Thorough Testing of Predictive Functional Control on a Well Validated In

Silico Model of Type I Diabetic Patients

German Campetellia , Mercedes Lombartec , Alfredo Rigallic , Marta S. Basualdoa,b
a

Computer Aided for Process Engineering Group (CAPEG). French-Argentine International Center for Information
and Systems Sciences (CIFASIS-CONICET-UNR-AMU). 27 de Febrero 210 bis, S2000EZP Rosario, Argentina
b
Universidad Tecnol
ogica Nacional (UTN). Facultad Regional Rosario (FRRo). Zeballos 1341, S2000BQA Rosario,
Argentina - campetelli, [email protected]
c
Bone Biology Laboratory, School of Medicine, Rosario National University (UNR-CONICET), Rosario, Argentina

Abstract
Diabetes technology has been focused since three decades ago on developing the Artificial Pancreas
(AP) through several closed-loop control algorithms linking glucose measurements and insulin delivery. This work presents the first well testing capabilities of the Predictive Functional Control (PFC)
algorithm for deciding about the correct insulin dosage under meals intake disturbances in the context
of AP.
The study is done by applying the PFC in a well validated model of the endocrine system, as
a substitute to animal trial. The tests were done in silico using the parameters of 30 patients: 10
children, 10 adolescents, and 10 adults. To realistically represent the full closed loop system, a model
of a subcutaneous glucose sensor was added and the constraints related to the insulin pump were taken
into account by the PFC design.
The performance of the controller is evaluated by means of the Control Variability Grid Analysis
(CVGA) technique under different scenarios and finally it is confronted with the Adaptive Predicted
Control with Robust Filter (APCwRF) strategy showing satisfactory results for all the studied patients.
Keywords: Artificial Pancreas, Diabetes Technology, PFC, 30 Virtual Patients, CVGA.
1. Introduction

total number of people with diabetes is projected

to rise from 382 million in 2013 to 592 million in
Diabetes Mellitus is a disorder of the metabolism 2035 (International Diabetes Federation, 2013).
where either insufficient insulin is produced by the
Nowadays, to treat this disease, diabetic pabeta cells in the pancreas, or the body is unable
tients measure their blood sugar content by prickto effectively utilize that insulin. As a result, glu- ing their fingers several times a day and inject
cose cannot be transported to the cells, leading to doses of insulin accordingly. From a control point
dangerously blood glucose incursions out of the of view, this is an open loop method that tries to
healthy levels. Untreated over time, high blood correct blood glucose no more than 6 times a day,
glucose levels can lead to costly complications and
usually before having a meal. The AP is thought
low blood glucose can lead to death. It is a very
to regulate glucose content in blood in real time,
frequent chronic disease that in the last years has trying to achieve a dynamic behavior as closed as
reached the proportion of an epidemy. The preva- possible to a healthy human pancreas would do.
lence of diabetes for all age-groups worldwide was
This could be done by means of a blood glucose
estimated to be 9.9% in 2030 by the International sensor accurate enough to give its measurement in
Diabetes Federation (IDF Diabetes Atlas). The real time, an insulin pump that delivers the corPreprint submitted to Industrial & Engineering Chemistry Research

September 16, 2014

rect dosage of insulin that the body needs, and the

control algorithm, which is responsible for calculating this amount of insulin in real time. It has
been shown that by the use of an AP during 10
years is possible to achieve savings of about USD
23 million (to Medicare in USA)and extrapolating
to 25 years can be saved USD 1.9 billion (OGrady
et al., 2011).
To date, many control algorithms have been
tested. PID (proportional integral derivative) (Ramprasad et al., 2004) and MPC (model predictive
control) (Hovorka et al., 2004; Magni et al., 2009;
Campetelli et al., 2013a) control laws are among
the most well-known methodologies proposed in
literature. However, model-based control strategies have been used with more encouraging outcomes in tighter regulation of blood glucose levels. The knowledge incorporated by the models
in these types of controllers is what makes them
more appealing. It is worth mentioning that other
types of control algorithms have been tested too
like robust H (Parker et al., 2000).
The PFC corresponds to the family of MPC
and it has been used many times in very different industrial applications with excellent results.
However, the first reported results were given by
(Daz et al., 2005) with promising results but employing the older version of Cobellis model (Carson et al., 1982). Based on these results, the authors propose here a most rigorous analysis taking
into account the last usable techniques for comparison purposes. Particularly, this type of controller is recognized in industry to be more intuitive for the users, has a great capacity to handle nonlinear systems, unstable and with large
dead times, these are inherent characteristics of
diabetic subjects. Another remarkable property
of PFC methodology is the control zone which
means that the closed loop time response parameter can vary according to the allowed error (%
- control zone) which is constrained to be inside
an expected range. This is very suitable to keep
the blood glucose concentration inside the desired
healthy zone because change the tuning parameter on line to achieve this important purpose because it is closely related with the health care of
diabetic persons. Hence, this property prevents
2

dangerous excursions outside the healthy range

and gives to this control algorithm more versatility for diabetes care. The preliminary results obtained with the validated model and considering
the CVGA for evaluating the PFC performance
was presented at (Campetelli et al., 2010) during
the preclinical trial. In the present work, the usefulness of this controller is in-depth proven facing
different adverse situations and confronted to the
specific MPC strategy given at (Campetelli et al.,
2013a).
This paper is organized firstly presenting the
subjects under study. After that, the sensor model
and the In Silico Trial are described. Then, a brief
description of the PFC algorithm is included. After that, the specific consideration to implement
PFC in the context of AP such as the development of the models for the controller and the adjustment of its parameters are explained. Finally,
the PFC is tested with and without the in silico
sensor, introducing errors in the internal models
and other realistic scenarios related to the meal
announcement of the patients in order to evaluate if it is robust enough to perform successfully in
the AP context. A rigorous comparison between
PFC and APCwRF on the 30 virtual patients under the same scenario is included at the end of
the paper to achieve a real dimension of the PFC
potentiality.
2. Subjects
The algorithm robustness is tested here on 30
in silico type 1 diabetes mellitus (T1DM) patients
through a dynamic simulation environment which
is called UVa/Padova Simulator (Kovatchev et al.,
2009). This well-known mathematical model of
the endocrine system is based on the pioneering
works of Dr. Claudio Cobelli, who has been focused on the modeling of the endocrine system
since more than 30 years Carson et al. (1982).
The full model was constructed based on information of 300 patients (100 adults, 100 adolescents and 100 children) and it was approved by
the FDA in 2008 to test control algorithms for the
artificial pancreas instead of using animal models
(Campetelli et al., 2013b). It is worth to men-

Adult #3

tion that the authors have previously tested the

APCwRF strategy on the same platform.
Plasma Glucose [mg/dl]

300

3. The Subcutaneous Sensor Model

In order to simulate more realistically the behaviour of a diabetic patient using an AP (Kovatchev et al., 2009), a model of the sensor was
incorporated. The in silico subcutaneous sensor
used in the simulations is the one developed by
(Breton and Kovatchev, 2008). After generating a
random calibration error, the components of sensor error can be modeled as:

200
Loop Closed

150
100
50
18

21

24

03

06
09
12
Time [hr]

15

18

21

24

Figure 1: Nominal scenario for adult patient. It can be

seen the upper bound (300 mg/dl) and the lower bound
(70 mg/dl) in accordance with the CVGA and the sensor
noise superimposed in gray.

1. Blood-to-interstitium glucose transport described by the equation:

1
IG
=
(IG BG);
t

250

3. At 7:30 Day 2, the patient has breakfast

lasting about 2 min with a carbohydrate
(CHO) content of 50 grams.
4. At approximately noon (12:00) Day 2, the
patient takes a lunch meal containing 65
grams CHO. Meal duration is 15 min.
5. At 18:00 Day 2, the patient takes a dinner
meal containing 80 grams CHO. Meal duration is 15 min.

(1)

where IG is the interstitial and BG is plasma

glucose concentration, and represents the
time lag between the two fluids;
2. Noise of the sensor, which is non-white (Gaussian). We therefore use ARMA process for
its modeling:
This scenario changes for adolescents and children just in the amount they eat (Adolescents:
e1 = v1
(2)
40/50/65 grams; Children: 25/30/40 grams).
en = 0.7(en1 + vn )
(3)
5. Predictive Functional Control (PFC)
en
);
(4)
n = + sinh(

with vn (0,1), i.i.d.. The sensor noise is n ,

which is driven by the normally distributed time
series en . The parameters , , , and are the
Johnson system (SU - unbounded system) parameters corresponding to empirical noise distributions established in accuracy trials.

In this section only a brief summary of the

main concepts of PFC algorithm, related to its
adaptation to be useful in the context of the AP,
is presented. A more deep version about theoretical aspects and practical implementation of
PFC, can be found at Richalet and ODonovan
(2009). The PFC technique is the third generation of a family of Model Algorithmic Control.
PFC basically consists of four main elements such
as a process dynamic model, a reference trajectory
yr (n), a self-compensation of the predicted error
and a specific structure for the manipulated variable (Fig. 2). The future error between yr (n) and
the predicted output over the coincidence horizon
[H1 , H2 ] is estimated. A self compensation is done
accounting for the actual mismatch between real
data and model output. The estimation of the

4. The In Silico Preclinical Trial

The performance of PFC was tested during
24 hours by a virtual protocol (Fig. 1) based on
Patek et al. (2009). For an adult:
1. Admit state: Patient blood glucose steady
at 100 mg/dl at 18:00 Day 1.
2. Control loop is closed at 21:00 Day 1.
3

In the canonic form:

y(k + 1) = m xmi (k) + d xmd (k) + . . .

+ Kmi (1 m )u(k) + Kdi (1 d )d(k) (9)

ym (k) = xm (k)
Calling the inputs of the manipulated variable
u(n) (insulin from the pump) and the perturbation d(n) (a meal), the FOTD model response at
the coincidence point (n + H) becomes:
H
y(n + H) = m
xmi (n) + dH xmd (n) + . . .

H1
X

H1j
m
Kmi (1 m )u(j + n) + . . .

j=0

Figure 2: PFC principles for design for one coincidence

point.

H1
X

dH1j Kdi (1 d )d(j + n)

j=0

(10)
future error at the coincidence horizon by specific kind of extrapolation, allows to improve the
The model output is the sum of the free (L)
model prediction. Within PFC, feedforward and and forced (F) responses:
feedback control actions can be jointly designed
and constraints are taken into account in a very
ym (n + H) = yF (n + H) + yL (n + H)
(11)
natural way.
The control law for first order with time delay
H
(FOTD) models on both process between blood
ym (n + H) = m
ymi (n) + dH ymd (n) + . . .
H
glucose concentration (controlled variable) and in + Kmi (1 m
)u(n) + Kdi (1 dH )d(n) (12)
sulin infusion (manipulated variable)(Gmi ) and blood
glucose concentration affected by meal intake (perTherefore the control equation is obtained by
turbation) (Gdi ) will be developed. Meanwhile,
the following steps:
the glucose entering to the system in any standard
(n + H) = (n)H
(13)
way such as oral glucose tolerance test (OGTT),intra
3T
s
venous glucose tolerance test (IVGTT) or meal
(14)
= e CLT R
intake are considered as typical disturbances.
The reference trajectory, defines an expected
closed loop response to track the set point, it is
usually selected as a first order dynamic given by:

mi s

Kmi e
1 + Tmi s
Kdi edi s
Gdi =
1 + Tdi s

Gmi =

Applying Z-transform:
(
(1m )
Gmi (z) = Kmi (z
m)
m = exp(Ts /Tmi )

(5)

C(n+j)yr (n+j) = j (C(n)yp (n)), 0 j H

(15)
where C(n) is the set-point, yp (n) the real process output and is a parameter that represents
the exponential convergence of the algorithm, and
thus fixes the closed-loop behavior. Ts is the sampling time which was adopted as 5 minutes because it is a typical frequency of measure given
by the commercial sensor.

(6)

(7)

(1d )
Gdi (z) = Kdi (z
d)
d = exp(Ts /Tdi )

(8)
4

H
ymi (n)
m
Free responses
dH ymd (n)

The control algorithm is given by:

(16)

u(n) = K0 (n) + K1 ymi (n) + ...

(23)
... + K2 ymd (n) + K3 d(n) + K4 ym (n)


H
Kmi (1 m
)u(n)
Forced responses (17)
H
Kdi (1 d )d(n)

(n) = C(n) yp (n)
(18)
(n + H) = C(n) yr (n + H)

K0 =
K1 =

C(n) yr (n + H) = H (C(n) yp (n)) (19)

K2 =

and considering:
p (H) = (C(n) yp (n))(1 H )


(20)

K3 =

p (H) = m (H)
(C(n) yp (n))(1 H ) = ym (n + H) ym (n)
(21)
Then,

K4 =

(1 H )
H)
Kmi (1 m
H
m
H)
Kmi (1 m
dH
H)
Kmi (1 m
Kdi (1 dH )
H)
Kmi (1 m
1
H)
Kmi (1 m

(24)
(25)
(26)
(27)
(28)
(29)

Finally, if it is assumed that d(n) = ymd (n) =

0,

H
(C(n) yp (n))(1 H ) = m
ymi (n) + . . .

U (z)
K0 (z m )
=
(z)
(z 1)

H
+ dH ymd (n) + +Kmi (1 m
)u(n) + . . .

+ Kdi (1 dH )d(n) ym (n) (22)

(30)

As can be seen from eq. 30 the controller

transfer function has an implicit integrator which
guarantees zero tracking error for step inputs.
The forced response is calculated by assuming
that the input to the system is related to base
functions like the ones shown in Fig. 3. Typically, these functions are steps B1 (i) = 1; ramps
B2 (i) = i or parables B3 (i) = i2 . The choice of
the base functions defines the input profile and
can assure a predetermined behavior.

Figure 3: Base function for input and forced output calculations.

u(n + i) =

Nb
X

K (n)U BK (i)

(31)

K=1

For a system with delay, the error between real

data from the endocrine system and the model
predictions at instant (nTdm ) must be estimated
as is shown in Fig. 4. Tdm is the time delay, hence
the predicted output of the biological system is
given by:
yp (n) = ym (n) + e0ret (n)

(32)

e0ret = yP ret yM ret

(33)

(n) = (C(n) yp (n))

(34)

where:
5

Adult #3
25

Insulin infusion [U/h]

20

15

10

0
18

21

24

03

06
09
12
Time [hr]

15

18

21

24

Figure 5: External insulin infused to Adult #3.

compartment. As a consequence of this, the meal

disturbance has to be given as a filtered response
into the glucose compartment and not as a step response into the stomach compartment (Ellingsen,
2008).
To obtain a representative model useful for implementing the meal announcement, the mean of
all model parameters for each group of patients
was taken and the glucose rate of appearance of
each group was saved in a matrix. Then, the controller receives information of a mean absorption
profile. Another way of solving this problem could
be detecting when a patient receives a meal as
shown in Dassau et al. (2008).
In our case, the relationship between insulin
infusion (manipulated variable) and blood glucose
(controlled variable) is named Gmi . Meanwhile
Gdi refers to the relationship between exogenous
glucose (Glucose rate of appearance Ra from a
meal) and blood glucose. Both models were found
to be well represented by first order with time delay. They were defined using identification techniques by exciting the inputs such as insulin delivery and in meal ingestion close to a nominal
condition. The step used depends on the group
studied. For the manipulated variable, having the
information of the Total Daily Insulin (TDI [U])
consumed by each patient, the mean value of all
patients was takenPinto account. For the pertur = Ra
i was calculated.
bation, the Ra

Figure 4: Prediction of the delayed response.

Taking into account equations 32 to 34 the

control action of the algorithm for the system with
delay is given by:
u(n) = K0 (n) + K1 ymi (n) + K2 ymd (n) + . . .
+ K3 d(n) + K4 ym (n) (35)
The parameters to be tuned for the PFC are:
number of coincidence points, in this case only
one is adopted (H), closed loop time response
(CLTR) of the reference trajectory, the control
zone considered so that CLTR moves linearly between two extremes values recognized as CLT R L
(low) and CLT R H (high), the transition zone
[%] that set the allowed zone for the controlled
variable expressed as Delta [%] with respect to
the set point value and constraints to the manipulated variable are also included by fixing maximum (Umax ), minimum (Umin ) and variations for
it [(dU/dt)max ] are adopted according to the insulin pump characteristics.
5.1. Models for PFC
For the specific case of the diabetes problem
treatment the PFC has three inputs, the glucose
measurement, the glucose set point (100 mg/dl
in our case) and the glucose rate of appearance
into the glucose compartment (Ra). The last input is only taken into account if the meal is previously announced. To avoid the nonlinearities
in the stomach compartment, the model for the
controller was linearized without considering this

5.2. PFC Tuning

First, the controller was adjusted to each individual patient according to the procedure described later in this section. It allows to see the
6

optimal performance achievable of PFC. Then,

the controller was tested over three groups of patients: 10 children, 10 adolescents, and 10 adults.
A set of parameters for the PFC was proposed to
each group of patients (See Table 1). This set of
parameters were obtained as follows:
1. for each patient, two first order models with
time delay were proposed (the insulin dosage
and meal intake as the disturbance):
Kmi emi s
1 + Tmi s
Kdi edi s
Gdi =
1 + Tdi s

Gmi =

Figure 6: Case Study 1: CVGA of PFC adjusted to each

patient using the sensor model.

(36)
(37)

an important role in the tuning of closed-loop glucose control algorithms and also in the comparison of their performances (Hinshaw et al., 2013).
So, it is a method for visualization of the extreme
glucose excursions caused by a control algorithm
in a group of subjects, with each subject represented by one data point for any given observation period. To differentiate among the groups of
patients, in all the CVGA plots they are represented as follows: adults with circles, adolescents
with diamonds, and children with pentagrams. In
this work, six case studies are presented where the
information at the top of each diagram is useful
to compare the achievements with each scenario
and strategy.

2. for each patient the CLTR L, CLTR H, H1

and H2 were computed as:
CLT R L = Tmi /2
CLT R H = 10Tmi
H1 = mi
H2 = mi + 3Tmi
3. then, the mean value of these parameters
were taken for each group (children, adolescents and adults) and the rest of the parameters were fixed as shown in Table 1. The
only parameter to be changed among the
patients was [(dU/dt)max ].
The parameter ([(dU/dt)max ]) represents the aggressiveness of the controller. If it is set in a low
value, the response of the controller is very soft.
Increasing it, the controller becomes more and
more aggressive. The adopted parameters for the
simulations shown in this work are included in Table 1. This table shows the parameters adopted
in cases study 3 (columns 2, 3 and 4) and 4 (fifth
column) as will be shown in the following section.

6.1. Case Study 1

In Figs. 1 and 5 the glucose and insulin profiles, obtained in the nominal scenario with the
PFC including the subcutaneous sensor model,
are reported for a patient of the group of adults
(#3). There can be seen superimposed the sensor signal in gray. In Fig. 6 the CVGA is constructed with the controller adjusted to each patient. As can be seen, the 30 subjects are in the
green zone, which means that they are not running major risks.

6. Results
The Control-Variability Grid Analysis (CVGA)
is a graphical representation of min/max glucose
values in a population of patients either real or
virtual. The CVGA provides a simultaneous assessment of the quality of glycaemic regulation in
all patients. As such, it has the potential to play

6.2. Case Study 2

In another study, the CVGA shown at Fig.
7is constructed under the assumption of using a a
perfect sensor. The controller tuning here is the
7

Table 1: Controller Parameters

Parameter
CLT R L,
CLT R H,
Delta,
Umin ,
Umax ,
( dU
) ,
dt max
Kmi ,
Tmi ,
mi ,
Kdi ,
Tdi ,
di ,
H,
H1 ,
H2 ,
Ts ,

Children
100.5
2010
30
0
150
dU
( dt )max
-1.899
201
104
42.243
146
7.9
1
104
707
5

Adolescents
184.5
3690
30
0
150
dU
( dt )max
-0.573
369
121
44.753
164.5
6.7
1
121
1228
5

Adults
167.5
3350
30
0
150
dU
( dt )max
-1.283
335
137.5
53.471
174.5
6.3
1
137.5
1142.5
5

same used case study 1. It can be seen that, although the 30 patients are in the safe zone, they
have all gone closer to the lower bound (to the
right - 70 mg/dl of blood glucose), with a bigger
risk of undergoing an hypoglycaemic episode. It
is remarkable that Case Study 1 is misleading because it seems that the controller performed better than in this case which is just a matter of
randomness (because of the sensor noise and error some patients crossed the line between zone
B and A). Hence, is very important to do firstly
the analysis of the sensor characteristics before
performing the experiments to have a real picture
about the quality of data is provided by this device. Particularly, a fault diagnosis and identification (FDI) study plays an important rol for diabetes care as it was demonstrated at (Campetelli
et al., 2011).

Mean
150.8
3016.7
30
0
150
dU
( dt )max
-1.2517
301.7
120.8
46.822
161.7
6.97
1
120.8
1025.8
5

Units
[min]
[min]
[%]
[pmol/min]
[pmol/min]
[pmol/min2 ]
[(mg/dl)/(pmol/min)]
[min]
[min]
[(mg/dl)/(mg/kg/min)]
[min]
[min]
[dimensionless]
[min]
[min]
[min]

Figure 7: Case Study 2: CVGA without the use of the

sensor model.

6.4. Case Study 4

In this case study, the controller parameters
were set the same to all patients. The mean value
of all tuning parameters, shown in the fifth column of Table 1 was used. Then, only the value
of [(dU/dt)max ] was tuned to each individual patient, as done before, achieving that all the patients were sent to the safety zone as can be seen
in Fig. 9. Even though the number of patients in
zone A is smaller than in the other case studies,
everybody is still in the green zone, which means
that they are under control.

6.3. Case Study 3

In this case study, the same controller parameters were adopted for each group of patients as
explained in Section 5.2. In Fig. 8 a CVGA performed with this PFC tuning is showed. It results
in a poorer performance of the controller than in
case study 1 but everybody is still in the safe region.
8

(a) CVGA with an advanced meal

announcement announcement

Figure 8: Case Study 3: CVGA using mean parameters in

the controller for each group of patients.

(b) CVGA with a delayed meal

announcement
Figure 10: Case Study 5: CVGA study of meal announcements.

6.5. Case Study 5

In this case study the effect of meal announcement is being examined. First, the case where the
patient gives the meal announcement 60 minutes
in advance to the controller can be seen in Fig. 10
(a). The summary given at the top of the graph
indicates that this is the best of all scenarios because 40 % of the patients are in zone A. On the
other hand, in Fig. 10 (b) is shown the case when
the patient forget to announce a meal on time and
do it with a delay of 60 minutes. Even though the
controller is capable of dealing with all the subjects correctly, keeping them on the healthy range,
the result is worse than the other previous cases.

Figure 9: Case Study 4: CVGA using mean parameters in

the controller for all patients.

6.6. Case Study 6

Finally the importance of giving the correct information to the control algorithm regarding the
quantity of food ingested is going to be evaluated.
In this case study amounts of food intake by the
patients do not match to the announced meals.
In the first picture, Fig. 11 (a), the patients eat
25 % less than the announced meal informed to
the controller. Then, the patients eat 25 % more
(Fig. 11 (b)), 50 % less (Fig. 11 (c)) and 50 %
more food than that announced (Fig. 11 (d)). As
can be seen, better results are obtained when the
given information is more accurate allowing the
controller deals better than in the opposite situation. This fact is reflected by the percentages
shown at the top of the corresponding CVGA diagrams.

(a) CVGA when the patients eat 25 % less

food

6.7. Comparison with the improved APCwRF

It was considered important to do a rigorous
comparison with another efficient MPC strategy
presented previously, APCwRF (Campetelli et al.,
2013a). The main control structure of APCwRF
involves a commutation between a linear timevarying robustness filter (RF) in the feedback path
of the control loop and an adaptive predictive
controller (APC). The decision of which of both
modes has to work is based on specific indicators
which are closely related to the state of the subject which is checked every sampling time.
According to the characteristics of the diabetes problem, it was necessary to introduce some
important modifications to the original algorithm
of APCwRF. The controller uses FIR models for
the predictions which are adapted on-line. Because of the nonlinearity and variability of the
patients endocrine system, three internal models were used, linearized on three different conditions taking into account hypoglycaemic state
(70 mg/dl), normal state (95 mg/dl) and hyperglycaemic state (151.3 mg/dl).
The adopted set-point trajectory for the APCwRF
is based on the blood glucose evolution of a healthy
subject. A control zone is proposed through a
variable weight factor (). Hence, each time the
diabetic subjects blood glucose is in the hyperhypoglycaemic (hazardous) range, changes its

(b) CVGA when the patients eat 25 % more

food

(c) CVGA when the patients eat 50 % less

food

10
(d) CVGA when the patients eat 50 % more
food
Figure 11: Case Study 6: CVGA study of wrong intakes
information.

7. Conclusions
This work presented a deep study to give the
first results about the PFC implementation for diabetes care, tested in a well-known mathematical
model of the glucose-insulin system, accounting
for the physiological variability of different patients of distinct age groups. The PFC demonstrated to be able of giving very good performance
in all the scenarios proposed here and behaved
better than a more sophisticated MPC strategy
like the APCwRF.
From the proposed tests for comparison purposes different realistic scenarios were analyzed.
In addition, the use of CVGA gives the opportunity to rigorously comparing with other control
strategies since it is helpful to handle the information of the 30 patients and to present a picture about the achievements of each one. The
fifth case studied, corresponding to the meal announcement given in advance, reported the highest performance since zone A achieved 40 %. As
an example of the dynamic behavior obtained by
the PFC implementation in Fig. 1 was shown the
evolution of a type 1 diabetic adult after the virtual protocol defined in Section 3. The insulin
infusion rate calculated by the PFC controller to
keep this patient within the normal glucose range
is shown in Fig. 5.
It must be remarked that the PFC results very
intuitive for selecting all the adjusting parameters
involved in this technique as was shown in Section 5.2. Additionally, since PFC has been widely
proved in industries its design is mature enough
to be a good candidate for AP implementation.
Hence, even though a more sophisticated implementation of this kind of MPC philosophy could
be applied, the results obtained here demonstrate
the PFC potentiality for these types of complex
problems. It has been proved that the robustness of the control algorithm is capable of dealing
with inter-patient variability and, to a certain extent, typical mistakes of the user of an AP. Taking the same parameters within the same group
of patients resulted in a very good performance
of the control algorithm. So, it gives more simplicity and seems to be robust enough to be use-

Figure 12: CVGA corresponding to 60 mins in advance

meal announcement using APCwRF. Picture taken from
Campetelli et al. (2013a).

value to reduce quickly the error between the desired trajectory and the blood glucose measure.
The APCwRF was also improved with Meal Announcement (60 minutes in advance) which is handled by a family of 3 FIR models generated by
the information of food intake (perturbation) as
input and the knowledge of blood glucose impact
(output).
In Fig. 12, the CVGA corresponding to the
application of APCwRF using a Meal Announcement strategy in advance, the same scenario analyzed for Case Study 5, can be seen. At the top of
Fig. 12 the summary of the results indicates that
APCwRF presents 30% of the patients inside zone
A (the best) meanwhile Fig. 10 reports 40% for
PFC in the same zone. Even though both results
are encouraging PFC reports a clear advantage
respect to APCwRF. Probably, the most important contribution on the extended APCwRF was
given through the introduction of a control zone
which was part of the PFC philosophy since its
original conception. This capability present on
both strategies could explain the successful comparison reported by PFC. As a summary of this
last test, the very good performance of PFC can
be achieved with the associated benefit of its simpler way to tune and more intuitive for the users
who are not trained on adjusting MPC strategies.
In addition, the PFC algorithm is already developed for industrial applications, so it is prepared
to demonstrate enough robustness to be implemented in the context of the artificial pancreas.
11

ful for a group of models corresponding to different patients. Moreover, even when the subjects
gave wrong information related to the amount of
food ingested and when the announcement was
given with delay, the PFC technique gave very
good results. Probably this good result obeys to
the control zone implementation. From the set of
30 patients considered here, the results are very
promising, however in future would be useful to
check this statement with more patients.

Carter, R.E., Cobelli, C., Kudva, Y.C., Basu, A., 2013.

Diurnal pattern of insulin action in type 1 diabetes. implications for a closed-loop system. Diabetes 62, 2223
2229.
Hovorka, R., Canonico, V., Chassin, L.J., Haueter, U.,
Massi-Benedetti, M., Orsini Federici, M., Pieber, T.R.,
Schaller, H.C., Schaupp, L., Vering, T., Wilinska, M.E.,
2004. Nonlinear model predictive control of glucose concentration in subjects with type 1 diabetes. Physiol.
Meas. 25, 905920.
International Diabetes Federation, 2013. Diabetes atlas
6th edition.
Kovatchev, B.P., Breton, M., Dalla Man, C., Cobelli, C.,
2009. In silico preclinical trials: A proof of concept in
closed-loop control of type 1 diabetes. J. Diabetes Sci.
Technol. 3, 4455.
Magni, L., Raimondo, D.M., Dalla Man, C., De Nicolao,
G., Kovatchev, B., Cobelli, C., 2009. Model predictive control of glucose concentration in type I diabetic
patients: An in silico trial. Biomed. Signal Process.
Control .
OGrady, M.J., John, P.M., Winn, A., 2011. Changes in
Medicare Spending for Type 1 Diabetes With the Introduction of the Artificial Pancreas. Technical Report.
OGrady Health Policy LCC.
Parker, R.S., Doyle III, F.J., Ward, J.H., Peppas, N.A.,
2000. Robust H glucose control in diabetes using a
physiological model. AIChE 46, 25372549.
Patek, S.D., Bequette, B.W., Breton, M., Buckingham,
B.A., Dassau, E., Doyle III, F.J., Lum, J., Magni, L.,
Zisser, H., 2009. In silico preclinical trials: Methodology
and engineering guide to closed-loop control in type 1
diabetes mellitus. J. Diabetes Sci. Technol. 3, 269282.
Ramprasad, Y., Rangaiah, G.P., Lakshminarayanan, S.,
2004. Robust PID controller for blood glucose regulation in type I diabetics. Ind. Eng. Chem. Res. 43,
82578268.
Richalet, J., ODonovan, D., 2009. Predictive Functional Control - Principles and Industrial Applications.
Springer.

References
Breton, M., Kovatchev, B., 2008. Analysis, modeling, and
simulation of the accuracy of continuous glucose sensors. J. Diabetes Sci. Technol. 2, 853862.
Campetelli, G., Lombarte, M., Basualdo, M.S., Rigalli, A.,
2013a. Extended adaptive predictive controller with robust filter to enhance blood glucose regulation in type i
diabetic subjects. Computers & Chemical Engineering
59, 243 251. Selected papers from ESCAPE-22 (European Symposium on Computer Aided Process Engineering - 22), 17-20 June 2012, London, UK.
Campetelli, G., Lombarte, M., Basualdo, M.S., Rigalli,
A., 2013b. Scaling the endocrine system from rats to
humans, in: Kraslawski, A., Turunen, I. (Eds.), 23rd
European Symposium on Computer Aided Process Engineering, pp. 145150.
Campetelli, G., Zumoffen, D., Basualdo, M., 2011. Improvements on non-invasive blood glucose biosensors using wavelets for quick faults detection. Journal of Sensors 2011, 11.
Campetelli, G., Zumoffen, D., Basualdo, M., Rigalli, A.,
2010. Testing PFC controller on a well validated in silico
model of a type I diabetic patient, in: 9th International
Symposium on Dynamics and Control of Process Systems, Leuven, Belgium.
Carson, E.R., Cobelli, C., Finkelstein, L., 1982. The Mathematical Modeling of Metabolic and Endocrine Systems. Wiley-Interscience.
Dassau, E., Zisser, P.H.H., Jovanovi`e, B.A.B.L., Dalla
Man, C., Cobelli, C., Veh, J., Doyle III, F.J., 2008. Implications of meal library & meal detection to glycemic
control of type 1 diabetes mellitus through MPC control, in: 17th IFAC World Congress, Seoul, Korea.
Daz, R.G., Aguerre, S., Roldan, G., Basualdo, M., 2005.
A decision support system based on a closed loop PFC
applied for type 1 diabetes, in: 16th IFAC World
Congress, Praha. Czech Rep.
Ellingsen, C., 2008. Safety constraints in an artificial cell: an implementation of Model Predictive Control
MPC with Insulin-on-Board IOB. Masters thesis. Norwegian University of Science and Technology.
Hinshaw, L., Dalla Man, C., Nandy, D.K., Saad, A.,
Bharucha, A.E., Levine, J.A., Rizza, R.A., Basu, R.,

12