Advanced Closed-Loop Control System Improves Postprandial Glycemic Control Compared With a Hybrid Closed-Loop System Following Unannounced Meal

Closed-loop control (CLC) systems for insulin management in individuals with type 1 diabetes (T1D) have provided improved time-in-target-range (TIR) and lower HbA_1c (1–5). These “artificial pancreas” (AP) systems use data from continuous glucose monitoring (CGM) devices and the user’s insulin sensitivity to inform complex algorithms that determine the on-going need for insulin, which is then delivered by continuous subcutaneous insulin infusion pumps. These systems thus achieve automated insulin delivery, with a goal to maintain blood glucose (BG) in a desirable range. However, while CLC systems perform very well at managing glycemia in the absence of prandial glucose excursions (such as overnight), it has been difficult to prevent prolonged hyperglycemia following consumption of carbohydrates that are not announced to the system (and bolused for) (6). This is partly because CLC systems experience inherent delays in CGM sensing of rising prandial glucose levels and in the initiation of insulin action following its infusion, while closed-loop algorithms must also avoid hypoglycemia from overaggressive insulin administration. Because of these considerations, all current commercially available CLC systems (1,2,7–9) are actually hybrid closed-loop (HCL) systems that require the user to enter the quantity of carbohydrate ingested, receiving a prandial insulin-to-carbohydrate ratio to avoid significant hyperglycemia.

Unfortunately, it is common for individuals with T1D to omit bolusing for carbohydrates. This affects 65% of adolescents at least once weekly (10), with 38% missing at least 15% of their boluses (11). Adolescents who miss four boluses weekly experience an increase of 1% in their HbA_1c (10), which may contribute to the large number of adolescents who fail to meet recommendations for HbA_1c levels (12).

A key obstacle for developing fully closed-loop (FCL) systems that can automate adequate insulin delivery in response to unannounced carbohydrate is the need to detect BG excursions that identify that a meal ingestion has occurred. For the current study, we developed a bolus priming system (BPS) that is tuned to changes in CGM glucose patterns corresponding to theoretical probabilities that an unannounced meal has been consumed. The system is designed to respond to these probable meals by providing a priming dose of insulin intended to prevent significant BG elevation before the CLC system can provide additional corrective insulin. This system is embedded in a model-predictive control (MPC) system called RocketAP (RCKT) that continually predicts future glycemia and calculates optimal insulin doses to maintain the user in a desired glucose target.

The primary objective of this study was to assess the safety and efficacy of the automatic BPS in RCKT and compare the glycemic performance, with and without meal announcement in RCKT, with our legacy AP control system, the Unified Safety System Virginia (USS) system (1,2,13). We hypothesized that RCKT with the BPS would be safe and would increase glucose TIR compared with USS in the absence of carbohydrate announcement. We tested these hypotheses among adolescents, the age-group of individuals with T1D who arguably stand to benefit the most from this type of improvement in diabetes care (12).

The University of Virginia Institutional Review Board and the U.S. Food and Drug Administration (IDE#G200206) approved this randomized controlled clinical trial (ClinicalTrials.gov NCT04545567). Written informed consent was obtained from at least one parent and written assent was obtained from each adolescent participant. Twenty-one subjects aged 12–20 years were recruited to participate in a 5-night hotel-based study, described further below. Inclusion criteria consisted of a documented diagnosis of T1D and insulin pump therapy for ≥6 months. Exclusion criteria included diabetic ketoacidosis or a severe hypoglycemic event (defined as seizure or loss of consciousness) in the past 6 months, use of an oral glucose-lowering agent, including metformin, any medical condition deemed high-risk by the investigator, or recent exposure to someone known to be infected with severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2).

The study design is outlined in Fig. 1. Enrollment visits were performed via phone or secure internet video connection, and medical history and insulin use parameters were obtained. Documentation of a physical examination by a medical professional within the prior year were reviewed. For female participants, a urine pregnancy test was sent to the family and documentation of negative result was confirmed. Subjects and their parents received training on the Dexcom G6 CGM (Dexcom, San Diego, CA) as needed. Subjects were randomized 1:1 using a block design to begin on USS or RCKT. A pediatric endocrinologist reviewed each participant’s home insulin pump parameters at the time of enrollment and made adjustments with the family as needed. All participants provided at least 14 days of Dexcom G6 sensor data and insulin records from home use to individualize the RCKT system.

At study admission, all participants were started on a Tandem t:ap insulin pump set (Tandem Diabetes Care, San Diego, CA) with their home insulin parameters. Study participants also used the Dexcom G6 CGM. Both devices were connected to study-provided Android cell phones, which both reported back to the University of Virginia (Charlottesville, VA) remote monitoring system (DiAs Web Monitoring) (14) and ran the DiAs system (15) for either of the study AP systems described further below. A study blood glucose meter (ContourNext Link; Ascencia Diabetes Care, Parsippany, NJ) and study blood ketone meter (Precision Xtra; Abbott, Alameda, CA) were provided to all subjects for use as necessary in adherence to the glycemic guidelines listed below. Subjects were also fitted with a physical activity tracker (Fitbit Charge 3), the data from which was not used by either AP system.

This study compared two CLC systems, an investigational version of a commercially available HCL system (USS, control condition) that has been extensively tested in prior clinical trials (1,2,6,16–21), and a new, advanced hybrid CLC (RCKT, experimental condition), for their ability to control glycemia following unannounced meals. The USS was used as previously described (19); the “sleep mode” feature was programmed for participants at enrollment from 10:00 p.m. to 7:00 a.m.; the “exercise mode” was not needed in this study. The RCKT system integrates a novel BPS module within an MPC framework. MPC, rather than a single-control strategy, encompasses a general control paradigm. It incorporates an individualized mathematical model of the user’s glucose-related metabolism and finds the optimal insulin sequence that minimizes deviations from predetermined glucose targets and constraints. The user’s model is individualized into a linear time-invariant framework by using the data collection period, a well-established method in the literature (22).

In RCKT, the MPC part commands optimized microbolus doses through the pump every 5 min to react to internal metabolic disturbances while the BPS triggers escalating safe amounts of insulin (fractions of the total daily insulin) as the observed probability of a meal-like disturbance increases. Although RCKT is originally designed to function as a CLC system with the ability to recognize an unannounced meal and act on it, RCKT also has the capacity of functioning as a standard HCL. Once the user announces a meal, the meal bolus is computed as the 50% of the bolus calculated from the subject’s carbohydrate-to-insulin ratio and correction factor, overriding any simultaneous BPS action.

Following the previous ideas, RCKT uses model predictions from a personalized model and CGM measurements to determine the basal insulin dose (microbolus) that minimizes a cost function that includes 1) a term to correct the participant’s glucose concentration to the target, 2) a term penalizing low glucose values, and 3) a regularization term to weight the difference between two consecutive microboluses. In addition to the previous elements, RCKT was integrated with the legacy Safety System (23) and Hyperglycemia Mitigation System (24), enabling safe doses limited by the perceived risk of hypoglycemia and prevailing hyperglycemia, respectively. Due to the short duration of this trial, we did not allow online self-adaptation of the internal model but solely relied on the personalized data collection during the 2 weeks prior to the trial.

The BPS system runs every 5 min by estimating the probability that a meal-like disturbance has occurred in the prior 30 min. A priming bolus will then be requested when the disturbance probability is greater than a specific set of thresholds: for P = 0.3, 0.5, 0.7, and 0.9, the bolus requests will be of 3%, 5%, 6%, and 7% of total daily insulin, respectively. Every new dose will account for the insulin-on-board from preceding BPS doses. Further details of the BPS module are provided in the Supplementary Material.

Upon arrival to the study hotel, participants were started on the first AP system to which they had been randomized (Fig. 1). Data collection toward overall glycemic control on either of the systems began the day of arrival at ∼12:00 p.m. Participants were switched between alternate AP systems at ∼11:00 a.m. on hotel day 4, with data collection beginning immediately and continuing until discharge at ∼10:00–11:00 a.m. on hotel day 6. This design resulted in four 22- to 23-h segments to analyze (noon to 10:00–11:00 a.m. the next day), two with RCKT, and two with USS, each segment containing three meals, with dinner being announced (first segment of each algorithm) or not (second segment).

On hotel days 2–5, participants ate dinner at ∼6:00 p.m. Each participant had chosen a specific meal that was identical between these study days and contained 44–62 g of carbohydrate, 35–42 g of protein, and 27–41 g of fat. On hotel days 2 and 4, the carbohydrate amount of this dinner was entered into the CLC system, and the participant received the corresponding insulin bolus (in USS, this was the full bolus according to the home insulin-to-carbohydrate ratio; in the RocketAP system, this was half of the insulin from the home insulin-to-carbohydrate ratio). On hotel days 3 and 5, no dinner information was entered into the CLC system. Data for the primary outcome were then collected until 6 h after dinner. Additional meals during study days contained approximately 47 ± 9 g of carbohydrate and were always entered into the CLC system, and the corresponding insulin bolus was given. For activity, participants had a supervised light walk of 1–2 miles daily on each of the study days and were otherwise in their hotel rooms or engaging in social activities with the other study subjects.

Members of the study team remotely monitored participants’ real-time CGM data through DiAs Web Monitoring for the entirety of hotel portion of the study and alerted study nurses, physicians, and technicians regarding glycemic concerns and device connection issues. In addition, at least one member of the study team was present at all times during activities and monitored participants via a mobile device. Most treatment decisions were based on CGM data, with self-monitoring of BG performed at the discretion of the medical team. The study team intervened with ∼5–15 g of fast-acting carbohydrate if the CGM reading was <70 mg/dL throughout the study admission.

For hyperglycemia, the study team intervened if the participant was noted to be >300 mg/dL for ≥2 h. At that point, the participant was asked to check the infusion site and check a blood ketone level. If a problem with the pump site was suspected, the pump site was changed, and a correction dose of insulin was given along with oral hydration. Ketones were also checked regardless of BG for vomiting, fever, or significant illness. Blood ketones were treated if elevated per our glycemic treatment guidelines until the participant’s CGM read between 70 and 250 mg/dL and ketones were <0.6 mmol/L. Participants were required to stop the study and transition back to their home pump for ketones ≥1.5 mmol/L.

All participants and study staff interacting personally with participants were required to have negative SARS-CoV-2 testing 2–3 days before the study and again on the day of study arrival. To minimize the chance of viral exposure, individuals not associated with the study were not permitted on the floor of the hotel where participants and study staff stayed.

All glycemic outcomes were computed based on CGM records (25). The primary outcome was percentage time-in-range between 70 and 180 mg/dL (TIR) from dinner until 6 h later. Secondary glycemic outcomes included time-in-tight-range 80–140 mg/dL (TTR), percentage of time in hypoglycemia (<50, <60, <70 mg/dL), percentage of time in hyperglycemia (>180, >250, >300 mg/dL), average CGM number of hypoglycemia events (CGM glucose <70, regardless of symptoms), and total amount of insulin used—all for both dinner time to 6 h later and 6:00 p.m. to 7:00 a.m. Additional outcomes were the percentage of time in closed-loop, coefficient of variation, total daily injected insulin, and number of hypoglycemic treatments (26) All of these were prepared for presentation via standard guidelines (25). Outcomes were further divided into segments of the day: dinner–dinner + 6 h, dinner–dinner + 12 h, and overnight (11:00 p.m.–7:00 a.m.). Paired t tests or nonparametric Wilcoxon signed rank tests were used to compare the control and experimental admissions (both RCKT and USS comparisons and announced and unannounced dinners within each of these) in terms of glycemic outcomes in the case of normally or nonnormally distributed samples, respectively (Shapiro-Wilk test, quantile-quantile plot). The significance level was set at P < 0.05. No correction for multiple analysis was performed. Data are reported as mean ± SD if normally distributed and as median and interquartile range (IQR) if nonnormally distributed. Data formatting and preparation, as well as the statistical analysis, were done with MATLAB R2019b (MathWorks, Natick, MA) and SPSS 27 (IBM, Armonk, NY) software.

In total, 21 adolescents were enrolled in the study, and 18 participants completed the study and were included in the analysis. Excluded participants included one participant who did not respond to ongoing contact for study arrangements, one participant who tested positive for SARS-CoV-2 during the isolation period, and one participant who tested positive for SARS-CoV-2 upon arrival at the hotel. Baseline data for the 18 participants are listed in Supplementary Table 1. The average for age was 15.6 ± 1.6 years (range 13–20) and for HbA_1c was 7.4% ± 1.5 (range 5.8–12.6 [57 ± 4.5 mmol/mol; range 40–114]). Sex distribution was 9 females and 9 males.

For the following results, statistical comparisons are set as RCKT versus USS unless, stated otherwise. TIR 70–180 mg/dL for the period 6:00 p.m.–12:00 a.m. after the unannounced dinner (primary outcome) was significantly higher in RCKT than with USS (83% [IQR 64–93] vs. 53% [40–71]; P = 0.004) (Table 1). TTR 80–140 mg/dL was also higher in RCKT (49% [IQR 41–59] vs. 27% [22–36]; P = 0.002). Mean CGM and percentage of time >180 mg/dL were significantly lower in RCKT (141 ± 21 mg/dL vs. 166 ± 26 mg/dL; P = 0.001 and 17% [IQR 1.3–34] vs. 47% [28–60]; P = 0.01), respectively (Table 1). The difference in glycemia during this period is further depicted in Fig. 2 (shown for individual participants in Supplementary Fig. 1). Improvements in glycemia for RCKT versus USS extended in the 12-h period after the unannounced dinner (6:00 p.m.–6:00 a.m.) (Supplementary Table 2): TTR (73% [IQR 70–76] vs. 52% [42–64]; P < 0.001), TIR (90% [79–96] vs. 75% [58–83]; P = 0.004), mean CGM (123 ± 11 mg/dL vs. 145 ± 25 mg/dL; P < 0.001), and exposure to hyperglycemia percentage of time >180 mg/dL (8.7% [3–17] vs. 25% [14–37]; P = 0.003).

The RCKT system also outperformed USS following the announced dinners, showing a higher TIR and lower percentage of time <70 mg/dL (100% [IQR 99–100] vs. 93% [85–99], P = 0.004; and 0% [0–0] vs. 0.7% [0–0.8], P = 0.04, respectively) (Table 1). Time <70 mg/dL after the announced dinner was lower for RCKT versus USS (1.3 ± 2.8% vs. 4 ± 6%).

Overall control during ∼46 h revealed that RCKT had higher TTR and TIR (72.3 ± 7.9% vs. 63.7 ± 13%, P = 0.01; and 87 ± 6.6% vs. 80 ± 9.6%, P = 0.007, respectively) and lower mean BG and percentage of time >180 mg/dL (122 ± 7.5 mg/dL vs. 128 ± 15.5 mg/dL, P = 0.05; and 9.4 ± 5.6% vs. 13.4 ± 8.7%, P = 0.03, respectively) (Supplementary Table 2). Compared with baseline control, there was an increase in TIR on USS for 15 of 18 participants and on RCKT for 17 of 18 participants (Supplementary Table 4 and Fig. 3). Comparing RCKT to USS, there was an increase in TIR for 13 of 18 participants.

It is worth noting that remarkably tight control was achieved overnight by RCKT versus USS for TTR 70–140 mg/dL (95.3% [IQR 90.4–100] vs. 76.3% [58.5–87.4], P < 0.001) as well as TIR 70–180 mg/dL (99.2% [95.7–100] vs. 92.2% [81.2–96], P < 0.001) and mean BG (106.4 ± 7.3 mg/dL vs. 123 ± 20 mg/dL, P = 0.002) (Supplementary Table 5). With notably higher exposure to hyperglycemia, the percentage of time >180 mg/dL with the legacy controller was 0% (IQR 0–0) vs. 5% (0–9.8, P = 0.002).

During the 6-h (postprandial) and 12-h (postprandial and overnight) periods after dinner, there was almost no time in hypoglycemia for either RCKT or USS, <1% of the time in both systems (Table 1 and Supplementary Table 2), and the controllers did not differ in the number of hypoglycemic events (Supplementary Table 6).

During the overall study period (compared with the 6-h and 12-h periods after study dinners), more time was spent in hypoglycemia, 2.2% for RCKT and 2.9% for USS (P = 0.22) (Supplementary Table 3). Similarly, the overall number of hypoglycemia events per person was greater during the entire study, 3 for RCKT and 4.5 for USS. This was also a greater amount of time spent <70 mg/dL than was seen at baseline care for all participants (Supplementary Table 4 and Fig. 3). The number of these hypoglycemia events that occurred in the time around the light activity (1–2 mile walks daily, from 3:00–6:30 p.m.) was 36 (1 per person per day) for RCKT and 40 (1.1 per person per day) for USS, representing ∼56% of the total number of hypoglycemia events for the study.

The infused insulin amounts following dinners were similar for both systems (Table 1), while overall insulin amounts were lower on RCKT versus USS (78 ± 23 units vs. 85 ± 20 units) (Supplementary Table 3). The BPS in RCKT stimulated the administration of priming insulin doses in response to unannounced meals 83.4% of the time (15 of 18 dinners). This system was never triggered by BG fluctuations at any other portion of the trial. The mean of priming doses delivered was 2.4 units (range 0–6) (Supplementary Table 7).

There were no device-related adverse events during the study.

While HCL systems in long-term use provide improvements in glycemia (1–4), there remains a need for FCL systems that can offer adequate BG control in the absence of user meal input. We demonstrated improved TIR following an unannounced meal for RCKT, a novel MPC system with a dedicated BPS to address meal-like disturbances compared with a legacy CLC system. The RCKT system had a TIR of 83% during the 6 h after a dinner of 34–62 g of carbohydrate that was unannounced—supporting the potential for this type of system to improve overall control among individuals who occasionally omit announcement of carbohydrate ingestion. Importantly, this study assessed efficacy in adolescents, an age-group particularly notorious for missing snack and meal announcements, contributing to high HbA_1c levels (10–12).

The central feature of RCKT in improving glycemia after unannounced carbohydrate was the pairing of the BPS with an advanced control system (MPC), which allowed safe insulin priming that only occurred when glycemic excursions were derived from carbohydrate intake and not during infusion set failures, for example. These priming doses appeared effective at reducing the spike in BG after unannounced meals, with a lower peak BG and more rapid return to BG near target. Furthermore, the priming system appeared safe, with no postprandial lows following the unannounced dinner and no accidental triggers from BG fluctuations (e.g., dawn phenomenon) at other parts of the trial. The priming system is designed to deliver increasingly higher insulin doses depending on the probability the system has determined that a meal took place. The highest cumulative priming dose delivered in this study was 6 units, and the highest amount delivered as a proportion of total daily insulin was 7%. In this way, the system is intended to avoid hypoglycemia if the BPS determines a high probability of a meal having occurred when in fact none has. We did not observe such an erroneous trigger of the meal detection system in this limited trial, although clearly, longer studies are needed to confirm the safety of this systems in common use.

While the priming system did not result in hypoglycemia, we noted a higher-than-expected number of hypoglycemia events for both RCKT and USS systems overall (3 and 4.5 episodes per participant per day, respectively). These did not occur during the 6-h or 12-h after dinner but instead were other portions of the trial and were of unclear cause. Compared with other diabetes camp studies, this study included very little activity, walking a total of 1–2 miles over the course of each day, although this may still have represented more activity than some adolescents were accustomed to during the pandemic. One other possible explanation is that the home insulin parameters for some of the adolescents may have included relatively high basal rates as a response to missing some insulin for carbohydrates during home care. Still, another possibility is that the CLC systems were overly aggressive in their response to higher BG levels. Knowing that this was a possibility, for announced meals the RCKT system was designed to only deliver half of the usual insulin dose calculated from the participant’s insulin-to-carbohydrate ratio; however, even using this lower prandial insulin dose after announcing carbohydrate ingestion, there were occasional episodes of hypoglycemia. Finally, it should be noted that hypoglycemia tends to be underrepresented during supervised studies as CGM is monitored and hypoglycemic treatment rapidly administered. Future iterations of the system may require either further reductions in the insulin-to-carbohydrate dose delivered or a less robust response to hyperglycemia.

This system compares favorably to systems tested in settings of unannounced carbohydrate, acknowledging potential differences in conditions for participants in each trial. Weinzimer et al. (27) assessed adolescents using a Medtronic closed-loop system, reporting mean glucose levels of 141 mg/dL during FCL use compared with 135 mg/dL in HCL; TIR data were not reported. Cameron et al. (28) assessed CLC among adults in hospital and hotel settings with a system that included accelerometry-based exercise detection, finding 73–78% TIR. Dovc et al. (29) assessed adults in FCL using the DreaMed GlucoSitter in a setting that included structured exercise and unannounced meals, reporting TIR 57.9% and 53.3% when using standard aspart and faster aspart, respectively. Blauw et al. (30) assessed FCL in a bihormonal system (with a separate pump delivering glucagon) among adults at home, reporting TIR of 86.6% compared with 53.9% in open loop. Turksoy et al. (31) evaluated an automatic meal bolus algorithm integrated to an integrated multivariable adaptive AP based on the online estimation of the glucose R_a, reporting TIR of 71% compared with 63% for the continuous subcutaneous insulin infusion (control) condition in a small pilot study. Samadi et al. (32) assessed the benefits of a fuzzy-based meal detector system on a retrospective analysis using data collected from 11 subjects with T1D, reporting 93.5% and 68% detection sensitivity for meals and snacks, respectively, with a detection time of 35 ± 23 min. Future side-by-side assessments of systems in similar populations and settings would be helpful for more direct comparison of efficacy.

In this system, we used changes in CGM glucose patterns to trigger the priming meal bolus. This approach allows for sensor glucose disturbances to trigger insulin dosing, which only occur in the absence of carbohydrate announcement. The safety of a late user-initiated meal bolus would depend on whether the BPS has already provided insulin, in which case the system would provide a message regarding the prior BPS dose; in deciding on the need for a further user-initiated bolus, users would then have to be mindful of how much insulin the BPS has already delivered. Alternative means of addressing unannounced carbohydrate in a CLC system include features such as entrained anticipation of meals or activity, where individuals who eat or exercise at predictable times most days cause the system to adjust insulin delivery in the time leading up to that period (33). Other approaches toward meal detection could also be achieved via external sensing of meal-type movement, such as through a wrist-worn accelerometer such as a smart watch.

Even in the absence of the BPS, the MPC algorithm in RCKT achieved better TIR than USS, as seen following the announced dinner. This may relate to this being a more advanced controller including an individualized glucose prediction directly into the decision-making process. Apart from reacting to the difference with respect to the target, RCKT modulates the aggressiveness as a function of the rate of glucose change, providing an opportunity to inject higher microboluses when glucose levels suddenly rise and to back off promptly when glucose levels trend downward. This study benefitted from a randomized crossover design, allowing direct comparison of four CLC dinner approaches in a group of adolescents, the age range likely most affected by unannounced carbohydrate ingestion.

Limitations of early supervised clinical trials, including this study, are that they do not allow extrapolation of the system’s safety and efficacy outside of this restrictive setting; of note, fixed meal schedules, restrained carbohydrate intake, lack of intensive exercise, and absence of daily stressors may affect glycemic control. High-intensity activity may be a particular challenge for fully closed-loop systems (19,33,34), in which preexercise adrenaline could result in elevations in glucose that may trigger inappropriate insulin dose escalation by the BPS. Additionally, while the order of the controller was randomized, there was a fixed order of the announced and unannounced dinners and a lack of a significant wash out period between controller sessions, all of which could have influenced glycemic control, particularly the days after unannounced carbohydrate at dinner. Therefore, further tests will be needed in free-living conditions.

In conclusion, this system, with automatic prandial bolus priming, achieved 83% in the 6-h period around a dinner without carbohydrate announcement among adolescents. While HCL overcomes some of the inherent delays in CLC via calculating insulin needed for carbohydrate and delivering the dose before eating, automated meal detection for prandial glycemic management may help achieve an acceptable degree of control for individuals who miss opportunities to announce carbohydrate ingestion during use of a CLC system.

Acknowledgments. The authors thank the study volunteers and the research staff at the University of Virginia Center for Diabetes Technology.

Funding. This project was funded by generous grants from the University of Virginia’s LaunchPad Program (to M.D.D. and M.D.B.) and the Strategic Investment Fund project “Precision Individualized Medicine for Diabetes (PriMeD).”

Duality of Interest. M.D.D. reports receiving grant or material support from Tandem, Medtronic, and Dexcom. M.D.B. has received honoraria and travel reimbursement from Dexcom and Tandem, and research support from Dexcom, Novo Nordisk, and Tandem. No other potential conflicts of interest relevant to this article were reported.

Author Contributions. J.G.-T. designed the core of the experimental CLC system, was involved in the study design, provided technical support remotely and on-site in all the admissions, analyzed data, performed statistical analyses, and was involved in writing and editing the manuscript. J.L.D., R.E.-Z., and C.L.B., provided on-site support during screenings and admissions and were involved in writing and editing the manuscript. C.L.K.K. maintained and checked the functionality of the DiAs system, the functionality of the closed-loop systems in the smartphone, conducted the testing of the systems before the study, provided technical support remotely and on-site in all the admissions, and reviewed and edited the manuscript. J.P.C. designed the BPS logic for the experimental controller and reviewed and edited the manuscript. M.D. and C.W. coordinated the study monitoring, processed data, and were involved in writing and editing the manuscript. M.C.O assisted with the development of the protocol, coordinated institutional review board submission and investigational device exemption submission to the U.S. Food and Drug Administration, contributed to the study planning and preparation, and reviewed and edited the manuscript. H.M and K.K. were the research coordinators for the study and responsible for all participant interaction and operations during the trial, and reviewed and edited the manuscript. M.D.B. contributed to the algorithm and study design, supervised its implementation, performed statistical analyses, and was involved in writing and editing the manuscript. M.D.D. was the study physician responsible for all participant activities and was the University of Virginia site principal investigator, contributed to the study design, answered queries from the regulatory boards, including the U.S. Food and Drug Administration, and was involved in writing and editing the manuscript. M.D.D. 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.

Prior Presentation. Parts of this study were presented in abstract form at the 14th International Conference on Advanced Technologies & Treatments for Diabetes (ATTD 2021), 2–5 June 2021.