LAUSR

Abstract Observers are essential in artificial pancreas systems, either for states or disturbance estimation. Kalman filters based on the Hovorka model are largely applied for this purpose. However, simpler approaches can be used too. We intend to analyze whether the observer structure, the applied model and the individualization of the model parameters affect the estimation accuracy. We perform an in-silico comparison between two Kalman filters and a sliding mode observer, as observer structures previously proposed in the field. All observers are implemented in population and individualized versions of the Hovorka and the simpler Identifiable Virtual Patient models. To tune the Kalman filters, a genetic algorithm based framework was developed. The results indicate that the choice of the model has a larger effect on the outcome than the choice of the observer structure. Finally, the observer based on the Hovorka model does not always perform the most accurate estimation given the higher difficulty during the identification of its parameters.