Type 1 diabetes mellitus (T1DM) is a metabolic disorder characterized by autoimmune destruction of pancreatic beta-cells that result in a lack of endogenous insulin production. As a result, millions of people with T1DM throughout the world require exogenous insulin delivery to survive. Unfortunately, determination of the proper insulin delivery is fraught with highly difficult challenges and dangers, leading to long term complications that decrease the mean life expectancy of people with T1DM by a decade or more. An Artificial Pancreas (AP) system holds great promise in saving these lives by automating insulin delivery for people with T1DM.

An AP combines an automated glucose sensor (continuous glucose monitor; CGM) and an automated insulin pump (continuous subcutaneous insulin infusion pump; CSII) with an insulin delivery algorithm. One of the greatest challenges for AP systems is the wide variability in intra- and inter-individual insulin sensitivities, which when combined with the low therapeutic index (TI; the ratio between therapeutic to lethal dosages) of insulin, means that even a slight miscalculation may have lethal consequences. A safe and efficient personalization scheme for the AP algorithms that does not require individual model identification has been developed and validated in both in silico and clinical trials. Both model predictive control (MPC) and proportional-integral-derivative (PID) controllers that incorporated this personalization showed excellent performance in maintaining subjects' blood glucose concentrations within the safe glycemic range.

Another monumental challenge in blood glucose control of people with T1DM is the asymmetry inherent in the human blood glucose scale -- that is, low blood glucose concentrations (hypoglycemia) carry acute consequences leading to death potentially in a matter of minutes, while high blood glucose concentrations (hyperglycemia) carry chronic consequences over longer timescales. An exponential asymmetric formulation of the cost function to penalize high and low glucose excursions differently has been developed and implemented in an MPC controller. This eMPC significantly decreases incidences of hypoglycemia and hyperglycemia in comparison to standard formulations under in silico clinical trials, and also shows promise in increasing subject safety under advisory mode traces.

Finally, the concept of personalization can be extended from algorithms to CGM performance. As CGM insertion requires the placement of a foreign object within the subcutaneous space of the body, each person's response to CGMs can vary significantly. A run to run (R2R) algorithm that utilizes previously discarded data as part of a CGM's natural product cycle of one week can be utilized to identify individual calibration parameters personalized to each subject, rather than estimating a population average calibration set for each new insertion. The proposed R2R algorithm shows statistically significant reduction in residuals between CGM sensor readings and reference blood glucose values over standard care during in silico testing.