A Fully Data-Driven Predictive Controller With Incremental Learning for Automated Insulin Delivery Systems

Automated insulin delivery (AID) systems have been proven to be a promising solution to control the glucose levels of patients with type 1 diabetes. However, the implementation of the closed-loop control method of AID systems is still challenged by the prior requirement for patient information, such as accurate physiological parameters, to formulate a personalized control scheme. In this work, we propose a fully data-driven predictive controller for AID systems without the need for prior knowledge. Specifically, a predictive control framework constructed by subject-specific insulin-glucose data is employed to develop an initial controller, and incremental learning is utilized to achieve the dynamic adaptation of the glycemic metabolism model using the individualized data selectively generated along the control process. Besides, an insulin-sensitivity (IS) estimation method suitable for AID systems is presented, which is utilized to perceive the personalized response of glucose to insulin and adjust the aggressiveness of the decision in real time by integrating the estimated IS conditions into the designed input penalty term of the optimization problem. To guarantee the safety of the patient, by analyzing the impact of accumulated insulin in the body on glucose fluctuation, a data-driven insulin-on-board (IOB) constraint estimation approach is introduced. The effectiveness of the proposed method is evaluated through in silico experiments using the FDA-accepted UVa/Padova type 1 diabetes mellitus (T1DM) simulator and advisory-mode analysis utilizing the clinical data.