Algorithm-Driven Initiation and Adaptation of Hybrid Closed-Loop in Young Children with Type 1 Diabetes: A Pilot Study

Abstract

Introduction:

Glucose regulation in young children is complicated by higher glycemic variability, unpredictable behaviors, and low insulin needs. While the benefits of automated insulin delivery (AID) for this population are established, how to initiate and adjust pump settings still represents a challenging task for health care providers. In this study, we investigate the safety and efficacy of using algorithm-driven initiation and adjustments of AID parameters in children aged 2–6 years.

Methods:

Participants used AID at home for 8 weeks. Initial settings and periodic adjustments of therapy profiles (basal rates, insulin-to-carbohydrate ratios, insulin-correction factors, and sleep schedules) were provided through a cloud-based investigational software. Investigators reviewed therapy recommendations and could adjust if necessary. Primary safety endpoints included the percentage of time <54 mg/dL and >250 mg/dL, tested for noninferiority with respect to baseline. Primary efficacy endpoints (tested in a hierarchical manner) were the percentage of time in 70–180 mg/dL, mean glucose, the percentage of time >250 mg/dL, <70 mg/dL, and <54 mg/dL.

Results:

Thirty-two participants (age range: 2.0–5.9 years) were recruited for the study; 29 had sufficient data for the analysis. Investigators overrode 15% of software recommendations. The percentage of time <54 mg/dL and >250 mg/dL was noninferior in the 8-week follow-up with respect to baseline (P < 0.001). Statistically significant improvements were observed in the percentage of time in 70–180 mg/dL (P = 0.005), >250 mg/dL (P = 0.003), and mean glucose (P = 0.02). No difference was observed in the percentage of time <70 mg/dL (P = 0.34). Furthermore, no difference was observed with respect to a similar study cohort (same age range, n = 86) with expert pediatric endocrinologists modifying pump settings.

Conclusions:

Findings from this pilot study suggest that the use of AID with algorithm-driven initiation and adjustment of pump parameters is safe and effective in young children with type 1 diabetes. Further study of the algorithm in a larger cohort is indicated. Clinical Trials Registration number: NCT06017089

Introduction

Management of type 1 diabetes (T1D) is challenging for people of all ages, but can be especially difficult for young children due to significantly more variability in insulin sensitivity, activity, eating habits, and sleeping habits.^1,2 While neither insulin injections nor continuous subcutaneous insulin delivery are as precise as endogenous insulin secretion, newer automated insulin delivery (AID) systems attempt to more closely mimic the natural process with less active involvement by the user. In the last decade, several pivotal studies have shown the benefit of different AID systems in children aged 6 or older.^3

–10 More recently, the efficacy and safety of these systems were demonstrated for pediatric patients down to 1 year of age.^11

–14 Despite the clear efficacy of AID in youth, current hybrid closed-loop systems (AID systems optimized to function with meal announcements) still require highly educated professionals to determine initial pump settings and make dose adjustments throughout the use of the system—tasks that are made even more challenging by young children’s high glycemic variability, behavioral uncertainty, and inability to communicate their symptoms.

Artificial intelligence (AI), that is, the set of technologies mimicking human learning and decision-making, has been increasing in its use and applicability in health care.¹⁵ AI can be used to analyze health-related signals (e.g., medical images or wearable sensor data) and identify trends, patterns, and abnormalities, allowing health care providers rapid and in-depth understanding of medical conditions for focused care; furthermore, AI-based advisors can assist clinical decision by providing personalized treatment recommendations. The use of AI in T1D research has followed the increasing availability of data through continuous glucose monitoring (CGM) sensors and insulin infusion pumps, with applications ranging from daily operations (e.g., smart bolus calculators, hypoglycemia prevention) to personalization of AID settings.^16,17

The pairing of AI utilizing large amounts of data with an AID system to program initial insulin pump settings and suggest adjustments over time is a novel path for moving the field of diabetes technologies forward and may be on the horizon.¹⁸ These AI programs may be beneficial in assisting health care providers, regardless of their expertise, in the initiation and use of AID; in turn, this could significantly increase the access and use of AID beyond tertiary centers and allow rapid and frequent dose adjustments based on glycemic trends.

This study evaluated the safety and efficacy of the t:slim X2 insulin pump with Control-IQ technology (Tandem Diabetes Care) in conjunction with a novel AI Advisor-Driven pump initiation and parameter adaptation program in young children with T1D. This is the first study of its kind in pairing AI with AID in this challenging population.

Materials and Methods

Study conduct and design

This study was approved and overseen by the JAEB Center for Health Research Institutional Review Board and an independent data and safety monitoring board. Funding for conducting the study was provided by the National Institute of Diabetes and Digestive and Kidney Disease (NIDDK). Tandem Diabetes Care (San Diego, CA) provided the automated insulin pumps (t:slim X2 insulin pumps with Control-IQ technology); insulin pumps were paired with Dexcom G6/G7 CGM sensors, provided by Dexcom, Inc. (San Diego, CA). An investigational device exemption for the cloud-based software used in the study was approved by the Food and Drug Administration.

The PEDAP-AI (PEDiatric Artificial Pancreas with Automated Initiation; the complete list of the PEDAP-AI study group is reported in the Supplementary Data) study was coordinated by the JAEB Center for Health Research and conducted at three institutions: the Center for Diabetes Technology at the University of Virginia (UVA), the Barbara Davis Center for Diabetes (BDC), and Stanford University. Informed consent for each participant was provided electronically by a legally authorized representative (e.g., a parent). Eligibility criteria included diagnosis of T1D for at least 1 month before enrollment, age ≥2 and <6 years, use of a Dexcom CGM sensor at the time of enrollment, total daily insulin (TDI) of at least 5 units and a body weight of at least 20 lbs (9.1 kg). A complete list of eligibility criteria is provided in the Supplementary Data (Section S.1).

After obtaining informed consent, if the enrollment criteria were met, participants’ legal guardians were trained in the use of the study pump within 10 days of completion of the screening visit. Participants’ legal guardians installed and ran Tandem’s t:connect app on a smartphone.

Participants were either using multiple daily injections (MDI) or pump before being enrolled in the study. At study start, participants were switched to a study-provided closed-loop insulin pump (t:slim X2 insulin pump with Control-IQ technology, Tandem Diabetes Care, San Diego CA). Initial pump settings for previous MDI users were provided by the UVA Clinical Portal, which housed the automated initiation and adjustment advisor logic based on current glycemic metrics, TDI, and body weight. Settings for pump users were instead initiated to replicate their personal pump therapy profiles, unless investigators decided otherwise. Participants then proceeded to use the closed-loop system for 8 weeks. A safety visit was conducted on day 3; if the percentage of time <70 mg/dL was >4%, the AI-based advisor recommended a safety adjustment to reduce the system aggressiveness. Subsequent visits took place at weeks 1, 2, 4, and 6. At each of these visits, AI-based recommendations for therapy profiles were provided. Study investigators reviewed all recommendations before making any adjustments to the pump settings and could either accept these recommendations as they were or modify them based on clinical judgment. Off-schedule AI-based recommendations were also available upon investigator’s request through the UVA Clinical Portal. Participant families could also discuss with the investigator and study staff their concerns regarding the recommendations; however, before any changes, the settings were reviewed by the investigator. No recommendations were sent to participants without investigator review.

A final study visit was conducted at week 8, during which the study participants were transitioned back to their preferred poststudy insulin therapy. A follow-up safety visit occurred 3 days after the final visit to confirm a safe transition to the poststudy insulin therapy.

Trial visits, including enrollment and training on how to use the system, could be conducted either virtually or in person. A scheme of the study design is reported in Figure 1.

FIG. 1.

Design of the PEDAP-AI study.

The UVA clinical portal

During the conduct of the study, investigators used the UVA Clinical Portal to monitor participant data (through ambulatory glucose profiles and estimated insulin-on-board derived from insulin pump records) and view recommended adjustments to pump settings. This investigational, cloud-based software system consisted of a front-end website for data review and receipt of pump setting adjustment recommendations, and a back-end logic that periodically retrieved participant data and generated AI-based adjustments for pump parameters. The UVA Clinical Portal collected participants’ data every few hours from a third-party commercial system, Tandem’s t:connect mobile app, which was connected to the participant’s insulin pump.

On fixed days as per protocol, the UVA Clinical Portal generated recommendations on how to adjust Control-IQ therapy profiles, including basal rates (BR), insulin-to-carbohydrate ratios (ICR), correction factors (CF), and the start and end times for “sleep activity” (a control mode in Control-IQ characterized by a lower glucose target, increased basal limits, and no automatic delivery of corrective insulin boluses). These recommendations could advise modulating the complete BR, ICR, or CF profiles by up to ± 10%, while a modulation of up to ± 20% was allowed on each 4-h time segment. Investigators reviewed these recommendations on the Titration page in the UVA Clinical Portal (see Fig. 2). This page also provided a console that allowed physicians to simulate slight changes (up to ± 20% on each hour segment) to the recommended profiles. In addition, on the Titration page, physicians could request off-schedule AI-based adjustments of therapy profiles, with a limit of one recommendation per day.

FIG. 2.

Screenshot of the Titrate page from the UVA Clinical Portal. The calendar on the top left allows physicians to select which dates should be considered in the simulations. The console in the middle allows the simulation of small changes in the recommended therapy profiles (e.g., changing the start/end of the sleep schedule or increasing/decreasing some hour segments of the other therapy profiles). The plot on the bottom of the page shows the distribution of the glucose traces over the selected days for that patient and the time in ranges (TIRs) when using the nominal therapy profiles (gray), the artificial intelligence (AI)-based recommendation (blue), or by applying modifications to the AI-based recommendation through the console (red). When hovering the mouse over it, the TIR bars report (bottom to top) the estimated percentage of time below 54 mg/dL, between 54 and 70 mg/dL, between 70 and 180 mg/dL, between 180 and 250 mg/dL, and above 250 mg/dL.

The portal generated simulated outcomes and AI-based recommendations by using the UVA replay simulator.^19,20 This data-driven tool integrated a virtual representation of the participant (a so-called digital twin), which consisted of a minimal model of glucose-insulin dynamics tailored on participants’ data, and an insulin net-effect signal, aimed at capturing unmodeled phenomena. For more details on the generation of the digital twins, see previous studies.^20
–22 By using the digital twins, the UVA Clinical Portal could simulate the glycemic response to what-if scenarios, that is, to changes in Control-IQ therapy profiles.

When <5 days of data were available, the portal used a heuristic rule to determine profile adjustments, recommending a 10% decrease in BR and a 10% increase in ICR and CF whenever CGM-measured time below range was above 4%. When >6 days of data were available, therapy recommendations were generated using the digital twins as a test bed to simulate the effect of small changes in the therapy profiles. Optimal modulation factors for BR, ICR, and CF and the start/end of sleep mode were found by using a gradient-based optimization algorithm to minimize a cost function based on low and high blood glucose indices in the replay simulations. For more details, the reader is referred to Diaz et al.²²

Statistical analysis

The study was not formally powered to test any a priori hypotheses. The goal was to have at least 30 participants completing the study, equally distributed among the three sites.

The primary endpoints for this study were safety outcomes. CGM-measured percentage of time <54 mg/dL and >250 mg/dL at baseline versus the 8-week follow-up was tested for noninferiority, with noninferiority margins of 0.5% and 3%, respectively (P < 0.05). Baseline values were computed from the most recent 28 days of CGM data before the screening visit. Severe hypoglycemia (SH) events and diabetes ketoacidosis (DKA) events were documented.

If the two primary CGM-derived safety outcomes were determined to be noninferior, changes from baseline in other CGM-based metrics were tested in a hierarchical manner ( $P < 0.05$ ). To preserve the overall type 1 error among the formal comparisons, tests were performed in the following order: percentage of time in range (TIR) 70–180 mg/dL, mean glucose, and percentage of time >250 mg/dL, <70 mg/dL, and <54 mg/dL. If any P value results >0.05, the hierarchical analysis would stop, and P values would not be computed. Secondary outcomes included percentage of time in 70–140 mg/dL (tight range), >180 mg/dL, >300 mg/dL, <60 mg/dL, standard deviation (SD), coefficient of variation, low and high blood glucose indices, and weekly hypoglycemic and hyperglycemic event rate. Secondary insulin endpoints include TDI and % of insulin delivered via basal.

Results from this PEDAP-AI study were compared with those from a sample of participants from the prior PEDAP study,¹⁴ which evaluated initiation of and diabetes management with the Control-IQ system in the same age group (2 to <6 years old). All hierarchical and secondary efficacy metrics were compared between baseline and a 2-week follow-up period (weeks 7–8 from closed-loop initiation). To meet the criteria for inclusion in this analysis, the PEDAP participants had to have available at least 168 h of CGM data during baseline and the 2-week follow-up period; have HbA1c measurements at baseline within 0.2 percentage points of the range of the PEDAP-AI participants’ measurements; and remain in closed loop for at least 50% of the 8-week period after closed-loop initiation. To be included in this analysis, PEDAP-AI participants had to qualify for the primary analysis and have available at least 168 h of CGM data during weeks 7 and 8 of the study. Differences have been adjusted for the baseline value of the outcome, age at closed-loop initiation, MDI or pump use before closed-loop initiation, and clinical site.

Results

Study population

Thirty-three individuals were screened and assessed for eligibility at three clinical study sites. One participant did not meet the inclusion criteria and was excluded. Baseline characteristics and demographics for the 32 participants who initiated closed-loop therapy are reported in Table 1. At baseline, participants’ ages ranged from 2 to 5.9 years, with 28% being younger than 4 years. The time since the diagnosis of T1D ranged from 1 to 48 months. Most of the participants were male (66%), White (69%), and non-Hispanic (84%). Before closed-loop initiation, 63% of participants were on MDI and the remaining were on pump therapy, with an HbA1c ranging between 5.6 and 9.6%.

Table 1.

Characteristics of the PEDAP-AI Patients at Baseline

	n = 32
Age (years)
2 to <4	9 (28%)
4 to <6	23 (72%)
Median (IQR)	4.9 (4.0, 5.4)
Range	2.0 to 5.9
Diabetes Duration (months)
1 to <3	3 (9%)
3 to <6	9 (28%)
6 to <12	5 (16%)
12 to <24	3 (9%)
24 to <48	8 (25%)
≥48	4 (13%)
Median (IQR)	10 (4, 36)
Range	1 to 48
BMI Percentile
Median (IQR)	76% (58%, 92%)
Range	7% to 99%
Sex—Female	11 (34%)
Race
White	22 (69%)
Black/African American	3 (9%)
Asian	3 (9%)
More than one race	4 (13%)
Hispanic ethnicity	5 (16%)
Parent Education
High school graduate/diploma/GED	4 (13%)
Technical/Vocational	1 (3%)
Associate degree (AA)	4 (13%)
College graduate (bachelor’s degree or equivalent)	9 (28%)
Advanced degree (e.g., master’s, PhD, MD)	14 (44%)
Annual Household Income
$35,000 to <$50,000	1 (3%)
$50,000 to <$75,000	5 (16%)
$75,000 to <$100,000	4 (13%)
$100,000 to <$200,000	10 (31%)
≥$200,000	11 (34%)
Unknown	1 (3%)
Health Insurance^a
Private	26 (81%)
Medicare	2 (6%)
Medicaid	4 (13%)
HbA1c—Central Lab at Closed-Loop Initiation (n = 31)^b
<7.0%	12 (39%)
7.0% to <8.0%	12 (39%)
≥8.0%	7 (23%)
Mean ± SD	7.3 ± 0.9
Range	5.6 to 9.6
Insulin Modality
MDI	20 (63%)
Pump^c	12 (38%)
Total Daily Insulin (U/kg/day)
Mean ± SD	0.61 ± 0.21
Range	0.28 to 0.98
Number of Pump Boluses or Injections of Short-Acting Insulin per Day (n = 31)^d
0 to 3	12 (39%)
4 to 6	13 (42%)
7 to 9	5 (16%)
≥10	1 (3%)
Median (IQR)	5 (3, 6)
Range	2 to 12
DKA in Last 12 Months
None	24 (75%)
1^e	8 (25%)
SH in Last 12 Months
None	32 (100%)

Two participants classified as having private insurance also had Medicaid.

The participant who withdrew shortly after the closed-loop initiation visit did not provide an HbA1c sample.

Personal pumps were Tandem t:slim X2 with Control-IQ (n = 10) and Tandem t:slim X2 (n = 2).

One participant’s number of short-acting injections was unknown.

All instances were associated with initial diagnosis of T1D.

DHA, diabetes ketoacidosis; MDI, multiple daily injections; SH, severe hypoglycemia; SD, standard deviation; T1D, type 1 diabetes.

Four of these participants did not complete the 8-week follow-up: one participant withdrew due to discomfort with sensor (after completing week 4 visit), one due to discomfort with pump/infusion set (after closed-loop initiation), and two because families did not agree with the AI-driven therapy adjustments (both after completing week 4 visit). Of these latter two, one participant dropped out due to episodes of nocturnal hypoglycemia; the other withdrew because parents were used to making frequent adjustments to pump settings and the AI-based recommendations were different from their expectations, even though no concerns for hypo- or hyperglycemia were expressed. Three of these four participants are excluded from the primary safety and efficacy analysis of CGM-measured metrics due to lack of data (final n = 29); however, all enrolled participants (n = 32) are considered in the primary safety analysis of adverse events. Excluding the participant who withdrew right after closed-loop initiation, 17 participants used the Dexcom G6 and 14 used Dexcom G7 sensor.

Safety and efficacy outcomes

Table 2 reports the results for the noninferiority tests on the CGM-measure safety endpoints. Mean percent time below 54 mg/dL was 0.3% ± 0.2% during follow-up versus 0.3% ± 0.3% at baseline (mean difference −0.05%, 95% confidence interval −0.2% to 0.1%, P < 0.001). Percent time above 250 mg/dL decreases from 13% ± 10% at baseline to 9% ± 5% (mean difference −5%, 95% confidence interval −8% to −2%, P < 0.001). Eight adverse events were reported during the study from six participants (Table 3). Of these, one SH event was determined to potentially be related to the investigational device (i.e., the UVA Clinical Portal). During this event, the participant experienced a CGM low with a finger-stick glucose of 45 mg/dL that required assistance to treat. Glucose gel was utilized with rapid recovery and no glucagon was required. On pump review, the participant had been administered an autocorrection shortly before a carbohydrate bolus and the event occurred shortly after setting adjustments were made. The remaining adverse events (one DKA, one serious event involving hypoglycemia, and five nonserious adverse events) were unrelated to the investigational device.

Table 2.

Safety Endpoints, Tested for Noninferiority, and Efficacy Endpoints, Tested for Superiority, in a Hierarchical Order

	Baseline n = 29	Weeks 1–8 n = 29	Mean change from baseline (95% CI)	P value
Noninferiority tests
% CGM <54 mg/dL^a,b	0.3% ± 0.3%	0.3% ± 0.2%	−0.05% (−0.2%, 0.1%)	<0.001
% CGM >250 mg/dL^a,c	13% ± 10%	9% ± 5%	−5% (−8%, −2%)	<0.001
Superiority tests
% CGM in range 70–180 mg/dL^a	63% ± 15%	70% ± 8%	6% (2%, 11%)	0.005
Mean glucose (mg/dL)	168 ± 28	157 ± 14	−10 (−19, −2)	0.02
% CGM >250 mg/dL^a	13% ± 10%	9% ± 5%	−5% (−8%, −2%)	0.003
% CGM <70 mg/dL^a	2.0% ± 1.4%	1.7% ± 0.9%	−0.3% (−0.9%, 0.3%)	0.34
CGM <54 mg/dL^a	0.3% ± 0.3%	0.3% ± 0.2%	−0.05% (−0.2%, 0.1%)	−^d

Metrics are reported as mean ± standard deviation.

Estimates from skewed outcomes were derived from linear models that specified a t-distribution for the residuals.

Noninferiority limit: +0.5%.

Noninferiority limit: +3%.

P value not computed due to the previous outcome in the hierarchy being nonsignificantly different.

CGM, continuous glucose monitoring.

Table 3.

Adverse Events During the Study

	n = 32
All adverse events	8
SH events^a	1
DKA events^b	1
Other serious adverse events^c	1
Other adverse events N Events/N Participants	5/5
Hyperglycemia with or without ketosis^d	3/3
Gastroenteritis	2/2

Event was possibly related to investigational device; a severe hypoglycemic event is defined as a hypoglycemic event that (a) required assistance of another person due to altered consciousness, and (b) required another person to actively administer carbohydrate, glucagon, or other resuscitative actions.

Event was unrelated to the UVA Clinical Portal; DKA events had to meet DCCT criteria.

Event was unrelated to the UVA Clinical Portal: the participant had hypoglycemia after tubing fill without set disconnection that was determined not to be meet SH criteria.

All three instances were unrelated to the UVA Clinical Portal: two participant had infusion set failure; one participant contacted clinical site for guidance on management of hyperglycemia/ketosis related to illness. None of the instances included a ketone reading ≥ 1.0 mmol/L.

Table 2 also reports the results from the hierarchical analysis of the efficacy outcomes. With respect to efficacy comparing change from baseline to 8-week follow-up, TIR increased from 63 $%$ ± 15% to 70 $%$ ± 8% (mean difference 6%, 95% confidence interval 2% to 11%, P = 0.005), average glucose decreased from 168 ± 28 mg/dL to 157 ± 14 mg/dL (mean difference −10 mg/dL, 95% confidence interval −19 mg/dL to −2 mg/dL, P = 0.02) and the percentage of time above 250 mg/dL decreased from 13% ± 10% to 9% ± 5% (P = 0.003). No statistically significant difference was found when testing the percentage of time below 70 mg/dL (2.0% ± 1.4% vs. 1.7% ± 0.9%, mean difference −0.05%, 95% confidence interval −0.9% to 0.3%, P = 0.34); hence, the hierarchical testing procedure stopped, and no P value was computed for the percentage of time below 54 mg/dL.

Improvements were also observed for the percentage of time in tight range 70–140 mg/dL, above 180 mg/dL, above 300 mg/dL, and for glucose SD. A complete report of the secondary and exploratory outcomes is reported in Supplementary Tables S1 to S4.

Figure 3 reports a box plot of the TIR distributions, stratified by each biweekly period. The improvement in the percentage of TIR is evident right after closed-loop initiation, and it remains consistent throughout the weeks of the study. A complete report on the efficacy metrics, stratified by adaptation period, is reported in Supplementary Table S5.

FIG. 3.

Percentage of time in target range 70–180 mg/dL at baseline and at each biweekly period until study end.

During the study, investigators reviewed the AI-driven recommendations and could override them based on their judgment or families’ concerns. Table 4 summarizes the overrides to the recommendations, grouped by each adaptation cycle. The proportion of “off-schedule” therapy recommendations increased from 3% in the first 3 days to 28% in the last biweekly period. Recommendations occurred two to three times more often in the second half of the study than in the first half (27% in the last 2 weeks of the study), reaching a total of 33 overrides out of 211 total recommendations (16% over the total). Out of these overrides, 13 (39%) were directed toward more insulin infusion, while 16 (48%) were expected to deliver less insulin; the remaining 4 (13%) of the overrides had a mixed effect on insulin delivery. Information regarding reasons for overrides was not available.

Table 4.

Therapy Profile Recommendations and Override per Adaptation Cycle

Adaptation period	n = 211Recommendations	Percentage of recommendations that were protocol-scheduled	n = 33 Overrides	Percentage of overrides expected to deliver more insulin
Closed-loop initiation to day 3	33 (16%)	97%	3 (9%)	67%
Day 3 to week 1	26 (12%)	96%	3 (9%)	0%
Week 1 to week 2	37 (18%)	81%	3 (9%)	33%
Week 2 to week 4	38 (18%)	76%	5 (15%)	60%
Week 4 to week 6	38 (18%)	79%	10 (30%)	50%
Week 6 to week 8	39 (18%)	72%	9 (27%)	22%

Comparison with the PEDAP study

A total of 86 (84%) participants from the original PEDAP study qualified for comparison analysis, with 27 (93%) participants from the PEDAP-AI study. The two samples differ in size due to differences in protocols and enrollment targets.

Table 5 compares the results achieved for the primary efficacy metrics between these two cohorts at baseline versus weeks 7–8 after closed-loop initiation. No statistically significant differences are observed for any of the endpoints (Supplementary Tables S6–S9). Supplementary Figure S1 demonstrates the overlapping and similar 24-h glucose profiles for PEDAP and PEDAP-AI participants.

Table 5.

Comparison of the Primary Efficacy Metrics Between the PEDAP and PEDAP-AI Study

	BaselinePEDAP n = 86PEDAP-AI n = 27	Weeks 7–8^a PEDAP n = 86PEDAP-AI n = 27	Adjusted difference (95% CI)^a,b	P value
% CGM in Range 70–180 mg/dL
PEDAP	57% ± 16%	69% ± 10%	−3% (−9%, 2%)	0.37
PEDAP-AI	64% ± 14%	69% ± 10%	−3% (−9%, 2%)	0.37
Mean Glucose (mg/dL)
PEDAP	172 ± 30	156 ± 18	4 (−5, 13)	0.55
PEDAP-AI	167 ± 28	157 ± 16	4 (−5, 13)	0.55
% CGM >250 mg/dL^*
PEDAP	15% ± 11%	9% ± 6%	0.2% (−3.0%, 3.3%)	0.94
PEDAP-AI	13% ± 10%	8% ± 5%	0.2% (−3.0%, 3.3%)	0.94
% CGM <70 mg/dL^*
PEDAP	3.3% ± 2.5%	2.5% ± 1.6%	−0.2% (−1.2%, 0.7%)	0.82
PEDAP-AI	2.1% ± 1.4%	1.8% ± 1.0%	−0.2% (−1.2%, 0.7%)	0.82
% CGM <54 mg/dL^*
PEDAP	0.8% ± 0.8%	0.5% ± 0.5%	0.02% (−0.2%, 0.3%)	0.94
PEDAP-AI	0.3% ± 0.3%	0.3% ± 0.3%	0.02% (−0.2%, 0.3%)	0.94

P values and confidence intervals are False Discovery Rate (FDR-adjusted).

Weeks 7–8 for PEDAP study refers to study phase for CLC group (n = 55) and study extension phase for standard care group (n = 31).

Difference is PEDAP-AI minus PEDAP. The difference has been adjusted for the baseline value of the outcome, age at automated insulin delivery (AID) initiation, MDI or pump use before AID initiation, and clinical site.

Baseline and follow-up estimates from skewed outcomes were derived from linear models that specified a t-distribution for the residuals.

Discussion

The AI advisor-driven pump settings in conjunction with the AID system (Tandem t:slim X2 with Control-IQ technology) resulted in improved TIR and other CGM metrics reflective of hyperglycemia in young children with T1D when compared with baseline, with minimal safety events. In addition, our data indicated similar outcomes when comparing AI-developed pump settings (PEDAP-AI) to pump settings set and modified by highly trained pediatric endocrinologists (PEDAP). The study design allowed for investigator review and overriding the recommended settings if there were concerns for safety from either the investigator or the parent. Overall, 15% of recommendations were overridden, due to both concerns about recommendations being too aggressive or too conservative. The study protocol did not include a detailed reason for each override, so this was not available for this pilot study, but will be important for future use of the algorithm. While the number of overrides was relatively small, as 85% of recommendations were accepted, it indicates the potential “human factor,” such as fear of hypoglycemia and feelings of loss of control, that may be challenging for AI systems to recognize or overcome. It is well acknowledged how personal preferences and worries can affect therapeutic decisions in T1D care—especially for caregivers of individuals with T1D in a very young age. These findings prompt the development of engineering solutions aimed at further involving caregivers in the decision process, for example, by providing more than one viable recommendation so that families can decide which one they feel more comfortable with.

The use of AI-driven technologies is currently increasing across the world, including in health care. Clinical evidence shows the safety and efficacy of these tools for different tasks related to T1D management, fostering their application in real clinical practice. Recent studies are showing promising results when using digital twins for therapy adaptation in different AID systems.^23,24 Digital twins can be particularly beneficial for young children with T1D, as they provide a “virtual sandbox” to safely and quickly test therapy changes on such a vulnerable population. While the use of AI-driven insulin pump setting recommendations is still under development and being used in research settings only, the data from this study with young children have shown that it can be safely used with similar glycemic outcomes to physician-derived pump settings.

A significant limitation for use of diabetes technologies in all people with T1D continues to be due to limited understanding around the current AID systems, including how to identify the system most appropriate for the patient as well as understanding how to initiate an AID system and how to manage it following initiation to optimize the algorithm. While many systems are becoming more simplified in their initialization and use, there continue to be a large number of people who discontinue AID systems due to the system not meeting their expectations, of which optimization of pump settings plays a significant role in achieving the best outcomes possible. Unless caregivers are particularly savvy, pump settings are often only changed at the time of a health care provider visit, typically every 3–4 months. However, insulin needs are continuously changing in children with T1D and frequent dose adjustments are necessary for optimal control. AI-driven modification of settings could occur every 2 weeks with suggestions provided directly to the person with diabetes leading to best outcomes for the user in both the short term and long term without creating a significant burden on the health care provider, whether a community provider or a specialized endocrinologist.

This study was limited to a population of young children and therefore cannot be generalized at this stage to the population of people with T1D using AID systems. More research is needed to determine if similar AI-driven programming is compatible for an older population of people with T1D using AID. Data on reasons for overriding the algorithm-driven recommendations were not collected as part of this study and will be necessary for future research using this or similar algorithms. In addition, this study utilized a single automated system, the Tandem t:slim X2 pump with Control-IQ technology. Additional AI programming will need to be adapted to the unique algorithms in each AID system. The PEDAP-AI protocol was developed as a pilot trial and therefore had a small number of participants. Comparison analysis to the original PEDAP study was limited due to the relatively small PEDAP-AI cohort size but was able to provide some insight into AI pump setting recommendations in comparison with pump settings determined by pediatric endocrinologists with a research focus on AID systems in their clinical care. It is also important to note that all of the participants in the PEDAP group were on MDI therapy at baseline, while 38% of the PEDAP-AI cohort was already using AID before the start of the study.

Longer use of AI programming needs to be studied to determine a sustained impact or if long-term glycemic management with AI programming can reduce the variability seen in glycemic outcomes and insulin sensitivity throughout the course of childhood, adolescence, and into adulthood.

Conclusions

AI advisor-driven pump initiation and adjustments can be successfully used in young children with T1D using the Tandem t:slim X2 pump with Control-IQ technology. The future role of AI in the management of T1D is still to be determined, but our data indicate promising results when utilizing AI to initialize and adjust AID settings in a challenging population over an 8-week period with minimal safety concerns. Additional research is needed to determine whether the current programming can be used in an older population or with different AID systems taking into account different modifiable parameters with each system’s algorithm.

Footnotes

Acknowledgments

The authors thank Dexcom, Inc. and Tandem Diabetes Care for providing the study material.

Authors’ Contributions

J.P.: Software, investigation, data curation, and writing—original draft. E.C.: Investigation and writing—original draft. Z.W.R.: Investigation, statistical analysis, and writing—review and editing. M.F.V.-T.: Methodology, software, investigation, and writing—review and editing. J.L.D.C.: Methodology, software, investigation, writing—review and editing. M.D.D.: Conceptualization, investigation, supervision, and writing—review and editing. M.S.: Investigation and writing—review and editing. E.J.: Investigation, project administration, and writing—review and editing. R.K.: Investigation, supervision, project administration, and writing—review and editing. V.H.: Investigation, project administration, and writing—review and editing. J.W.L.: Investigation, supervision, data curation, and writing—review and editing. C.L.K.K.: Investigation, software, and writing—review and editing. B.B.: Investigation and writing—review and editing. R.B.: Investigation, writing—review and editing, and supervision. R.P.W.: Conceptualization, investigation, supervision, and writing—review and editing. M.D.B.: Grant PI, study design, IDE sponsor, engineering/software development, and writing.

Author Disclosure Statement

J.P. reports research support from Dexcom. E.C. conducts research with Tandem, Dexcom, Insulet, Medtronic, Beta Bionics, Eli Lilly, Abbott, and Luna Health. Z.W.R. reports no personal financial disclosures. M.F.V.-T. reports research support from Dexcom. J.L.D.C. is at present an employee of Insulet Co. and this work was performed during her postdoctoral program at UVA, during which she received research support and royalties from Dexcom handled by the University of Virginia’s Licensing and Ventures Group. M.D.D. reports research funding from Dexcom, Medtronic, and Tandem Diabetes Care. M.S. reports research funding from Tandem Diabetes Care and Insulet. E.J. reports no disclosures. R.K. reports no disclosures. V.H. reports no disclosures. J.W.L. reports no personal financial disclosures but reports that his institution has received funding on his behalf as follows: grant funding, study supplies, and consulting fees from Tandem Diabetes Care; grant funding and study supplies from Dexcom; study supplies from Novo Nordisk; consulting fees from Ypsomed. C.L.K.K. reports no disclosures. B.B. has received consulting fees from Medtronic, Ypsomed, and Arecor. R.B. reports no personal financial disclosures but reports that his institution has received funding on his behalf as follows: grant funding, study supplies, and consulting fees from Insulet, Tandem Diabetes Care, and Beta Bionics; grant funding and study supplies from Dexcom; grant funding from Bigfoot Biomedical, Embecta, Sequel Med Tech, and MannKind; consulting fees and study supplies from Novo Nordisk; consulting fees from Vertex, Hagar, Ypsomed, Sanofi, and Zucara; and study supplies from Medtronic, Ascensia, Roche, and Eli Lilly. R.P.W. reports research funding from Dexcom, Eli Lilly, and Tandem Diabetes Care, consulting for Dexcom and Tandem Diabetes Care, and serving as an advisory board member for Sequel and Provention Bio. M.D.B. reports research grants handled by the University of Virginia from the National Institutes of Health, Novo Nordisk, Dexcom, and Tandem Diabetes Care. In addition, M.D.B. has a number of patents with royalties licensed to Dexcom, Sanofi, Tandem Diabetes Care, and Novo Nordisk, including technologies used in this study. M.D.B. finally reports consulting/speakership activities with Dexcom, Tandem Diabetes Care, Roche, Portal Insulin LLC, BOYDSense, and Vertex.

Funding Information

Funding for conducting the study was provided by the NIDDK under grant number U01DK127551.

Supplementary Material

Supplementary Data

References

Monaghan

, Bryant

, Inverso

, et al. Young children with type 1 diabetes: Recent advances in behavioral research. Curr Diab Rep, 2022; 22(6):247–256; doi: 10.1007/s11892-022-01465-0

Schoelwer

, DeBoer

, Breton

. Use of diabetes technology in children. Diabetologia, 2024 Oct;67(10):2075–2084; doi: 10.1007/s00125-024-06218-0

Bergenstal

, Garg

, Weinzimer

, et al. Safety of a hybrid closed-loop insulin delivery system in patients with type 1 diabetes. JAMA, 2016; 316(13):1407–1408; doi: 10.1001/jama.2016.11708

Tauschmann

, Thabit

, Bally

, et al.; APCam11 Consortium. Closed-loop insulin delivery in suboptimally controlled type 1 diabetes: A multicentre, 12-week randomised trial. Lancet, 2018; 392(10155):1321–1329; doi: 10.1016/s0140-6736(18)31947-0

Brown

, Kovatchev

, Raghinaru

, et al.; iDCL Trial Research Group. Six-month randomized, multicenter trial of closed-loop control in type 1 diabetes. N Engl J Med, 2019; 381(18):1707–1717; doi: 10.1056/NEJMoa1907863

Breton

, Kanapka

, Beck

, et al.; iDCL Trial Research Group. A randomized trial of closed-loop control in children with type 1 diabetes. N Engl J Med, 2020; 383(9):836–845; doi: 10.1056/NEJMoa2004736

Brown

, Forlenza

, Bode

, et al.; Omnipod 5 Research Group. Multicenter trial of a tubeless, on-body automated insulin delivery system with customizable glycemic targets in pediatric and adult participants with type 1 diabetes. Diabetes Care, 2021; 44(7):1630–1640; doi: 10.2337/dc21-0172

Carlson

, Sherr

, Shulman

, et al. Safety and glycemic outcomes during the MiniMed™ advanced hybrid closed-loop system pivotal trial in adolescents and adults with type 1 diabetes. Diabetes Technol Ther, 2022; 24(3):178–189; doi: 10.1089/dia.2021.0319

Russell

, Beck

, Damiano

, et al.; Bionic Pancreas Research Group. Multicenter, randomized trial of a bionic pancreas in type 1 diabetes. N Engl J Med, 2022; 387(13):1161–1172; doi: 10.1056/NEJMoa2205225

10.

Ware

, Boughton

, Allen

, et al.; DAN05 Consortium. Cambridge hybrid closed-loop algorithm in children and adolescents with type 1 diabetes: A multicentre 6-month randomised controlled trial. Lancet Digit Health, 2022; 4(4):e245–e255; doi: 10.1016/s2589-7500(22)00020-6

11.

Sherr

, Bode

, Forlenza

, et al.; Omnipod 5 in Preschoolers Study Group. Safety and glycemic outcomes with a tubeless automated insulin delivery system in very young children with type 1 diabetes: A single-arm multicenter clinical trial. Diabetes Care, 2022; 45(8):1907–1910; doi: 10.2337/dc21-2359

12.

Forlenza

, Ekhlaspour

, DiMeglio

, et al. Glycemic outcomes of children 2–6 years of age with type 1 diabetes during the pediatric MiniMed™ 670G system trial. Pediatr Diabetes, 2022; 23(3):324–329; doi: 10.1111/pedi.13312

13.

Ware

, Allen

, Boughton

, et al.; KidsAP Consortium. Randomized trial of closed-loop control in very young children with type 1 diabetes. N Engl J Med, 2022; 386(3):209–219; doi: 10.1056/NEJMoa2111673

14.

Wadwa

, Reed

, Buckingham

, et al.; PEDAP Trial Study Group. Trial of hybrid closed-loop control in young children with type 1 diabetes. N Engl J Med, 2023; 388(11):991–1001; doi: 10.1056/NEJMoa2210834

15.

Jiang

, Jiang

, Zhi

, et al. Artificial intelligence in healthcare: Past, present and future. Stroke Vasc Neurol, 2017; 2(4):230–243; doi: 10.1136/svn-2017-000101

16.

Tyler

, Jacobs

. Artificial intelligence in decision support systems for type 1 diabetes. Sensors (Basel), 2020; 20(11):3214; doi: 10.3390/s20113214

17.

Qian

, Schumacher

. Latest advancements in artificial intelligence-enabled technologies in treating type 1 diabetes. J Diabetes Sci Technol, 2021; 15(1):195–197; doi: 10.1177/1932296820949940

18.

Cambuli

, Baroni

. Intelligent insulin vs. artificial intelligence for type 1 diabetes: Will the real winner please stand up? Int J Mol Sci, 2023; 24(17):13139; doi: 10.3390/ijms241713139

19.

Hughes

, Gautier

, Colmegna

, et al. Replay simulations with personalized metabolic model for treatment design and evaluation in type 1 diabetes. J Diabetes Sci Technol, 2021; 15(6):1326–1336.

20.

Villa-Tamayo

, Colmegna

, Breton

. Validation of the UVA simulation replay methodology using clinical data: Reproducing a randomized clinical trial. Diabetes Technol Ther, 2024; 26(10):720–727; doi: 10.1089/dia.2023.0595

21.

Villa-Tamayo

, Pavan

, Breton

. Analysis on the practical identifiability of the subcutaneous oral glucose minimal model. IFAC-PapersOnLine, 2024; 58(24):269–274; doi: 10.1016/j.ifacol.2024.11.048

22.

Diaz C

, Villa-Tamayo

, Moscoso-Vasquez

, et al. Simulation-driven optimization of insulin therapy profiles in a commercial hybrid closed-loop system. Comput Methods Programs Biomed, 2023; 242:107830; doi: 10.1016/j.cmpb.2023.107830

23.

Colmegna

, McFadden

, Fabris

, et al. Adaptive biobehavioral control: A pilot analysis of human-machine coadaptation in type 1 diabetes. Diabetes Technol Ther, 2024; 26(9):644–651; doi: 10.1089/dia.2023.0399

24.

Sehgal

, De Bock

, Grosman

, et al. Use of a decision support tool and quick start onboarding tool in individuals with type 1 diabetes using advanced automated insulin delivery: A single-arm multi-phase intervention study. BMC Endocr Disord, 2024; 24(1):167; doi: 10.1186/s12902-024-01709-y