A Relevant Glucose-Insulin Model: Validation using Clinical Data

International audienceA relevant glucose-insulin model with realistic asymptotic properties provide the computation of the tools for functional insulin therapy like basal rate, insulin sensitivity factor, carbo ratio insulin on board. The objective is to validate this model using clinical data of a large number of patients and to prove its ability to predict the behavior of glycemia and insulinemia dynamics.Data (CGM, insulin injections, carbohydrates) from 12 type-1 diabetic subjects were collected. An unconstrained optimization algorithm estimated the model’s parameters on n samples. Then, the estimated model was simulated on a set of n/2 samples as a cross-validation. The estimation was performed over at least 24 hours. The model-fit was very satisfactory as the mean standard deviation (SD) was 26 mg/dl. The cross-validation (performed over 16 hours on average) provides accurate glucose prediction with a mean SD of 40 mg/dl. Moreover, the computed tools for functional insulin therapy are consistent with the values determined by the clinical protocol.Realistic behavior of this model was already mathematically proven. In this study the relevant model was validated using clinical data as it accurately fitted the blood glucose of real patients. It was also shown that future values were well predicted. Thus it is a good candidate for a model based controller. As its parameters provide the computation of the tools for functional insulin therapy it will be useful for patient’s education