Hindawi Publishing Corporation

EURASIP Journal on Wireless Communications and Networking

Volume 2008, Article ID 495185, 10 pages
doi:10.1155/2008/495185

Research Article
Continuous Drug Infusion for Diabetes Therapy:
A Closed-Loop Control System Design

Jiming Chen,1 Kejie Cao,1 Youxian Sun,1 Yang Xiao,2 and Xu (Kevin) Su3
1 State
Key Laboratory of Industrial Control Technology, Institute of Industrial Process Control, Zhejiang University,
Hangzhou 310027, China
2 Department of Computer Science, The University of Alabama, Tuscaloosa, AL 35487, USA
3 Department of Computer Science, University of Texas at San Antonio, San Antonio, TX 78249, USA

Correspondence should be addressed to Jiming Chen, [email protected]

Received 15 July 2007; Revised 25 November 2007; Accepted 5 December 2007

Recommended by Hui Chen

While a typical way for diabetes therapy is discrete insulin infusion based on long-time interval measurement, in this paper, we
design a closed-loop control system for continuous drug infusion to improve the traditional discrete methods and make diabetes
therapy automatic in practice. By exploring the accumulative function of drug to insulin, a continuous injection model is proposed.
Based on this model, proportional-integral-derivative (PID) and fuzzy logic controllers are designed to tackle a control problem
of the resulting highly nonlinear plant. Even with serious disturbance of glucose, such as nutrition absorption at meal time, the
proposed scheme can perform well in simulation experiments.

Copyright © 2008 Jiming Chen et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

1. INTRODUCTION functions of the pancreas are replaced with external devices.

A typical external insulin pump is an electronic medical de-
Diabetes mellitus is a metabolic disorder in which insulin, vice that delivers insulin through a narrow and flexible plas-
a kind of hormone which promotes the uptake of glucose tic tubing that ends with a needle inserted just under the skin
into cells, cannot properly perform its role. Diabetes Mel- near the abdomen. The pump releases doses of insulin peri-
litus affects more than 100 million individuals throughout odically, at meals or at time when blood glucose is too high
the world and this number may be expected to double by based on measured values of glucose sensors.
2010 [1, 2]. The healthcare costs are estimated to be over 100 A patient’s glucose concentration may change dynami-
billion dollars annually for the 16 million people who suf- cally depending mostly on his/her physical activities and nu-
fer from this debilitating disease and its complications in the trition, and therefore, the amount of insulin needed varies
US [3]. Based on 2002 death certificate data, diabetes was the from time to time. A number of diseases may occur, possibly
sixth-leading listed cause of death in the US [4]. This ranking resulting in life-threatening health conditions if the supply of
is based on the 73 249 death certificates in which diabetes was insulin is not in time or not correctly dosed or fails for some
listed as the underlying cause of death. According to death reasons. For example, sustained hyperglycemia (blood glu-
certificate reports, diabetes accounted for a total of 224 092 cose exceeding 120 mg/dL) may lead to most of the long-term
deaths [4]. complications associated with diabetes, such as nephropathy
For patients with diabetes, especially, type I insulin- and retinopathy according to the Diabetes Control and Com-
dependent diabetes, tight control of glucose level is essen- plications Trial (DCCT) Research Group [5]. But to date, a
tial. Regulating blood glucose concentration using insulin in- common current method of therapy is a series of 3–5 daily
fusion pumps is important for these patients, because they in insulin injections with quantities of insulin based on 4–8
have deficiency of insulin production by pancreas that pre- daily invasive glucose measurements. In this sense, infusion
vents appropriate metabolism of glucose. For many patients of insulin is discretely controlled by users based on the feed-
under diabetes therapies, insulin injection using needles and back of several blood glucose measurements. It is obvious
syringes under skin is adopted to deliver insulin, so that the that such treatment lacks a reliable continuous monitoring,
2 EURASIP Journal on Wireless Communications and Networking

Disturbance
Required blood Plant
glucose level Blood
Drug rate Drug Insulin Glucose glucose
+ Error Drug- Insulin-
+
level
Controller Pump glucose +
− insulin

Measured
Measure

Figure 1: Closed-loop framework of blood glucose concentration control.

which may cause glucose concentration being out of a per- sulin dose based on short historical discontinuous blood glu-
mitted range because of control delay. This kind of therapy cose measurements and insulin doses’ settings. Using fuzzy
may not prevent glucose fluctuations occurring in many pa- logic controllers to regulate blood glucose level also had been
tients. Therefore, designing a continuous closed-loop control proposed in [17, 18]. In [17], fuzzy reasoning method was
system is needed for insulin infusion. The continuous con- used to monitor and help to detect hypoglycemia in diabetic
trol would be a great improvement in the daily treatment of patients. In [18], knowledge about patient treatment was in-
diabetes, especially, in some cases that medical persons are corporated, and inner-loop and outer-loop controllers using
not presented or the patients have less knowledge about the a Mamdani-type fuzzy scheme were designed. Other strate-
disease. Such an automatic control will benefit patients and gies like optimal control with quadratic cost function and
avoid some mistakes during injections and operations. artificial systems had been applied to blood glucose control
To implement the continuous closed-loop control, three [19, 20]. But these strategies were all discrete control, and the
primary components are needed for such therapy: an im- injection only happened at meal time.
plantable glucose sensor, a pump, and a control algorithm. In this paper, we present reviewed results of discrete con-
Implantable glucose sensors have been developed to con- trol to blood glucose concentration of type I diabetes. An in-
tribute to the interest in feedback-type insulin infusion tegral form is proposed to model the time course of plasma
pumps [6, 7], including needle-type sensors and extracor- insulin concentration along with continuous injection of a
poral sensors, coupled with iontophoresis [8], microdialysis typical dose of insulin preparations. Although the resulting
[9], or microperfusion [10] system. A pump mechanism has plant model is highly nonlinear, PID and Fuzzy logic con-
been studied extensively and is available. The most important trollers are carefully framed to tackle the corresponding non-
is a control algorithm to design the continuous closed-loop linear control problem. Simulation experiment results show
control system, which is the focus of this paper. The detailed that with such controllers the real-time blood glucose could
implementation of the system is beyond the scope of the pa- be restricted in the permitted bound, and the unexpected
per. concentration fluctuation due to the patients accidental dan-
In recent years, significant efforts have been made in gerous behaviors could also be dealt with.
the development of glucose control algorithms. Model-based The remainder of the paper is organized as follows. In
predictive control (MPC) algorithms have been recently re- Section 2, a continuous model is proposed to explore the ac-
ported in literature to successfully tackle constraints posed cumulative effect of drug on insulin, which is involved in
by several biomedical control problems, not only in blood the design of a closed-loop control system. Both PID con-
glucose concentration control in diabetic patients [11, 12], troller and fuzzy logic controller are designed, and the per-
but also in mean arterial pressure and cardiac output control formance is evaluated by extensive simulation experiments
during anesthesia [12, 13]. Parker et al. proposed a model in Section 3. Section 4 concludes this paper.
predictive control for type I diabetic patient blood glucose
control and adopted an asymmetric objective unction to ad- 2. A CONTINUOUS CLOSED-LOOP MODEL
dress the inherent performance requirement of the physio-
logical problem [14]. However, we believe that the function To describe the complete metabolism process of glucose,
of drug should be cumulated if continuous injection is ap- three important parameters related to type I diabetes patients
plied, and this makes the plant highly nonlinear. Such a char- are involved: dose of drugs, insulin concentration, and glu-
acteristic of the plant will weaken the performance of MPC. cose concentration. Therefore, we need two different models
The authors in [15] considered a considerable amount of un- to describe the relationships among them. One is the time
certainty of the parameters in a mathematical model of blood course function of drug to insulin while the other is the in-
glucose dynamics and proposed an H∞ controller for robust sulin to glucose. The process flow chart can be described in
closed-loop regulation. However, their approach took a sim- Figure 1. The object of the controller design is to minimize
ple glucose absorption model from food into account. Some the error, which makes the output track the required blood
intelligent advance control strategies are applied to the blood glucose level. The controller drives the pump to implement
glucose control system. A neural-network controller was de- continuous insulin infusion. Before the controllers are de-
veloped in [16], which suggests an appropriate next-time in- signed, the plant should be modeled firstly.
Jiming Chen et al. 3

90
Brain
80 Lispro
Insulin concentration (μU/ml)

70
Glucose
60 Heart measurement
Venous blood
50 f (t) &
Insulin lungs
40 infusion
30
Kidney
20
10 Hepatic artery

0 Meals
0 100 200 300 400 500 600 700
Time (min) Liver Gut
Portalvein
Figure 2: Time course of plasma insulin concentration after a sub-
cutaneous injection (10 U) of Lispro.
Periphery

2.1. A continuous model of drug to insulin

Tissue compartment
The development of the implantable sensor technology has
made it possible for continuous insulin injection. There have Direction of blood flow
been a number of results regarding related topics [21]. Phar-
maceutical research has produced various types of insulin to Figure 3: Detailed compartmental model of glucose-insulin inter-
glucose, through their combined injection and subcutaneous actions.
injections. The typical insulin preparations with faster dy-
namics, namely, Lispro and regular insulin, and delayed ac-
tion, such as neutral protamine hagedorn (NPH), are used
to meet the basal insulin requirement. In this paper, it is concentration after a subcutaneous injection (10 U) of Lispro
assumed that only the drug Lispro is used to produce in- in continuous closed-loop control for diabetes therapy. Al-
sulin [22], whose plasma insulin concentration is shown in though such accumulative effect has been neglected in, for
Figure 2 after subcutaneous injection of a typical dose of the instance, the traditional three meal-time discrete control
Lispro. methods, it is crucial for our continuous control. The trade-
Function f (t, d) is defined to describe the time course of off is that the nonlinearity of this model makes the control
plasma insulin concentration after a subcutaneous injection strategy hard to realize.
d dose of Lispro. It is assumed that there is a linear relation-
ship as in (1) when d is not so large:
2.2. Model of insulin to glucose
f (t, c∗d) = c∗ f (t, d). (1)
The relationship between insulin and glucose has been well
When d = 10 U, we can get the function f (t) with the
investigated in recent years [23–25]. There are mainly two
curve described in Figure 2. Because of the continuous injec-
different proposed models: Sorensen model and minimal
tion, the effects on insulin concentration each time should be
model.
cumulated. Therefore, the integral form is used as follows to
Sorensen model: as a physiologically based compartmen-
calculate the insulin concentration:
tal model of glucose and insulin dynamics, Sorensen model
 t
d(τ)∗ f (t − τ) [23, 24] is a six compartment model, as shown in Figure 3,
I(t) = dτ, (2) where the compartments are physiological representations of
t −360 10
brain, heart and lungs, liver, gut, and kidney peripheral tis-
where d(t) is defined as the dose of drug injected at time t. sue.
The duration from time point of injection of Lispro to invali- As a physiologically based compartmental model,
dation equals 360 minutes, and its effect should be taken into Sorensen model is described vividly. Furthermore, it also
account and calculated in active term of Lispro. This is the represents glucagon to complete the glucose-insulin system
most distinct characteristic of continuous control of blood model. However, it is difficult to get the parameters since pa-
glucose. In the previous works [18, 23], the interval of injec- rameters of 6 compartments for different patients are differ-
tion is long enough and the effects will not be overlapped, so ent.
that there is no accumulative result on insulin concentration. Minimal model: as a typical minimal model, Bergman
Taking the accumulative effect into consideration, (2) model [24, 25] is widely used in the blood glucose-level con-
stands for a new time course model for plasma insulin trol. It offers a good benchmark for testing the relationship
4 EURASIP Journal on Wireless Communications and Networking

Table 1: Insulin-glucose model parameters. The effect of PID controller

350
Parameters Value
300
p1 0.0337 min−1
p2 0.0209 min−1

Glucose (mg/dL)
250
p3 0.00000751 min−2 (μU/mL)−1
n 0.214 min−1 200
T 5 min
150
Gb 0.811 mg/mL
M 0.012 mg/mL/min 100
G 0.81 mg/mL
X 0.0054 μU/mL 50
0 500 1000 1500
Time (min)
250
Before PID controller
200 After PID controller

Figure 5: The output of system with PID controller.

Glucose (mg/dL)

150

100 3. CONTROLLER DESIGN AND SIMULATION

EXPERIMENTS
50
Before the controller design, we would like to parameterize
0
the whole plant model as follows. The function f (t) is ap-
proached by a high-order binomial. For the model of insulin
to glucose, by applying Laplace transform for (4), the trans-
−50
0 500 1000 1500 fer function from insulin to glucose is presented with the pa-
Time (min) rameters shown in Table 1. The simulations are carried out
in Matlab Simulink, and the data are collected by “scope” in
Figure 4: The meal disturbance of blood glucose. sink:
−0.00003
G(s) = . (5)
between insulin and glucose. The model can be depicted as s3 + 0.274s2 + 0.013656s + 0.00018
follows: The main disturbance caused by three meals can be de-
  picted as a curve shown in Figure 4 [23]. Under the distur-
Ġ(t) = − p1 + X(t) G(t) + p1 Gb + m(t),
bance, we want to design a controller to meet the require-
Ẋ(t) = − p2 X(t) + p3 I(t), (3) ments of normal person’s glucose concentration of about
˙ = τ ·u(t) − n·I(t),
I(t) 60–100 mg/dL before the meal and less than 140 mg/dL af-
ter meal. As the meal disturbance is much higher than nor-
where G(t) is the concentration of glucose, I(t) is the concen- mal glucose concentration level, it demands a controller with
tration of insulin, X(t) is the dynamic insulin response, Gb is good performance of disturbance rejection.
the basal level of glucose, and m(t) is the rate of exogenous Considering the possible unmodeled nonlinearities, two
glucose infusion; p1 , p2 , p3 , τ, and n are parameters defined control strategies are introduced to design corresponding
in Table 1. controllers, that is, the PID controller and fuzzy logic con-
In order to simplify the nonlinearity, the idea of linearity troller.
of Bergman model is proposed in [26]:
  3.1. PID controller design for continuous closed-loop
Ġ(t) = − p1 + X(t) G(t) − G·X(t) + G·X + p1 Gb + m(t),
control system
Ẋ(t) = − p2 X(t) + p3 I(t),
˙ = τ ·u(t) − n·I(t), The standard PID controller could be stated as
I(t)
   
(4) Kc ∗ 1 + s/Ti ∗ Td ∗s + 1
PID = , (6)
Td ∗s/Ad + 1
where G and X are the average values of G(t) and X(t), which
are defined in Table 1. where Kc = P, Ti = I, Td = D, and Ad = 10.
The above parameters in Bergman model can be ob- For this PID controller, there are three important param-
tained easily [26, 27]. The Bergman model ignores the effects eters: P, I, and D, where P is a proportional feedback in
of glucagon, which is not important in this paper. which the stronger it is, the more powerful the feedback is;
Jiming Chen et al. 5

0.4 Kc = 0.0045, Td = 61

0.35 150

0.3
Controller’s output (10 U)

100

Glucose (mg/dL)
0.25
50
0.2
0
0.15
−50
0.1
−100
0.05
−150
0 0 500 1000 1500
0 500 1000 1500
Time (min) Time (min)

Figure 6: The corresponding output of PID controller. Ti = 10

Ti = 100
Ti = 100000

Ti = 100000, Td = 61 Figure 8: The effect of parameter of Ti in PID controller.

250

200

proportional parameters of Kc . When Kc = 0.001, it is rea-

Glucose (mg/dL)

150
sonable that the glucose (dash line) is much higher than ex-
pected since K c is so small that the injected drug is deficient.
100
On the contrary, when Kc = 0.01, which is stronger than ex-
pected, too much drug is injected which leads to the fluc-
50
tuations of glucose (dotted line). Furthermore, the concen-
tration of glucose is also lower (about 60 mg/dL) than floor
0 level of the expected bound. By a large scale simulation, it is
found that when Kc = 0.0045, the glucose (solid line) meets
−50 the required range better.
0 500 1000 1500
With given Kc = 0.0045, the effect of the integral pa-
Time (min)
rameter Ti in PID controller is illustrated in Figure 8. When
Kc = 0.01 Ti = 100000, which means that there is almost no inte-
Kc = 0.0045 gral effect, the output (solid line) meets the demands well.
Kc = 0.001 When Ti = 100, which enriches the integral effect a little,
Figure 7: The effect of parameters of Kc in PID controller. the controller considers the effect in a long duration, for ex-
ample, from beginning to current time point, the effect of
more doses of drug is taken into account for type I diabetes
patients, and the output (dotted line) of blood glucose con-
I is an integral role feedback, which benefits the steady per- centration is lower than the required. When Ti = 10, which
formance but does not contribute to the dynamic perfor- means the integral effect has been quite large, the output of
mance of the system; and D is a differential role of feedback. blood glucose concentration (dash line) is much worse and
Appropriate differential role of the plot can improve dynamic far lower than expected floor limitation.
performance significantly. Figure 9 shows the effect of parameter of Td under given
The parameters are adjusted and chosen based on the re- Kc and Ti . When Td = 30, which means the differential ef-
sults in [28]. Figure 5 depicts the control effects with P = fect is weak, the controller considers the rate of glucose little,
.0045, I = 10000, and D = 61. Obviously, the effect of PID the hysteresis is too strong, which makes the system react too
controller meets output requirements that the concentration late, and the output (dash line) cannot meet the requirement
of blood glucose is between 60 mg/dL and 140 mg/dL while and goes beyond the upbound sporadically. When Td = 61,
the output of the controller is shown in Figure 6 which is very the output (solid line) performs better within the expected
small. bound than during a period of 24 hours. When Td = 100,
We have also done a lot of experiments to explore the differential role is enhanced, the controller considers the rate
control effects of tuning the controller parameters Kc , Ti , and of glucose change, and the output of controller is too strong,
Td , as well as the sample time. Some of them are analyzed so that the fluctuations of glucose (dotted line) are more se-
as follows. Figure 7 shows the effect caused by changing the rious.
6 EURASIP Journal on Wireless Communications and Networking

Kc = 0.0045, Ti = 100000 Table 2: Rules for fuzzy logic controller.

160
150 Glucose
Rate
140 Overlow Good High Overhigh
130 Overlow Zero Zero Zero Zero
Glucose (mg/dL)

120 Low Zero Zero Zero Little

110 High Zero Norm Most Most
Highest Zero Norm Most Most
100
90
80 there exists a tradeoff between wireless communication cost
70 and the performance of control system [29].
60
50 3.2. Fuzzy logic controller
0 500 1000 1500
Time (min) Fuzzy logic control is also an advanced process control,
which imitates the logic of human thought, and much less
Td = 30
rigid than the calculations computers generally perform [30].
Td = 61
Td = 100 There are three steps for the process of a fuzzy logic algo-
rithm: fuzzification, rules, and defuzzification.
Figure 9: The effect of parameter of Td in PID controller.
(1) Fuzzification: the input of a controller is an exact
number, for example, the concentration of glucose is
100 mg/dL. What the fuzzification does is to fuzzify the
150 concentration such as low concentration, high concen-
140 tration, and proper concentration. Every exact number
130
has the weight of all these low, high, and proper con-
centrations.
120 (2) Rules: After defining the fuzzy concept of input, we
Glucose (mg/dL)

110 should make rules to decide what the output should

100 be: more drug, a little drug, or no drug. For example,
we define the following rule: if the concentration of
90
glucose is high and the rate of glucose is rising, then
80 the drug should be more.
70 (3) Defuzzification: After the rule, we get the output of
60 fuzzy concept, for example, more of 0.8 and little of
0.2. But the output which is the object model’s input
50
0 500 1000 1500 must be an exact number that needs to be defuzzified.
Time (s) By defuzzification, the output gets an exact number.

Sample time = 0.5 In this paper, it is assumed that there are two different
Sample time = 1 inputs of the concentration of glucose and the change rate of
Sample time = 5 concentration, and one output of the dose of drug. “overlow,”
“good,” “high,” and “overhigh” are defined for the concentra-
Figure 10: The effect of parameter of sample time in PID controller tion. The rate is “overlow,” “low,” “high,” and “highest.” The
under the best parameters of controller like in Figure 5. dose of drug is defined as “zero,” “little,” “norm,” “more,” and
“most.” Ten rules are defined such as
(1) if (rate is overlow) then (dosage is zero);
(2) if (concentration is overhigh) and (rate is low) then
The effect of parameter of sample time in PID controller
(dosage is little);
is also analyzed in Figure 10 under the best parameters of
(3) if (concentration is overhigh) and (rate is highest) then
controller like in Figure 5. When the sample time equals to 5
(dosage is most).
minutes, the data cannot describe the system characteristics
properly and in time, and the output (dotted line) also has For rule 1, when the rate is overlow, the injection dose
fewer errors. When the sample time equals to 0.5 (dash line) should be chosen as zero. If we still inject some drug, the
and time equals to 1 (solid line), the result is much better. concentration will decrease so fast that it may reach below
A higher sample rate makes the data approach much closer 60 mg/dL. Therefore, we choose rule 1. For rule 2, when the
to real system, but if there is a wireless monitoring station concentration is overhigh and rate is low, the little dose is
for patients of diabetes, the data of glucose concnetration chosen. If the rate is low, we do not need more drug, while
of patines should be transmitted to the central station, and if the concentration is overhigh, we need drug to avoid the
Jiming Chen et al. 7

The effect of PID and fuzzy logic controller

350

300
0.5

Glucose (mg/dL)
0.4 250
Dosage

0.3
0.2 200
0.1
150
0
0.5 50 100
Rat 0 0
e tion
−0.5 −50 ncentra 50
Co 0 500 1000 1500
Figure 11: The surface viewer of fuzzy logic controller. Time (min)

Fuzzy logic controller

PID controller
concentration staying high for too long. Therefore, we Without controller
choose little for this situation. For rule 3, when the concen- Figure 12: The system output of glucose concentration under PID
tration is overhigh and the rate is high, the dose will be given and fuzzy logic controller.
with the most quantity, since we need more insulin to de-
crease the concentration.
The ten rules also can be described in detail in Table 2. 0.7
We can then get the surface viewer which depicts the re-
lationship between the input and the output in Figure 11. 0.6
After discussing the fuzzy logic controller, we can see the
Controller’s output (10 U)

effect in Figure 12. As illustrated in this figure, the concen- 0.5

tration of blood glucose is too high (may ≥ 300 mg/dL) for
0.4
the diabetes patients without any control strategies. When
the strategies of continuous control are applied to the closed-
0.3
loop system, the output value of PID is bounded at about 66–
135 mg/dL while the output of fuzzy logic control is about 0.2
71–128 mg/dL. It is shown that fuzzy logic controller is simi-
lar to, but gets a little better performance than, PID. 0.1
As for the output of fuzzy logic controller that is corre-
sponding to the dose of drug, it can be found in Figure 13 0
0 500 1000 1500
that the dose of drug is less than 10 U and can be injected
Time (min)
automatically from time to time.
Figure 13: The controller output under fuzzy logic controller.
3.3. The effect of unexpected disturbance
Unexpected disturbance may happen, for example, a patient
might eat an apple in nonmeal time, and this should be con- shown in Figure 15. Therefore, we suggest that the patients
sidered but obviously is difficult to deal with by using the can get a little food when the concentration of glucose is at
traditional discrete time methods. Taking this into account, the point of vale, while it is dangerous to eat any more food
simulation experiments, shown in Figures 14-15, are pre- when the concentration is the peak point.
sented to test the robustness of the two control strategies. Above all, the robustness of the two control strategies is
In Figure 14, we put the unexpected disturbance at the vale quite desirable for the system with small unexpected distur-
point of concentration of glucose. Both the PID controller bance. While the unexpected disturbance happens at the vale
and fuzzy logic controller perform better when the amplitude point of the concentration of glucose for both two strategies,
of the unexpected disturbance is not too large. But when the the confined amplitude is about 30 mg/dL which meets the
amplitude reaches around 30 mg/dL, the output of closed- required concentration of glucose, that is, between 60 mg/dL
loop system will be out of the expected bound, especially, and 140 mg/dL, as shown in Figure 14. On the other hand,
when the amplitude is lower than 60 mg/dL. the confined amplitude would decrease to about 15 mg/dL
It should also be observed that if the unexpected distur- for the PID controller and 3.7 mg/dL for the fuzzy logic con-
bance occurs at a local peak concentration of glucose, the troller while the disturbance happens at the peak point of
output becomes worse, especially, for the fuzzy logic con- the concentration of glucose, as shown in Figure 15. There-
troller. It is quite reasonable that the larger the amplitude of fore, the robustness of PID controller seems a little better
the disturbance becomes, the worse the result is obtained, as than fuzzy logic controller. In other words, the PID and Fuzzy
8 EURASIP Journal on Wireless Communications and Networking

The unexpected disturbance The unexpected disturbance

Glucose (mg/dL)

30 30

Glucose (mg/dL)
20 20

10 10
0 0
0 500 1000 1500 0 500 1000 1500
Time (min) Time (min)
Amplitude 7.5 Amplitude 3.7
Amplitude 22 Amplitude 15
Amplitude 30 Amplitude 22
(a) (a)

The output of PID control strategy with unexpected disturbance The output of PID control strategy with unexpected disturbance
140 150
Glucose (mg/dL)

120

Glucose (mg/dL)
100
100
80
50
60
40
0 500 1000 1500 0
0 500 1000 1500
Time (min)
Time (min)
Amplitude 0 Amplitude 22
Amplitude 0 Amplitude 15
Amplitude 7.5 Amplitude 30
Amplitude 3.7 Amplitude 22
(b)
(b)
The output of fuzzy logic control strategy The output of fuzzy logic control strategy
with unexpected disturbance with unexpected disturbance
140 150
Glucose (mg/dL)

Glucose (mg/dL)

120
100
100
80
50
60
40 0
0 500 1000 1500 0 500 1000 1500
Time (min) Time (min)
Amplitude 0 Amplitude 22 Amplitude 0 Amplitude 15
Amplitude 7.5 Amplitude 30 Amplitude 3.7 Amplitude 22
(c) (c)
Figure 14: The robustness test while the unexpected disturbance Figure 15: The robustness test while the unexpected disturbance
occurs at the vale point of the concentration of glucose. occurs at the peak point of the concentration of glucose.

logic controllers have their advantages for different kinds of between insulin and glucose, a general plant model is pre-
patients, and the PID controller is more appropriate for those sented. In order to deal with the resulting nonlinear control
who often take some saccharated food unexpectedly, while problem, two different control strategies, PID controller and
the fuzzy is a better one for those taking food on schedule. Fuzzy logic controller, are presented and well analyzed. Based
on our simulation experiments, both strategies meet the ex-
4. CONCLUSIONS pected objective, that is, maintaining the blood glucose in the
permitted bound.
In this paper, the problem of continuous drug infusion for
diabetes therapy was considered. Firstly, a continuous drug- ACKNOWLEDGMENTS
insulin model for closed-loop control system was proposed,
exploring the accumulative effects from drug to insulin. Then The work of Zhejiang University is partially supported by the
based on the classical Bergman model depicting the relation Nature Science Foundation of China (NSFC) under Grant
Jiming Chen et al. 9

60604029, Nature Science Foundation of Zhejiang Province cose control,” in Proceedings of the American Control Confer-
under Grant Y106384, Joint Funds of NSFC-Guangdong un- ence, vol. 5, pp. 3483–3487, Chicago, Ill, USA, June 2000.
der Grant U0735003, the Specialized Research Fund for the [15] K. H. Kienitz and T. Yoneyama, “A robust controller for in-
Doctoral Program of Higher Education (SRFDP) (Grant no. sulin pumps based on H-infinity theory,” IEEE Transactions on
20050335020) and 111 Projects. The work of Professor Yang Biomedical Engineering, vol. 40, no. 11, pp. 1133–1137, 1993.
Xiao is partially supported by the US National Science Foun- [16] F. Andrianasy and M. Milgram, “Applying neural networks
dation (NSF) under Grant CNS-0716211. to adjust insulin-pump doses,” in Proceedings of the 7th IEEE
Workshop on Neural Networks for Signal Processing (NNSP ’97),
pp. 182–188, Amelia Island, Fla, USA, September 1997.
REFERENCES [17] N. Ghevondian and H. Nguyen, “Using fuzzy logic reason-
ing for monitoring hypoglycaemia indiabetic patients,” in Pro-
[1] T. Mandrup-Poulsen, “Recent advances: diabetes,” British ceedings of the 19th Annual International Conference of the IEEE
Medical Journal, vol. 316, no. 7139, pp. 1221–1225, 1998. Engineering in Medicine and Biology Society (IEEE/EMBS ’97),
[2] D. Takahashi, Y. Xiao, F. Hu, and M. Lewis, “A survey of vol. 3, pp. 1108–1111, Chicago, Ill, USA, October-November
insulin-dependent diabetes part I: therapies and devices,” to 1997.
appear in International Journal of Telemedicine and Applica- [18] D. U. Campos-Delgado, M. Hernández-Ordoñez, R. Femat,
tions. and A. Gordillo-Moscoso, “Fuzzy-based controller for glucose
[3] American Diabetes Association, “Standards of medical care regulation in type-1 diabetic patients by subcutaneous route,”
for patients with diabetes mellitus,” Diabetes Care, vol. 26, pp. IEEE Transactions on Biomedical Engineering, vol. 53, no. 11,
S33–S50, 2003. pp. 2201–2210, 2006.
[4] “National diabetes clearinghouse: diabetes statistics,” NIH [19] M. E. Fisher and K. L. Teo, “Optimal insulin infusion resulting
Publication no. 063892, November 2005. from a mathematical model of blood glucose dynamics,” IEEE
[5] The Diabetes Control Complications Trial Research Group, Transactions on Biomedical Engineering, vol. 36, no. 4, pp. 479–
“The effect of intensive treatment of diabetes on the develop- 486, 1989.
ment and progression of long-term complications in insulin- [20] C. Cobelli and A. Mari, “Control of diabetes with artificial sys-
dependent diabetes mellitus,” The New England Journal of tems for insulin delivery - algorithm independent limitations
Medicine, vol. 329, no. 14, pp. 977–986, 1993. revealed by a modeling study,” IEEE Transactions on Biomedi-
[6] A. L. Aalders, F. J. Schmidt, A. J. M. Schoonen, I. R. Broek, cal Engineering, vol. 32, no. 10, pp. 840–845, 1985.
A. G. F. M. Maessen, and H. Doorenbos, “Development of [21] R. Dudde, T. Vering, G. Piechotta, and R. Hintsche,
a wearable glucose sensor: studies in healthy volunteers and “Computer-aided continuous drug infusion: setup and test of
in diabetic patients,” International Journal of Artificial Organs, a mobile closed-loop system for the continuous automated in-
vol. 14, no. 2, pp. 102–108, 1991. fusion of insulin,” IEEE Transactions on Information Technol-
[7] M. J. Muehlbauer, E. J. Guilbeau, B. C. Towe, and T. A. Bran- ogy in Biomedicine, vol. 10, no. 2, pp. 395–402, 2006.
don, “Thermoelectric enzyme sensor for measuring blood [22] R. Bellazzi, G. Nucci, and C. Cobelli, “The subcutaneous
glucose,” Biosensors and Bioelectronics, vol. 5, no. 1, pp. 1–12, route to insulin-dependent diabetes therapy: closed-loop and
1990. partially closed-loop control strategies for insulin delivery
[8] M. J. Tierney, J. A. Tamada, R. O. Potts, L. Jovanovic, and S. and measuring glucose concentration,” IEEE Engineering in
Garg, “Clinical evaluation of the Gluco Watch Biographer: a Medicine and Biology Magazine, vol. 20, no. 1, pp. 54–64, 2001.
continual, noninvasive glucose monitor for patients with dia- [23] C. Owens, H. Zisser, L. Jovanovic, B. Srinivasan, D. Bonvin,
betes,” Biosensors and Bioelectronics, vol. 16, no. 9–12, pp. 621– and F. J. Doyle III, “Run-to-run control of blood glucose con-
629, 2001. centrations for people with type 1 diabetes mellitus,” IEEE
[9] K. Jungheim, K. J. Wientjes, I. Heinemann, V. Lodwig, T. Transactions on Biomedical Engineering, vol. 53, no. 6, pp. 996–
Koschinsky, and A. J. Schoonen, “Subcutaneous continuous 1005, 2006.
glucose monitoring,” Diabetes Care, vol. 24, pp. 1696–1697, [24] R. S. Parker and F. J. Doyle III, “Control-relevant modeling in
2001. drug delivery,” Advanced Drug Delivery Reviews, vol. 48, no. 2-
[10] Z. Trajanoski, G. A. Brunner, L. Schaupp, et al., “Open-flow 3, pp. 211–228, 2001.
microperfusion of subcutaneous adipose tissue for on-line [25] R. N. Bergman, L. S. Phillips, and C. Cobelli, “Physiologic
continuous ex vivo measurement of glucose concentration,” evaluation of factors controlling glucose tolerance in man.
Diabetes Care, vol. 20, no. 7, pp. 1114–1121, 1997. Measurement of insulin sensitivity and β -cell glucose sensitiv-
[11] R. S. Parker, F. J. Doyle III, and N. A. Peppas, “A model-based ity from the response to intravenous glucose,” Journal of Clin-
algorithm for blood glucose control in type I diabetic pa- ical Investigation, vol. 68, no. 6, pp. 1456–1467, 1981.
tients,” IEEE Transactions on Biomedical Engineering, vol. 46, [26] F. Chee, A. V. Savkin, T. L. Fernando, and S. Nahavandi, “Op-
no. 2, pp. 148–157, 1999. timal H∞ insulin injection control for blood glucose regula-
[12] P. Dua, V. Dua, and E. N. Pistikopoulos, “Model based control tion in diabetic patients,” IEEE Transactions on Biomedical En-
for drug delivery systems,” in Proceedings of the 3rd IASTED gineering, vol. 52, no. 10, pp. 1625–1631, 2005.
International Conference on Biomedical Engineering, pp. 691– [27] C. Ionescu, S. Zlate, and R. D. Keyser, “An approach to con-
696, Innsbruck, Austria, February 2005. trol the blood glucose level in diabetic patients,” in Proceedings
[13] M. Mahfouf, A. J. Asbury, and D. A. Linkens, “Uncon- of the IEEE International Conference on Automation, Quality
strained and constrained generalised predictive control of and Testing, Robotics, vol. 1, pp. 48–53, Cluj-Napoca, Roma-
depth of anaesthesia during surgery,” Control Engineering nia, May 2006.
Practice, vol. 11, no. 12, pp. 1501–1515, 2003. [28] G. F. Franklin, J. D. Powell, and A. Emami-Naeini, Feed-
[14] R. S. Parker, E. P. Gatzke, and F. J. Doye II, “Advanced model back Control of Dynamic Systems, Prentice-Hall, Upper Saddle
predictive control (MPC) for type I diabetic patient blood glu- River, NJ, USA, 4th edition, 2002.
10 EURASIP Journal on Wireless Communications and Networking

[29] E. Çayirci, T. Çöplü, and Ö. Emiroğlu, “Power aware many to

many routing in wireless sensor and actuator networks,” in
Proceedings of the 2nd European Workshop on Wireless Sensor
Networks (EWSN ’05), pp. 236–245, Istanbul, Turkey, January-
February 2005.
[30] W. Pedrycz, Fuzzy Control and Fuzzy Systems, Research Studies
Press, Baldock, UK, 1989.