Model Predictive Control of Blood Glucose for Type 1 Diabetic Rats in a Cyber-Physical System

Abstract

Advances in sensing and communication technologies have enabled a tight integration between computational and physical elements that is known as cyber-physical systems (CPS). This integration has been also applied to medical devices and healthcare systems and referred to as medical CPS (MCPS). This paper presents development of patient models and algorithms for a MCPS to control blood glucose level (BGL) in type 1 diabetic rats, which is known as artificial pancreas system (APS). An APS consists of continuous glucose monitor (CGM), insulin infusion pump, and control algorithm that makes autonomous decisions about insulin injection to maintain BGL within a normal range. As the control method, we adopted model predictive control (MPC) due to its robustness, flexibility, and use of constraints. Cyber-physical interactions in APS occur by physiological feedback (BGL read by CGM) to the controller, and developing MPC requires identification of such a model of glucose-insulin interactions to predict BGL and take an appropriate action accordingly. In this study, a physiological patient model of diabetic rats is developed and fitted to the data collected from a subject. Using the patient model, we train an artificial neural network (ANN) that predicts BGL based on time-series input data. Considering BGL predicted by the ANN, the NN-MPC controls insulin injection so that BGL can be maintained within the normal range. A simulation study showed the NN-MPC yielded a good performance in a simulated environment. The study showed the potential of the proposed approach for developing a fully closed-loop MCPS for BGL control.