The impact of non-model-related variability on blood glucose prediction

Physiological models are frequently used to predict blood glucose values from insulin and meal data of people with diabetes. Obviously, errors in the input data used result in prediction errors. A more complex problem is that no model may include all factors influencing the blood glucose level in any given situation. We have analyzed the influence of five parameters on prediction accuracy with respect to the time horizon.

A physiological model, consisting of an insulin model, a meal model, and a glucose metabolism model in combination with a Monte Carlo simulation, was used for this investigation. It was used to examine the change in blood glucose following the intake of carbohydrate and insulin. The intra-individual variability, which was studied, included pharmacokinetic variability of insulin aspart and estimation error of carbohydrate intake, as well as the accuracy of blood glucose meters and insulin pens.

Simulations showed how the coefficient of variance for the different model compartments changes over time. For average people with diabetes the inaccuracies of blood glucose meters and carbohydrate estimates contribute to more than half of the variance.

We showed how blood glucose prediction is severely affected by the inaccuracy in the input variables. Metabolic fluctuations, causing variability in insulin dynamics, also display important effects, but these are difficult to change. The inaccuracy of carbohydrate counting and the use of blood glucose meters appear to be the two main sources of error, which can be reduced through better patient education.