Parameter Estimation of a Physiological Diabetes Model Using Neural Networks

Diabetes Mellitus is a chronic disease characterized by elevated glucose levels in the blood due to deregulated insulin levels. The management of diabetes is greatly based on self-management of the patient by insulin injections, diet and exercise. Because of this, the study of mathematical models capable of describing the glucose-insulin metabolism in patients with diabetes can be a great tool to help this management. An example of such model, which has been extensively used, is the Eindhoven Diabetes Education Simulator (E-DES) model. Systems biology is a field of study focused on the structure and dynamics of biological systems, and mathematical modelling of these systems through non-linear ordinary differential equations. Parameter estimation, the process of determining the values of undetermined parameters in a mathematical model, is important for achieving reliable predictive models. Many parameter estimation methods exist, including traditional optimization methods and newer techniques such as artificial neural networks, in particular Systems Biology Informed Neural Networks (SBINNs). The goal of this study was to evaluate the performance of SBINNs in estimating E-DES parameters and in fitting to simulated glucose and insulin plasma measurements. Considering different variations of SBINNs, we were able to find a network capable of estimating some of the most important parameters of the E-DES model using simulated plasma glucose and insulin data with a good performance on simulated data.