“Learning” Can Improve the Blood Glucose Control Performance for Type 1 Diabetes Mellitus

Abstract

The learning-type artificial pancreas system achieved good glycemic regulation and provided increased effectiveness over time. It showed a satisfactory performance even when the blood glucose was challenged by exercise or alcohol.

ISPAD Clinical Practice Consensus Guidelines 2018: Diabetes technologies

“…Such adjustments may be beneficial for patients on MDI therapy, conventional pump therapy, and may even help tune emerging artificial pancreas systems. Wang et al conducted a pilot study assessing a learning‐type artificial pancreas, which showed significant improvement in the time spent in target range with use of this learning system vs open loop control ( P = 0.02) . Dassau et al conducted a randomized crossover trial where algorithmic adjustment of open‐loop settings was conducted prior to artificial pancreas (AP) control and was compared to AP control without algorithmic adjustment, similar TIR was noted between the two groups .…”

Two‐layer model predictive control for chemical process model with integrating controlled variables

“…Thus, some approaches are proposed to treat uncertainty in the framework of robust optimization . In the literature, a learning‐type MPC algorithm is employed to track the setpoint, which could be time‐varying within‐day. The quadratic approximation of the comprehensive nonlinear steady‐state process model is used to develop the steady‐state predictive model in the SSTC layer by Ławryńczuk et al The integration of RTO with MPC in the presence of plant model mismatch and constraints is investigated by Marchetti et al In the work, the authors used a post‐optimality approach to analyze the effect of the uncertainty on the optimal solution of SSTC and determined their stability limits.…”

Multivariate learning framework for long‐term adaptation in the artificial pancreas

“…A closely related approach that can exploit CGM measurements and continuous insulin delivery is iterative learning control, which can be regarded as a two‐timescale enhancement of run‐to‐run methods . In Reference , iterative learning model predictive control (MPC) was proposed to adapt the reference trajectory of the closed‐loop controller used for glucose regulation, which was tested in a pilot study recently …”

“…13 In Reference 14, iterative learning model predictive control (MPC) was proposed to adapt the reference trajectory of the closed-loop controller used for glucose regulation, which was tested in a pilot study recently. 15 Despite the aforementioned progress, the problem of systematically adapting multiple key AP parameters (e.g., segments in BR and CR profiles, controller parameters) under lifestyle disturbances while explicitly enforcing safety considerations remains to be explored. In this work, a two-phase data-driven multivariate parameter learning framework for long-term adaptation of AP is developed.…”