A Hybrid Neural Network Model Predictive Control with Zone Penalty Weights for Type 1 Diabetes Mellitus

Abstract

In this paper, a hybrid neural network model is developed to predict and control the blood glucose (BG) of the patient who has type 1 diabetes mellitus (T1DM). The proposed model consists of two parts: a linear finite impulse response (FIR) model and a nonlinear autoregressive exogenous input (NARX) network. A recently developed and well-acknowledged meal simulation model of the glucose-insulin system for T1DM is employed to create virtual subjects. Data from virtual subjects are used to identify an intermediate physiological model, and then our proposed hybrid model is trained and validated based on this intermediate model. The key features of the resulting hybrid model are that it reveals satisfactory accuracy of long-term prediction and does not require an immeasurable state for model initialization. The developed hybrid model is then embedded in a nonlinear model predictive control (MPC) controller with zone penalty weights, and this closed-loop controller is implemented on these virtual subjects for simulation-based preclinical testing. The results show that promising glycemic control performance can be achieved. Moreover, this overall BG control methodology is easily portable and has the ability to arbitrarily start the therapeutic control at any initial point.