(443c) An Integrated Multivariable Artificial Pancreas Control System

Patients with Type 1 diabetes (T1D) must administer exogenous insulin in order to regulate their blood glucose concentration (BGC). Hyperglycemia (usually defined as >180 mg/dl) is a common post-prandial problem in patients with T1D. Multiple daily injections of insulin or continuous subcutaneous insulin infusion by a pump are the most common approaches for administering insulin.  Commercial glucose concentration measurement systems and insulin infusion pumps are available and are gaining acceptance by patients. Continuous glucose monitoring (CGM) systems enable collection of frequent (sampling times in the range of 1-5 minutes) subcutaneous glucose concentration (GC) data and monitoring, and prediction of BGC. Automating the decisions on insulin infusion rates based on GC measurements to develop an artificial pancreas (AP) system seems to be a simple problem. Most AP systems proposed are based on subcutaneous glucose measurement and subcutaneous insulin infusion. This deceivingly simple control problem has many challenges. The diffusion of glucose from the blood stream to subcutaneous tissue introduces a measurement lag.  Hence, excessive insulin can be infused based on subcutaneous GC measurement. Subcutaneous infusion of insulin has diffusion limitations to the blood stream as well. The delay of insulin absorption and insulin action can also cause excess insulin delivery to the body. Once insulin is infused it cannot rapidly be removed from the body, and excess insulin subsequently causes hypoglycemia (usually defined as ≤ 70 mg/dL). Hypoglycemia may lead to unconsciousness, seizures and diabetic coma if untreated, and provides a major challenge for AP systems. Since patients with T1D do not have natural means to reduce plasma insulin levels and their glucagon response is often impaired, they are often unable to prevent low BGC. Hypoglycemia is prevented or treated by eating and/or reducing or stopping insulin infusion.  Hence, early detection and prevention of hypoglycemia is essential for the implementation of fully automated closed-loop control for AP systems. Several approaches have been proposed to accommodate measurement and insulin infusion delays.  What makes the system even more challenging is the variations of the human body to glycemic changes, physical activity and stress.  There are significant differences between the responses of different patients with T1D as well as variations of the dynamic behavior from one day to another for the same person. We have developed hypoglycemia early warning systems and AP control systems based on subject-specific recursive models that address many of the challenges listed.

Subject-specific time series models can capture a subject’s daily glucose variations and predict his/her future glucose excursions. A subject’s metabolic and physical activity and emotional state are known to have a significant effect on glucose metabolism and daily glucose excursions. We also use physiological signals measured continuously with a multi-sensor body monitor along with a subject’s recent glucose history from a CGM to develop the proposed subject-specific multivariable models and control algorithms. A model that can predict the insulin concentration in the blood also called insulin on board (IOB) is also used in the control algorithm as a limiting factor for prevention of hyperinsulinemia.

The objective of this this research is to develop an integrated multivariable system that that can deal with both hyperglycemia and hypoglycemia. For this purpose a system consisting of hypoglycemia early alarm system and closed-loop controller was developed.

Subject-specific glucose prediction models are developed by using measurements from a CGM and physiological signals from a multi-sensor armband. The frequent data from sensors are analyzed and modeled by time-series methods. Adaptive system identification consists of on-line parameter identification by using the weighted constrained recursive least squares (WCRLS) method and a time-varying forgetting factor method that changes the importance of recent measurements adaptively. The modeling system estimates model parameters that enable the adaptation of the model to inter-/intra-subject variation and glycemic disturbances. The models developed are linear, low-order, and easy to identify which make them good candidates for use in closed-loop control. The WCRLS and the time-varying forgetting methods enable dynamic adaptation of the models to inter-/intra-subject variation and glycemic disturbances.  A constrained optimization assures that every model developed with recursive identification is stable.  Generalized predictive control (GPC) methods are used to determine insulin infusion flow rates. IOB predictions are used as constraints in the control algorithm. A hypoglycemia early alarm system that uses glucose concentration from CGM, IOB, and physical activity information to predict future BGC is integrated with the control system.  Both the hypoglycemia early warning system and the AP controller developed do not receive any announcements from the user such as meal or physical exercise information.

The control algorithms based only on glucose measurement information are tested by using the University of Virginia / University of Padova simulator. Multivariable control systems are tested in clinical studies. Errors in glucose concentration predictions are reduced by building multivariable models that use information from the armband, when compared to predictions done solely on glucose measurements. With a time-varying forgetting factor, a better identification performance with lower computational load is obtained. When glucose predictions are determined to be low as early as 20 minutes before the potential onset of the hypoglycemia episode, supplemental carbohydrates are provided. The integrated early warning and control system is successful in regulating glucose concentrations in response to various meal and exercise disturbances without hypoglycemia episodes.

(Corresponding author: A. Cinar cinar@iit.edu)