Adaptive Model Predictive Control for Nonlinearity in Biomedical Applications

An adaptive model predictive control (MPC) formulation is proposed in this work for optimal insulin dosing decisions in artificial pancreas (AP) systems. To this end, a recursive subspace-based system identification approach is used to characterize the transient dynamics of biological systems, specifically the metabolic processes involved in diabetes. Subsequent to system identification, an adaptive MPC algorithm is designed using the recursively identified models to effectively compute the optimal insulin delivery for AP systems. A feature extraction method based on glucose measurements is used to detect rapid deviations from the desired set-point caused by significant disturbances and subsequently modify the constraints of the optimization problem for negotiating between the aggressiveness and robustness of the controller to suggest the required amount of insulin. The efficacy of the proposed adaptive MPC is demonstrated using simulation case studies.