Enhancing Glucose Sensor Models: Modeling the Drop-Outs

Abstract

Background:

Computer simulation environments have been used in the development of many artificial pancreas systems. A glucose sensor model is an essential part of these environments, and different models have been proposed. However, not one of these models accounts for drop-outs of sensor readings, a well-known phenomenon caused by physical pressure on the sensor site. In this work, we have proposed an enhanced model that accounts for drop-outs and demonstrated its improvement over the existing one-compartment model.

Materials and Methods:

Potential drop-outs were augmented to the existing model, and their incidences and magnitudes were estimated simultaneously with the model parameters using the Bayesian approach. Drop-outs and model parameters were estimated from data collected from 15 subjects with type 1 diabetes who underwent an artificial pancreas study. Model fitting and parameter estimates were contrasted between the enhanced model and the existing one-compartment model.

Results:

Both models achieved similar parameter estimates (P=not significant) and were all physiologically plausible. The enhanced model further estimated 1.71 drop-outs per day, which improved model fit (weighted residual reduced from [minimum −4%, maximum 3%] to [−3%, 2%]) and reduced significantly the deviance information criteria from 2739.72 to 1456.00.

Conclusions:

The enhanced model improves fitting of glucose levels and should allow more realistic simulations that assesses artificial pancreas systems.