Diabetes is a disease characterized by degradation of insulin secretion from pancreas and consequently failure in blood glucose regulation in the body. All patients with type 1 diabetes and some patients with type 2 diabetes depend on exogenous insulin. The main objective of diabetes management is to reduce glucose variability and to keep the blood glucose concentration within tight control (normoglycemic range). Currently, insulin-dependent patients deal with this life-long problem by taking 3-7 daily fingerstick blood glucose measurements and 3-5 daily insulin injections or administering insulin with a manually controlled pump. Patients generally experience prolonged hyperglycemia or hypoglycemic episodes as it is a difficult task to decide on the required insulin amount/rate and correct timing of injection or bolus insulin infusion, with changing daily life conditions (e.g. diet, exercise, stress or illness). Closing the glucose control loop with a fully automated artificial pancreas will definitely improve quality of life for insulin-dependent patients. In this research, we propose an adaptive model-based control algorithm that keeps glucose concentrations within normoglycemic range and dynamically responds to disturbances like physiological or external glucose perturbations. The algorithm handles delays associated with insulin absorption, transient difference between plasma and interstitial glucose, and variations in inter/intra-subject glucose-insulin dynamics. The adaptability of the controller is further assured with subject-specific linear models developed from patient's continuous glucose monitoring (CGM) device data.

The performance of model-based control algorithms is highly dependent on the accuracy of the process-model used to represent the true dynamics of the system. In literature, model-based control strategies employing physiological glucose-insulin dynamics models have been previously proposed for closing the loop for insulin-dependent patients. Differently from the physiological models that are generally representative of only an average subject under specific disturbance-free conditions, we make use of frequently sampled glucose concentration data for development of patient-specific models that can dynamically adapt to inter-/intra-subject variations or disturbances like meal consumption or exercise. Time-series analysis is utilized to develop linear and low-order autoregressive moving-average models (ARMA) from patient's own continuous glucose monitoring data. At each sampling step, the model parameters are recursively tuned using the weighted recursive least squares (RLS) method. The RLS algorithm is further integrated with a change detection method to dynamically capture the unpredicted glucose fluctuations due to physiological or external perturbations and to provide a faster response under such conditions. The proposed glucose prediction strategy is then integrated with adaptive model-based control methods for closing the glucose control loop for patients with diabetes.

Depending on the site of glucose measurement and/or insulin delivery, one of the major challenges of the closed-loop glucose regulation is the variable time-delay introduced to the system due to insulin absorption into the bloodstream and time-lag between subcutaneous glucose and blood glucose concentrations. Subcutaneous space appears to be the minimally invasive site and therefore is the most preferred site by CGM devices and insulin pumps currently available in the market. Differently from many of the proposed glucose control algorithms in the literature that assume the minimally delayed intravenous site for either glucose sensing or insulin delivery (or both), we focus on the most delayed site (subcutaneous) for both glucose measurements and insulin administration. To cope with delays associated with insulin absorption, we introduce a Smith predictor like structure to the closed-loop algorithm. Even tough almost all of the available CGM sensors monitor glucose in the subcutaneous interstitial fluid rather than blood, management of diabetes is still defined in terms of blood glucose (not subcutaneous) by clinicians and therefore the standard choice for control performance is still the blood glucose regulation. Therefore, we also introduce a lag-filter to account for the time-lag between subcutaneous and blood glucose to the closed-loop structure.

In summary, variations in glucose-insulin dynamics are monitored by online identification of the model. At each step with the new coming subcutaneous glucose measurement, the current blood glucose concentration is estimated using the lag-filter. Model parameters are updated, and estimation of future time course of blood glucose is performed. These parameters are then used in the model-based control algorithm for calculation of the appropriate control action (insulin infusion rate). The required insulin infusion rate is determined by solving an optimization problem that minimizes the deviation of the predicted glucose values from a reference glucose trajectory. Two well-known model-based control strategies, generalized predictive control (GPC) and linear quadratic control (LQC) are used for control-law calculations, and are modified to include a delay compensator (Smith predictor like structure) associated with insulin absorption and a time-varying reference glucose trajectory.

Assuming subcutaneous glucose measurements, the closed-loop performance is evaluated on a simulated patient with type 1 diabetes and results are reported for subcutaneous delivery of rapid-acting insulin. The algorithm is able to keep blood glucose concentrations within normoglycemic range in presence of multiple-meal consumptions with simultaneous physiological glucose variations (e.g. changes in insulin sensitivity). The proposed closed-loop algorithm for blood glucose regulation provides robust control to a wide range of inter-/intra-subject variability of glucose-insulin dynamics and it further accounts for delays associated with insulin absorption or time-lag between plasma and interstitial glucose concentrations.