||Prediction and Control of Blood Glucose in T1DM Patients
Johnsen, Charlotte Aagaard
||Jørgensen, John Bagterp (Scientific Computing, Department of Informatics and Mathematical Modeling, Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark)
Finan, Daniel Aaron (Scientific Computing, Department of Informatics and Mathematical Modeling, Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark)
||Technical University of Denmark, DTU, DK-2800 Kgs. Lyngby, Denmark
||In this study, the possibility of predicting glucose levels in type 1 diabetes mellitus patients using simple black box models was investigated. More specifically,
two types of linear models were developed: The autoregressive exogenous input
(ARX) and autoregressive moving average exogenous input (ARMAX).
When identified from simulated data from a prominent physiological model of
type 1 diabetes, the models performed well, indicating that the dynamics are well
captured by the simple linear models. Promising results were also obtained for
reducing the quantitative meal information to binary information corresponding
to a push-button, 'meal announcement', scenario. The models showed robustness to both errors in the meal information, and variation in insulin sensitivity
levels. The latter was realized by manipulating parameters in the physiological
Using clinical data obtained from type 1 diabetes subjects in ambulatory conditions, the potential use of activity as exogenous inputs was investigated. Both
classical linear least squares identification as well as robust identification techniques such as Huber-regression were explored. None of these potential means
of improvement gave any convincing results. Overall, the models produced reasonable predictions but (especially the ARMAX models) showed inconsistencies.
However, based on the ambulatory nature and limited amount of the available
clinical data, little could be concluded from the obtained results and the potential benefits need further exploration.
The overall conclusion regarding the system identification results, was that an
ARX model is the best candidate of the two for predicting glucose levels.
Various model predictive control (MPC) algorithms using an ARX model as the
internal model were investigated. The ARX model was identified using data
from the physiological model, which also acted as the subject for the MPC. The
degree of input information included for system identification and control gave
rise to six different MPC cases. As a minimum, only insulin information was
used and in addition to this, either binary or quantitative meal information was
used. Despite the promising results from system identification, using binary
meal information for MPC lead to inconsistent performance. The potential for
using meal announcement instead of full meal knowledge was investigated and
showed promising results.
The cases that resulted in the best BG control, without requiring full meal
knowledge for control, used either no meal information or a meal announcement
15 min in advance. Both of these control scenarios are feasible to implement in
a closed-loop artificial pancreas device.
Creation date: 2009-09-21
Update date: 2009-09-21