End-to-end offline reinforcement learning for glycemia control.

The development of closed-loop systems for glycemia control in type I diabetes relies heavily on simulated patients. Improving the performances and adaptability of these close-loops raises the risk of over-fitting the simulator. This may have dire consequences, especially in unusual cases which were not faithfully - if at all - captured by the simulator. To address this, we propose to use model-free offline RL agents, trained on real patient data, to perform the glycemia control. To further improve the performances, we propose an end-to-end personalization pipeline, which leverages offline-policy evaluation methods to remove altogether the need of a simulator, while still enabling an estimation of clinically relevant metrics for diabetes.

Insulin Requirements According to Sex and Weight in a Population of 9036 Adult Persons With Type 1 Diabetes Using Closed-Loop Insulin Delivery.

BACKGROUND: The prediction of the individual insulin needs may facilitate the initiation of insulin therapy. Our aim was to explore the relationships between body weight, sex, and daily amounts of insulin delivered by a hybrid closed-loop system. METHODS: We performed a retrospective data collection of all consenting adult patients with type 1 diabetes who were equipped in Europe with the Diabeloop Generation 1 (DBLG1) hybrid closed-loop insulin delivery device between March 1, 2021 and February 28, 2023. RESULTS: A total of 9036 users (59% females, age 45.6 ± 14.3 years) were included, reaching a mean follow-up of 320 ± 143 days, an overall 2 887 188 days of data. We observed a mean insulin-weight ratio of 0.617 ± 0.207 U/kg (0.665 ± 0.217 for males and 0.584 ± 0.193 for females, P < .001). Exploratory analysis of a subset of 4066 patients reaching >70% Time in Range (70-180 mg/dL) showed a mean insulin-weight ratio of 0.55 ± 0.17 U/kg (P < .001) (0.59 ± 0.18 for the 1438 males and 0.53 ± 0.16 for the 2628 females). CONCLUSION: This large real-world analysis provides a quantitative estimation of the daily insulin requirements in adult patients with type 1 diabetes and shows significant differences between sex. These findings have relevant implications in the practical management of insulin therapy.

First Generation of a Modular Interoperable Closed-Loop System for Automated Insulin Delivery in Patients With Type 1 Diabetes: Lessons From Trials and Real-Life Data.

BACKGROUND: DBLG1 (Diabeloop Generation 1) stands as one of the five commercially available closed-loop solution worldwide for patients with type 1 diabetes as of 2023. Our aim was to provide an overview of all data obtained with this system regarding outcomes and populations, with an emphasis on interoperability. METHODS: This report includes all available sources of data (three randomized control trials and five surveys on real-life data). Collection ran from March 3, 2017 to April 30, 2022. RESULTS: We gathered data from 6859 adult patients treated with closed-loop from three to 12 months. Overall, all sources of data showed that time in range (TIR) 70 to 180 mg/dL, starting from 47.4% to 56.6%, improved from 12.2 to 17.3 percentage points. Time in hypoglycemia was reduced by 48% in average (range: 26%-70%) and reached a level of 1.3% in the largest and most recent cohort. In patients with excessive time in hypoglycemia at baseline (≥5%), closed-loop allowed a reduction in time below range (TBR) by 59%. The comparison of days with declared physical activity versus days without physical activity did not show differences in TBR. The improvement in TIR observed with three different pump systems (Vicentra Kaleido, n = 117; Sooil Dana-I, n = 84; and Roche Insight, n = 6684) ranged from 15.4 to 17.3 percentage points. DISCUSSION: These data obtained in different European countries were consistent throughout all reports, showing that this closed-loop system is efficient (high improvement in TIR), safe (remarkably low level of TBR), and interoperable (three pump settings so far).

One-year real-world performance of the DBLG1 closed-loop system: Data from 3706 adult users with type 1 diabetes in Germany.

AIM: The Diabeloop Generation 1 (DBLG1) system is an interoperable hybrid closed-loop solution that was commercialized in Germany in March 2021. We report the longitudinal glycaemic outcomes among the first 3706 users in a real-world setting. METHODS: We performed a retrospective data collection of all consenting adult patients with type 1 diabetes who were equipped in Germany with the DBLG1 system before 30 April 2022, and with a minimum 14 days of closed-loop usage. RESULTS: In total, 3706 users (41% women, age 45.1 ± 14.5 years) met the inclusion criteria, reaching a mean follow-up of 131.0 ± 85.1 days, an overall 485 600 days of continuous glucose monitoring data, and a median time spent in closed-loop of 95.0% (IQR 89.1-97.4). The median percentage time in range (70-180 mg/dl) was 72.1% (IQR 65.0-78.9); the time below 70 mg/dl was 0.9% (0.5-1.7), the time below 54 mg/dl was 0.1% (0.1-0.3), and the median Glucose Management Index was 7.0% (6.8-7.3). Exploratory analysis of a subset of 2460 patients in whom baseline glycated haemoglobin (HbA1c) was available [7.4% (IQR 6.9-8.0)] showed that the achieved mean time in range was influenced by baseline HbA1c, ranging from 65.8 ± 9.9% (A1c ≥8.5%) to 81.3 ± 6.8% (A1c <6.5%). CONCLUSION: This large real-world analysis confirms the relevance of the DBLG1 automated insulin delivery solution for the achievement of standards of care in adult patients with type 1 diabetes.

Efficacy of a Hybrid Closed-Loop Solution in Patients With Excessive Time in Hypoglycaemia: A Post Hoc Analysis of Trials With DBLG1 System.

BACKGROUND: Automated insulin delivery is an efficient treatment for patients with type 1 diabetes. Little is known on its impact on patients with excessive time in hypoglycaemia. METHODS: We performed a post hoc analysis of three randomized control trials that used the DBLG1 (Diabeloop Generation 1) hybrid closed-loop solution. Patients whose time below 70 mg/dL during baseline, open-loop phase exceeded 5% were selected. The outcomes were the differences between the closed-loop and the open-loop phases in time in various ranges and Glycemia Risk Index (GRI). RESULTS: We identified 45 patients exhibiting ≥5% of time below 70 mg/dL during the open-loop phase. Under closed-loop, the time in hypoglycaemia (54 to <70 mg/dL) dropped from 7.9% (SD 2.4) to 3.2% (SD 1.6) (difference -4.7% [-5.3; -4.1], P < 10-4). The time below 54 mg/dL decreased from 1.9% (SD 1.3) to 0.8% (SD 0.7) (difference -0.9% [-1.4; -0.8], P < 10-4). The time in range (TIR 70-180 mg/dL) improved from 63.3 (SD 9.5) to 68.2% (SD 8.2) (difference 5.1% [2.9; 7.0], P < 10-4). The GRI improved from 51.2 (SD 12.4) to 38.0 (SD 10.9) (difference 13.2 [10.4; 16.0], P < 10-4). CONCLUSION: DBLG1 decreased time in hypoglycaemia by more than 50% even in patients with excessive time in hypoglycaemia at baseline, while also improving both TIR and GRI, under real-life conditions. The improvement in GRI (13.2%) exceeded that of the improvement in TIR (5.1%) indicating that in this data set, GRI was more sensitive than TIR to the improvement in glycaemia achieved with closed-loop. These results support the safety and efficacy of this treatment.