Model predictive control (MPC) is the "forefront of current research into closed-loop systems" (Hovorka 2011). However, none of the approaches described in the literature seems to provide a thorough solution to the following challenges: i) account for subcutaneous hormone application, ii) minimize the overall consumption of exogenous hormones (e.g. by avoiding unnecessary counteracting, or by containing BG in a sensible interval rather than forcing it to a specific equilibrium), iii) adopt a sophisticated physiological model to enable consideration of known upcoming disturbances like announced meals.
We employ the UVA/Padova T1DMS simulator (The Epsilon Group, Charlottesville, Virginia USA) to provide proof of concept for a novel multivariable nonlinear MPC algorithm. It bases upon a model similar to the latest Padova model to reconstruct the current state of the virtual patient from past measurements (moving horizon estimation) and to find the optimal control signal in order to avoid both hyper- and hypoglycemia. For both tasks, one has to define expedient cost functions that provide a mathematical formulation of the conflicting goals. The controller receives continuous BG and accepts meal announcements. While most existing bihormonal algorithms use glucagon merely as a rescue system, our controller defines both insulin and glucagon dosage at the same time and thus in a coordinated way, such that the pending effect of any already administered hormone is accounted for in the prediction.
We found the controller to exhibit highly desirable behavior in most situations; the example of a spontaneous meal is shown in the figure.

02 - USING MULTIVARIABLE NONLINEAR MODEL PREDICTIVE CONTROL FOR A BIHORMONAL CLOSED-LOOP ARTIFICIAL PANCREAS