Randomized Crossover Comparison of Automated Insulin Delivery Versus Conventional Therapy Using an Unlocked Smartphone with Scheduled Pasta and Rice Meal Challenges in the Outpatient Setting

Abstract

Background: Automated Insulin Delivery (AID) hybrid closed-loop systems have not been well studied in the context of prescribed meals. We evaluated performance of our interoperable artificial pancreas system (iAPS) in the at-home setting, running on an unlocked smartphone, with scheduled meal challenges in a randomized crossover trial.

Methods: Ten adults with type 1 diabetes completed 2 weeks of AID-based control and 2 weeks of conventional therapy in random order where they consumed regular pasta or extra-long grain white rice as part of a complete dinner meal on six different occasions in both arms (each meal thrice in random order). Surveys assessed satisfaction with AID use.

Results: Postprandial differences in conventional therapy were 10,919.0 mg/dL × min (95% confidence interval [CI] 3190.5–18,648.0, P = 0.009) for glucose area under the curve (AUC) and 40.9 mg/dL (95% CI 4.6–77.3, P = 0.03) for peak continuous glucose monitor glucose, with rice showing greater increases than pasta. White rice resulted in a lower estimate over pasta by a factor of 0.22 (95% CI 0.08–0.63, P = 0.004) for AUC under 70 mg/dL. These glycemic differences in both meal types were reduced under AID-based control and were not statistically significant, where 0–2 h insulin delivery decreased by 0.45 U for pasta (P = 0.001) and by 0.27 U for white rice (P = 0.01). Subjects reported high overall satisfaction with the iAPS.

Conclusions: The AID system running on an unlocked smartphone improved postprandial glucose control over conventional therapy in the setting of challenging meals in the outpatient setting. Clinical Trial Registry: clinicaltrials.gov NCT03767790.

Introduction

Automated insulin delivery (AID) systems have been extensively studied in persons with type 1 diabetes (T1D).¹ In a number of recent studies of hybrid closed-loop AID, the largest improvements were in overnight glucose control and reductions in hypoglycemia,^2–7 but postprandial hyperglycemia still remains a significant challenge.⁸ The challenges for postprandial glycemic control include relatively slower onset of action for subcutaneous rapid-acting insulin compared to the normal physiological response, inter- and intrasubject variability, and the fact that many AID systems require mealtime insulin boluses to be initiated by users, which may constrain postprandial insulin delivery to prevent later hypoglycemia.

In addition to insulin delivery, meal choices are essential to successful glycemic control for persons with T1D.^9,10 While many different bolus strategies have been explored to improve postprandial glycemic outcomes,^11,12 we have shown that there remains a significant difference in the postprandial glycemic response in persons with T1D between different foods when consumed in equal amounts of carbohydrate.¹³ Most AID studies do not mandate and do not report meal choices made during the studies, despite the well-recognized importance of meal choices, different carbohydrate sources of foods, differing meal macronutrient content, and the importance of overall dietary management in impacting glycemic results in all forms of diabetes.^10,14,15

We performed the first 24/7 at-home AID study requiring specific meal choices during the study to compare postprandial, as well as overall, glucose control related to AID use versus conventional therapy (sensor-augmented pump [SAP] to include predictive low-glucose suspend [PLGS]). In addition, to facilitate ease of use and user acceptance of the AID system, the system ran on an unlocked smartphone. We hypothesized that the AID system would be easy to use, would improve overall safety and efficacy over conventional therapy, and would improve glycemic control for the specified study meals.

Research Design and Methods

Study design

Adult subjects with T1D first completed a 2-week conventional therapy run-in period with a Dexcom G6 continuous glucose monitor (CGM) if they were not current Dexcom CGM users. After the run-in, subjects were randomized to each treatment arm, and meal order was randomized within each treatment arm, using permuted block randomization with the pseudorandom number generator of Wichmann and Hill as modified by McLeod.^16,17 Subjects were assigned 2 weeks (±3 days) of AID use (intervention arm) and 2 weeks (±3 days) of conventional therapy using SAP/PLGS (control arm) at home in random order, after which they crossed over to the other treatment arm (Supplementary Fig. S1). In the control arm of the study, subjects were allowed to continue use of the PLGS feature of their home pump if they were already using this and it was compatible with the study CGM, otherwise they used SAP.

During each 2-week period after randomization, subjects consumed preweighed portions of either (1) a pasta made from 100% semolina flour or (2) extra-long grain white rice (Supplementary Table S1) as part of a complete dinner meal on six different occasions in both arms (each meal thrice in random order, Supplementary Fig. S1). Subjects were specifically instructed how to prepare the assigned study meals when they had time to cook the meal and prepare a salad to compliment the meal, have minimal insulin on board from prior boluses, and could assure their best to try not to give further carbohydrate intake or extra insulin over the next 5 h after the meal started. Subjects were instructed to consume one to two meal portions per assigned study meal (same amount for each study meal) and, specifically, match their premeal insulin bolus with their carbohydrate ratio accordingly. This allowed subjects flexibility to incorporate the study protocol into their daily lives for those who preferred smaller or larger portions, as long as they used the same portion size for each study meal. Extended boluses were not used for the study meals. Subjects were instructed to wait for 5 h after the start of the study meal to give any correction bolus. Subjects were also instructed to perform their usual activity level before and after each study meal.

Eligible participants were ≥18 years of age with T1D for at least 1 year, using an insulin pump for at least 3 months, and had an HbA1c <10.5%. Key exclusion criteria were pregnancy, one or more episodes of hypoglycemia or hyperglycemia requiring an emergency room visit in the past 6 months, gastrointestinal disease such as celiac disease, history of gastroparesis, multiple food allergies, any form of gluten sensitivity or wheat allergy, or allergies to any ingredients present in the study meals.

The study was conducted by the Sansum Diabetes Research Institute (Santa Barbara, CA). Design of the control algorithms and engineering of the AID device were done at the Harvard John A. Paulson School of Engineering of Applied Sciences, Harvard University (Cambridge, MA). The nutrition component of the protocol was developed in collaboration with the Harvard T.H. Chan School of Public Health (Boston, MA). The protocol was approved by the U.S. Food and Drug Administration (FDA) and the Advarra Institutional Review Board and was registered at clinicaltrials.gov (NCT03767790). Informed consent was obtained before all study procedures.

AID system

We have previously reported on the design, development, and initial clinical evaluation of our smartphone-based AID application platform, the interoperable artificial pancreas system (iAPS).¹⁸ The iAPS runs on an unlocked smartphone and interfaces wirelessly with leading CGMs, insulin pump devices, and decision-making algorithms. In this study, subjects used the Tandem t:AP insulin pump and Dexcom G6 CGM wirelessly connected to the iAPS running on the unlocked Google Pixel 2 smartphone when in the closed-loop arm and used their personal pump and study Dexcom G6 CGM in the SAP/PLGS conventional therapy arm. For subjects who preferred not to have to carry two phones (their personal phone and the study smartphone), they were allowed to place their personal SIM card into the study phone and use it as their personal phone while it ran the AID system concurrently with any apps they chose to run. To the best of our knowledge, this is the first time an unlocked smartphone has been used in an FDA approved investigational device exemption (IDE) outpatient study.

Boluses for all meals under closed loop were given at mealtime based on each subject's personal carbohydrate ratio as previously described,⁵ although the iAPS allowed modification of these boluses by each user at their discretion. Boluses for the six study meals in each arm were determined ahead of time, and subjects were instructed to give the meal bolus at the start of the meal without any correction dose. For nonstudy meals, participants could select the type and carbohydrate content of their meals and snacks according to their dietary preferences and adjust their meal bolus accordingly.

The Zone Model Predictive Control (MPC) algorithm¹⁹ running on iAPS used a target glucose range of 90–120 mg/dL during the day and 100–120 mg/dL during the night (12 am to 4 am). The MPC insulin delivery was constrained to four times for each subjects' nominal basal insulin profile from 10 pm to 4 am as a safety constraint.

Safety monitoring during AID use

Similar to our prior studies,^3,5,18,20 the AID system included a Health Monitoring System (HMS) that was used to monitor CGM glucose independent from the AID control algorithm.²¹ The HMS advised users to ingest fast-acting carbohydrate to prevent impending hypoglycemia (CGM <65 mg/dL) that could not be prevented by controller action alone.

As this was the first 24/7 at-home study with the iAPS, remote monitoring alerts to study staff were used to notify for prolonged hyper- or hypoglycemia during AID use, as well as for prolonged connectivity interruptions. Subjects or a predesignated emergency contact was contacted, if necessary, in response to these alerts.

Outcome measures

The primary outcome for the overall study was percent time in glucose range 70–180 mg/dL measured by CGM to determine safety and efficacy of the integrated system in this outpatient exploratory study on an unlocked smartphone. Secondary outcomes included percent time CGM glucose <70 mg/dL, percent time CGM glucose <54 mg/dL, percent time CGM glucose >180 mg/dL, percent time CGM glucose >250 mg/dL, mean glucose, median glucose, standard deviation (SD) and coefficient of variation of glucose as a measure of glycemic variability, low blood glucose index (LBGI), high blood glucose index, and percent time CGM was active, as per recent CGM outcome consensus report guidelines.^22–24 Additional outcomes included insulin delivered, percent time closed loop was active, and tracking of adverse events and adverse device effects.

Glycemic outcomes specific to the study meals included the following: total postprandial glucose area under the curve (AUC, mg/dL × min) for the 6-h postprandial period after the start of the study meals, time to peak CGM glucose (in minutes from the start of the study meals), peak CGM glucose (mg/dL) during the 6-h postprandial period, glycemic variability percentage (GVP, %)²⁵ as a measure of glucose variability in the 6-h postprandial period, and total glucose AUC when glucose was less than 70 mg/dL (AUC under 70 mg/dL, mg/dL × min) for the 6-h postprandial period. Insulin outcomes specific to the study meals included the following: meal bolus (U) at the start of the study meals and postprandial sum of insulin microbolus (around basal) for the 6-h postprandial period after the study meal boluses. The glucose AUC and insulin sum outcomes were also analyzed over 2 h periods (0–2, 2–4, and 4–6 h) to investigate dynamic changes during a given meal.

Validated surveys^26–30 were given before and after the 2 weeks of AID use to assess hypoglycemic awareness, diabetes distress, and behaviors related to hyper- and hypoglycemia avoidance and treatment. The Technology Acceptance Questionnaire, given at the end of the 2 weeks of iAPS use, was used to assess acceptance and ease of use of the system, with a score of 1 (strongly disagree) to 5 (strongly agree) for each question.³¹ The System Usability Scale (SUS) assessed overall usability.³²

Sample size calculation

This was an exploratory study using the iAPS for the first time in the 24/7 at-home setting, with the addition of specific study meals during open and closed-loop use. Sample size was based on detecting a 25% reduction of peak glucose rise (maximum rise from baseline glucose) from the start of each study meal to the peak CGM glucose reading (mg/dL) during the 6-h postprandial window, compared between white rice and regular pasta. Based on our prior study comparing rice and pasta meals in the open loop setting, which showed a peak glucose difference of 51.4 mg/dL,¹³ we estimated that having at least eight subjects performing six meals in each arm would have a 95.6% power to detect a significant difference in mean glucose peak between the study meals (G-Power 3.1.9.4).

Statistical analysis

Data analysis was performed using the intent-to-treat approach. All data for ten subjects were analyzed for overall glycemic metrics. For study meals, one subject's meal data could not be included in the analysis for both open and closed-loop because of correction insulin boluses given within 2 h of the study meals on almost all occasions and unclear written meal logs that were inconsistent with downloaded pump bolus records. Two other subjects only completed 11 and 9 meals, respectively, and 2 meals were missing more than 30% of the CGM data postprandially. All other study meals that were completed were included in the analysis, which was a total of 102 meals. Any rescue or correction events during study meals were used as termination points of the intervention (19 out of 102 meals).

Outcomes were evaluated using the linear mixed effect model (LMEM) in which treatment was included as primary fixed effect for AID versus conventional therapy comparison, while for meal challenges, meal, treatment, and its interaction were included as primary fixed effects. To account for excess zeros in glucose AUC under 70 mg/dL data, a two-part model was fit with a logistic mixed effects model for the binary outcome and a LMEM for the logarithmic transformation of the nonzero responses conditional on the binary model.^33,34 The random effects on both intercept and slope with subjects as grouping factor were considered using a maximal model structure justified by the experimental design to minimize Type I error rates.³⁵

Since the effect of a prior meal and its bolus could potentially be evidenced in the glycemic response of the dinner study meal,¹⁴ we adjusted the outcomes for covariates such as starting CGM value and the rate of glucose change at the start of the study meals as previously reported.¹³ The outcomes were also adjusted for amount of carbohydrate consumed and for period effect related to treatment, as well as meals. The meal types were blinded for statistical analysis. Further details on models are provided in Additional Information on Statistical Analysis section of the Supplementary Data.

Data are reported as mean ± SD, and significance is reported at 0.05 level. The P-values from mixed effects model individual estimates and their linear combination are reported using the Satterthwaite approximation for degrees of freedom. The outcomes were based on consensus guidelines and were declared before conducting the study and the subsequent data analysis, and there were no corrections made for multiple comparisons. The data were analyzed using Matlab 2018b (The MathWorks, Inc.) and R Language (R Development Core Team, 2019).

Results

Patients and duration of AID use

Ten adults with T1D were enrolled and completed the study. Mean age was 45.2 ± 17.0 years, and HbA1c at randomization was 6.8% ± 0.7%. Baseline demographics are summarized in Supplementary Table S2. One subject used the PLGS feature of their personal Tandem insulin pump in the control arm of the study. Two subjects chose to move their personal SIM card to the unlocked study phone and use the study phone as their personal phone during the AID arm of the study, which they completed successfully without issue. Overall, subjects were in closed loop for 85.2% ± 11.2% of the time during the 2-week closed-loop period, which equates to mean of 11.93 ± 1.6 days of closed-loop use per subject. During AID use, the percentage of time CGM was available was 98.5% ± 1.8%, while percentage of time a valid CGM reading between 40 and 400 mg/dL was available was 90.7% ± 4.0%.

Study meal results

Glucose and microbolus insulin data from the study meals, stratified by pasta and white rice, are shown in Figure 1, where it can be observed that CGM glucose was elevated for white rice relative to pasta, while with AID use insulin delivery was frequently reduced below basal for pasta while it was above basal for white rice in the later postprandial period.

FIG. 1.

Sensor glucose in mg/dL (top row) and microbolus insulin in units (bottom row) from start of the study meals (Time Zero) through 360 min, stratified by pasta and white rice. Data are shown as median (IQR) across subjects for each time point. The CGM profiles are adjusted at baseline by subtracting the CGM value at the start of the meal at each CGM time point for each subject. CGM, continuous glucose monitor; IQR, interquartile range.

Complete study meal results of difference between meals for each treatment and difference between treatments for each meal for glucose and insulin outcomes are shown in Table 1. Descriptive statistics for raw glucose and insulin outcome stratified by each combination of meal and treatment are shown in Supplementary Table S3, and estimates of individual parameters of the study meal result model for glucose and insulin outcomes are shown in Supplementary Table S4. Main observations on effect of meals and treatments are as follows. The meal bolus did not significantly change between meals and treatments, confirming that subjects generally followed the meal bolusing instructions. For AID-based control, the insulin microbolus sum over 6 h trended toward increase for white rice over pasta, but not statistically significantly, by 0.79 U (P = 0.09). Furthermore, for pasta, AID-based control decreased insulin delivery over conventional therapy in the 0–2 h period by 0.45 U (95% confidence interval [CI] −0.69 to −0.21, P = 0.001), while for white rice, AID-based control decreased insulin delivery over conventional therapy in the 0–2 h period by 0.27 U (95% CI −0.48 to −0.07, P = 0.01). AID-based control increased insulin delivery for white rice over pasta by 0.73 U (95% CI 0.26–1.20, P = 0.01) in the 2–4 h period.

Table 1.

Estimates of Difference Between Meals (White Rice – Pasta) for Each Treatment and Difference Between Treatments (Automated Insulin Delivery–Conventional Therapy) for Each Meal During Meal Challenges for Glucose and Insulin Outcomes Using Linear Mixed Effects Models

Outcomes	White rice–pasta in conventional therapy (${\beta }_1$)	White rice–pasta in AID (${\beta }_1+{\beta }_3$)	AID–conventional therapy for pasta (${\beta }_2$)	AID–conventional therapy for white rice (${\beta }_2+{\beta }_3$)
Glucose outcomes
AUC (0–6 h), mg/dL × min	10,919 (3190.5 to 18,648), P = 0.009a	6522.1 (−3615.9 to 16,663.2), P = 0.24	5651.3 (−2842 to 14,145), P = 0.17	1254.3 (−5975.8, 8486.7), P = 0.73
AUC (0–2 h), mg/dL × min	1634.4 (−598.81 to 3867.6), P = 0.14	2423.5 (−198.7 to 5046.4), P = 0.1	98.8 (−1596.6 to 1794.4), P = 0.9	888 (−2021.4 to 3799.4), P = 0.56
AUC (2–4 h), mg/dL × min	505.35 (−1145.1 to 2155.8), P = 0.52	−79.15 (−1444.5 to 1287.1), P = 0.91	711.19 (−703.99 to 2126.4), P = 0.29	126.68 (−1057 to 1312.5), P = 0.83
AUC (4–6 h), mg/dL × min	181.12 (−1891 to 2253.3), P = 0.84	−1260.55 (−2864.7 to 343.3), P = 0.15	376.76 (−1797.3 to 2550.8), P = 0.68	−1064.91 (−2463.7 to 333.7), P = 0.15
AUC under 70 mg/dL,b mg/dL × min	−1.47 (−2.50 to −0.45), P = 0.004a	−0.37 (−2.15 to 1.4), P = 0.68	−0.20 (−1.05 to 0.64), P = 0.64	0.9 (−0.59 to 2.4), P = 0.23
Meal peak, mg/dL	40.91 (4.55 to 77.28), P = 0.03a	16.02 (−11.12 to 43.18), P = 0.27	20.62 (−14.24 to 55.50), P = 0.21	−4.25 (−28.82 to 20.3), P = 0.73
Meal peak time, min	28.0 (−67.41 to 123.41), P = 0.52	−78.72 (−136.9 to −20.53), P = 0.01a	53.19 (−30.86 to 137.26), P = 0.18	−53.53 (−121.62 to 14.56), P = 0.15
GVP, %	9.39 (−1.07 to 19.86), P = 0.07	−0.94 (−7.60 to 5.95), P = 0.78	8.99 (0.74 to 17.23), P = 0.03a	−1.35 (−10.77 to 8.24), P = 0.79
Insulin outcomes
Meal bolus, U	−0.12 (−0.42 to 0.17), P = 0.37	0.008 (−0.25 to 0.27), P = 0.94	0.2 (−0.16 to 0.56), P = 0.24	0.33 (−0.03 to 0.71), P = 0.1
Microbolus sum (0–6 h), U	−0.12 (−0.67 to 0.46), P = 0.69	0.79 (−0.02 to 1.61), P = 0.09	−0.47 (−1.81 to 0.86), P = 0.44	0.43 (−1.04 to 1.92), P = 0.57
Microbolus sum (0–2 h), U	−0.02 (−0.20 to 0.17), P = 0.86	0.15 (−0.11 to 0.41), P = 0.3	−0.45 (−0.69 to −0.21), P = 0.001a	−0.27 (−0.48 to −0.07), P = 0.01a
Microbolus sum (2–4 h), U	0.04 (−0.30 to 0.38), P = 0.81	0.73 (0.26 to 1.20), P = 0.01a	−0.23 (−1.17 to 0.70), P = 0.56	0.47 (−0.15 to 1.07), P = 0.18
Microbolus sum (4–6 h), U	−0.01 (−0.34 to 0.31), P = 0.93	0.07 (−0.34 to 0.48), P = 0.75	0.24 (−0.82 to 1.30), P = 0.55	0.32 (−0.66 to 1.31), P = 0.53

The table reports the main parameter point estimates and their linear combination, 95% CI in parentheses and P-values for model $Y\sim {\beta }_0+{\beta }_{1}Rice+{\beta }_2{T}_{AID}+{\beta }_{3}Rice:T_{AID}$.

^aDenotes the estimated differences satisfying significance threshold of 0.05.

^bTwo-part model: results are from the log-normal mixed effects model on the nonzero part.

AID, Automated Insulin Delivery; AUC, area under the curve; CI, confidence interval; GVP, glycemic variability percentage.

Peak glucose increased for white rice over pasta by 40.9 mg/dL (95% CI 4.6–77.3, P = 0.03) in conventional therapy while for AID-based control was 16.1 mg/dL (95% CI −11.1 to 43.2, P = 0.27), but not statistically significant. Time to peak decreased for white rice over pasta by 78.7 min (95% CI −136.9 to −20.53, P = 0.01) in AID-based control.

Glucose AUC increased for white rice over pasta by 10,919 mg/dL × min (95% CI 3190.5–18,648.0, P = 0.009) in conventional therapy, while for pasta it also clinically increased, but not statistically significantly, in AID-based control over conventional therapy by 5651.3 mg/dL × min (95% CI −2842.0 to 14,145, P = 0.17). Glucose AUC increased for white rice over pasta in AID-based control by 6522.1 mg/dL × min (95% CI −3615.9 to 16,663.2, P = 0.24), but not statistically significantly.

Glucose AUC in 4–6 h period decreased for white rice over pasta by 1260.5 mg/dL × min (95% CI −2864.7 to 343.3, P = 0.15) in AID-based control while for white rice it also decreased in AID-based control over conventional therapy by 1064.9 mg/dL × min (95% CI −2463.7 to 333.7, P = 0.15), although neither reached statistical significance.

Regarding AUC under 70 mg/dL, white rice resulted in lower estimate over pasta by a factor of 0.22 (95% CI 0.08–0.63, P = 0.004) in conventional therapy. Glucose variability measured by GVP increased for white rice over pasta by 9.39% (95% CI −1.1 to 19.9, P = 0.07) in conventional therapy but not statistically significantly. For pasta, GVP also increased in AID-based control over conventional therapy by 8.99% (95% CI 0.7–17.2, P = 0.03). The difference between meals in AID-based control was −0.94% (95% CI −7.60 to 5.95, P = 0.78), although not statistically significant.

Overall glycemic outcomes

Overall glycemic outcomes are summarized in Table 2. Compared to SAP/PLGS, AID-based control clinically increased, but not statistically significantly, the primary outcome of percent time in range 70–180 mg/dL by 3.4% (95% CI −2.3 to 9.1, P = 0.22), while significantly decreasing time spent below 70 mg/dL by −1.2% (95% CI −2.1 to −0.2, P = 0.02) and LBGI by −0.3 (95% CI −0.5 to −0.1, P = 0.01). For overnight period, AID-based control clinically decreased, but not statistically significantly, glucose variability as measured by glucose SD by −9.3 mg/dL (95% CI −19.8 to 1.2, P = 0.08). The period effect was not found to be significant for the primary outcome.

Table 2.

Glycemic Metrics Mean (Standard Deviation) Comparing Performance of the Automated Insulin Delivery System with Conventional Therapy (Sensor Augmented Pump/Predictive Low-Glucose Suspend) for the Entire Study Period, with Estimated Effect Calculated Using a Linear Mixed Effects Model with Adjustment for Treatment Period

Metric	Conventional therapy (n = 10)	AID (n = 10)	Estimate (95% CI)	P
Day and night
Mean CGM glucose	150.7 (21.3)	147.7 (18.5)	−3.0 (−11.9 to 6.0)	0.47
Median CGM glucose	142.9 (20.1)	139.2 (18.8)	−3.8 (−11.6 to 4.1)	0.31
SD CGM glucose	54.2 (12.2)	50.6 (10.5)	−3.6 (−10.2 to 2.9)	0.25
CV CGM glucose (%)	35.7 (4.4)	34.2 (5.0)	−1.6 (−4.1 to 0.9)	0.18
%Time <54 mg/dL	0.6 (0.7)	0.4 (0.4)	−0.2 (−0.5 to 0.1)	0.20
%Time <60 mg/dL	1.3 (1.4)	0.9 (1.0)	−0.4 (−0.9 to 0.0)	0.07
%Time <70 mg/dL	3.5 (2.8)	2.3 (1.9)	−1.2 (−2.1 to −0.2)	0.02a
%Time 70–140 mg/dL	46.3 (12.0)	50.2 (11.5)	3.9 (−1.5 to 9.3)	0.14
%Time 70–180 mg/dL	70.6 (11.7)	74.0 (11.2)	3.4 (−2.3 to 9.1)	0.22
%Time >180 mg/dL	25.9 (13.1)	23.7 (11.9)	−2.2 (−7.9 to 3.5)	0.40
%Time >250 mg/dL	7.1 (7.2)	4.9 (6.1)	−2.1 (−5.8 to 1.6)	0.23
%Time >300 mg/dL	2.1 (3.0)	1.4 (2.5)	−0.7 (−2.6 to 1.3)	0.48
LBGI	1.0 (0.7)	0.7 (0.5)	−0.3 (−0.5 to −0.1)	0.01a
HBGI	6.0 (3.2)	5.3 (2.8)	−0.7 (−2.2 to 0.8)	0.31
Overnight (12 am to 6 am)
Mean CGM glucose	159.6 (23.6)	153.9 (22.7)	−5.8 (−20.2 to 8.6)	0.39
Median CGM glucose	150.2 (25.1)	148.6 (24.7)	−1.6 (−15.1 to 12.0)	0.80
SD CGM glucose	57.8 (14.3)	48.5 (14.4)	−9.3 (−19.8 to 1.2)	0.08
CV CGM glucose (%)	36.2 (6.9)	31.8 (8.9)	−4.5 (−9.7 to 0.8)	0.09
%Time <54 mg/dL	0.6 (1.2)	0.6 (1.1)	−0.0 (−0.9 to 0.9)	NS
%Time <60 mg/dL	1.3 (1.6)	1.2 (2.1)	−0.1 (−1.5 to 1.3)	0.87
%Time <70 mg/dL	2.6 (2.5)	2.8 (4.0)	0.1 (−2.3 to 2.5)	0.92
%Time 70–140 mg/dL	42.0 (13.9)	43.4 (16.7)	1.4 (−8.7 to 11.5)	0.76
%Time 70–180 mg/dL	67.0 (15.5)	70.4 (14.9)	3.4 (−7.7 to 14.4)	0.51
%Time >180 mg/dL	30.3 (16.8)	26.9 (16.2)	−3.5 (−15.2 to 8.3)	0.52
%Time >250 mg/dL	10.6 (10.0)	5.5 (7.3)	−5.1 (−10.7 to 0.6)	0.08
%Time >300 mg/dL	3.7 (5.5)	1.9 (3.9)	−1.8 (−5.7 to 2.1)	0.33
LBGI	0.8 (0.6)	0.7 (0.9)	−0.1 (−0.6 to 0.4)	0.60
HBGI	7.3 (3.8)	5.9 (3.4)	−1.4 (−3.7 to 0.9)	0.21

^aDenotes the estimated differences satisfying significance threshold of 0.05.

CGM, continuous glucose monitor; CV, coefficient of variation; HBGI, high blood glucose index; LBGI, low blood glucose index; SD, standard deviation; NS, not significant.

The paired comparison of mean glucose and time in range during intervention and control arms of the study is shown in Figure 2. For 8 out of 10 subjects time in hypoglycemia was reduced with AID use, as shown by the size of the bubble plots in Figure 2, as mean glucose decreased for 6 out of 10 subjects and time in range 70–180 mg/dL increased for 6 out of 10 subjects.

FIG. 2.

Paired comparison of mean glucose and time in 70–180 mg/dL range during SAP/PLGS and AID arms. The solid lines connect individual subjects, with the bubble size proportional to time in hypoglycemia below 70 mg/dL, while dashed lines show mean change with annotation. Use of AID resulted in decrease of percent time spent below 70 mg/dL in 8 out of 10 subjects. AID, Automated Insulin Delivery; PLGS, predictive low-glucose suspend; SAP, sensor-augmented pump.

Adverse events

There were no protocol-related or device-related adverse events during the study. There were no adverse device effects. Only two remote monitoring notifications for glycemia (one for prolonged hyperglycemia and one for prolonged hypoglycemia) required investigator intervention to call the subject as per protocol. In both cases the subjects were already aware of the issue and had already made an appropriate intervention to restore euglycemia.

Summary of survey results

Mean SUS score for use of the iAPS was 73 ± 14.3, correlating to a score of “Good.”³² With regards to the questions on the Technology Acceptance Questionnaire “How easy to use was the iAPS?”, “How useful in managing your diabetes was the iAPS?”, and “How much did you trust the device?”, mean scores were 3.8, 3.9, and 4.4, respectively, indicating generally high user satisfaction with the system.

Subjective comments written by participants to the investigators during the trial were very positive. One subject wrote to the clinical investigator during the study, “My glucoses have never been better with this system. Any way I could disappear into the night and keep this pump and sensor?” Another subject wrote, “I am amazed by the iAPS algorithm. Hats off to the team.” Some subjects wrote on their surveys to change the results scale. On the Technology Acceptance Questionnaire, in response to the question “How useful in managing your diabetes was the iAPS? (1–5)”, one subject handwrote in a “10” in the margin. In response to “How much did you trust the device (1–5)” the same subject answered with a handwritten “100”.

Additional survey results comparing pre and post iAPS use are summarized in Supplementary Table S5.

Conclusions

This pilot clinical trial assessed the performance of an AID system with specific meal choices in the unsupervised outpatient setting. This study offers new insights into important areas of diabetes management, as few previous studies have examined the effects of different meal content in the context of closed-loop insulin delivery. Gingras et al. reported that the addition of both fat and protein was associated with a 40-min delay in time to glycemic peak and 39% higher 5-h postmeal basal insulin requirements compared with a standard meal under closed loop.³⁶ We have also previously reported on the challenges of higher fat content in meals, with a higher glucose peak and a higher percent time glucose above 180 mg/dL in a closed-loop study with no meal bolus and higher fat content for a fixed amount of carbohydrate.³⁷

Previously we reported on the differences in glycemic outcomes comparing pasta and white rice meals in people with T1D.¹³ In that in-clinic study, we showed that pasta cooked al dente and prepared in a healthy manner can reduce glycemic peaks and glucose AUC relative to other foods in open loop. In this study, however, we specifically examined the source of carbohydrates in the context of both open and closed loop, not different macronutrient content, as all the study meals were equal amounts of carbohydrate with the rest of the meal content fixed. To make the study realistic for what subjects would normally eat, we did not heavily restrict the meal carbohydrate content, but instead allowed one to two portions per study meal, as long as each subject ate the same number of carbohydrates per study meal throughout both arms of the study.

In the current study, AID-based control decreased insulin delivery for pasta in 0–2 h, suggesting that the meal bolus was on average too high for pasta, while AID-based control increased insulin delivery for white rice in the 2–4 h period, suggesting that meal bolus was on average insufficient for white rice. Significant changes in insulin delivery did not take place during the 4–6 h period. While the rice and pasta meals had the same number of carbohydrates and were of similar macronutrient content, their insulin requirements were different. For glucose AUC under 70 mg/dL, change to white rice from pasta had a stronger effect than change from conventional therapy to AID-based control. We had previously reported that regular pasta has higher odds of hypoglycemia than white rice for the same amount of meal bolus.¹³ Peak glucose and time to peak glucose were not well defined for the prescribed meals, especially for pasta, as seen in Figure 1 in contrast to our previous study.¹³

In conclusion, the postprandial difference in study meals on glycemic outcomes such as glucose AUC, AUC under 70 mg/dL, and peak glucose was clearly differentiated in conventional therapy, while with AID-based control these differences were clinically reduced and not statistically significant, highlighting the importance of meal choices in influencing glycemia, as well as the potential of the AID system to minimize glycemic disturbances postprandially. Future investigations with higher power may be needed to detect additional differences with AID use.

This study suggests that the source of carbohydrates (both low and high glycemic index) is important in affecting blood glucose levels, and AID-based control in its current form can improve upon these results, allowing people with diabetes to better control glycemic excursions with different types of meals and improve quality of life, although there is still room for significant improvement in how AID systems handle these challenging meals. In this study, pasta had lower glucose AUC than white rice, but this was at a cost of increased risk of hypoglycemia, while white rice had increased risk of prolonged hyperglycemia. While the controller could have delivered more insulin in response to white rice overall and in the last 2 h for pasta, in the case of excessive meal bolus, the only option was prolonged pump suspension following the meal bolus. Future AID systems can achieve a better balance by incorporating a carbohydrate ratio specific to each meal type, in this case by increasing the carbohydrate ratio for pasta and decreasing it for white rice and by further tuning model predictions and zone MPC weights for each meal type.

This study also used an unlocked smartphone in an FDA approved IDE study in the outpatient at-home setting. Use of the AID system reduced exposure to hypoglycemia over the conventional arm throughout the 2 weeks of use (Table 2). The duration of iAPS and CGM use was comparable to our previous pilot iAPS study¹⁸ and to the recently reported international diabetes closed loop (iDCL) outpatient study.² Subjects were allowed to place their personal phone SIM card into the study phone and use the unlocked study phone as their personal phone while it ran the AID system concurrently, reducing the need to carry multiple devices. Survey results indicated that subjects found the system easy to use, had high satisfaction with the system, and were ready to use it in longer clinical trials. We believe that these types of simplifications to research AID systems and reduction of device burden should be considered in all future AID studies when possible, as they make systems easier to use, facilitate appropriate use, and increase user satisfaction. Furthermore, an AID system that can be easily used by research subjects in the outpatient setting and that does not require constant oversight and supervision from clinical and engineering staff facilitates better data collection and does not interfere with the study intervention compared to when the study team has to actively manage the devices.

A strength of this study was that it was an outpatient, randomized crossover, remote monitored study with scheduled meal challenges while putting a very low device burden on the subjects. However, we note some limitations of our study. First, since the study was conducted on a small number of subjects with tight baseline glycemic control, this limits scope of improvement with AID-based control and overall generalizability. Second, as the meal challenges were conducted at home in free living conditions, we cannot completely isolate the effect of study meal from other causes of glycemic disturbances (meals or otherwise). Finally, effect of one subject's use of a PLGS system on insulin outcomes was not analyzed.

From these data, we draw two main conclusions. First, the unlocked smartphone-based AID system can be safely used in the outpatient setting without any adverse events. Second, AID-based control can improve the postprandial glycemic response to challenging meals, although there remains a large effect of the meal type that is still difficult to overcome. The next step in improving the postprandial meal response is optimizing AID systems for different meal types (both different sources of carbohydrate and different macronutrient content), as these issues are still of utmost importance in improving glucose control for people with diabetes.

Authors' Contributions

S.D., J.E.P., M.M.C., M.P., C.A., J.M., F.J.D., D.M.E., and E.D. helped design the study protocol and ensured the regulatory approval of the study; analyzed data; and authored the article. S.D., J.E.P., F.J.D., and E.D. helped construct the controller infrastructure, analyzed data, and edited and revised the article. J.E.P., M.M.C., C.A., and M.P. performed data acquisition for the clinical study, analyzed data, and edited and revised the article. J.M. and D.M.E. designed research, analyzed data, and edited and revised the article. E.D. edited and reviewed the final article and was the principal investigator of the project. E.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.

Author Disclosure Statement

J.E.P. reports receiving grant support, provided to his institution, consulting fees, and speaker fees from Tandem Diabetes Care, grant support, provided to his institution, and advisory board fees from Medtronic, grant support, provided to his institution, and consulting fees from Eli Lilly, grant support and supplies, provided to his institution, from Insulet, and supplies, provided to his institution, from Dexcom, Inc. (San Diego, CA). F.J.D. reports product support from Dexcom, Inc., and Tandem Diabetes Care, as well as patent royalties from Insulet, Inc., Dexcom, Inc., Mode AGC, and Roche, and is a Scientific Advisor to Mode AGC. D.M.E. is a Scientific Advisor to the Health and Wellness Advisory Committee of the Barilla Center for Food & Nutrition (Italy). E.D. reports consulting fees from Eli Lilly, speaker bureau fees from Roche Diabetes Care, and product support from Dexcom, Inc., and Tandem Diabetes Care, as well as patent royalties from Insulet, Inc., Dexcom, Inc., Mode AGC, and Roche. No conflicts of interest relevant to this project are reported for the rest of the authors.

Funding Information

The study was supported by an unrestricted gift from the Barilla Foundation to the Harvard John A. Paulson School of Engineering and Applied Sciences in support of collaborative nutrition research. The pasta and sauce used in the trial were provided by Barilla. Additional support was provided by the National Institutes of Health (DP3DK104057 and DP3DK113511) and the Harvard Accelerator. Product support was provided by Dexcom, Inc. who provided research discount pricing on continuous glucose monitoring sensors, transmitters, and receivers (IIS-2018-019). Tandem t:AP insulin pumps were purchased from Tandem Diabetes Care at full price. J.E.P. was also supported, in part, by a grant from the William K. Bowes, Jr. Foundation (WKB-2017-22754). The funders and device manufacturers had no influence on the design or conduct of the trial and were not involved in data collection or analysis, the writing of the article, or the decision to submit it for publication.

Supplementary Material

Supplementary Data

Supplementary Figure S1

Supplementary Table S1

Supplementary Table S2

Supplementary Table S3

Supplementary Table S4

Supplementary Table S5