| Type 1 diabetes mellitus (T1DM) is a disease characterized by an absence of endogenous insulin and chronically elevated blood glucose. Health problems, such as microvascular diseases, are prevalent in individuals with T1DM, but preventable with intensive insulin therapy. The global prevalence of T1DM is estimated at 20 million people.;Using glucose sensors and insulin pumps in a closed-loop fashion could improve therapeutic efficacy. This is challenging because disturbance effects are faster to appear in blood glucose than those of exogenous subcutaneous insulin. The processes governing the effects of insulin and carbohydrate are complex and time-varying. The challenge for a successful control algorithm is therefore to deliver insulin aggressively, in order to normalize glycemia, while also remaining robust to the dynamic variation of glucose-insulin kinetics in an individual with T1DM.;Clinical trials were conducted to develop personalized models for a controller. Proportional-integral-derivative control and model predictive control (MPC) were compared to establish how these controllers could be tuned to consider robustness explicitly. Hardware trials demonstrated the practical issues of closed-loop control using devices designed for open-loop use. Advanced features of MPC were investigated in combination with clinical safety features. Computational cycles and power in a portable ambulatory device are at a premium, thus the control law was reformulated using multi-parametric programming so that optimal control was obtained from a lookup table evaluated a priori.;Clinical trials showed that low-order models, with parameters determined heuristically, captured the critical bandwidth frequencies of measurement data. Simulation studies showed that the performance of higher-order models diminished as the degree of nonlinearity in the virtual subject model increased. Safety constraints derived from clinical expertise were valuable for avoiding hypoglycemia. The multi-parametric MPC implementation was shown to be feasible for this control law. The ramifications of these results are that tradeoffs between experimental costs in model development, robustness of controller performance, and the practicalities of algorithm implementation should be considered simultaneously when developing an algorithm for closed-loop insulin delivery. |
