Maximizing performance of linear model predictive control of glycemia for T1DM subjects

Abstract

The primary objective of this paper is the custom design of an effective, yet relatively easy-to-implement, predictive control algorithm to maintain normoglycemia in patients with type 1 diabetes. The proposed patient-tailorable empirical model featuring the separated feedback dynamics to model the effect of insulin administration and carbohydrate intake was proven to be suitable for the synthesis of a high-performance predictive control algorithm for artificial pancreas. Within the introduced linear model predictive control law, the constraints were applied to the manipulated variable in order to reflect the technical limitations of insulin pumps and the typical nonnegative nature of the insulin administration. Similarly, inequalities constraints for the controlled variable were also assumed while anticipating suppression of hypoglycemia states during the automated insulin treatment. However, the problem of control infeasibility has emerged, especially if one uses too tight constraints of the manipulated and the controlled variable concurrently. To this end, exploiting the Farkas lemma, it was possible to formulate the helper linear programming problem based on the solution of which this infeasibility could be identified and the optimality of the control could be restored by adapting the constraints. This adaptation of constraints is asymmetrical, thus one can force to fully avoid hypoglycemia at the expense of mild hyperglycemia. Finally, a series of comprehensive in-silico experiments were carried out to validate the presented control algorithm and the proposed improvements. These simulations also addressed the control robustness in terms of the intersubject variability and the meal announcements uncertainty.