ARTIFICIAL PANCREAS: AN IN SILICO STUDY OF ADVANCED MODEL PREDICTIVE CONTROL STRATEGIES

The International Diabetes Federation estimates that, nowadays, more than 537 million people in the world have diabetes, a chronic disease that has spread rapidly in recent decades. Future forecasts are not promising: the number is set to grow to 643 million by 2030. Diabetic patients must necessarily constantly pay attention to the diet and signals of their own body through measurements, which can be invasive by imposing important limitations in the quality of life. All these reasons have led research to introduce a device capable of administering a quantity of insulin according to the level of plasma blood glucose, to keep it as close as possible to the optimal one. This device, called Artificial Pancreas (AP), would substantially improve the quality and life expectancy of a diabetic patient. The main objective of this thesis is to evaluate algorithms presented in the literature on a in silico population never previously tested and propose improvements for each of them. The new population, released in 2017, is composed of 100 adult patients who have totally different biological parameters compared to those considered in the previous studies. In order to deal with the inter- and intra-variability of patients, the objective function was individualized by applying a calibration procedure, based on a performance index, present in the literature and then a new one based on clinical parameters. Using the UVA/Padova metabolic simulator, simulations were carried out, both in a nominal scenario and in robustness tests. The control laws applied are based on different MPC techniques present in the literature: without constraints (UMPC), with constraints (CMPC) and with a posteriori saturation constraints (SMPC). In addition, a new evaluation index has been introduced, which allows to evaluate the best strategy through a compact value summarizing the statistical analysis indices. Comparing the control performance obtained by the various MPC techniques, it has been shown that CMPC is the best control strategy in terms of blood glucose level in the euglycemic target, at the cost of some case of patient in hypoglycemia, which is however negligible. Finally, the Impulsive Response (IR) identification technique was implemented to adapt the model to a specific patient, obtaining satisfactory results, both in terms of prediction capabilities and when the model is used in the CMPC. The results obtained open to interesting future developments that would allow to a significant progress on the Artificial Pancreas research.