Risk-Aware Control for Insulin Delivery via Nonlinear MPC with Safety Barrier Functions and Probabilistic Learning of Uncertainties

Abstract

Maintaining blood glucose within a physiologically safe range is critical for people with diabetes, as deviations can lead to acute or chronic complications. Hypoglycemia, in particular, represents an immediate threat and requires prioritized mitigation in autonomous insulin delivery systems. This paper introduces a risk-aware hybrid nonlinear model predictive control (NMPC) framework that combines data-driven uncertainty quantification with formal safety assurance through control barrier functions (CBFs). To account for key uncertainties, such as physiological time delays, unannounced meals, and stress-induced glucose variability, Gaussian processes (GPs) are employed as probabilistic estimators. The proposed method dynamically monitors glucose and regulates insulin injection to enforce safe glucose level control by preventing hypoglycemia. The proposed framework is evaluated using validated physiological simulators for various realistic scenarios. The results show a robust performance in maintaining safety under high uncertainty, preparing a foundation for translation into next phase of our research as safe autonomous diabetes management systems.