A Constrained Model Predictive Controller for an Artificial Pancreas

Abstract

A Constrained Model Predictive Control (CMPC) approach for regulating blood-glucose levels in people with type 1 Diabetes is proposed. The controller uses the past suggested insulin, the subcutaneous glucose level, and an estimation of the carbohydrate amount of the future meals provided by the patient as inputs to decide the quantity of insulin to inject by a subcutaneous pump. This strategy achieves good control performance by keeping into account a series of bounds which allow the control law to be as conservative as possible to avoid hypoglycemia phenomena without increasing the risk of hyperglycemia. The constraints definition is based on the knowledge of in-vivo clinical trials performed with an unconstrained MPC. In order to avoid the solution of the constrained optimization problem, a saturated MPC (SMPC), where all the constraints are applied as saturations, is also considered. The controller performance is evaluated in an in-silico study on 100 virtual patients of the UVA/Padova simulator. In order to underline the robustness of CMPC and SMPC in presence of model uncertainties, the simulations are performed both in nominal and in perturbed scenarios.