(184l) Feedback Predictive Control Versus Model Predictive Control for Automatically Controlling Blood Glucose Concentration

Type 1 diabetes is the condition that features the loss of endogenous insulin production ability. Artificial pancreas (AP) is the system that can automatically control glucose level by determining insulin infusion rate via control algorithms. Model predictive control (MPC) is a widely used control algorithm in AP for the automatic control of glucose concentration. However, it has drawbacks due to long time lags associated with inputs (e.g. carbs and insulin) and inaccurate predictions over long prediction horizons. The proposed algorithm, feedback predictive control (FBPC), has the potential to avoid those issues to achieve fully closed-loop control of glucose and significantly reduce the variations of glucose around its target.

In this work, FBPC is first demonstrated in a first order plus dead time (FOPDT) process to explicitly showcase its superiority over MPC mathematically. Then the results of FBPC is compared to MPC for the 30 in-silico subjects in a “head-to-head” comparison with all things being equal except for the differences in the control algorithms. Lastly, to further evaluate the capability of handling unmeasured disturbances in control of glucose, meals are treated as unmeasured disturbances for first five (5) cases of virtual subjects.

For the standard deviation of glucose about its mean (Stdev), FBPC was smaller than MPC for all 30 cases. On the average, relative to FBPC, the Stdev for MPC was 33% higher. With meals as unmeasured disturbances, the Stdev was 133% larger for MPC relative to FBPC. For FBPC, its Stdev was less than 10% larger for unmeasured meals versus measured.

The proposed FBPC algorithm has the ability to produce significantly tighter variability of glucose about its target and has the capability in handling unmeasured disturbances. Thus, FBPC should be considered as an alternative to MPC for fully automatic control of glucose.