(696g) An in-Silico study of Feedforward Predictive Control in Blood Glucose Concentration for People with Type 1 Diabetes

Rollins, D., Iowa State University
Mei, Y., Iowa State University
Artificial pancreas (AP) has the potential to tightly control glucose concentration for people with type 1 diabetes. AP has seen rapid development in semi-automatic control in recent years. To further advance AP, fully automatic/closed-loop control that can determine meal bolus infusion automatically is the next step. The core component in fully closed-loop AP is a control algorithm that can determine insulin infusion rate. Model-based control algorithms are wide used in various AP applications. However, a critical drawback in them is the need for accurate glucose prediction far into the future due to long time lags for system inputs (e.g. carbs, insulin). In this context, the proposed method, feedforward predictive control (FFPC), does not require future prediction of glucose and has the potential for perfect control, which are major advantages with great potential to significantly tighten glucose control in this application.

This work first demonstrates theoretical performance limits of FFPC in first order plus dead time (FOPDT) processes with two inputs under ideal conditions (i.e. manipulated variables (MVs) with the ability to both raise and reduce the controlled variable). This FOPDT study presents the proposed algorithm mathematically and demonstrates its strengths under ideal conditions. Next, the proposed FFPC is applied and evaluated using 30 virtual subjects from a FDA approved diabetes simulator in a comparison study against model predictive control (MPC), with insulin infusion rate as the MV that only has one-directional effects on glucose (i.e. insulin being only capable of reducing glucose level).

For the FOPDT study, FFPC shows it can reach perfect control under ideal conditions, while MPC failed to reach the same level of performance under the same conditions. For simulation study in diabetes simulator, in terms of the standard deviation (Stdev) of glucose about its mean, FFPC was about 4% smaller than MPC on the average, which indicates the potential of FFPC with restrictions on inputs (i.e. insulin as the only MV).

This study shows FFPC has the potential of perfect control given hormones that can raise glucose can be manipulated in the future (e.g. dual-hormone infusion systems). This prospect of perfect control and not using glucose predictions will be a major advantage in AP research.