Developed by: A. Santopaolo (2019).
Supervisor: prof. D. Iacoviello.
Achievement: Optimal Control exam.
Artificial Pancreas was implemented using an ANN that maintains the glucose levels of patients with type 1 diabetes in an optimal range. The neural network takes as input many parameters of the patient's Glucose-Insulin system and calculates the controller output (insulin flow rate) based only on the current values of the Glucose-Insulin system. In other words, the neural network does not take into account the history of the parameters of the patients. First an accurate representation of the Glucose - Insulin system of human body is modelled. The mathematical used here is the Hovorka model. Hovorka utilises a compartment model which represents the glucoregulatory system and includes submodels representing absorption of subcutaneously administered insulin and gut absorption. To generate the data for training our neural network, an reference controller was used. The reference controller employed was Model Predictive Control. The controller samples the glucose system parameters every 15 minutes and gives the insulin flow rate as output. A custom cost function is used that not only minimises the controller effort but also the rate of change to provide a smoother controller output. Moving target trajectory facilitates slow, controlled normalization of elevated glucose levels and faster normalization of low glucose values. Data is then collected from this control scheme for different initial values of glucose concentrations and different meal times and quantities so that the ANN controller can be effective under diverse conditions. The neural network trained uses multilayer feed-forward back-propagation that relates the output to the inputs by hyperbolic tangent sigmoid transfer function and optimized by Levenberg-Marquardt, the training of the ANN was successfull with R is approaching to 1 (R is a measure of the ANN performance. 1 is best possible result). The generated ANN was then used as a controller and the Insulin - Glucose system was simulated, the ANN was successfully able to maintain normoglycemia.
In conclusion, ANN is viable for use as an Artificial Pancreas. The model can manage in a very good manner the highly non linearity of glucoregolatory system.
The code has been written in Matlab. The code is available, and can be opened with Matlab software 2018b.