Modeling, identification and nonlinear model predictive control of type I diabetic patient

Abstract 

Patients with type I diabetes nearly always need therapy with insulin. The most desirable treatment would be to mimic the operation of a normal pancreas. In this work a patient affected with this pathology is modeled and identified with a neural network, and a control strategy known as Nonlinear Model Predictive Control is evaluated as an approach to command an insulin pump using the subcutaneous route. A method for dealing with the problems related with the multiple insulin injections simulation and a multilayer neural network identification of the patient model is presented. The controller performance of the proposed strategy is tested under charge and reference disturbances (setpoint). Simulating an initial blood glucose concentration of 250 mg/dl a stable value of 97.0 mg/dl was reached, with a minimum level of 76.1 mg/dl. The results of a simulated 50 g oral glucose tolerance test show a maximum glucose concentration of 142.6 mg/dl with an undershoot of 76.0 mg/dl. According to the simulation results, stable close-loop control is achieved and physiological levels are reached with reasonable delays, avoiding the undesirable low glucose levels. Further studies are needed in order to deal with noise and robustness aspects, issues which are out of the scope of this work.

Keywords: Diabetes mellitus, Glucose, Infusion pumps, Close-loop control, Model predictive control, Neural networks, System identification