Validation of the UVA Simulation Replay Methodology Using Clinical Data: Reproducing a Randomized Clinical Trial

Abstract

Background:

Computer simulators of human metabolism are powerful tools to design and validate new diabetes treatments. However, these platforms are often limited in the diversity of behaviors and glycemic conditions they can reproduce. Replay methodologies leverage field-collected data to create ad hoc simulation environments representative of real-life conditions. After formal validations of our method in prior publications, we demonstrate its capacity to reproduce a recent clinical trial.

Methods:

Using the replay methodology, an ensemble of replay simulators was generated using data from a randomized crossover clinical trial comparing the hybrid closed loop (HCL) and fully closed loop (FCL) control modalities in automated insulin delivery (AID), creating 64 subject/modality pairs. Each virtual subject was exposed to the alternate AID modality to compare the simulated versus observed glycemic outcomes. Equivalence tests were performed for time in, below, and above range (TIR, TBR, and TAR) and high and low blood glucose indices (HBGI and LBGI) considering equivalence margins corresponding to clinical significance.

Results:

TIR, TAR, LBGI, and HBGI showed statistical and clinical equivalence between the original and the simulated data; TBR failed the equivalence test. For example, in the HCL mode, simulated TIR was 84.89% versus an observed 84.31% (P = 0.0170, confidence interval [CI] [−3.96, 2.79]), and for FCL mode, TIR was 76.58% versus 77.41% (P = 0.0222, CI [−2.54, 4.20]).

Conclusion:

Clinical trial data confirm the prior in silico validation of the UVA replay method in predicting the glycemic impact of modified insulin treatments. This in vivo demonstration justifies the application of the replay method to the personalization and adaptation of treatment strategies in people with type 1 diabetes.

Introduction

People with type 1 diabetes (T1D) require a lifelong insulin replacement therapy to compensate for the insulin secretion deficiency due to the autoimmune destruction of pancreatic beta-cells. Over the past few decades, there has been a substantial investment in diabetes technology driven by the increased availability of precise and interconnected medical devices, such as continuous glucose monitoring (CGM) devices and continuous subcutaneous insulin infusion (CSII) pumps, and the advances in simulation platforms that leverage mathematical models of glucose-insulin dynamics.^1,2 These metabolic simulators are powerful tools to design and validate new treatment paradigms, greatly accelerating and de-risking the path to clinical validation.^1,3

Among the simulation platforms used in T1D treatment research, the FDA-accepted UVA/Padova simulator stands out.⁴ The simulator has a large in silico cohort of 100 children, 100 adolescents, and 100 adults. Other well-known platforms include the Oregon Health and Science University simulator,⁵ and the Cambridge simulator.⁶ In addition, a minimal model describing glucose kinetics was developed by Grosman et al.,³ which also served as a test bed to predict the insulin therapy outcomes of different clinical trials.^7,8

These simulators rely on specific mathematical functions that relate carbohydrate ingestion and insulin administration to generate blood glucose time series.⁹ However, present physiological models do not encompass the entirety of human biological systems. They are limited in accounting for various phenomena affecting the glycemic profile of an individual, such as glycemic disturbances and patient behavior.^10,11

Recent works have focused on enhancing T1D simulators to fill the lack of real-life scenarios and capture variability observed in the field. Ahmad et al.¹¹ used clinical data from 14 subjects to generate a new virtual patient cohort by means of parameter optimization of the UVA/Padova model to capture day-to-day variations in glycemia and represent the glycemic outcomes of the data collected. However, this method did not fully capture intraday variability caused by factors unrelated to meals and exercise present in real data. In another recent study, a method for determining the parameters of a simplified version of the UVA/Padova simulator based on the Markov chain Monte Carlo was presented.¹² This strategy seeks to mimic the glucose response to different changes in insulin therapy observed in real-world data, although it has only been evaluated within single-meal scenarios and for specific time frames without meals or boluses in the four hours preceding the period of interest.

Our approach aims to reproduce the variability observed in the real-world blood glucose traces for daylong streams of data, regardless of meal intakes and/or the presence of unmodeled phenomenon (e.g., mixed physical activity, stress, or illness).^13,14 This methodology allows for the creation of ad hoc simulators of 24 h of field-collected T1D data, enabling the quantification of the glycemic response to “what-if” scenarios. It is based on the identification of a simplified glucose-insulin model and the estimation of an additive disturbance signal, aka the net-effect signal, that represents unmodeled phenomena. The domain of validity of this approach has been assessed with experimental scenarios generated from with the UVA/Padova simulator by comparing the true and replayed glucose traces when changing meals and insulin therapy.¹³ The present work presents a step forward on the validation of the UVA replay simulator by reproducing real-life glucose traces under different insulin regimens from subjects undergoing a clinical trial.

Materials and Methods

The objective of this study is to showcase the potential of the UVA replay simulator as a benchmark for testing and validating treatment strategies for T1D within the context of daily scenarios experienced by real-life patients. To that end, data from the DCLP6 clinical trial (clinicaltrials.gov NCT04877730), which explored different automated insulin delivery (AID) modalities, hybrid closed loop (HCL) and fully closed loop (FCL), are used as baseline for comparison.¹⁵ The intra- and intersubject variability observed in field is captured using the replay methodology, generating a cohort of representative virtual subjects per control modality. Subsequently, each virtual cohort is set in closed loop with the different control modes, simulating glucose response to therapy variations. The resulting outcomes from these simulations are compared with those observed in real subjects.

Data set

Data from 32 people with T1D were collected during the DCLP6 clinical trial conducted by UVA and approved by the UVA Institutional Review Board for Health Sciences Research (210035; Charlottesville, VA).¹⁵ The study aimed to assess the glycemic impact of different modalities of the UVA AID system for T1D treatment. Participants completed 3 study days. Each day on a different control modality randomly assigned at 16:00 h.¹⁵ Three meals were provided every day, dinner ranging from 18:00 h to 21:00 h, lunch fixed at 13:00 h for all subjects, and breakfast from 7:00 h to 9:00 h. Despite not being used by the controllers under the FCL mode, the information about the carbohydrate content and time of each meal was manually recorded by the study team.

For simplicity, data from the HCL and FCL modalities were selected for this analysis: CGM and insulin delivery (basal rate (BR) modulation, automatic bolus, and manual bolus) with the corresponding time stamps were downloaded from the DiAs Web Monitoring (DWM) tool used during the clinical trial.¹⁶ In addition, time and value data pertaining to BR, carbohydrate ratio (CR), and correction factor (CF) profiles, hypoglycemia treatments (HTs), self-monitoring blood glucose (SMBG), insulin-on-board (IOB), estimated model states, and log messages were also downloaded from DWM for unit test purposes. Body weight (BW) was obtained from the participants’ medical history.

Details on data preprocessing can be found in the Supplementary Data S1, along with the description of the unit tests performed to identify and address potential discrepancies in control actions and glucose traces when applying the strategy to real-world data. Following preprocessing, the data were separated in 24-h segments corresponding to the use of a specific AID modality by an individual participant (a subject/modality pair).

Virtual diabetic subject

The methodology to generate virtual subjects is based on the replay simulation tool developed at UVA.^13,14 This is a data-driven tool that integrates a glucose–insulin minimal model personalized to represent each day of data for each subject, along with a net-effect signal obtained through regularized deconvolution intended to capture unmodeled phenomena.

Glucose-insulin minimal model

The metabolic model used to construct the virtual patients is built around a discrete-time linearized version of the subcutaneous oral minimal model^14,17 updated by Hughes et al.¹³ The model can be described by the following: $x (k + 1) = A x (k) + B_{i} u (k) + B_{m} m (k) + B_{w} w (k),$ (1) $y (k) = C x (k),$ where $k$ is the discrete time stamp and $x$ is the metabolic state, including blood glucose concentration, the proportion of insulin in the remote compartment, the amount of nonmonomeric and monomeric insulin in the subcutaneous space, plasma insulin concentration, and the glucose mass in the first and second compartment of the oral glucose transport subsystem. The model inputs are the exogenous insulin infusion rate $u$ (mU/min), carbohydrate intake $m$ (mg/min), and a net-effect signal $w$ . The output of the model $y$ corresponds to the blood glucose concentration. $A$ , $B_{i}$ , $B_{m}$ , $B_{w}$ , $C$ are state-space matrices describing the interaction between glucose and the system inputs. A detailed description of the model and its parameters can be found in Hughes et al.¹³ and Diaz et al.¹⁸

Model identification

For each subject/modality pair (the 24 h of data where a study participant used a specific AID modality), model (1) is individualized on patient available data (CGM, insulin, meals, and HTs), while setting the net-effect to zero. Parameters related to insulin sensitivity and fractional glucose effectiveness are estimated on a daily basis, while parameters related to the carbohydrate absorption rate are fitted to each individual meal to account for intermeal absorption differences. Parameters are identified by maximizing the posterior probability of observing the set of parameters conditioned on the day’s data, with population averages used as priors. Population parameters are used for the remaining parameters.¹⁸

Residual metabolic signal estimation

The net-effect signal is estimated following the procedure described by Patek et al.¹⁴ In summary, the residual metabolic signal for the daily data is estimated through regularized deconvolution by inverting the model previously identified. Once the net-effect signal is obtained, it is fed back into model (1) with the insulin/meal/HT records to reproduce the data, encompassing all its inherent variability.

Closed loop simulation

Once each virtual subject model is obtained, it is integrated with the control algorithm, that is, glucose and insulin signals are regenerated through replay simulation instead of using data retrieved from the DWM. As shown in Figure 1, the first simulation step, denoted replay, consists of simulating each virtual patient under the sequence of control modalities tested in the study. This step aims to provide an idea of how well we are reconstructing the signal when factors occurring in real life are closely reproduced. The second step is resimulation, where the insulin therapy applied to the virtual subject is modified to predict the glucose impact. Here, aiming to compare the predicted glucose with data available, resimulation is performed with the alternate control modality as the one evaluated in the clinical trial for this 24-h data segment (see Fig. 1).

FIG. 1.

Process diagram of the replay validation methodology.

Hypoglycemia treatments

Rescue carbohydrates are an important simulation factor to consider. Although virtual subjects are generated using data on HT timing and carbohydrate amount, HTs must be automated based on the glucose response in the replay and resimulation phases. This is because insulin dose changes may result in different glucose conditions requiring new or omitting previous HTs. In this regard, conditions aiming to emulate each virtual subject’s glucose response to fast-absorbing carbohydrate ingestion are considered as in Diaz et al.¹⁸ The glucose absorption submodels corresponding to HTs are averaged and used to simulate the response to the new rescues. In the absence of recorded HTs, an average submodel from meals is computed considering accelerated carbohydrate absorption rates. Furthermore, the subject’s behavior in treating hypoglycemic events is approximated to determine carbohydrate amount and timing. Carbohydrate amount is the average from the HTs provided, and timing is determined by the average glucose value and glucose derivative when HTs were ingested, as well as the time between consecutive HTs.

Outcomes and statistical analysis

Assessment of the replay phase is done by computing the root mean square error (RMSE) between CGM data and the simulated glucose traces for each subject/control-modality pair. Glucose outcome metrics are computed from both the clinical trial source data files and the simulated blood glucose values during both the replay phase and the resimulation phase when alternating the control mode. Considered glycemic outcomes are the percentages of time between 70 and 180 mg/dL (TIR), above 180 mg/dL (TAR), and below 70 mg/dL (TBR). In addition, we computed the high and low blood glucose indices (HBGI and LBHI) and reported the total daily insulin (TDI). An equivalence test is performed to assess whether the outcomes from data-replay and data-resimulation are similar under predefined equivalence margins.^19,20 For each outcome metric, the margins are defined as follows: $δ_{TIR} = 5$ , $δ_{TBR} =$ 1, $δ_{TAR} = 5$ , $δ_{LBGI} =$ 0.5, $δ_{HBGI} = 1$ , and $δ_{TDI} = 5$ .²¹ Paired t-tests are performed to address the two sides of the test with significance level set at 0.05. Outcomes are reported as mean ± standard deviation (SD) with corresponding larger of the two one-sided P-values and 90% confidence intervals (CIs) as recommended in Dixon et al.²⁰

Results

Replay of clinical trial results

Each day of data collected during the clinical trial is reproduced with the replay methodology to obtain an associated virtual subject representation and then set in closed loop with the corresponding simulated control modality. The cumulative distribution of the RMSE between CGM data and replayed glucose traces for the cohort of subjects using each of the control modalities is shown in Figure 2. An RMSE of less than 20 mg/dL was obtained except for one subject/modality pair. In addition, Table 1 reports outcome metrics obtained from collected data and from the replay stage for both control modalities. Equivalence with P < 0.001 was shown for all metrics except for TBR, not being able to show TBR equivalence under the HCL mode. All 64 subject/modality pairs were selected for analysis of the resimulation phase.

FIG. 2.

Cumulative distribution of RMSE between CGM reported from the clinical trial and replay glucose with each control modality. RMSE, root mean square error; CGM, continuous glucose monitoring.

Table 1.

Comparison of Data Versus Replay Overall Outcome Metrics When Using Each Control Mode During the Clinical Trial

	HCL modality
	HCL-data	HCL-replay	P value	CI	Equivalence margin
TIR (%)	84.89 ± 10.27	84.39 ± 10.33	<0.0001	[−1.51, 0.49]	5
TBR (%)	2.42 ± 3.25	3.00 ± 3.37	0.1742	[−0.19, 1.33]	1
TAR (%)	12.68 ± 10.11	12.62 ± 10.55	<0.0001	[−0.75, 0.62]	5
LBGI	0.93 ± 0.72	0.99 ± 0.78	<0.0001	[−0.05, 0.17]	0.5
HBGI	2.70 ± 1.89	2.66 ± 1.98	<0.0001	[−0.15, 0.06]	1
TDI (U)	48.95 ± 24.44	48.72 ± 24.04	<0.0001	[−0.94, 0.47]	5

	FCL modality
	FCL-data	FCL-replay	P value	CI	Equivalence margin
TIR (%)	76.58 ± 11.72	77.30 ± 11.07	<0.0001	[−0.50, 1.95]	5
TBR (%)	2.50 ± 2.77	2.77 ± 3.08	0.0283	[−0.35, 0.90]	1
TAR (%)	20.92 ± 11.71	19.93 ± 11.43	<0.0001	[−2.02, 0.03]	5
LBGI	0.82 ± 0.56	0.89 ± 0.57	<0.0001	[−0.05, 0.17]	0.5
HBGI	4.19 ± 2.25	4.03 ± 2.21	<0.0001	[−0.29, 0.02]	1
TDI (U)	44.63 ± 19.44	45.10 ± 19.64	<0.0001	[−0.05, 1.00]	5

Resimulation with different control modalities

Each virtual subject/day pair is simulated using a different control modality, so that subjects generated in HCL mode are simulated using FCL therapy, and subjects generated in FCL are simulated using HCL (as in Fig. 1). Individual subject glucose and insulin traces comparing collected data, replay phase, and resimulation phase under both control modes are presented in Supplementary Figure S1. The Figure illustrates how the effects on glucose levels during the clinical trial are approximated when the therapy switch is applied to each virtual subject/day. Specifically, it shows faster drops of glucose levels during prandial periods due to manual boluses when using the HCL mode, and the corresponding increase in glucose levels aligning with original data collected during the FCL mode.

The distributions of TIR, TBR, and TAR for each control modality are depicted in Figure 3, comparing the outcomes computed from collected data and those from the resimulation phase. Population outcome metrics are also reported in Tables 2 and 3, for HCL and FCL, respectively. Equivalence was demonstrated for all glycemic metrics except for TBR. The CI for the TIR, TAR, LBGI, HBGI, and TDI is completely contained within the defined equivalence margin. The simulated TBR was higher than TBR obtained in the clinical trial in both HCL (data: 2.42% vs. simulation: 2.94%) and FCL (data: 2.50% vs. 2.68%) modalities. In addition, outcome metrics during overnight and daytime periods are reported in Supplementary Table S1. In HCL mode, data versus simulation for daytime TIR is 81.93% versus 81.28%, TBR is 2.42% versus 3.15%, and delivered TDI is 41.36U versus 40.76U, while during the overnight period, TIR is 94.14% versus 93.41%, TBR is 2.14% versus 2.27%, and delivered TDI is 7.66U versus 6.86U. In the case of FCL, daytime TIR is 72.50% versus 72.70%, TBR is 2.47% versus 3.08%, and delivered TDI is 38.11U versus 37.66U, and overnight metrics correspond to a TIR of 88.91% versus 91.70%, TBR is 2.53% versus 1.54%, and delivered TDI is 6.53U versus 7.42U.

FIG. 3.

Comparison of the distribution of glucose outcome metrics obtained from data and resimulation phase.

Table 2.

Comparison of Glycemic Outcomes When Using HCL Mode During the Clinical Trial and When Resimulating the Virtual Patient Cohort Generated with FCL

	HCL modality
	HCL-data	HCL-resimulation	P value	CI	Equivalence margin
TIR (%)	84.89 ± 10.27	84.31 ± 8.84	0.0170	[−3.96, 2.79]	5
TBR (%)	2.43 ± 3.25	2.94 ± 3.29	0.2832	[−0.89, 1.93]	1
TAR (%)	12.68 ± 10.11	12.75 ± 8.88	0.0064	[−3.11, 3.24]	5
LBGI	0.93 ± 0.72	0.97 ± 0.60	0.0015	[−0.20, 0.28]	0.5
HBGI	2.70 ± 1.89	2.68 ± 1.85	0.0072	[−0.66, 0.61]	1
TDI (U)	48.95 ± 24.44	47.62 ± 21.27	0.0046	[−3.57, 0.90]	5

Table 3.

Comparison of Glycemic Outcomes When Using FCL Mode During the Clinical Trial and When Resimulating the Virtual Patient Cohort Generated with HCL

	FCL modality
	FCL-data	FCL-resimulation	P value	CI	Equivalence margin
TIR (%)	76.58 ± 11.72	77.41 ± 11.63	0.0222	[−2.54, 4.20]	5
TBR (%)	2.50 ± 2.77	2.68 ± 3.42	0.1305	[−1.02, 1.39]	1
TAR (%)	20.92 ± 11.71	19.91 ± 10.62	0.0220	[−4.23, 2.20]	5
LBGI	0.31 ± 0.35	0.36 ± 0.53	0.0001	[−0.12, 0.23]	0.5
HBGI	3.34 ± 2.27	3.46 ± 2.57	0.0296	[−0.64, 0.88]	1
TDI (U)	44.63 ± 19.44	45.10 ± 21.39	0.0000	[−1.16, 2.09]	5

Discussion

There are at present numerous platforms for generating virtual cohorts of T1D subjects with the goal of evaluating new treatments and allowing rapid iteration and improvement of therapies in a safe, low-cost environment. Even though some platforms are based on maximal models, they still fail to capture all the glucose variabilities observed in data collected from real subjects. To address this limitation, a replay methodology that captures glucose variability as a net-effect signal was previously proposed and validated in silico. In this work, we detail the features to consider when applying the strategy to real-world data and showcase its domain of validity to the resimulation of daily glucose traces by comparing the glucose outcomes with those obtained during a clinical trial.

The validation process comprises two stages. The first is the creation of the cohort of virtual subjects to replay data by closing the loop with the control strategy used during data generation. Following that, a resimulation stage is performed in which insulin therapy is altered to assess its effect on glucose. Here, the control mode was switched from HCL to FCL or from FCL to HCL, and glycemic metrics were compared with those obtained during the clinical trial.

Importance of the replay stage relies on providing an idea of how well the method is able to reproduce real-world data, as nonsimulated events that occur in real-life conditions such as pump disconnection, pump occlusions, or bolus override cause differences in the control actions that ultimately lead to a discrepancy in CGM traces when using simulation strategies. Available signals from the pump or monitoring devices can be used to mitigate these differences in control action, as an example, here the information from the time stamps of CGM is used to reproduce asynchronism occurring in real data (see Supplementary Data S1). An equivalence test such as the one presented in Table 1 helps to have a better idea about the ability to reproduce the cohort metrics obtained in the collected data and possible aspects to consider when drawing conclusions in the resimulation stage. In this study, equivalence was demonstrated for all metrics except TBR under the HCL mode.

Replay simulations have been used to suggest insulin dosing parameters in a personalized manner to individuals with T1D. Works related encompass optimization of BR, CR, and CF in multiple daily insulin injection, CSII, and AID therapies.^18,22,23 It has been acknowledged that, as any simulation strategy, replay simulations have a domain of validity, and so, a threshold based on RMSE can be imposed to consider a valid subject/day to move forward to the resimulation stage.

The obtained results show TIR, TAR, LBGI, HBGI, and TDI equivalence between real and resimulated data with both control modes. This highlights the replay methodology’s potential to predict the performance of an untested control strategy in a clinical setting. In turn, suggesting a valid use as a new platform for the validation of new therapy designs in scenarios that include glucose variability observed in real subjects. When performing the resimulation with both control modes, a TBR greater than that of the collected data was observed. We attribute this to two main factors as follows: first, the evaluation of the TIR with the hard threshold imposed at 70 mg/dL, as for some virtual subject/day pairs, despite having similar responses, the glucose remained close to but below 70 mg/dL for longer periods, inflating the TBR (this possible explanation is supported by the equivalence of LBGI, a different hypoglycemia metric that is not affected by the threshold effect); and second, the change of HT inputs from those reported in the data to the automated version in the resimulation, as both the conditions to trigger an HT (CGM value and its derivative) and the glucose response to fast-absorbing carbohydrates were averaged from daily data. To counteract this second factor, better personalized rules for determining the timing, amount, and effect on glucose of HT could be sought.

Additional outcome reports during overnight and daytime periods provide insights into how the resimulation stage adequately reflects glucose outcomes according to the effect each control mode had during the trial. In the case of HCL mode, insulin dosage increased during the day to similar amounts as those collected in the data (TDI data: 41.36U ± 19.93U, TDI resimulation: 40.76U ± 18.62U), which consequently dropped glucose to similar levels than those observed during the trial (TAR data: 15.55% ± 10.53%, TAR resimulation: 15.57% ± 10.48%). For the case of FCL, insulin dosing during the day was reduced (TDI data: 38.11U ± 17.22U, TDI resimulation: 37.66U ± 16.76U) increasing TAR and reducing TIR as observed in data (TAR data: 25.03% ± 11.48%, TAR resimulation: 24.22% ± 10.56%, TIR data: 72.50% ± 11.14%, TIR resimulation: 72.70% ± 12.20%).

To further show the potential of the replay methodology, results were compared with those obtained with the UVA/Padova simulator, which was used to gain FDA approval for the DCLP6 clinical trial. The simulator was set to mimic the clinical trial conditions as follows: 24 h per control mode, 3 fixed meals at 7:00 h, 13:00 h, and 19:00 h, with intraday variability to account for down phenomenon and variations of insulin sensitivity. The comparison is reported in Supplementary Table S2. The results suggest that the outcomes of the clinical trial are better reflected with the replay simulator. For example, in the case of HCL mode, the TIR observed in the trial data was 84.89%, which is closer to the replay simulator’s TIR 84.31% compared with 79.04% obtained with the UVA/Padova simulator. Similarly, the replay simulator’s TAR was 12.75%, compared with 20.72% for the UVA/Padova simulator, which is again closer to the clinical trial data (12.68%).

This work represents an important advancement in the evaluation of the UVA replay methodology in particular, and the first clinical validation of a replay methodology in general, by reproducing the outcomes of a full day of a clinical study under controlled conditions. While an essential next step, we recognize that this singular case alone is not sufficient to confidently assert a broadened domain of validity. Rather, validation of in silico modeling methodologies requires an iterative approach as new real-world data sets become available. This work may also provide a blueprint for the validation of similar replay environments, leveraging controlled clinical trials to allow for the in silico treatment switch.

Future research should assess the strategy’s validity using data from free-living conditions, as hotel-based AID trials still impose restrictions on behaviors and therefore glycemic variability and allow for data quality that may be hard to achieve in real life. Furthermore, while comprehensive analyses of model identification and input independence were outside the scope of this validation study, we continue to work on improving the present method, particularly as it pertains to baseline accuracy of the core physiological model identification process and the independence of the additive signal to inputs that may be changed in the application of our simulation methodology. In addition, the effect of metabolic changes over longer time periods and data handling in an uncontrolled environment remains to be investigated.

Conclusions

The ability of the UVA replay methodology to generate realistic virtual T1D subjects and predict clinical outcomes is shown in this study. Resimulating cohorts with different therapies replicated key glycemic metrics from a closed-loop trial, establishing validity as a standard for in silico evaluation. The presented platform should allow for the testing of new treatments under realistic conditions, thereby accelerating therapeutic optimization and delivery to improve diabetes management.

Footnotes

Author Disclosure Statement

The authors declared the following potential conflicts of interest. M.F.V. and P.C. receive research support and royalties from Dexcom handled by the University of Virginia’s Licensing and Ventures Group. M.D.B. declares research support handled by the University of Virginia by Dexcom, Novo Nordisk, Tandem Diabetes Care; patent royalties handled by the University of Virginia by Dexcom, LifeScan, Novo Nordisk, and Sanofi. Honoraria: Tandem and Sanofi. Consulting: Roche, Portal Insulin LLC, and Dexcom.

Funding Information

This work was supported by an NIH grant 5R01DK129553.

Supplementary Material

Supplementary Data

Supplementary Figure S1

Supplementary Table S1

Supplementary Table S2

Supplementary Data S1

References

Cobelli

, Man

, Sparacino

, et al. Diabetes: Models, signals, and control. IEEE Rev Biomed Eng, 2009; 2:54–96; doi: 10.1109/RBME.2009.2036073.

Colmegna

, Wang

, Garcia-Tirado

, et al. Mapping data to virtual patients in type 1 diabetes. Control Eng Pract, 2020:103; doi: 10.1016/j.conengprac.2020.104605

Grosman

, Wu

, Miller

, et al. Sensor-augmented pump-based customized mathematical model for type 1 diabetes. Diabetes Technol Ther, 2018; 20(3):207–221; doi: 10.1089/dia.2017.0333

Visentin

, Campos-Náñez

, Schiavon

, et al. The UVA/Padova type 1 diabetes simulator goes from single meal to single day. J Diabetes Sci Technol, 2018; 12(2):273–281; doi: 10.1177/1932296818757747

Resalat

, Youssef

, Tyler

, et al. A statistical virtual patient population for the glucoregulatory system in type 1 diabetes with integrated exercise model. PLoS One, 2019; 14(7):e0217301; doi: 10.1371/journal.pone.0217301

Wilinska

, Chassin

, Acerini

, et al. Simulation Environment to Evaluate Closed-Loop Insulin Delivery Systems in Type 1 Diabetes, 2010; 4; www.journalofdst.org

Grosman

, Parikh

, Roy

, et al. In silico evaluation of the medtronic 780G system while using the GS3 and its calibration-free successor, the G4S sensor. Ann Biomed Eng, 2022; 51(1):211–224; doi: 10.1007/s10439-022-03079-9

Grosman

, Wu

, Parikh

, et al. Fast-acting insulin aspart (Fiasp^®) improves glycemic outcomes when used with MiniMedTM 670G hybrid closed-loop system in simulated trials compared to NovoLog^® . Comput Methods Programs Biomed, 2021; 205:106087; doi: 10.1016/j.cmpb.2021.106087

Noguer

, Contreras

, Mujahid

, et al. Generation of individualized synthetic data for augmentation of the type 1 diabetes data sets using deep learning models. Sensors (Basel), 2022; 22(13):4944; doi: 10.3390/s22134944

10.

Mujahid

, Contreras

, Beneyto

, et al. Conditional synthesis of blood glucose profiles for T1D patients using deep generative models. Mathematics, 2022; 10(20):3741; doi: 10.3390/math10203741

11.

Ahmad

, Ramkissoon

, Beneyto

, et al. Generation of virtual patient populations that represent real type 1 diabetes cohorts. Mathematics, 2021; 9(11):1200; doi: 10.3390/math9111200

12.

Cappon

, Vettoretti

, Sparacino

, et al. ReplayBG: A digital twin-based methodology to identify a personalized model from type 1 diabetes data and simulate glucose concentrations to assess alternative therapies. IEEE Trans Biomed Eng, 2023; 70(11):3227–3238; doi: 10.1109/tbme.2023.3286856

13.

Hughes

, Gautier

, Colmegna

, et al. Replay simulations with personalized metabolic model for treatment design and evaluation in type 1 diabetes. J Diabetes Sci Technol, 2020; 15(6):1326–1336; doi: 10.1177/1932296820973193

14.

Patek

, Lv

, Ortiz

, et al. Empirical representation of blood glucose variability in a compartmental model. In: Prediction Methods for Blood Glucose Concentration Design, Use and Evaluation; 2016:133–157; doi: 10.1007/978-3-319-25913-0_8

15.

Garcia-Tirado

, Colmegna

, Villard

, et al. Assessment of meal anticipation for improving fully automated insulin delivery in adults with type 1 diabetes. Diabetes Care, 2023; 46(9):1652–1658; doi: 10.2337/dc23-0119

16.

Place

, Robert

, Ben Brahim

, et al. DiAs web monitoring: A real-time remote monitoring system designed for artificial pancreas outpatient trials. J Diabetes Sci Technol, 2013; 7(6):1427–1435; doi: 10.1177/193229681300700603

17.

Dalla Man

, Caumo

, Cobelli

. The oral glucose minimal model: Estimation of insulin sensitivity from a meal test. IEEE Trans Biomed Eng, 2002; 49(5):419–429; doi: 10.1109/10.995680

18.

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

19.

Tamayo-Sarver

, Albert

, Tamayo-Sarver

, et al. Advanced statistics: How to determine whether your intervention is different, at least as effective as, or equivalent: A basic introduction. Acad Emerg Med, 2005; 12(6):536–542; doi: 10.1197/j.aem.2005.01.010

20.

Dixon

, Saint-Maurice

, Kim

, et al. A primer on the use of equivalence testing for evaluating measurement agreement. Med Sci Sports Exerc, 2018; 50(4):837–845; doi: 10.1249/MSS.0000000000001481

21.

Battelino

, Danne

, Bergenstal

, et al. Clinical targets for continuous glucose monitoring data interpretation: Recommendations from the international consensus on time in range. Diabetes Care, 2019; 42(8):1593–1603; doi: 10.2337/dci19-0028

22.

El Fathi

, Fabris

, Breton

. Titration of long-acting insulin using continuous glucose monitoring and smart insulin pens in type 1 diabetes: A model-based carbohydrate-free approach. Front Endocrinol (Lausanne), 2021; 12:795895; doi: 10.3389/fendo.2021.795895

23.

Fabris

, Gautier

, Breton

. Automated Adaptation of Insulin Treatment in Type 1 Diabetes. In: Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS. Institute of Electrical and Electronics Engineers Inc.; 2021:5039-5042. doi: 10.1109/EMBC46164.2021.9630191

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.92 MB

0.56 MB

0.46 MB

0.53 MB

0.51 MB