Comparative Study of Two Adaptive Predictive Control Strategies to be Implemented in Artificial Pancreas Technology

This work presents a comparative study of two control algorithms based on specific characteristics of innovative adaptive predictive control. One of them consists on including in the algorithm a set of rules to choose between two alternatives: Adaptive Predictive part and a Robust Filter (APCwRF). This last version was called enhanced (E) APCwRF. The second methodology is the Predictive Functional Control (PFC) where the parameters of its internal model are continuously adapted to follow the real patient evolution by a recursive estimation based on UD factorization. This last option was called “adaptive PFC (APFC)”. The main objective of both strategies is to provide improvements to the Artificial Pancreas (AP) technology for patients with Type 1 Diabetes Mellitus (T1DM) disease. The EAPCwRF can be implemented with two alternative models the Finite Impulse Response (FIR) and the nonlinear Wiener-type block-oriented. In both cases the model parameters are updated by performing recursive identification using a Kalman Filter (KF). Hence, the control algorithm checks the quality of the prediction of the two options (FIR and Wiener) and decides which of them is more adequate to be adopted. The robust filter part is able to choose between the two models to be implemented. The tests are performed with four strategies: APCwRF, EAPCwRF, PFC and APFC together with real data of the patient without control. We used a previously developed global data-driven model (GM) for defining the final tuning (commissioning). The closed loop responses are tested with the in silico patient obtained through dynamic simulations of the GM for different meals intake and insulin dosages. The new strategies, EAPCwRF and APFC, included the insulin pump shut down when the GM predicts hypoglycemic events and exceed the recommended insulin on board (IOB) magnitude for the patient N° 5041 from the Center of Diabetes Technology (UVa/USA). The Control Variability Grid Analysis (CVGA) and the percentage of time with the patient inside the healthy range are considered to evaluate the control algorithm performances. Based on them, we present the conclusions and discuss some future works.