Effect of Model, Observer and Their Interaction on State and Disturbance Estimation in Artificial Pancreas: An In-Silico Study

Abstract

The state and disturbance estimations are an indispensable part of the state-of-the-art model-based controllers as related to the artificial pancreas, supporting the decision-making and self-tuning of the algorithms. They are not just important when state-feedback kind of controller structure is applied, but also play a crucial role in the estimation of, for example, the amount of the acting drug (insulin) in blood or meal intake estimation which has determining role in the short and long term effectivity of the given therapy. This information is also important for physicians to support them in knowledge-based decision-making to be sure a given therapy or device works well. This article compares three observers – a linear-parameter-varying (LPV) dual Kalman filter (KF), a LPV joint KF, and a nonlinear sliding mode observer (NSMO) – designed with two individualized models – Hovorka and Identifiable Virtual Patient model (IVP). The article also statistically quantifies the effect of the observer algorithm and model structure on the accuracy of the estimation of plasma insulin, rate of glucose appearance, and glucose. Data for the analysis was generated by the UVa-Padova simulator. Results indicated that, for the rate of glucose appearance and the plasma insulin, the type of model and the observer structure explain less than 10% of the variability in the error, while the inter-patient variability contributes to the error more than 50%. This reveals a limiting factor in the estimation accuracy that might be improved by model parameter adaptation.

Keywords

• Kalman filters
• sliding mode observers
• artificial pancreas
• model identification
• linear parameter varying model
• estimation