Individualized model predictive control for the artificial pancreas: In silico evaluation of closed-loop glucose control

Despite the continuous efforts devoted to AP development in the last decades, an artificial pancreas (AP) system is not yet available on the market. One of the major issues involves the inter-subject variability affecting type 1 diabetes (T1D) patients, which makes the definition of a single controller suitable for any patient practically impossible. Moreover, a state-of-the-art, noninvasive, and portable AP system is composed of subcutaneous hardware components, and the control algorithm must be properly designed to reside on a standalone device with limited battery life and computational power. These characteristics make the design of a safe and effective AP system even more challenging, due to the inherent delays affecting the subcutaneous insulin delivery route and the tradeoff between control performance and computational power expenditure. As a result of the model predictive control's (MPC's) ability to address inherent delays of the process under control, it is one of the most promising control approaches in the context of an AP. However, the achievable control performance is strictly related to the prediction capabilities of the model included in the controller, which, in general, can be highly nonlinear. The currently used MPC in clinical experiments relies on a linear average glucose-insulin model designed to represent the average dynamics of a subject with diabetes. This non-individualized MPC is not designed to cope with patient-specific dynamics but is designed to be non-computationally demanding and robust enough to result in a safe and effective control law.