Miniaturized Neural Networks for Deploying Fully Closed Loop Insulin Delivery Systems: A Pilot Study Featuring Flexible Meal Announcement Options

Abstract

Background:

Automated insulin delivery (AID) has revolutionized glucose management. Next-generation AID systems focus on reducing user input, particularly for mealtime dosing, aiming for fully closed loop (FCL) control. Our goal was to assess the safety and feasibility of the next iteration of FCL control, using a miniature neural network to enable implementation within existing hardware capabilities.

Methods:

In a randomized crossover trial, six adults with type 1 diabetes completed seven days of usual care and seven days using AIDANET in free-living conditions. AIDANET is designed to enable FCL control, but carbohydrate counting and a novel easy-bolus strategy were enabled for one day each to test the system in hybrid closed loop modalities.

Results:

The mean glucose during usual care was 168 ± 24.3 mg/dL, compared to 161.3 ± 16.7 mg/dL using the AIDANET system. Time-in-range (TIR) 70 to 180 mg/dL was 63.3% ± 14.9% in usual care compared to 66.4% ± 8.3% using AIDANET, while time-below-range (TBR) 70 mg/dL remained within acceptable margins (0.9 ± 1 vs 1.6 ± 1.8). There were no serious adverse events during the study. The hybrid bolusing options provided safe glycemic control, with carbohydrate counting achieving 57.1% TIR with 0.6% TBR, and Easy Bolus achieving 70.5% TIR with 1.5% TBR.

Conclusion:

This pilot-feasibility study demonstrates that the AIDANET system provides safe glycemic control. The small sample size (n = 6) limits overall generalizability, and further larger, statistically powered trials to validate these results are warranted.

Introduction

Automated insulin delivery (AID) systems have significantly improved the management of type 1 diabetes (T1D) by continuously adjusting insulin infusion based on data from continuous glucose monitors (CGMs).^1-5 However, current AID systems still require user-initiated meal announcements to achieve optimal glycemic control. This reliance on carbohydrate counting places a substantial daily management burden on users and requires advanced nutrition literacy, which not all patients possess.^6,7 To address these limitations, fully closed loop (FCL) technology—where insulin delivery is fully automated without mealtime intervention—has emerged as the next frontier in diabetes care. Despite the promise of FCL, some users may prefer retaining control over mealtime insulin administration, citing concerns over automation and glycemic control. ⁸ Others may want a balance between full automation and manual intervention to achieve tighter glucose regulation. Next-generation AID technologies should therefore accommodate a range of user preferences.

Several FCL systems have been developed and evaluated in clinical trials, demonstrating that insulin delivery without meal announcements can still achieve favorable glycemic outcomes.^9,10 Most of these systems rely on model predictive control (MPC) algorithms for basal insulin adjustments. However, MPC-based systems can be computationally demanding, making them difficult to implement in insulin pumps with limited processing power. A promising alternative is the use of neural networks (NNs) to approximate MPC behavior, enabling AID implementation on existing hardware.^11,12 Previous clinical trials have validated the feasibility of NN-based approaches, but the networks considered were large, requiring significant device memory, and therefore may not have been suitable for embedding in current insulin pumps. An in-silico analysis showed that smaller networks can reproduce insulin commands with very high fidelity while reducing memory usage and processing demands, ¹³ and another analysis indicates that these smaller networks may be more robust to unseen inputs than the previous large networks, suggesting an additional advantage. ¹⁴ While this study focuses on replacing MPC with a compact NN, other AID systems have incorporated NNs in clinical trials as supplementary modules for tasks like insulin suspension, bolus automation, or meal detection,^15-18 and other in-silico designs use NNs as the core decision-making element of the AID system.^19,20 The study presented here aims to demonstrate the feasibility of using one of these smaller NNs to command insulin delivery.

This work uses the UVA-AIDANET system, which has been described previously^10,11,21,22 and is designed to work both with and without meal announcements. When used in FCL, this system provided superior mean glucose, TIR, TITR, TAR 180 mg/dL, and TAR 250 mg/dL compared to available HCL systems over a week of at-home use in adults and adolescents with T1D.^23,24 AIDANET used with meal announcement was previously tested using carbohydrate counting, ¹¹ however, limited carbohydrate counting literacy may hinder the performance of the system and limit user adherence to this treatment method. In this work, we introduce a simplified meal announcement approach as an alternative to carbohydrate counting. This method allows users to signal meal occurrences without the need for precise carbohydrate estimation, thereby reducing numeracy requirements while still permitting user intervention at mealtime. This approach serves as an intermediate step between FCL and conventional carbohydrate counting.

The primary objectives of this work are to (1) demonstrate safety and feasibility of the AIDANET system run in a new smaller network version when used in full closed loop by adults with T1D and (2) assess the safety of AIDANET when used in HCL using carbohydrate counting and the proposed simplified announcement on specific days. We hypothesize that performances of this smaller version of AIDANET will provide safe and effective care when used at home.

Methods

This study is a randomized cross-over study comparing participants’ usual care with one week at home on AIDANET. AIDANET was used at home for one week following a two-night hotel stay when participants were taught how to operate the new system before the at-home portion of the study. The comparator usual care week was randomly assigned to being before or after use of AIDANET. This study also assessed differences in glycemia outcomes on assigned days when the participant (1) entered total carbohydrate amount and (2) announced when food was starting to be consumed.

Study AID System

This work uses the UVA-AIDANET system as the AID system which has been described in previous works.^10,11,21,22 The UVA-AIDANET system has four main components: the NN which modulates basal delivery every 5 minutes; a Bolus Priming System (BPS) which automatically provides an initial bolus for meal-like disturbances; ²¹ a hyperglycemia mitigation system which provides correction boluses for sustained hyperglycemia; and a performance-based adaptation system which modulates overall system aggressiveness based on glycemic metrics. ²⁵ The system was designed to primarily operate without user meal announcements.

This work features a new wrapper module, POGO, which rescales the insulin requests to better meet the needs of individuals. ²⁶ Using a ratio of estimated insulin needs, POGO translates a patient’s insulin inputs to a standardized nominal model used by the AID system, then adjusting the output dose to align with the patient’s actual insulin sensitivity: enabling personalization without modifying the AID algorithm. This simplifies the requirements for the NN by only having to learn outputs for a nominal individual while POGO manages individualization. Previous studies [NCT06041971] ¹¹ have used a network requiring 3500Kb of storage, and in this work we use only 11Kb: 300 times less memory usage. The NN is trained in the same manner as in Kovatchev et al ¹¹ using a dense, saturated dataset built from 1.4 billion input-output pairs collected from in-silico and in-vivo sources. ²⁷ Additionally, we implement an Easy Bolus feature, where a simple meal announcement (e.g., “I am eating now”) immediately triggers a bolus from the BPS module, simplifying user input.

Study Design

The University of Virginia Institutional Review Board (IRB-HSR#301828) and the Food and Drug Administration (IDE#G240236) approved this randomized controlled clinical trial (ClinicalTrials.gov NCT06633965). Eligible participants were adults (18-60 years) with T1D for at least one year, using an insulin pump (≥3 months, either open-loop or hybrid-closed-loop therapies), along with a Dexcom G6 or G7 CGM. They were required to have a supportive companion nearby and not be pregnant or breastfeeding. Exclusion criteria included recent severe hypoglycemia, DKA, long-term renal disease, adrenal insufficiency, seizure disorders, untreated thyroid conditions, coronary artery disease, recent steroid use, planned surgery, adhesive intolerance, or investigator-determined risk. There were no limits on the baseline glycated hemoglobin (HbA1c); detailed inclusion and exclusion criteria can be found on clinicaltrials.gov.

Enrollment and screening visits were performed in-person or via phone or secure internet video connection during which medical history and insulin used parameters were obtained and documentation of a physical examination within the prior year was reviewed. Female participants of childbearing potential were provided a urine pregnancy test.

Data were gathered from an approximately 1-week control period on glycemic control and insulin administration with the participants’ usual care therapy. Participants were randomized 1:1 to either Group A (control period prior to AIDANET use) or Group B (control period after AIDANET use).

The AIDANET period consisted of a two-night supervised hotel stay followed by six nights of home use with remote monitoring with an Automated Notification System to alert health care professionals of various glycemic events. Participants ate three meals per day, with post-meal glucose data recorded for analysis. Only the dinner on night 2 was announced using carbohydrate counting, all other meals were unannounced. A supervised 30-minute walk was conducted on Day 2, with a temporary control rate (TCR) set at the study physician’s discretion to manage glucose levels as desired for each participant. Data on exercise intensity, glucose trends, and hypoglycemia were collected. During the seven-day remote monitored at-home period, participants continued using the system while eating freely. On day 2, at-home participants were instructed to use the Easy Bolus strategy; and on day 4, at-home participants were instructed to use carbohydrate counting for all meals. They followed their usual diet and exercise routines, using TCR as needed, before returning to their standard diabetes therapy.

Patient-reported outcomes (PROs) were collected before and after the AIDANET intervention using two quantitative surveys: the Technology Expectation and Experience, ²⁸ and the INSPIRE questionnaire ²⁹

Study Devices

For the AIDANET period, participants were connected to the study devices when they arrived for the hotel period of the study. Participants switched therapy to the study pump (BLE connected t: slim pump, Tandem Diabetes Care, San Diego, CA) and Dexcom G6 CGM (Dexcom, Inc., San Diego, CA). These devices were connected to the DiAs platform³⁰ (University of Virginia, Charlottesville, VA) using a study-provided Android cellphone that allowed remote monitoring (UVA DWM system³¹). The AIDANET system was then initialized at 80% of the subjects’ total daily insulin (TDI, computed as an average of the previous 7 days) to promote conservative control over the first few days before adaptation, accounting for lifestyle changes in the controlled hotel conditions compared to the participants’ usual routines.

A study blood glucose meter (ContourNext One; Ascencia Diabetes Care, Parsippany, New Jersey) and study blood ketone meter (Precision Xtra; Abbott, Alameda, CA) were provided to all participants for use as necessary in adherence to the glycemic guidelines. On admission to the hotel, participants were also fitted with a physical activity tracker (Fitbit^® Charge 3; Fitbit, San Francisco, CA; data not used by the AID system).

Glycemic Treatment Guidelines and Remote Monitoring

During the hotel stay, participants were treated for hypoglycemia with 5 to 15 g of oral glucose if CGM readings dropped below 70 mg/dL, confirmed with a fingerstick, or if the participant reported hypoglycemia symptoms. Hyperglycemia over 300 mg/dL for two hours or 400 mg/dL at any point prompted hourly blood glucose and ketone checks, site evaluation, and corrective boluses as necessary. Throughout the trial (at the hotel and at home), participants were remotely monitored via the DiAs platform and diabetes web monitoring (DWM) system, with automated alerts for glycemic extremes. Providers contacted participants or companions as needed for timely intervention, following safety protocols for hypoglycemia and hyperglycemia.

Data Analysis

The total sample size of six participants was selected based on previous experience for safety and feasibility studies and was intended to support an initial safety evaluation before proceeding to a larger out-patient efficacy study. Due to the small sample size, no P values are reported. The primary outcomes are mean CGM and glucose management indicator (GMI) during the last 7 days of AIDANET compared to the 7 days of usual care. Additional secondary outcomes recommended by the international consensus on CGM: CGM standard deviation, CGM coefficient of variation, percent time <54 mg/dL, percent time <70 mg/dL, percent time in tight range 70 to 140 (TITR), percent time in range 70 to 180 mg/dL (TIR), percent time >180 mg/dL, and percent time >250 mg/dL. Outcomes were further segregated between FCL, CHO counting, and easy bolus days during the AIDANET period.

Continuous glucose monitor data management considerations: (1) saturated CGM values “high” and “low” were replaced by 401 mg/dL and 39 mg/dL, respectively; (2) CGM gaps shorter than 1 h were interpolated; (3) CGM data around recorded occlusion events were removed from 2 h before to 2 h after the time of the event; and (4) The 2 h of CGM data following a pump/DiAs communication interruption longer than 1 h were removed. These criteria were determined prior to the study and included in the protocol.

Results

Eight participants with T1D provided informed consent. Of these, one did not pass the screening process and another withdrew prior to the study intervention. The remaining six participants were randomized, all of whom (100%) completed the trial in December of 2024. The demographic characteristics are shown in Table 1. All participants used commercially available HCL AID technology for their usual care.

Table 1.

Baseline Characteristics of Participants.

A1c (%)	7.17 ± 0.41
Age (y)	40.3 ± 7.7
Weight (kg)	83.4± 14.6
BMI (kg/m2)	29.7 ± 3.9
Diabetes duration (y)	24.9 ± 8.7
Gender	Female	3
	Male	3
Ethnicity	Not Hispanic or Latino	6
Race	White	6
CGM use	Dexcom G6	3
	Dexcom G7	3
Pump use	Tandem / t: slim X2 with Control-IQ	4
	Insulet / Omnipod 5	2

Values are Presented as: Mean ± Standard Deviation, or Count (#).

Table 2 presents glycemic metrics of the last 7 days of home use of the AIDANET system compared to the 7 days of usual care. The mean CGM values showed a slight improvement with AIDANET, with an average reduction of 6.7 mg/dL (all days) and 9.5 mg/dL (FCL days), though we emphasize that n = 6 is too small a sample size to test these differences statistically. Similarly, the GMI and TIR showed minor improvements with AIDANET, but the differences were small and within a narrow range. Other metrics such as time below 70 mg/dL, time above 180 mg/dL, and time above 250 mg/dL also exhibited marginal changes between the interventions, with AIDANET showing slightly better outcomes in terms of reducing time spent in higher glucose ranges. Total daily insulin ranged from 22.9 to 93.4 units in usual care and 28.6 to 122.7 units with AIDANET, demonstrating the system’s ability, with the new POGO module, to manage a wide range of insulin requirements.

Table 2.

Glycemic Metrics of AIDANET Compared to Usual Care.

	Usual care	AIDANET	Difference	FCL Days	Difference
Mean CGM	168.0 ± 24.3	161.3 ± 16.7	−6.7 [−26.45, 13.05]	158.5 ± 18.0	−9.5[−32.61, 13.61]
GMI	7.3 ± 0.6	7.2 ± 0.4	−0.2 [−0.63, 0.31]	7.1 ± 0.4	−0.2 [−0.78, 0.33]
Time in range	63.9 ± 14.9	66.4 ± 8.3	2.5 [−8.29, 13.34]	67.7 ± 9.4	3.8 [−10.12, 17.77]
Time in tight range	38.5 ± 13.8	42.7 ± 10.8	4.1 [−8.32, 16.62]	43.3 ± 11.6	4.8 [−8.14, 17.66]
Time below 70	0.9 ± 1.0	1.6 ± 1.8	0.7 [−1.67, 3.07]	1.8 ± 2.0	0.9 [−1.59, 3.45]
Time below 54	0.1 ± 0.2	0.3 ± 0.6	0.2 [−0.46, 0.85]	0.3 ± 0.6	0.2 [−0.48, 0.86]
Time above 180	35.2 ± 15.2	32.0 ± 8.8	−3.2 [−13.78, 7.33]	30.5 ± 9.8	−4.8 [−18.31, 8.79]
Time above 250	10.3 ± 8.9	9.7 ± 6.3	−0.6 [−8.15, 6.86]	8.2 ± 8.0	−2.2 [−12.26, 7.96]
Coefficient of variation	34.3 ± 3.8	36.0 ± 3.0	1.7 [−3.67, 7.09]	34.7 ± 4.3	0.4 [−5.91, 6.74]
Standard deviation	57.9 ± 11.8	58.0 ± 6.8	0.1 [−0.34, 0.60]	55.1 ± 8.4	−2.8[−3.33,−2.22]
Total Daily Insulin	66.2 ± 26.5	75.3 ± 33.1	9.1 [−4.10, 22.28]	69.3 ± 29.1	3.1 [−8.20, 14.35]

Values are shown as: mean ± standard deviation, or mean [95% CI].

Figure 1 shows the ambulatory glucose profile (AGP) plot between the AIDANET and usual care. We notice the largest improvement during the overnight period with lower glucose values and tighter envelopes. This indicates consistently tighter control overnight using AIDANET compared to current HCL systems. Daytime glycemic levels remain very similar between the usual care and AIDANET periods.

Figure 1.

Ambulatory glucose profile (AGP) plot of AIDANET vs usual care. Red shows AIDANET system and blue shows usual care period.

Figure 2 shows the Ambulatory Glucose Report of the mean performance of the system using different bolus strategies. It is worth noting that different numbers of days are considered for each category: 7 UC, 5 FCL, 1 HCL, 1 Easy Bolus. HCL and Easy Bolus values may be less robust against outlier values. Easy Bolus has the highest TIR with 70.5%, followed by FCL then HCL with 67.2% and 57.1%, respectively. To analyze system response to meals, the daytime period (6:00 AM to midnight) was used to approximate the postprandial response. Exact postprandial periods could not be determined because participants were not required to log their meals, and existing meal announcements on non-FCL days may have been unreliable in free-living conditions. Results during this window mirror the overall findings: Easy Bolus had the highest TIR, followed by FCL and then HCL.

Figure 2.

Ambulatory glucose report. Show the mean time in specified ranges of the system when using different bolus strategies. Results show overall (left) and daytime (6:00 AM to midnight) (right) periods.

Due to inconsistencies in survey collection and small sample size, the full analysis of PROs is not included. However, it is worth noting participants rated the statement “I would very much like to keep using the program” highly after the intervention, with a median score of 5.00 on a 5-point Likert scale [interquartile range (IQR): 4.00-5.00], reflecting a strong positive reception of the FCL system.

In a retrospective analysis, the original dosing algorithm was evaluated with the same inputs as the NN. It was found that 99% of the NN outputs (the 5 minutes micro bolus) deviated by less than 31.7 mU from the original dosing algorithm it was trained on, and that no deviations occurred when glucose <70 mg/dL. Thus, the NN and the original dosing algorithm issued virtually identical commands.

A software bug related to 0% TCR and resulting in incorrect insulin history reconstruction was detected during the study. Participants were instructed not to use 0% TCR and no software changes were made. There were no serious adverse events or adverse events during the study.

Discussion

This initial feasibility study evaluated a mini NN to implement MPC in an AID system. We expect that this will constitute an important step toward having an AID system that has potential to be used embedded in insulin pumps. The study assessed system performance using different bolus strategies: FCL, hybrid closed-loop with carb counting (HCL), and the new Easy Bolus approach. The primary objective was to determine the safety and feasibility of the AIDANET system, providing a foundation for larger-scale clinical trials.

The results demonstrate good glycemic control when using AIDANET. Mean glucose levels appear lower when using AIDANET, including in FCL mode: without any meal announcements. Secondary outcomes, such as TIR and TITR, tended to increase with AIDANET, supporting the possible effectiveness of the system. In addition, time above range tended to decrease, highlighting possible improved hyperglycemia management without increasing hypoglycemia risks in a meaningful way. These findings were confirmed over both the overall and daytime periods.

Bolusing behaviors produced some unexpected findings. Even though Easy Bolus seemed to outperform FCL in TIR, suggesting that minimal meal announcements may enhance glycemic outcomes, HCL seemed to fall short of expected benefits. This is potentially due to conservative carbohydrate ratio and correction factor settings based on an initial 80% of total daily insulin estimate that was not updated during system adaptation. This may have resulted in under-bolusing, leading to suboptimal glycemic control. In addition, as we did not control or record participant diets when at home, meals in HCL may have differed from meals in FCL.

Future work could focus on refining meal announcement strategies to enhance glycemic outcomes further. Larger, statistically powered trials are needed to compare the efficacy of different bolusing strategies within AIDANET. Notably, this study was conducted with a small sample size (n = 6), which does not allow for the generalization of the findings. A larger trial is necessary to validate these results across a more diverse population. Ultimately, the findings indicate that AIDANET is safe and potentially effective, supporting its further study.

Conclusion

This pilot-feasibility study demonstrates that the AIDANET system provides effective glycemic control across different bolusing strategies. The use of a smaller NN proved effective, supporting its feasibility as a more efficient alternative to larger models. The Easy Bolus approach showed promising results, while carbohydrate counting may require further optimization of carbohydrate ratio and correction factor settings. While the study supports the safety and efficacy of AIDANET, the small sample size (n = 6) limits the generalizability of the findings.