Advanced Hybrid Automated Insulin Delivery System based on Successive Linearization Model Predictive Control: The UniBE System

TL;DR

This paper introduces UniBE, a hybrid automated insulin delivery system for type 1 diabetes that uses successive linearization model predictive control to cope with nonlinear glucose–insulin dynamics and physiological variability. It integrates basal modulation with an insulin bolus delivery module and personalizes the Hovorka-based glucose model via EKF and sensitivity analysis. In-silico evaluation using the UVa/Padova simulator across nine perturbation scenarios shows high time-in-range (roughly 75–93%), low hypoglycemia, and robust performance under persistent and time-varying errors, establishing proof-of-concept for clinical translation. The work highlights adaptive constraint handling and meals-related boluses in a clinically relevant hAID framework, supporting future preclinical/clinical validation.

Abstract

Background and objective: Hybrid automated insulin delivery (hAID) systems represent the most advanced therapy for type 1 diabetes (T1D). Current systems rely on linear or linearized models of glucose homeostasis, which may compromise prediction accuracy and, in turn, timely decision-making by the controller. Physiological variability further complicates insulin requirements, underscoring the need for controllers that adapt dynamically and reduce user burden. Methods: We introduce the University of Bern (UniBE) hAID system, a framework based on successive linearization model predictive control (MPC). The controller integrates basal insulin infusion with the insulin bolus delivery module for meal-related and corrective bolus dosing, adapting bounds in real time to glucose dynamics while accounting for both automated and user-initiated inputs. In-silico evaluation was conducted using the commercial version of the FDA-accepted UVa/Padova metabolic simulator across nine scenarios involving persistent and time-varying errors in meal timing, carbohydrate estimation, and basal insulin profiles. Results: In the baseline scenario, UniBE achieved a mean time in range of 92.0+-13.2%, with time below range at 0.1+-0.2% and time above range at 7.9+-13.2%. Across perturbation scenarios, time in range remained between 75.1 and 92.8%, with low hypoglycemia incidence, demonstrating resilience to clinically relevant disturbances.

Figures (2)

• Figure 1: Overview of training and test datasets. The training set is used for model parameter personalization, while the test set is used for control performance evaluation.
• Figure 2: Overview of the robustness analysis datasets. The training set is used for model parameter personalization, while the test set is used for control performance evaluation.