Stochastic Virtual Population of Subjects With Type 1 Diabetes for the Assessment of Closed-Loop Glucose Controllers

Abstract

Closed-loop glucose control is an emerging treatment approach to manage type 1 diabetes. Closed-loop systems consist of a continuous glucose monitor, an insulin infusion pump, and a dosing algorithm that directs insulin delivery based on sensor levels. Testing of dosing algorithms in computer simulations may replace animal testing, accelerates development, and saves resources. We propose here a novel approach to generate a virtual population, to be used in metabolic simulators, from routine experimental data through the process that we term "stochastic e-cloning." We build on a nonlinear physiologically motivated time-varying model of glucose regulation. We adopt the Bayesian approach to estimate model parameters and to obtain the joint posterior probability distribution of time-invariant and time-varying parameters with the use of the Markov chain Monte Carlo methodology. The estimation process combines prior knowledge and experimental data to generate a sample from the posterior distribution, which can be subsequently used to conduct in silico experiments reflecting population and individual variability, and associated uncertainty as closely as possible. The approach is exemplified using data collected in 12 young subjects with type 1 diabetes. We demonstrate unbiased fit to the data, physiological plausibility of parameter estimates, and results of in silico testing using a stochastic virtual subject.