in the PP group. This TIR corresponds to a mean of 0.9 more hours per day spent in range. There were also significant improvements in TIR in the CC or ITT groups.

Subgroup analyses are shown in Table 4. In the 22 individuals with type 2 diabetes, TIR improved by 6.5% (95% CI 3.0–10.1%), which corresponds to a mean of 1.6 hours more per day spent in the target range. The increase of TIR in the type 1 diabetes subgroup was not statistically significant. Study participants were also grouped according to the frequency of calculator use during the 30-day study period. There did not appear to be a linear relationship between calculator use and improvement in TIR.

Exploratory analyses were performed for four secondary end points: changes in time above range (TAR; glucose >180 mg/dL), TBR (glucose <70 mg/dL), glucose coefficient of variation (CV), and glucose SD. These secondary end points are shown in Table 5. Statistical significance was not tested for the exploratory secondary end points. TAR numerically decreased in the three study groups. TBR was low at baseline in the three groups and did not increase.

Glucose CV and glucose SD are measures of glucose variability, which has been associated with cardiovascular risk, mortality, severe hypoglycemia, and reduced quality of life (9). In this exploratory analysis, both glucose CV and glucose SD decreased during the study.

We also examined diurnal and nocturnal TIR. Both improved during the intervention period, mostly because of improvement in the subgroup with type 2 diabetes. In the type 1 diabetes subgroup, the magnitude of the improvement seemed greater during the day than at night. These data are shown in Table 6.

The DDS17 is a well-established questionnaire used in clinical practice and research. Each item in the questionnaire is rated on a six-point scale ranging from 1 = not a problem to 6 = a very significant problem. DDS17 scores <2.0 correspond to little or no distress, score of 2.0–2.9 correspond to moderate distress, and scores ≥3.0 indicate high levels of distress (7,8). This study was not designed specifically to improve diabetes distress, and baseline distress scores for all participants were low. The mean DDS17 score for all participants was 1.72 at baseline and 1.67 at the end of the study period (Table 7). The change in the mean DDS17 score was not statistically significant (P = 0.1). A subgroup of participants (25 of 54) was defined as those with a baseline DDS17 score greater than the median of the group. The subgroup of these distressed participants was found to have a significant reduction in diabetes distress during the study (P = 0.038).

The DDS17 instrument also reports on four subscores: regimen distress, interpersonal distress, physician distress, and emotional burden. In those participants who were in the distressed group at baseline, the subscores for regimen distress and interpersonal distress improved significantly.

The adverse event of hypoglycemia is the major concern for individuals who are prescribed bolus insulin. Because the mobile app potentially applied CGM trend arrow adjustments to insulin bolus dose recommendations, the concern about individuals taking a higher dose of insulin because of upward trend arrows was considered. In this prospective clinical trial, there were no hospitalizations, emergency department visits, emergency medical calls, or severe hypoglycemic events requiring assistance from others. As shown in Table 5, TBR was quite low and did not increase during the study period while the IBC was being used. Specifically, in the PP group, whose participants used the calculator the most frequently, mean TBR was 1.2 ± 1.1% compared with 1.1 ± 0.9 at baseline.

In addition to examining TBR, we identified specific hypoglycemia events in the dataset. We defined a hypoglycemia event as an episode during which the CGM glucose value dropped to <70 mg/dL for >15 minutes. Per participant, there was a mean of 10.7 such events in 30 days before the study period and 10.0 such events during the 30-day intervention period. This difference was not significant (P = 0.72).

The use of CGM in individuals with type 1 or type 2 diabetes who are prescribed bolus insulin has been shown to be effective at improving glycemia (10,11). Teaching these individuals how to adjust their insulin doses based on their current glucose value, carbohydrate intake, previous insulin boluses, and exercise, with adjustments for the CGM trend arrows, can be challenging. The development and use of IBCs that incorporate CGM-informed glucose values and trend arrow–adjusted insulin doses have been recommended by diabetes technology experts (12).

A CGM-informed IBC designed for use with an automated insulin delivery system recently received FDA clearance (13). However, out of the nearly 8 million individuals on insulin therapy in the United States, <20–30% of people with type 1 diabetes and <1% of people with type 2 diabetes use insulin pump therapy (14). This leaves a large population who could benefit from the availability of tools to support daily insulin dosing but who perceive the complexity, cost, and inconvenience of insulin pumps to outweigh the benefits. In this study, we describe the results of a clinical trial using a CGM-informed IBC embedded in a mobile app that can be used by people with type 1 or type 2 diabetes on intensive insulin therapy.

In the international consensus statement regarding CGM metrics for clinical trials, a change in mean TIR ≥3% is considered clinically meaningful (15). In this clinical trial, use of the mobile app for 30 days was associated with significant improvement in TIR (∼4%) without increasing TBR. Similar improvements were seen in both the CC and the PP groups.

A higher degree of improvement in TIR was noted among participants with type 2 diabetes. It has been demonstrated that providing people with diabetes a CGM device improves glycemia (10,11). This is likely because CGM users are aware of their glucose values at all times, which allows for improvements in self-management and more precise insulin dosing. Participants in this trial were already using CGM before enrolling in the study. Therefore, the results suggest that the improvement in TIR was likely the result of the study intervention and not merely the use of CGM alone. We hypothesize that the greater improvement in TIR among participants with type 2 diabetes could be attributed to less precise bolus insulin dosing at baseline compared with participants with type 1 diabetes. Furthermore, more precise insulin bolus dosing may lead to a decrease in glucose variability, which could further improve health outcomes for these individuals (9).

In an exploratory analysis, we examined diurnal and nocturnal TIR. We expected TIR to improve more during the day when insulin boluses are being administered. Among participants with type 2 diabetes, the improvement in TIR was similar in the day and at night; in those with type 1 diabetes, the improvement during the day seemed greater. This difference may be due to the absolute dependency of those with type 1 diabetes on precise insulin dosing with meals.

A major limitation of this study comes from its relatively small number of participants as well as its single-arm design primarily investigating safety. This design was selected to investigate noninferiority in TIR during the study period and was powered to collect safety data for the purposes of an FDA submission. In this case, a control group or cross-over analysis was not required. Although the improvement in TIR was statistically significant, the benefit may not have been entirely attributable to the use of the IBC. Effects of merely enrolling in the clinical trial and using the mobile app’s educational features cannot be excluded; however, the greater increase in TIR among those who used the IBC more frequently during the study period (Table 4) does suggest an effect of the IBC itself.

Another limitation of this study comes from the good baseline A1C of the study population. Ninety-five percent of participants had a baseline A1C <8%. Individuals with much higher A1C values could have potentially benefited more from the use of the IBC or could have developed more hypoglycemia when using it. However, this first clinical trial with a CGM-based insulin calculator was focused on safety; therefore, the intention was to recruit individuals whose baseline glycemia was in or near target. The disadvantage of this approach is that the potential magnitude of improvement would be limited.

The use of bolus insulin adds complexity to the burden of diabetes. A digital health solution that assists people with diabetes self-management and insulin dosing may reduce diabetes distress. Alternatively, requiring people with diabetes to perform additional tasks using a mobile app has the potential to increase burden. The results from the DDS17 questionnaire demonstrated no increase in diabetes distress. In fact, in the subgroup of participants who were distressed at baseline (n = 25), DDS17 scores improved. Interestingly, subscores related to treatment regimens and interpersonal relationships also improved. Thus, the mobile app may become a trusted partner for users, improving glycemic outcomes and reducing the stress related to computing insulin bolus doses or discussions with family members. Future research may help to reveal the mechanisms for this reduction in diabetes distress.

Editorial assistance was provided by Janice MacLeod, MA, RD, CDCES, FADCES, of Janice MacLeod Consulting and was paid for by Welldoc.

This study was funded by Welldoc.

M.S., A.K., and A.I. are employees of Welldoc. G.A. has been a consultant for Dexcom, Insulet, and Medscape and has received research funding from Fractyl Laboratories, Insulet, and Tandem Diabetes. No other potential conflicts of interest relevant to this article were reported.

M.S. wrote the first draft of the manuscript. M.S., C.K., and G.A. designed the protocol. J.P. and G.A. served as principal investigators at their respective sites. C.K. carried out most of the statistical analyses. A.K. and A.I. contributed to the analyses. All authors reviewed and edited the final manuscript. M.S. 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.