Integrating MPC with Learning-Based and Adaptive Methods to Enhance Safety, Performance and Reliability in Automated Insulin Delivery

Abstract In this work, an adaptive-learning model predictive control (AL-MPC) framework that integrates disturbance forecasting, uncertainty quantification, learning, and recursive subspace identification is developed. The proposed technique can be used for continuous systems affected by repetitive disturbances with unknown periods. The AL-MPC integrates online learning from historical data to anticipate impending disturbances and proactively counteract their effects to an adaptive MPC. This is done by using machine learning to quantify the significant disturbances from historical data and forecast their future evolution time series. Behavior patterns of the system can be identified from historical data, and the set-point, objective function weights, and constraints of the controller can be modified in advance for the anticipated time periods of the disturbance effects. AL-MPC is used to regulate glucose concentration (GC) in people with diabetes by automated insulin delivery. Simulation results indicate that the AL-MPC can regulate the BGC 75.4% of the simulation time in the target range (70-180) mg/dL without causing any hypoglycemia and hyperglycemia events.