A digital twin-enhanced decision support system improves time-in-range in type 1 diabetes: a randomized clinical trial

Abstract

Most individuals with type 1 diabetes (T1D) worldwide continue to be managed with multiple daily injections or sensor-augmented pumps. Decision-support systems (DSSs) have emerged as cost-effective tools to enhance treatment adherence and glucose control. We conducted a randomized, open-label, parallel-group study to evaluate STUDIA, a DSS incorporating a digital twin-enabled simulation-assisted bolus calculator. Twenty-eight participants with T1D used either the simulation-assisted calculator or traditional carbohydrate counting for prandial insulin dosing, with glucose monitored using Freestyle Libre. After four weeks, the group using the simulation-assisted calculator showed a 7% increase in time in the target glucose range (70–180 mg/dL) compared to the control (p < 0.001), along with a lower hypoglycemia incidence rate (RR 0.31, p = 0.022). Model performance yielded a mean absolute percentage error at 60 min of 19.2 ± 6.7%, increasing at longer horizons, with most discrepancies between simulation and sensor falling into no-risk or slight-risk zones. These findings support the safety, feasibility, and potential clinical utility of the STUDIA system in people with T1D.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-23165-x.

Introduction

Glycemic control in type 1 diabetes (T1D) remains a challenge with only ~ 20% of American and ~ 30% of European adults achieving a hemoglobin A1c (HbA1c) target of < 7%¹. In Colombia, where this study originated, only ~ 50% of adults achieve the recommended glycemic targets². Although automated insulin delivery (AID) is rapidly becoming the standard of care for T1D, treatment with multiple daily injections (MDI) remains the standard of care in developing countries due to barriers related to access, provider bias, and perceived complexity of new technologies^3,4.

People with T1D trained in carbohydrate counting to titrate insulin dosage had better glucose control than those who received traditional diabetes nutritional education, allowing for a less restricted diet⁵. However, adjusting prandial insulin based solely on carbohydrate estimation presents significant challenges due to potential errors arising primarily from identifying meal ingredients, concerns related to hypoglycemia, and unaccounted variations in insulin sensitivity⁶. The observed inaccuracies in estimating carbohydrate intake⁷ presents challenges in managing postprandial blood glucose levels effectively and hinders the ability to make well-informed decisions to achieve optimal postprandial control⁸.

Patient-oriented Decision Support Systems (DSS) are dedicated smartphone applications designed to help patients make informed and clinically relevant decisions about their medical condition(s)^9,10 These are designed to provide user-specific, relevant information in lay language, prioritized and presented to patients, caregivers, and clinicians at the right time to enhance health and healthcare¹¹. DSSs in the context of type 1 diabetes (T1D) hold promise as a cost-effective alternative for optimizing glucose control among individuals using Multiple Daily Injections (MDI) or Continuous Subcutaneous Insulin Infusion (CSII) when compared to automated insulin delivery (AID). However, clinically tested DSSs aiming at improving glucose control in T1D have only demonstrated marginal benefits to date¹².

In precision medicine, a digital twin (DT), also referred to as a medical digital twin (MDT), features a virtual replica of a person or part thereof that integrates multiscale and multimodal data to simulate and predict disease progression and therapeutic response¹³. DTs are often encapsulated in (i) non-interpretable machine learning architectures^14,15, (ii) integrated multi-physics, multiscale, probabilistic computer models^16–18, or a combination of the above^19,20. In the context of diabetes, DTs have the potential to modify therapy in real time and anticipate undesired events such as hypo- and hyperglycemia and diabetes ketoacidosis (DKA) in response to physical activity and meal states when integrated into decision support systems²¹. Additionally, they can provide clinicians with insights regarding both the short-term and long-term implications of a specific therapeutic approach in an individual²². However, model personalization is challenging due to the need for massive datasets (data-driven models) and reliable identification procedures (physiology-based models).

We conducted a parallel two-arm, randomized controlled clinical trial to assess a novel digital twin-enhanced simulation-assisted prandial bolus calculator over time in the target range of 70–180 mg/dL (or 3.9–10.0 mmol/L) in individuals with Type 1 Diabetes (T1D). This decision support system was integrated into a user-friendly Android application called STUDIA. The app utilizes a digital twin model of the user to deliver real-time predictions of postprandial glucose levels, taking into account current glycemic levels, carbohydrate estimates, and patient-specific therapy parameters. This intervention serves as a verification step before prandial insulin administration in individuals with T1D who utilize continuous glucose monitors (CGM) in conjunction with either multiple daily injections (MDI) or continuous subcutaneous insulin infusion (CSII). We hypothesized that this approach would facilitate behavior modification, leading to healthier food selection and optimized prandial bolus strategies among participants compared to controls.

Result

Participant characteristics

Between May and November 2023, we screened 68 individuals with type 1 diabetes (T1D), from whom 44 were excluded (see Fig. 1). Of the 28 participants enrolled, 14 were randomized to a digital twin-enabled simulation-assisted decision support system (SA-DSS) and 14 to the standard of care (bolus calculator, BC) control group. The mean age at enrollment was 37.4 years (range 18–68); 14 participants (50%) were female. Hemoglobin A1c was overall 7.28 ± 0.94% (range 5.4–9.0.4.0). Only three participants were on continuous subcutaneous insulin infusion (CSII, 11%). Two participants self-identified as Afro-Colombian, while the remaining participants did not report affiliation with any specific ethnic category. Other demographic and clinical characteristics are presented in Table 1.

Fig. 1.

CONSORT diagram of participant flow in the STUDIA trial. This figure presents a consort diagram detailing the participant flow including screening, randomization, allocation, follow-up, and analysis.

Table 1.

Participants characteristics.

	Control (N = 14)	STUDIA (N = 14)
Age (years)	26 (19.5)	39.5 (22.5)
Sex
Female	8 (57.1)	6 (42.9)
Male	6 (42.9)	8 (57.1)
Time in range	59.9 ± 21.0	63.4 ± 24.0
Time counting carbohydrates
Less than a year	1 (7.1)	2 (14.3)
More than a year	13 (92.9)	12 (85.7)
Education
Secondary	5 (35.7)	3 (21.4)
Post-Secondary	9 (64.3)	11 (78.6)
Insulin pump users	2 (14.28)	1 (7.14)
Hemoglobin A1c (%)	7.29 ± 1.00	7.27 ± 0.873
Coefficient of variation	36.4 (32.8, 40.2)	33.6 (28.4, 36.8)
DTSQ	28.5 (26.3, 29.8)	28.5 (23.3, 31.5)

Data are mean ± SD, n (%), n/N (%), or median (IQR). DTSQ Diabetes Treatment Satisfaction Questionnaire.

Effect of SA-DSS on time in the target range (TIR)

From baseline to week four, the percentage of time spent within the target glucose range (TIR, 70–180 mg/dL) in the SA-DSS arm increased from 59.2 ± 21.9% to 67.1 ± 16.3% as compared with those in the BC arm, whose TIR decreased from 63.4 ± 23.7% to 60.4 ± 24.4%. The crude mean difference between groups was 7.44% (CI95% 3.20 to 12.61), while the adjusted mean difference was 6.95% (CI95% 3.51 to 10.39; p < 0.001). Figure 2 (panel A) presents the evolution of TIR in both arms for the main portion of the study.

Fig. 2.

Panel A: Changes in time in range (TIR). Panel B: Changes in time below range (TBR). Over 28 days of follow-up for participants using a simulation-assisted decision-support system (SA-DSS) or a traditional bolus calculator (BC). Each point represents an individual observation, with slight horizontal jitter applied for visualization purposes. Lines represent LOESS-smoothed trends for each group over time, and shaded areas indicate the 95% confidence intervals around the estimated curves.

Effect of SA-DSS on time below range (TBR)

The percentage of readings below 70 mg/dL (Time Below Range - TBR) was inferior in the SA-DSS arm compared to the BC arm with a crude difference of 1.53% (CI95% −2.59 to −0.47) and an adjusted difference of 1.11% (CI95% −2.20 to −0.20; p = 0.02). Figure 2 (panel B) illustrates the progression of TBR in both groups throughout the study. There were no statistically significant differences in other CGM-based metrics. Primary and secondary outcomes are summarized in Table 2.

Table 2.

Primary and secondary outcomes.

Outcome	Control (N = 14)	STUDIA (N = 14)	Estimated adjusted difference (95%CI)	p value
Change in TIR	−3.71 ± 12	7.93 ± 11.3	6.95 (3.51 to 10.39)	< 0.001
Change in TBR level 1	1.93 ± 2.89	−1.64 ± 2.10	−1.10 (−2.20 to −0.20)	0.022
Change in TBR level 2	0.71 ± 0.80	−0.28 ± 0.56	−0.11 (−0.39 to 0.17	0.422
Change in TAR level 1	1.43 ± 13.3	−3.93 ± 13.6	−3.21 (−7.67 to 1.25)	0.396
Change in TAR level 2	2.61 ± 2.5	1.78 ± 1.68	−1.01 (−3.03 to 1.0)	0.303
Change in CV	−1.09 ± 3.4	0.0357 ± 7.7	0.38 (−2.41 to 3.17)	0.787

Adjusted differences were estimated using an adjusted mixed-effects model. TIR Time in range, TBR Time below range, TAR Time above range, CV Coefficient of variation. Levels 1 and 2 for TBR and TAR are defined in⁵⁷.

Effect of SA-DSS on safety and adverse events (AE)

Eight participants (six in the BC arm) experienced a total of 20 hypoglycemic events. Among these, 15 hypoglycemic events were reported in the BC arm and five in the SA-DSS arm. The incidence rate of hypoglycemia events in the SA-DSS arm was 0.31 (95% CI 0.10 to 0.81; p = 0.025). In terms of severe AEs, two level 3 hypoglycemic events were reported in participants in the BC arm. No episodes of diabetes ketoacidosis (DKA) were reported during the study. Table 3 summarizes the safety outcomes of the study.

Table 3.

Safety outcomes.

Outcome	Control (N = 14)	STUDIA (N = 14)	Incidence rate (95%CI)	p Value
Any hypoglycemia			0.31 (0.10 to 0.81)	0.025
No of events	15	5	-	-
No of participants (%)	6 (42.8)	2 (14.3)	-	-
No of events			-	-
Level 3 hypoglycemia	2	0	-	-
No of participants (%)	2 (14.3)	0 (0)	-	-

Glucose level predictions

A digital twin engine was carefully integrated into the app’s backend. Every time a participant used the bolus calculator, a simulation was performed to predict postprandial glucose levels over four hours. This simulation was presented to the participant in the SA-DSS arm while it remained concealed for the BC arm. Nonetheless, all outcomes were securely stored in the database for subsequent review and analysis.

We compared the predictive ability of our digital twin module with the collected CGM data, using data only from the control group, as most individuals in the SA-DSS group were expected to adjust their behavior based on the prediction. Table 4 summarizes the performance of the digital twin model with prediction horizons of 60, 120, 180, and 240 min. The assessment focuses on the Root Mean Square Error (RMSE), accumulated Root Mean Square Error (aRMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) metrics, which are commonly used in literature. Figures S1-S3 compare the model predictions to their respective FGM traces for three participants in the control group, showcasing the best, worst, and average results. Figure 3 displays a revised color-coded surveillance error grid for all prediction horizons, while Fig. 4; Table 5 illustrates the prediction errors over time. Figure 5 presents the aRMSE for all prediction horizons by participants in the control group. Of the differences observed between the two methods, 72% (95% CI, 70% to 73%) were below 50 mg/dL (refer to Figure S4 in the Supplementary Material).

Table 4.

Results of glucose level prediction (Mean ± SD) assessed in participants on the BC arm.

Prediction horizons	60 min	120 min	180 min	240 min
RMSE (mg/dL)	42.4 ± 17.8	40.1 ± 17.1	44.5 ± 17.7	53.9 ± 26.8
aRMSE (mg/dL)	23.9 ± 10.5	36.7 ± 14.8	39 ± 14.2	42.6 ± 14.1
MAE (mg/dL)	33.6 ± 17.0	32.8 ± 14.4	36.5 ± 15.3	41.9 ± 23.7
MAPE (%)	19.2 ± 6.7	21 ± 9.5	29.0 ± 16.0	34.2 ± 20.8

Fig. 3.

Surveillance error grid for predicted glucose levels vs. actual CGM values. Panels A–D reflect the different prediction horizons.

Fig. 4.

Four-hour prediction error of the DT engine in the BC arm. The dotted line corresponds to the zero error axis. The continuous blue line and shaded areas represent the median error, interquartile range (IQR), and the 5–95 percentile range.

Table 5.

Surveillance error grid.

SEG risk level	SEG risk category		Percent 60-minutes	Percent 120-minutes	Percent 180-minutes	Percent 240-minutes
0	None		61.9	57.5	50.6	55.6
1	Slight, Lower		29.4	33.4	31.6	25.6
2	Slight, Higher		8.4	6.9	9.7	10
3	Moderate, Lower		0.3	1.9	6.2	5.6
4	Moderate, Higher		0	0.3	1.9	2.8
5	Severe, Lower		0	0	0	0.3
6	Severe, Upper		0	0	0	0
7	Extreme		0	0	0	0

Fig. 5.

Accumulated RMSE from when the prediction was made until 60 (A), 120 (B), 180 (C), and 240 (D) minutes ahead.

Usability and patient-reported outcomes (PROs)

The compliance rate, calculated as the number of completed simulations divided by the recommended simulations, was 57.5% (IQR 16.8 to 80.3) in the control group and 56.5% (IQR 26.3 to 75.0) in the intervention group. The system usability score (SUS) was 101 ± 6.9 in the control group and 101 ± 10.8 in the intervention group. The Diabetes Satisfaction and Treatment Questionnaire (DSTQ) change scores were 10.0 (IQR 9.25, 13.5) and 11.5 (IQR 8.5, 14.8) in the control and intervention groups, respectively. Among the individuals assigned to SA-DSS, 80% of the simulations resulted in some behavioral change, the most common being a change in insulin dosage (72%).

Discussion

In this randomized, controlled trial in people with T1D who have previous experience in carbohydrate counting, the change in the CGM-measured percentage of time spent in the target range (70–180 mg/dL) from baseline to the end of the observation period was 7% points higher among participants using the SA-DSS than among those with BC, an increase of 1.7 h per day. Notably, most improvements in TIR were observed within the first two weeks of follow-up, as illustrated in Fig. 2 (panel A). If sustained in time, this could lead to a reduction of approximately 0.7% in HbA1c over three months²³. Participants in the SA-DSS arm also reduced the time spent in hypoglycemia by 1.11% (with a 70% reduction in hypoglycemia-perceived events). These changes became more pronounced toward the end of the study, as depicted in Fig. 2 (panel B).

The results of this study indicate that the system provides feasible and potentially informative real-time forecasts of the post-prandial state, which may support patient engagement and encourage more thoughtful decision-making. These forecasts created opportunities to fine-adjust prandial insulin and/or carbohydrate intake, which was reflected in a significant increase in TIR coupled with a decrease in TBR, without increasing exposure to hyperglycemia. While prediction errors were not negligible, the observed clinical benefits suggest that this type of intervention remains promising and warrants further investigation²⁴.

The STUDIA app in the SA-DSS arm uses a digital twin of the user, i.e., a first-principles, highly detailed physiological model personalized through four interpretable parameters^25,26to assist in the meal-related decision-making process. Our system compares favorably to other decision-support systems tested in T1D, acknowledging the potential differences in design and participants’ conditions in each study^27,28. To our knowledge, this is the first free-living study to demonstrate a clinically significant increase (> 5%) in time spent within the target range (70–180 mg/dL) using a decision-support system in type 1 diabetes under MDI or CSII.

Earlier modern DSSs leveraged Bluetooth connectivity to BG meters, cloud computing, and logging capabilities to track glucose control and provide titration recommendations^27,28. Subsequently, CGM-enhanced Decision Support Systems (DSSs) improved monitoring capabilities and provided more advanced analytics to recommend personalized insulin titration parameters such as basal rate, insulin-to-carbohydrate ratio (CR), and correction factor (CF). These adaptations were implemented automatically for users of Continuous Subcutaneous Insulin Infusion (CSII) systems or manually for those employing Multiple Daily Injections (MDI)^29–34. DeBoer et al. investigated the effect of the CloudConnect DSS on improving communication between adolescents and their caregivers and, hence, diabetes management³⁵. In a recent study conducted by Colmegna et al., participants with T1D were encouraged to interact with a personalized simulation and replay engine via an interactive web simulation tool to facilitate diabetes self-management³⁶. However, as documented in the literature, the impact of the above interventions on key diabetes care outcomes has yielded limited efficacy¹².

Although the utilization of the system by participants in the SA-DSS arm was approximately half as frequent as anticipated, a phenomenon well described in the literature³⁷, both groups reported similarly high satisfaction and usability scores. As the two versions of the application differed only in the predictive output, with identical interfaces, no significant between-group differences were expected. In fact, while the direction of change favored the SA-DSS arm, these results were not statistically significant and should be interpreted as indicative of overall high usability rather than conclusive evidence of superiority. We speculate that the high scores in both groups were driven by the simplicity of the app’s interface and the participants’ prior experience with CGM interpretation.

The retrospective analysis of model predictions showed an outstanding predictive capacity in real life and proved that highly detailed mathematical models can be used in routine clinical practice. Highly accurate and consistent mathematical models can help increase credibility, user engagement, and adherence through system trust. In the future, adding adaptation and learning capabilities to the app can help guide the user on healthy food selection and conscious insulin utilization.

The graphical analysis of model performance highlights essential aspects of its clinical applicability as depicted through Figs. 3, 4 and 5. The stability of the median prediction error (Fig. 4) over time, and the consistency of the percentile bands (Fig. 5) suggests that the model maintains reliable forecasting across the postprandial window, in contrast to models based on artificial intelligence. Furthermore, the participant-level variability observed across prediction horizons reflects the heterogeneity of individual glycemic responses, underscoring the importance of personalized approaches in decision-support systems (Fig. 5). Together, these visualizations reinforce the potential of the model to support insulin dosing decisions with a reasonable degree of precision while also unveiling areas where adaptive calibration or user-specific tuning could further enhance predictive accuracy.

Glucose prediction is challenging. Recent years have seen a growing body of research focused on predicting blood glucose levels using artificial intelligence (AI), particularly machine learning (ML) and deep learning (DL) approaches. Traditional methods such as linear regression and support vector machines have been employed to forecast short-term glucose trends, showing moderate success when provided with high-resolution data from continuous glucose monitors (CGMs)³⁸. More advanced techniques, including recurrent neural networks (RNNs) and long short-term memory (LSTM) networks, have also been proposed³⁹. Hybrid models that combine physiological modeling with AI have also shown promise, offering improved performance by integrating domain knowledge with data-driven insights⁴⁰. However, most AI-based models require significant amounts of data for training and are assessed at a single endpoint, usually within a 60-minute prediction horizon, without offering insights into the system’s evolution. In contrast, we challenged the proposed system by extending the prediction horizon to 240 min to account for most of the postprandial period.

This study is subject to some limitations. First, due to the nature of the intervention, it was not possible to implement a blind design. Additionally, while the study duration was chosen to detect changes in the time in range and hypoglycemia-related risk, four weeks is still a relatively short timeframe for evaluating long-term glycemic benefits. Longer trials are also necessary to determine the app’s retention rate. Another potential challenge is the high variability in glucose levels within subjects. Although a crossover design could have addressed this issue, establishing an appropriate washout period for an educational intervention has proven difficult⁴¹. Using an FGM instead of an rtCGM could have hindered the assessment of critical glycemic changes not captured by the FGM data. The participants in both groups were experienced CGM users. Therefore, further studies involving CGM-naive users are necessary to ensure the generalizability of the findings. Finally, although Colombia is one of the most ethnically diverse countries in the world (due to the historical admixture among European Caucasians of Spanish descent, Indigenous peoples, and individuals of African origin), approximately 90% of the population does not self-identify with any officially recognized ethnic groups. This makes race and ethnicity characterization unsuitable to recognize the different social determinants of health from different subpopulations.

In addition, some methodological aspects warrant specific consideration. The relatively small sample size (n = 28) limits the generalizability of the findings; although statistically significant differences were observed in TIR and hypoglycemia incidence, the possibility of type I error cannot be fully excluded, despite adjustments for known baseline variables in the regression models. The open-label design also introduces the risk of a Hawthorne effect, as participants may have modified their behavior due to study participation; however, both groups received the same intensity of follow-up and observation, which may have mitigated this bias. Recruitment feasibility was another challenge: of the 68 individuals screened, 32 declined participation, primarily due to the burden of frequent study visits, an expected difficulty in early feasibility trials of novel interventions. Future studies, building on the safety profile and potential benefits observed in this study could be designed with reduced participant burden, which may improve recruitment rates and mitigate selection bias. Finally, the personalization of the digital twin was conducted heuristically and once prior to study entry by a physician with experience in the model development and in silico testing. While the interpretability of the parameters facilitated this approach, it relied on expert knowledge and was not standardized, limiting reproducibility and scalability. This pragmatic procedure was acceptable for a proof-of-concept trial but underscores the need for algorithmic or semi-automated calibration strategies in future implementations. Taking together, these considerations indicate that our trial should be interpreted as a proof-of-concept and feasibility study, rather than a definitive efficacy evaluation.

An additional direction for future work is the refinement of nutrient estimation within the digital twin. In this study, protein and fat contributions were incorporated as fixed components derived from prior experimental work and in silico analyses, since practical tools for patient-level estimation of these macronutrients remain underdeveloped. However, emerging technologies based on artificial intelligence and volumetric image recognition already allow for semi-automated meal analysis. Integrating such tools into the STUDIA system could enable real-time, user-friendly estimation of carbohydrate, protein, and fat intake, thereby improving the accuracy of postprandial glucose forecasts and further reducing the burden of carbohydrate counting.

In summary, feasible post-prandial glucose forecasting appears safe and informative as a supportive tool for prandial bolus decision-making and may positively influence glucose control in individuals with T1D under MDI or CSII. Further studies are warranted to optimize predictive accuracy and confirm long-term clinical benefits.

Methods

Study design and participants

The STUDIA trial evaluated whether an advanced meal bolusing approach —a simulation-assisted decision support system (SA-DSS arm) would be superior in TIR to the current best practice, bolus calculation with carbohydrate counting (BC arm). The study was a four-week, single-center, parallel (two-arm), randomized, controlled, prospective trial that evaluated the superiority of SA-DSS over BC on the change in the percentage of time in the target range (TIR, 70–180 mg/dL, 3.9–10 mmol/L) from baseline to the end of the study.

The study was conducted at the Pablo Tobon Uribe Hospital (PTUH), a tertiary referral center in Medellin, Colombia, following the principles of the Colombian medical research regulations and the Declaration of Helsinki. The study protocol⁴² received approval from the institutional ethics committee under reference number 022021 and was registered on clinicaltrials.gov (NCT05181917 - First Posted date 10/01/2022). All participants provided written informed consent.

Participants were referred to by three university hospitals and ambulatory clinics dedicated to treating people with diabetes. Major eligibility criteria included a diagnosis of T1D with insulin treatment for at least one year, use of continuous subcutaneous insulin infusion (CSII) or multiple daily injections (MDI) for intensive insulin therapy (IIT), prior training in carbohydrate counting and CGM use, HbA1c levels not exceeding 10% in the three months preceding randomization, age of 18 years or older, and being deemed suitable by the investigator to use the app. Key exclusion criteria included pregnancy or the intention to become pregnant during the study period, a possible or confirmed diabetic gastroparesis, chronic renal disease, an estimated filtration rate of 65 ml/min/1.73m2 or less using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) equation, and receiving any glucose-lowering medication other than insulin.

Randomization and group allocation

Following confirmation of eligibility, participants were randomized to the SA-DSS intervention or control condition in a 1:1 ratio using a sequence of computer-generated random numbers via an online randomization service (Sealed Envelope Ltd, London, UK). The sequence was concealed from the investigators. Due to the nature of the investigation, neither participants nor investigators were blinded. Instead, the data analysts were masked.

Admissions

At the screening visit, baseline characteristics were collected, and eligibility was assessed. The study consisted of a two-week run-in period, during which participants continued their usual insulin treatment and used a flash glucose monitor (FreeStyle Libre; Abbott Diabetes Care, Alameda, CA, USA). After the run-in period, all participants were trained on using the STUDIA app and randomized to either the SA-DSS or BC arms for the main four-week protocol phase. During the main study phase, each participant received a weekly follow-up phone call from a study staff member, during which the data was revised and uploaded to a secure server. The study member also discussed continuous monitoring data, inquired about adherence to the app, and addressed any questions regarding the study app and its usage. Upon completing the study, participants were asked to uninstall the app from their phones or return the study device, if applicable.

Study devices and interventions

STUDIA is a non-commercial smartphone app that implements BC and SA-DSS using a straightforward user interface. The app’s dataset was stored in a secure cloud server implemented in the Amazon Web Services (AWS) environment. The study PI performed data quality checks and followed the participants using a web-based platform. The system architecture and components used in the study are shown in Fig. 6. Participants wore flash glucose monitors (FreeStyle Libre; Abbott Diabetes Care, Alameda, CA, USA) to track blood glucose levels.

Fig. 6.

The HPTU system architecture and components.

The STUDIA app under SA-DSS featured a digital twin (DT) of the user, composed of a set of detailed physiological models describing the rate of glucose appearance from a mixed meal²⁶, hepatic glucose production²⁵, and a glucose-insulin system model integrating the two former to produce a glucose forecast. The study physician, who had also participated in the development of the underlying physiological models, heuristically modified four parameters prior to randomization, using information collected during the run-in period: Fractional glucose effectiveness () from glucose dynamics, insulin sensitivity ( from insulin dynamics, and the time lag related to stomach emptying () and energy flow leaving the stomach () from the GI-tract model (refer to the Supplementary Material for additional information). The interpretability of these parameters, together with prior in silico testing of the model, facilitated this personalization. Adjustments were based on clinical judgment and visual inspection of glucose responses to mixed meals and were performed once before study entry. No formal optimization algorithm or standardized error metric was applied, reflecting a pragmatic proof-of-concept approach.

The input screen for the bolus calculator was identical in both arms. At every meal, participants were asked to enter their current CGM value and estimated carbohydrate intake manually, which enabled the automatic calculation of their recommended bolus based on their pre-meal glucose level, personal carbohydrate-to-insulin ratio (CR), correction factor (CF), and target glucose levels. In the SA-DSS arm, participants received a graphical representation of a four-hour post-prandial glucose prediction for the entered meal and bolus based on the user-specific DT. Simulations could be repeated with different carbohydrate estimates at the user’s discretion (Fig. 7, panel A). Participants received only the recommended bolus in the BC arm (Fig. 7, panel B), with the graphical prediction concealed. In either case, the digital twin-based four-hour post-prandial glucose prediction was sent to the server for future analyses.

Fig. 7.

App usage in the two arms. Panel A: simulation-assisted decision support system (SA-DSS). Panel B: Classic bolus calculator (BC).

In the present study, participants manually entered only their estimated carbohydrate intake, which was the sole input required from the user. Protein and fat contributions were incorporated indirectly through fixed proportions relative to carbohydrate intake. These fixed components were derived from previous experimental work and in silico analyses that preceded this trial, as practical methods for patient-level estimation of protein and fat remain insufficiently developed. No image-based food recognition tools were used in this study.

Data and preprocessing

CGM, carbohydrate estimates, and insulin records were collected during the run-in period to establish the baseline characteristics and modify the model parameters to match the postprandial responses of the personalized DT. This resulted in a time-series database for each variable. We preprocessed the data to guarantee unit homogeneity and eliminate common data artifacts, such as outliers and incomplete data entries.

Model structure

The computational model used in the DT engine consisted of two maximal models and a compartmental model integrated into a glucose dynamics equation. The two maximal models describe the macronutrient absorption through the gastrointestinal (GI) tract⁴³ and the role of the liver in glucose homeostasis²⁵, respectively; they provide a detailed description of glucose appearance in the bloodstream in response to a mixed meal and endogenous glucose production from gluconeogenesis and glycogenolysis (Supplementary Tables 1 to 9). The compartmental model describes the subcutaneous insulin delivery via a triangular insulin dynamic, demonstrating good agreement with current insulin analogs with PK/PD⁴⁴. Figure S5 shows the glucose dynamics and its integration with the complementary models.

Evaluation of the digital twin engine

Deidentified data prospectively collected were used to assess the prediction ability of the model. The data included inputted glucose levels (from the FGM reading) and carbohydrate intake and the returned insulin boluses and model predictions by the app. Only data from participants within the control arm were used for evaluation, since participants within the intervention arm were likely to change their behavior after using the app. Data was preprocessed as in the data collection period, with no excluded CGM per initiation day or any adjustment for bias in nutrient intake⁴⁵.

We used the following key performance indicators (KPI) to evaluate the prediction performance of the proposed metabolic model:

1

2

3

4

where , , , and stand for root mean square error, mean absolute error, mean absolute percentage error, and accumulated root mean square error, respectively, and are the actual FGM value and model estimate at the defined end-point (), respectively, with and the number of evaluated predictions. Equations (1)-(3) are end-point metrics, meaning they assess the metric only at the given endpoint (). Equation (4) evaluates the RMSE within the entire prediction horizon () to account for the physiological trajectory to reach the endpoint. is the number of samples within the with the sampling time.

Outcomes and statistical analysis

The primary outcome of the study was the change from baseline to the end of the main portion of the study in the time in the target range (70–180 mg/dL or 3.9–10.9.0mmol/L) as measured by an FGM (Libre; Abbott Diabetes Care, Alameda, CA). Secondary outcomes included FGM-measured time below the range level 1 (TBR, < 70 mg/dL or 3.9mmol/L), time below the range level 2 (TBR, < 54 mg/dL or 3.0mmol/L), time above the range level 1 (TAR, > 180 mg/dL or 10.0mmol/L), time above the range level 2 (TAR, > 250 mg/dL or 13.9mmol/L), and coefficient of variation.

Safety outcomes included hyperglycemic crises and hypoglycemia events, as defined by the American Diabetes Association⁴⁶. The Spanish version of the Diabetes Treatment Satisfaction Questionnaire was administered to all participants at randomization and the end of the study.

Exploratory outcomes included the agreement between the four-hour glucose prediction (DT engine) and the corresponding FGM data and the system’s usability scale⁴⁷. Usability was assessed using the Spanish adaptation of the Computer Systems Usability Questionnaire (CSUQ), which ranges from 19 to 133 points, with higher scores indicating greater perceived usability. Additionally, a decision assessment was conducted after each use of the bolus calculator to identify any changes in insulin dosage, modifications in planned meal content, or a lack of action.

We hypothesized that participants in the SA-DSS arm would have a superior change from baseline to the end of the main study phase in the percentage of sensor readings within the target range (70–180 mg/dL) compared to participants in the BC arm. We estimated that 28 participants provide 90% power to detect a difference in TIR of 3.75% in the four-week intervention, assuming a standard deviation of 4%, a type I error rate of 0.05, and intra- and inter-subject variability of 5% and 9%, respectively⁴⁸. A sample size adjustment to accommodate up to 10% losses to follow-up was planned^48–50, but it was unnecessary.

Statistical analyses were performed according to intention-to-treat (ITT) principles. Mean ± SD and median (IQR) are reported for primary and secondary endpoints for normal/near-normal and skewed distributions, respectively. For the primary analysis, the change in the TIR from baseline to the end of the four-week observation period was compared between groups using a linear mixed-effects regression model while adjusting for pre-randomization TIR. We explored additional adjustment by type of insulin therapy (MDI vs. CSII), but given the small number of CSII users and the absence of any impact on the results, this variable was not included in the final model to avoid overfitting. No participants were on semi-automated or fully autonomous insulin delivery. The episodes of hypoglycemia and the number of hyperglycemic crises were compared using a ratio of incidence rates calculated using a Poisson regression^51,52. For the agreement analysis, since the differences between methods did not follow a normal distribution, a non-parametric approach was used to estimate the limits of agreement⁵³. It is essential to note that, to the best of our knowledge, there has been no thorough discussion on assessing the accuracy of glucose predictions for real-time insulin dosing in T1D. We selected an arbitrary threshold of 50 mg/dL based on an idealized 180 mg/dL peak in the postprandial state⁵⁴. The potential clinical impact of the differences was evaluated using a surveillance error grid analysis⁵⁵. No adjustment was made for multiple testing. All analyses were conducted using the R statistical package⁵⁶. Some plots were generated in Matlab.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Carlos E. Builes Montaño, John J. Zuleta, and Jose Garcia-Tirado participated in the study’s concept and design. Carlos E. Builes-Montaño and Jose Garcia-Tirado in the writing and editing of the manuscript. Hernan Alvarez, Juan C. Restrepo-Gutierrez, Laura Lema-Perez, and Alex Ramirez-Rincón edited the manuscript’s final version and provided scientific and clinical support. John J. Zuleta and Carlos E. Builes Montaño designed the statistical plan. All authors contributed to the article and approved the submitted version.

Data availability

The datasets generated and analyzed during the current study are not publicly available due to restrictions from the sponsoring institution. However, they are available from the corresponding author upon reasonable request.

Declarations

Competing interests

The authors certify that they have no affiliation or are involved with any organization or entity with any financial interest (such as fees, financial aid for education, shares, employment contracts, work as consultants, or any other type of interest) or non-financial interest (such as personal, professional relationships, affiliations, or beliefs) in the topic of interest or any material discussed in this manuscript. Carlos E. Builes-Montaño has received consulting or speaker fees from Sanofi, Novo Nordisk, Novartis, and Boehringer Ingelheim. Jose Garcia-Tirado reports receiving royalties from Dexcom. All the other authors report no potential conflict of interest.