CHoKI-based MPC for blood glucose regulation in Artificial Pancreas

TL;DR

This work develops a data-driven CHoKI-based MPC to autonomously regulate basal insulin in type 1 diabetes within an artificial pancreas, addressing the absence of a universal physiological model by learning patient-specific dynamics from UVA/Padova simulator data. The CHoKI predictor enforces componentwise Hölder continuity and integrates with an MPC that includes Insulin On Board constraints, constraint tightening, and an ISS framework to ensure safe closed-loop operation. The control objective targets the euglycemic band $[70,180]$ mg/dL while minimizing hypoglycemia risk and managing postprandial variability, using a 60-minute look-ahead with a 12-step horizon and a 5-minute sampling time. Results show improved safety and time-in-range relative to constant basal strategies, with observable conservatism and opportunities for adaptive or multi-model enhancements to cope with intra-day insulin sensitivity variation. Overall, the approach demonstrates promising potential for personalized, learning-based AP control in clinical-like simulations.

Abstract

This work presents a Model Predictive Control (MPC) for the artificial pancreas, which is able to autonomously manage basal insulin injections in type 1 diabetic patients. Specifically, the MPC goal is to maintain the patients' blood glucose level inside the safe range of 70-180 mg/dL, acting on the insulin amount and respecting all the imposed constraints, taking into consideration also the Insulin On Board (IOB), to avoid excess of insulin infusion. MPC uses a model to make predictions of the system behaviour. In this work, due to the complexity of the diabetes disease that complicates the identification of a general physiological model, a data-driven learning method is employed instead. The Componentwise Hölder Kinky Inference (CHoKI) method is adopted, to have a customized controller for each patient. For the data collection phase and also to test the proposed controller, the virtual patients of the FDA-accepted UVA/Padova simulator are exploited. The proposed MPC is also tested on a modified version of the simulator, that takes into consideration also the variability of the insulin sensitivity. The final results are satisfying since the proposed controller reduces the time in hypoglycemia (which is more dangerous) if compared to the outcome obtained with the standard constant basal insulin therapy provided by the simulator, satisfying also the time in range requirements and avoiding long-term hyperglycemia events.

Figures (9)

• Figure 1: In each graph, the vertical dashed red line marks the end of the fixed regressor, when the inputs displayed in the titles are applied. The blue line is the glucose trend. In a) the glucose increases a bit, due to the absence of basal insulin. In b) the glucose remains stable since the insulin amount is the reference value and equal to the regressor values. In c) the glucose decreases due to the basal amount of 180 p. In d) the glucose increases due to the presence of the meal and no basal insulin.
• Figure 2: The estimated boluses IOB is represented as the orange line and the IOB computed by the simulator is in blue. This is an example of the virtual patient Adult 10.
• Figure 3: The upper plot displays BG trends for all patients. The green zone represents the safe range and the black triangles depict meals. The lower plot shows basal insulin injections computed by the proposed MPC.
• Figure 4: Comparison of the BG values: the simulations performed with IOB constraints are represented in blue, and without them are in red.
• Figure 5: TIR results of the simulations performed with (graph on the right) and without the IOB constraints (graph on the left). Each bar represents a specific subject.