Linear parameter-varying model to design control laws for an artificial pancreas

Abstract

The contribution of this work is the generation of a control-oriented model for insulin-glucose dynamic regulation in type 1 diabetes mellitus (T1DM). The novelty of this model is that it includes the time-varying nature, and the inter-patient variability of the glucose-control problem. In addition, the model is well suited for well-known and standard controller synthesis procedures. The outcome is an average linear parameter-varying (LPV) model that captures the dynamics from the insulin delivery input to the glucose concentration output constructed based on the UVA/Padova metabolic simulator. Finally, a system-oriented reinterpretation of the classical ad-hoc 1800 rule is applied to adapt the model's gain. The effectiveness of this approach is quantified both in open- and closed-loop. The first one by computing the root mean square error (RMSE) between the glucose deviation predicted by the proposed model and the UVA/Padova one. The second measure is determined by using the ν-gap as a metric to determine distance, in terms of closed-loop performance, between both models. For comparison purposes, both open- (RMSE) and closed-loop (ν-gap metric) quality indicators are also computed for other control-oriented models previously presented. This model allows the design of LPV controllers in a straightforward way, considering its affine dependence on the time-varying parameter, which can be computed in real-time. Illustrative simulations are included. In addition, the presented modeling strategy was employed in the design of an artificial pancreas (AP) control law that successfully withstood rigorous testing using the UVA/Padova simulator, and that was subsequently deployed in a clinical trial campaign where five adults remained in closed-loop for 36 h. This was the first ever fully closed-loop clinical AP trial in Argentina, and the modeling strategy presented here is considered instrumental in resulting in a very successful clinical outcome.