Genetic algorithm tuning of artificial pancreas MPC with individualized models

Diabetes is a growing chronic disease and a worldwide problem. Without any available cure in sight for the public other methods needs to be applied to increase the life quality of diabetic patients. Artificial Pancreas (AP), a concept of having a closed loop system to control the glucose level on Type 1 Diabetes (T1D) patients has been introduced and is under development. In this thesis, Model Predictive Control (MPC) has been re implemented from scratch in MATLAB/SIMULINK with associated Kalman filter and prediction function. It was implemented in the latest version of the UVA/Padova Simulator which is a tool approved by FDA for simulating diabetes treatment in order to speed up the AP development. Different MPC cost functions where tested together with integral action on a simplified system using a linear approximation of a population model. It was implemented and tuned with a new simulation tuning method using Genetic Algorithm (GA). It showed that the quadratic cost function without integral action was the best with respect to performance and time efficiency. 3 hours was the best prediction horizon and was used for the individualized tuning using the University of Virginia (UVA)/Padova simulator. For the individualized MPC, models identified by the University of Padova were used. These simulations showed that an individualized model could be used for improved T1D treatment compared to an average population model even though the results were mixed. Almost all of the patients got improved treatment with the closed treatment and non hypoglycemic event occurred. The identification of better models is a great challenge for the future development of the AP MPC due to the excitation problems