A nonparametric approach for model individualization in an artificial pancreas

The identification of patient-tailored linear time invariant glucose-insulin models is investigated for type 1 diabetic patients, that are characterized by a substantial inter-subject variability. The individualized linear models are identified by considering a novel kernel-based nonparametric approach and are compared with a linear time invariant average model in terms of prediction performance by means of the coefficient of determination, fit, positive and negative max errors, and root mean squared error. Model identification and validation are based on in-silico data collected from the adult virtual population of the UVA/Padova simulator. The data generation involves a protocol designed to produce a sufficient input excitation without compromising patient safety, compatible also with real life scenarios. The identified models are exploited to synthesize an individualized Model Predictive Controller (MPC) for each patient, which is used in an Artificial Pancreas to maintain the blood glucose concentration within an euglycemic range. The MPC used in several clinical studies, synthesized on the basis of a non-individualized average linear time invariant model, is also considered as reference. The closed-loop control performance is evaluated in an in-silico study on the adult virtual population of the UVA/Padova simulator in a perturbed scenario, in which the MPC is blind to random variations of insulin sensitivity in each virtual patient. © 2015, IFAC (International Federation of Automatic Control) Hosting by Elsevier Ltd. All rights reserved.