Comparison of parametrized models for computer-based estimation of diabetic patient glucose response

Abstract

Two mathematical models for the description of diabetic patient glucose behaviour are proposed. Unlike high order differential-equation based compartmental models, these models employ only the data typically available to a diabetic patient: the history of measured blood glucose concentrations and of insulin injections. The model structures are compared with a naive benchmark (zero-order hold) model in a computer simulation. It is demonstrated that, given four daily blood glucose measurements and two daily insulin injections, a parametrized model of patient blood glucose response to insulin can provide relevant data in the estimation of a patient's future blood glucose response in terms of past blood glucose measurements and insulin injections. Parametrized model root means squared errors of glycaemic predictions for 18 simulated patients ranged from 7–22 mg dl^-1, as compared with 19–42 mg dl^-1 for the benchmark model.

Key Words:

• Diabetes
• Modelling
• Estimation
• Least squares