(373b) Automatic Insulin Delivery for Type 1 Diabetes Mellitus Using Offline Learning-Type Model Predictive Control (L-MPC)

Type 1
diabetes mellitus (T1DM) is an autoimmune disease manifesting as an absolute
deficiency of insulin secretion. In 2000, the total number of people with T1DM was
approximately 17.1 million, and this number is predicted to reach 36.6 million
in 2030 [1, 2]. The chronic hyperglycemia (high blood glucose concentration)
that results from T1DM causes a number of serious complications, such as cardiovascular
disease, stroke, hypertension, retinopathy, nephropathy, and neuropathy.

Exogenous insulin is administered in order to reduce
glycemia and sustain life; however, excessive insulin delivery results in
hypoglycemia (low blood glucose concentration), which has an immediate danger
much greater than that of hyperglycemia. Therefore, managing insulin delivery to
achieve normoglycemia is a daily challenge.

The development of external insulin infusion pumps and
the introduction of rapid-acting insulin analogues have made automated
intensive insulin therapy feasible. Additionally, glucose sensor technology has
improved significantly over the last few years and the reliable duration of in
vivo sensors continues to increase [3, 4]. Continuous subcutaneous insulin
infusion (CSII) pumps and continuous glucose monitoring (CGM) sensors are
bringing a fully closed-loop artificial pancreas closer to reality.

Due to the time delay between insulin injection and
glucose response, model predictive control (MPC) [5, 6] is one of the most
promising candidates for closed-loop control of an artificial pancreas. However,
MPC cannot improve control performance from day to day, in spite of the repetitive
nature of glucose-meal-insulin dynamics over a 24-hour period. In literature [7],
scholars used iterative learning control (ILC) to exploit the repetitive nature
of glucose-meal-insulin dynamics for the first time, where the "updating law"
of ILC was designed using MPC. This combination is named model predictive
iterative learning control (MPILC). MPILC can learn an individual's lifestyle, thus
allowing the control performance to improve from day to day.

To further improve the robustness of the learning
schemes, another novel combination of ILC and MPC was proposed in literature [8,
9], where MPC determined the insulin delivery rate and ILC was used to adjust
the set-point for MPC. This combination is termed as L-MPC. To the best of the
authors' knowledge, these contributions are the first reported works on L-MPC.

These prior works on L-MPC involve some online optimization
to get the "optimal" solution. However, the first generation of artificial
pancreatic b-cells will have limited computational power; therefore online
L-MPC should be transformed into an offline form. In this work, an offline L-MPC
was proposed and applied in automatic insulin delivery.

To derive the model for control design, a 12-hour step
response simulation was performed. A single-input single-output auto-regressive
exogenous (ARX) model was used to approximate the relationship between insulin
and glucose, where meal intake was considered as a disturbance. For mathematical
convenience, the ARX model is transformed to a state-space model, and then an analytical
solution for MPC can be obtained when there is no input constraint. In practice,
there are constraints on the insulin delivery rate, e.g., it cannot be
negative. The boundary values were used if the "optimal" input based on the offline
solution exceeded the physical constraints.

To test the robustness of the proposed algorithm, it was
implemented on ten adult subjects in the UVa/Padova diabetes simulator [10].
All subjects consumed three meals each day at {7:00, 12:00, and 18:00} with
fixed amounts of carbohydrate {40g, 85g, and 60g}, respectively. The order of
the ARX model was fixed as {2 and 1} for all subjects. There were four
parameters in the proposed scheme: learning gain, prediction horizon, control horizon,
and input variation penalty weight. The first three parameters were fixed as 0.5,
50, and 5, respectively. The input variation penalty weight was proportional to
the subject's insulin sensitivity gain.

After ten days, L-MPC reduced the average tracking
error (ATE) by 42% compared with MPC. In addition, L-MPC was robust to random
variations in meal size within ±50%
of the nominal value and meal timings within ±40 min. Future work will be
clinical evaluation of the proposed scheme.

This work was
supported by the Juvenile Diabetes Research Foundation (JDRF) grant
22-2006-1115.

Correspondence to:

Department of Chemical Engineering,
University of California, Santa Barbara, Santa Barbara, CA 93106-5080

frank.doyle@icb.ucsb.edu

REFERENCES

[1]      Eiselein L, Schwartz
HJ, Rutledge JC. The challenge of type 1 diabetes mellitus. ILAR Journal, 2004;
45(3):231-236.

[2]      Wild S, Roglic G,
Green A, Sicree R, King H. Global prevalence of diabetes: estimates for the
year 2000 and projections for 2030. Diabetes Care, 2004; 27(5):1047-1053.

[3]      Chia CW, Saudek
CD. Glucose sensors: toward closed loop insulin delivery. Endocrinol Metab.
Clin. North Am., 2004; 33(1):175-195.

[4]      Heinemann L.
Continuous glucose monitoring by means of the microdialysis technique:
underlying fundamental aspects. Diabetes Techno. Ther., 2003; 5(4):545-561.

[5]      Parker RS, Doyle III FJ, Peppas NA. A model-based algorithm for blood glucose control in type I diabetic
patients. IEEE Trans Biomed Eng, 1999; 46(2):148¨C157.

[6]      Hovorka R, Canonico V, Chassin LJ, Haueter U,
Massi-Benedetti M, Federici MO, Pieber TR, Schaller HC, Schaupp L, Vering T,
Wilinska ME. Nonlinear model predictive control of glucose concentration in
subjects with type 1 diabetes. Physiol Meas, 2004; 25(4):905¨C920.

[7]     
Wang Y, Dassau E, Doyle III FJ. Closed-loop control of artificial pancreatic b-cell in type 1 diabetes mellitus using model
predictive iterative learning control. Accepted by IEEE Trans Biomed Eng, 2009.

[8]     
Wang
Y, Doyle III FJ. Indirect
iterative learning control: application on artificial pancreatic b-cell. 21^st Chinese Control &
Decision Conference, June 17-19, 2009, Guilin, China, in press.

[9]      Wang Y, Zisser H, Dassau
E, Jovanovič L, Doyle III FJ. Model predictive control with learning-type
set-point: application to artificial pancreatic b-cell. Submitted to AIChE Journal,
2009.

[10]   Kovatchev BP, Breton MD, Dalla Man C, Cobelli C. In silico preclinical
trials: a proof of concept in closed-loop control of type 1 diabetes. J
Diabetes Sci Technol, 2009; 3(1):44-55.