The Use of a Smart Bolus Calculator Informed by Real-time Insulin Sensitivity Assessments Reduces Postprandial Hypoglycemia Following an Aerobic Exercise Session in Individuals With Type 1 Diabetes

In type 1 diabetes (T1D), insulin therapy based on multiple daily insulin injections or continuous subcutaneous insulin infusion allows compensation of the practically absent internal insulin secretion that follows from the autoimmune destruction of pancreatic β-cells (1). As a consequence, the quality of glycemic control in T1D is heavily dependent on multiple daily treatment decisions by the patients that account for a wide variety of factors influencing insulin demand (2). In this context, despite the improving accuracy of glucose monitoring devices (3,4) and the growing development of decision support systems (5,6), suboptimal insulin replacement remains common in T1D, leading to excess mortality and complication rates that are still significantly higher when compared with the general population (7,8).

The main mediator of time-varying insulin requirements in individuals with T1D is insulin sensitivity (IS), a metabolic parameter quantifying the ability of insulin to stimulate glucose uptake and inhibit endogenous glucose production (9–11). Systematic intraday variations of IS are typically handled through optimal tuning of basal rate, insulin-to-carbohydrate ratio (CR), and correction factor (CF) profiles (5,12–17). However, superimposed IS fluctuations driven by acute, nonsystematic events happen frequently and further complicate insulin dosing (18–20). Among the main triggers of altered IS, physical activity has been shown to influence insulin requirements for several hours past the conclusion of the exercise effort, remaining a major challenge for individuals with T1D and oftentimes leading to imperfect insulin replacement, worsened quality of glycemic control, and higher glucose variability (21–27).

In the attempt to tailor insulin doses to the time-varying metabolic needs of individuals with T1D, “smart” bolus calculators have been developed that leverage continuous glucose monitor (CGM) trend information (28,29), machine-learning algorithms to predict the postprandial glycemic excursion (30), and model-based assessments of a person’s IS (31–34). The latter effort typically requires physiological models of glucose-insulin dynamics and parameter/state estimation techniques for the quantitative assessment of IS from CGM and insulin pump data. Schiavon et al. (31) proposed an algebraic formula that allows for the estimation of IS in individuals with T1D wearing a sensor-augmented insulin pump in response to a single meal under carefully controlled conditions. The method was used by Dassau et al. (32) to optimize basal rate and CR open-loop settings before initializing closed-loop control in individuals with T1D. Boiroux et al. (33) presented a technique that continuously tracks postprandial glucose dynamics and IS, and was used for adaptive basal-bolus calculation leveraging in silico data. In other work (34), we proposed the use of a technique for the real-time (RT) tracking of IS from CGM and insulin pump data in the framework of smart prandial insulin dosing. The algorithm relies on parameter identification techniques to model specific meal dynamics and a Kalman filter–based state estimation to track IS over the horizon of interest, and it showed promising results when used in simulation to modulate insulin boluses in the presence of increased/decreased IS levels (34).

In this work, we present the results obtained from the first clinical deployment of the IS tracking method combined with the IS-informed bolus calculator (34), in the control of one dinner meal following an early afternoon aerobic exercise session in 15 adults with T1D at the University of Virginia (UVA). Physical activity is a known trigger of heightened IS both during and following the exercise bout (21); therefore, we hypothesized that the proposed smart bolus calculator used in the hours after an aerobic exercise session could track a state of increased IS and correct the insulin dose accordingly, allowing mitigation of the occurrence of postprandial hypoglycemia and providing protection against low blood glucose levels.

The IS-informed bolus calculator was tested in a single-site, randomized, crossover clinical trial at UVA (ClinicalTrials.gov, no. NCT03709108) designed to compare the system to standard bolus therapy in the control of large and sustained IS fluctuations. Upon research protocol approval by the UVA institutional review board, study subjects (18–65 years old) were recruited by phone and advertisement. Inclusion criteria included diagnosis of T1D (>1 year), use of an insulin pump (>1 year) and CGM (>6 months), and hemoglobin A_1c (HbA_1c) levels <8.5% at screening to ensure treatment compliance. Exclusion criteria included recent history of severe hypoglycemia or diabetes ketoacidosis (within the last 6 months), pregnancy, and conditions incompatible with the practice of physical activity and aerobic exercise. After signing informed consent, subjects participated in a 4-week at-home data collection, during which they were asked to consistently use CGM and insulin pump, and to enter all consumed carbohydrates into their pump or dedicated smartphone app. At the end of data collection, subjects participated in two 24-h admissions in semicontrolled conditions at a local hotel. The admissions started around 9:00 a.m.; around 12:00 p.m., a standardized lunch meal was provided to the study participants, who were then accompanied to a local gym around 2:00 p.m. to engage into an aerobic exercise session starting at 2:30 p.m. (three consecutive 15-min bouts of static bicycle separated by 5-min resting periods). Upon completion of the exercise session, subjects were accompanied back to the hotel and asked to perform quiet activities until dinner. Around 7:00 p.m., a standardized dinner meal was provided. In the control admission, the dinner insulin dose was computed using the participants’ standard bolus calculator, while in the experimental admission, the IS-informed bolus calculator was deployed. After dinner, study participants were asked to remain quiet and retire to their room around 10:00 p.m. The following morning, subjects were discharged around 9:00 a.m., after being offered breakfast without dietary restrictions. During the admissions, all standardized meals were provided by a local chain restaurant and were designed to contain ∼0.75–0.85 g/kg of body weight of carbohydrates and 0.2–0.3 g/kg of body weight of protein and fat. No caffeinated beverages were allowed during the admissions. The goal of the exercise session was to create a state of immediate and delayed heightened IS (21). The order of the two admissions was randomized at the beginning of the study and they were scheduled to be at least 48-h apart to ensure proper washout of the effect of exercise. In line with this, subjects were instructed to avoid physical activity and alcohol consumption in the 48 h before each admission.

The IS tracking algorithm used in the study was introduced by Fabris et al. (34) and is outlined in the Supplementary Data. The algorithm defines a new IS index (S_I) that is dynamically estimated using CGM data, insulin pump records, and a model of glucose-insulin dynamics embedded into an optimally tuned Kalman filter. Relying on this algorithm, the proposed smart insulin bolus is computed by modulating a standard bolus with the ratio of usual IS estimated from historical data and an RT IS assessment computed on-demand at the time of the bolus administration. Specifically, in the study, by running the IS tracking algorithm on the CGM and insulin records gathered during the 4-week data collection, a 24-h median IS profile was built for each study participant that was considered to be representative of his/her usual intraday pattern of sensitivity to insulin action. The profile embeds IS fluctuations driven by circadian rhythms and systematic behavioral habits, and it is assumed to be compensated for by individual insulin treatment parameters routinely optimized by treating physicians. After computing the IS profile, indicating with B the standard bolus, the IS-informed insulin dose (B_IS) is calculated in RT as

where CHO is the amount of meal carbohydrates, BG is the prevailing CGM value, BG_TGT is the glycemic target, IOB_BOL/BAS is insulin on board (IOB) from bolus (BOL) and basal (BAS) injections, and IS_RT/IS_PRF are the IS index estimates from the IS tracking algorithm, computed in RT and drawn from the profile at a comparable time of day, respectively. On the basis of the design of the IS-informed bolus calculator, the system captures IS deviations with respect to the profile and administers a larger/smaller insulin dose if an IS level lower/higher than usual is detected. Of note, if the RT IS estimate equals the estimate from the profile, then B_IS naturally converges to B.

Subjects were instructed to wear a study provided CGM system (Dexcom G6; Dexcom, San Diego, CA) and their personal insulin pump during the entire duration of the trial (data collection and hotel admissions). If deemed necessary by the study team, subjects were asked to insert a new sensor up to 48 h before each admission. No CGM calibration was required, in accordance with the glucose sensor manufacturer’s guidelines. During the hotel admissions, all subjects used the Diabetes Assistant (DiAs) mobile platform, a smartphone-based infrastructure developed at UVA for the implementation and clinical testing of closed-loop and open-loop glucose management strategies (35–37). At the beginning of both admissions, each participant’s CGM system was connected to a DiAs device, which reported data back to the UVA remote monitoring system (DiAs Web Monitoring) for supervision by the study team. Study participants were asked to use DiAs for the computation of the lunch and dinner bolus during both admissions. In the control admission, DiAs implemented a standard bolus calculator based on the participants’ usual insulin treatment parameters, a 110 mg/dL glycemic target, and a 4-h IOB curve for both meals. In the experimental admission, DiAs was programmed to implement the standard bolus calculator at lunch and the IS-informed bolus calculator at dinner, using the same settings for CR, CF, glycemic target, and IOB as the control admission. After calculating the suggested bolus amount in DiAs, subjects were asked to administer the indicated insulin dose through their personal insulin pump. When the IS-informed bolus calculator was being used, safety was guaranteed by saturating the bolus modulation to ±30% of the standard insulin dose and requiring study physician approval before communicating the insulin amount to the subjects.

During the hotel admissions, participants were remotely monitored by the study team using the DiAs Web Monitoring system (37). The study team intervened if 1) CGM values were <80 mg/dL during the exercise session or <60 mg/dL any other time and 2) CGM values were >300 mg/dL for at least 1 h or >400 mg/dL at any time. The first condition triggered the prompt administration of ∼15 g of fast-acting rescue carbohydrates; after receiving a first hypoglycemia treatment, CGM rise was closely monitored by the study team, and the administration of an additional 15 g was considered if the CGM was still <80 mg/dL after 20 min. If the second condition occurred, correction boluses deemed appropriate by the study physician were administered to the study participants, and the concentration of blood ketones was checked. Loss of remote monitoring or CGM disconnection also triggered study team interventions. Loss of remote monitoring for >2 h was a stopping criterion.

All glycemic outcomes were computed based on CGM records. The primary outcome was exposure to hypoglycemia in the 4 h following dinner, as measured by postprandial low blood glucose index (LBGI) (38). Secondary outcomes included postprandial high blood glucose index (HBGI) (39) and percent time in hypoglycemia <70 mg/dL, in range 70–180 mg/dL, and in hyperglycemia >180 mg/dL and >250 mg/dL. In addition, the total number of hypoglycemia rescue treatments administered to the study participants in the 4 h following dinner was compared between the admissions.

As a safety and feasibility study, no power analysis was computed for the design of the clinical trial. The desired enrollment was 15, which is in line with previous sample sizes used in our pilot clinical trials. Paired t tests or nonparametric Wilcoxon signed-rank tests were used to compare the control and experimental admission in terms of glycemic outcomes in the case of normally or nonnormally distributed samples, respectively (Shapiro-Wilk test). The significance level was set at a P value <0.05. Data are reported as mean ± SD. Data formatting/preparation and computation of the IS profile were executed in Matlab R2019a (MathWorks), and all statistical analyses were computed in SPSS Statistics 26 (IBM).

Fifteen adults with T1D completed the study. One participant was not compliant with the study procedures during the hotel admissions and was therefore excluded from the analysis. The remaining 14 subjects were predominantly males (four females), experienced pump users, and, on average, well controlled with an HbA_1c of 6.9 ± 0.9%. Complete characteristics of the study participants are outlined in Table 1.

Personal insulin pumps used in the study included a variety of Medtronic pumps (670G, 530G, Paradigm; N = 6), the Insulet Omnipod (N = 5), and the Tandem t:slim ×2 (N = 3).

We recorded eight adverse events (AEs) during the study. Five AEs were deemed unrelated to the subjects’ participation in the clinical trial: 1) diagnosis of torticollis upon emergency room visit after discharge from hotel admission; 2) two events of mild fever immediately prior to and after discharge from hotel admission; 3) sustained hyperglycemia with elevated blood ketone concentration following dinner during hotel admission (see below); and 4) significant hypoglycemic event after discharge from hotel admission. Three AEs were deemed related to the study: 1) bruising at the glucose sensor insertion site during data collection; and 2) two instances of headache possibly because of caffeine withdrawal during hotel admission.

Because of unexpected events that altered the consistency of the data being collected, two participants were asked to repeat one hotel admission for the following reasons: 1) at the second admission, nonavailability of the standardized meals consumed during the first admission; and 2) sustained exposure to hyperglycemia following the dinner bolus (more than 1 h >300 mg/dL) that did not resolve despite the administration of an insulin correction dose, which was possibly driven by a condition of psychological stress of the subject due upcoming work-related commitments or insulin infusion site malfunction (unrelated AE number 3).

Overall, the IS tracking algorithm consistently detected a state of increased IS in the hours following the exercise session and preceding the dinner meal (IS increase at dinnertime: 30.47 ± 0.31%, P = 0.0002) (Fig. 1), and it allowed for substantial mitigation of the occurrence of postprandial hypoglycemia (Fig. 2).

A summary of glycemic outcomes is presented in Table 2 and Fig. 3. As compared with standard therapy, the use of the IS-informed bolus calculator significantly reduced postprandial LBGI (2.02 ± 2.63 vs. 3.31 ± 2.45, P = 0.006) and percent time <70 mg/dL (8.48 ± 11.88% vs. 15.18 ± 12.21%, P = 0.049), with a trend of increased HBGI (3.13 ± 4.42 vs. 2.09 ± 3.48, P = 0.075), and no change in percent time >180 mg/dL (13.24 ± 24.45% vs. 10.42 ± 22.27%, P = 0.5) and >250 mg/dL (2.08 ± 5.36% vs. 1.19 ± 4.45%, P = 0.317). No significant increase in terms of time in 70–180 mg/dL was achieved with the use of the system (78.27 ± 26.39% vs. 74.41 ± 20.97%, P = 0.469). However, the total number of hypoglycemia rescue treatments needed by the study participants was decreased from 12 to 7 when the IS-informed bolus calculator was deployed, even though this difference was not significant (P = 0.238).

No difference in terms of blood glucose levels at dinnertime was detected between the admissions (control versus experimental: 91.29 ± 49 mg/dL vs. 90.64 ± 41.09 mg/dL, P = 0.925) (Fig. 4).

Suboptimal insulin dosing in T1D is at the root of individuals’ inability to reduce glucose variability and achieve good glycemic control, therefore representing a source of long-term diabetes complications. Periodical review of glycemic profiles, computed manually by treating physicians or automatically through learning methods and ad hoc algorithms (5,12–17), allows for the optimization of insulin treatment parameters to account for systematic intraday IS patterns. However, insulin requirements in T1D are highly variable and impacted by several physiological and psychobehavioral factors affecting IS (18–20). Among these, physical activity remains a major challenge for individuals with T1D (22–27), with aerobic exercise having a delayed impact on individuals’ sensitivity to insulin action lasting for several hours past the conclusion of the activity bout (21).

In this work, we presented the first clinical validation of an IS tracking algorithm and IS-informed bolus calculator in 15 adults with T1D, in semicontrolled conditions, using CGM data and insulin pump records to estimate IS in RT (34). The smart bolus calculator was used to control glycemia during a dinner meal following an early afternoon aerobic exercise session designed to increase IS. The IS tracking algorithm informing the bolus calculation detected a 31% exercise-induced IS increase at the time of dinner. The corresponding bolus modulation allowed to significantly reduce postprandial exposure to hypoglycemia [as measured by LBGI (38) and percent time <70 mg/dL] with a nonsignificant shift toward increased HBGI (39), percent time >180 mg/dL, and percent time >250 mg/dL. The comparison of postprandial control was particularly robust because of the comparable blood glucose levels at dinnertime achieved in the two sets of admissions. In addition, the need for hypoglycemia rescue treatments in the hours following dinner was almost halved when the IS-informed system was used.

In the literature, smart bolus calculators informed by CGM trends or machine learning algorithms have been proposed and validated (28–30). Cappon et al. (29) compared the in-silico performance of three popular techniques for insulin bolus calculation that account for the dynamic information provided by glucose rate of change. On the basis of slightly different rules, the three methods increase the insulin dose in the presence of rising glucose, while they decrease the dose if glucose is falling. In the presence of a negative rate of change (i.e., falling glucose), for a preprandial blood glucose between 80 and 100 mg/dL (comparable to what was observed in our study), Cappon et al. showed that the average ratio between postprandial LBGI obtained with the use of the smart systems and postprandial LBGI obtained with standard bolus calculators ranges between 0.76 and 0.87 (depending on the actual glucose rate of change and the specific algorithm used). If the same analysis is applied to our data, the average postprandial LBGI ratio we obtain is 0.54. Despite cautiousness being necessary in comparing the two studies as Cappon et al. were presenting in-silico results and their postprandial horizon was longer than ours, these results suggest that the IS-informed system can achieve better performance than bolus calculators informed by glucose rate of change. Furthermore, the IS-modulated calculation is more flexible in that the insulin dose is adjusted based on the evaluation of metabolic features also in the absence of obvious CGM trends.

As a pilot clinical trial, the study presented here had several limitations. These included the relatively small number of study participants and the highly controlled conditions in which the technology was tested. In addition, on the basis of its definition, the IS-informed bolus calculator relies on optimally tuned insulin therapy parameters that ensure adequate postprandial glycemic control in the absence of deviations with respect to the profile. While it was not done as part of this study, the benefit of using the IS-informed bolus calculator may be maximized by combining it with an algorithm for CR/CF/basal rate optimization.

In conclusion, results from this randomized, crossover pilot clinical trial showed that the IS-informed bolus calculator developed at UVA is safe and feasible in a group of 15 adults with T1D exposed to triggers of large and sustained IS fluctuations in a semicontrolled environment. The system was able to capture acute IS changes and showed improved protection against hypoglycemia following a moderate-intensity aerobic exercise session. These results suggest that the use of a smart bolus calculator informed by RT IS assessments may benefit prandial insulin dosing by following the time-varying metabolic needs of individuals with T1D.

Future work will be directed toward extending the testing of the IS-informed bolus calculator over several days of use in the presence of repeated triggers of IS fluctuations in semicontrolled conditions and over several weeks of unsupervised home use. Outcomes from these clinical studies, which will involve both adolescents and adults with T1D, will determine the generalizability of the results presented in this manuscript and will possibly validate the effectiveness of our smart bolus calculator over standard therapy in the improvement of glucose control and reduction of glucose variability in individuals with T1D.

Acknowledgments. The authors thank the study volunteers and the research staff at the UVA Center for Diabetes Technology, particularly Emma Emory, Jose Garcia-Tirado, Jacob Hellman, Matthew Kime, Laura Kollar, Helen Myers, Chase Phillips, Jessica Robic, and Christian Wakeman. The authors also thank the anonymous reviewers for the comments and constructive criticism that allowed improvement of the quality of the article.

Funding. The study was funded by JDRF (Advanced Postdoctoral Fellowship 2-APF-2017-378-A-N) and the UVA Center for Diabetes Technology.

Duality of Interest. C.F. consults for Epsilon (Abbott). S.M.A. has received research support from Medtronic. M.D.B. has received honoraria from Dexcom and research support from Sanofi, Dexcom, and Tandem. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. C.F. researched data and wrote the manuscript. R.M.N. and S.M.A. researched data and reviewed the manuscript. J.P., K.A.C., C.L.K.K., C.L.B., and M.C.O. researched data and contributed to writing the manuscript. D.R.C. and M.D.B. contributed to the study design and discussion and reviewed the manuscript. C.F. 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 analysis.