Available online at www.sciencedirect.

com
ScienceDirect

Mathematics and Computers in Simulation 194 (2022) 198–209

www.elsevier.com/locate/matcom

Original articles

Application of model reference adaptive control and modified Smith

predictor to control blood glucose in type 1 diabetic patients
Sina Taherinasaba ,∗, Saedeh Soleimaniaslb , Saleh Taherinasaba
a Department of Electrical and Computer Engineering, Fakhrodin Asad Gorgani Institute of Higher Education, Gorgan, Iran
b Department of Automatic Control and Robotics, École Centrale de Nantes, Nantes, France

Received 7 January 2020; received in revised form 18 May 2021; accepted 25 May 2021
Available online 4 June 2021

Abstract
In this paper, a new adaptive control structure for time-delayed systems has been proposed, combining the model reference
adaptive control with the modified Smith predictor (SP). Due to the extensive variability of the patients’ metabolism, an in
silico clinical trial consisting of 30 patients is utilized to simulate the personal variability in the glucose control system. All the
patients have random changes and sinusoidal oscillations in their parameters of the Dalla Man glucose-insulin model. Based
on the quantitative and qualitative indicators, the performance of the proposed algorithm is compared to a PID controller of
the Smith predictor structure. The results of the simulation show that the proposed control scheme is successful against the
interpatient variation in fasting conditions, meal disruption rejection, and in terms of robustness. Insulin therapy often leads
to high fluctuations in blood glucose and hypoglycemic events in patients with type 1 diabetes. On the other hand, a closed
blood glucose loop control with an artificial pancreas can improve the patients’ quality of life. In this paper, the physiological
behaviour of the system is modelled inversely using the daily patients’ data acquired by the GIM simulator.
⃝c 2021 International Association for Mathematics and Computers in Simulation (IMACS). Published by Elsevier B.V. All rights
reserved.
Keywords: Type 1 diabetes mellitus; Model reference adaptive control; Time-delayed system; Smith predictor; Artificial pancreas; Hypoglycemic

1. Introduction
Blood glucose regulation in diabetic patients is one of the most important control issues in the field of medical
systems. In the absence of a reliable glucose sensor, the control strategy cannot go beyond the open-loop strategy. In
recent years, with the development of subcutaneous glucose sensors, a new path has been addressed for closed-loop
control strategies. As it is known, the artificial pancreas consists of a set of sensors, a control algorithm, and a
subcutaneous insulin pump, which enables it to regulate the blood glucose concentrations in diabetic patients like
a healthy pancreas. Also, there is a kind of hybrid closed-loop system using artificial intelligence to analyse data
in real-time, equipped with a locked-down smartphone providing the main algorithm and the user interface [13].
In the case of glycemic control, control strategies are based on adaptive modelling [2,11], optimal control [1,12],
predictive control [5,8], and infinite soft control [6,7], which are evaluated by simulation and experimental studies.
In these methods, the main limitation in the performance of the controller depends on the accuracy of the utilized
∗ Corresponding author.
E-mail address: [email protected] (S. Taherinasab).

https://doi.org/10.1016/j.matcom.2021.05.033
0378-4754/⃝ c 2021 International Association for Mathematics and Computers in Simulation (IMACS). Published by Elsevier B.V. All rights
reserved.
S. Taherinasab, S. Soleimaniasl and S. Taherinasab Mathematics and Computers in Simulation 194 (2022) 198–209

model in representing the insulin/glucose dynamics. Most of the insulin/glucose models presented in the articles
are physiological chamber models which mostly represent a moderate patient under certain conditions [4,19,20,24].
High levels of nonlinearity and a large number of identified parameters make it difficult to determine a unique model
for each patient. Many studies have been done on nonlinear systems with or without time delays. For example, some
research provided the abstract characterization of a class of delay-free nonlinear systems [3]. Another limitation of
the model-based controllers is the fact that they require detailed models of perturbation [9,20] and food consumption
information of the patients [10,25].
In healthy individuals, pancreatic beta is responsible for insulin secretion. The PID control strategy (which lacks
a model) mimics the pancreatic phase I and phase II responses. Steele et al. tested a discrete-time PID controller
on 10 patients with type 1 diabetes. Clinical test results showed a significant decrease in hypoglycemia (Blood
glucose lower than 70 mg/dl) incidence in patients. However, hyperglycemia (Blood glucose more than 120 mg/dl)
incidence was reported after eating. The reason can be the fact that the PID algorithm, like many other algorithms,
cannot compensate for the inherent system delays (i.e., subcutaneous insulin absorption and subcutaneous glucose
measurement) [26]. In this work, for the first time, a novel adaptive control structure has been proposed for time-
delayed systems which is a combination of the model reference adaptive control with modified Smith predictor due
to extensive variability among patients in metabolism.
The remainder of the paper is organized as follows: Section 2 describes all the material and methods used in this
work. This is followed by the overview of a physiological model of type 1 diabetes and a model-reference adaptive
control strategy with the Smith predictor in Sections 2.1 and 2.2. The modelling concept is presented in Section 2.3.
MRAS design using the Lyapunov theory is explained in 2.4. Improving perturbation repulsion by modified Smith
predictor and also simulation procedure are investigated in Sections 2.5 and 2.6 respectively. In Section 3, the final
results are explored. Finally, the conclusion part is presented in Section 4.

2. Materials and methods

Due to the inherent dynamic properties of the system, a control algorithm with the following capabilities is
required:
• Ability to anticipate the time delays in the blood glucose problem and provide a control signal to prevent
excessive fluctuations in blood sugar levels.
• Adaptive performance to resist the dynamic variability of the system.
Many studies have been performed to find an appropriate control algorithm to guarantee hypoglycemia avoidance.
For example, Resalat et al. [17] have proposed a solution that consists of a state feedback control law with a
real-time hyperglycemia correction, benefiting from two algorithms of Proportional–Integral–Derivative (PID) and
Model-Predictive Control (MPC). However, in this paper, the modified Smith Predictor structure and the Model
Reference Controller (MR) are combined to provide the requirements. In the Smith Predictor structure, the system
model is used to predict the delayed response of the system and to generate a feedback signal corresponding to
the error between the system output and the delay model output. The results have been obtained using the black
box identification model. This method does not require a detailed display of insulin/glucose dynamics, modelling of
perturbation, or entering the system inputs manually. The blood glucose is only predicted and controlled based on
the patients’ inputs, outputs, and black box identification. To compensate for the PID approach weakness against the
inherent system delays, the controller is merged with Smith predictor. Also, the overall performance is compared
with the open-loop controller and Smith predictor reference model controller.

2.1. A physiological model of type 1 diabetes

Modelling type 1 diabetes has begun in the 1980s. In general, the most effective and widely used models during
the last five years include Dalla Man [4], Hovorka [20], and the modified Bergman model [19]. Also, Liu et al. [16]
have introduced a glucose-insulin model that fits the clinical data from in-and-outpatients for two days. Most virtual
models are validated among a small number of individuals and only via measuring the plasma glucose and insulin
concentrations which is a big limitation. However, the Dalla Mann model, in addition to the plasma insulin and
glucose concentrations, has been validated among a large statistical population (204 people) based on different
amounts of insulin and glucose such as intrinsic glucose production rate, glucose uptake rate, glucose uptake, and
199
S. Taherinasab, S. Soleimaniasl and S. Taherinasab Mathematics and Computers in Simulation 194 (2022) 198–209

Fig. 1. The combination of the model reference adaptive control structure and the Smith predictor.

insulin release. Also, the specific salient feature of the Dalla Man model is the accurate physiological description of
glucose passage through the stomach and intestine, its digestion and its absorption, which is incompletely described
in other models. The FDA has adopted this model as an alternative to animal tests for glucose controllers [22]. In
general, a model is needed to predict system dynamics for these two reasons:
1. It makes it possible to create a virtual patient to test the controller virtually before a clinical trial.
2. The control algorithm can use the process model as an internal model to predict the system output.
In this paper, the Dalla Man model is used to test the performance of a controller to create a virtual disease.
However, instead of using this model, inverse modelling is used in the controller design process. This is because
that the Dalla Man model identification for any particular patient is not easy in the real world and is only possible
through expensive tests. Also, the actual patient parameters change throughout the day and the parameters of this
model have to be updated because of its complexity.

2.2. Model-reference adaptive control strategy with the Smith predictor

The model-reference adaptive control is one of the most important adaptive controllers in which the desired
process and the controller’s behaviour are formed separately and another loop changes the feedback of the controller
parameters. The parameters are updated based on the error log which is the difference between the system output
and the reference model output. The parameter adjustment mechanism in the model reference adaptive system is
determined by two methods of gradient or MIT law controller parameter adjustment based on the line gradient
relative to the parameters. But there is no guarantee that an adaptive controller based on this method yields a stable
closed-loop system. With the increment of the matching efficiency, the system becomes unstable. To suppress this
drawback, the Lyapunov theory of stability is used. The Smith predictor is a predictive controller for systems with
a net delay. As shown in Fig. 1, there are two parallel paths in the Smith predictor structure for the control signal:
1. The first path crosses the actual (patient) process
2. The second path crosses the identified model of the process (Ĝ P (s)e−t0 s )
This model predicts the process response without delay (y p ). Then, this prediction is compared to the optimal
set point to determine the needed control action. To eliminate the perturbation, Smith predictor compares the output
of the actual process to the prediction of the process that takes the time delay into account (y1 ) and the error (e1 )
is recorded to be contributed in the overall error signal delivered to the controller.

2.3. Modelling

As it is shown in Fig. 1, a predictive model Ĝ P (s) of the delay in the process (t0 ), is needed in the Smith
Predictor structure. So, linear system identification is performed. Blood glucose level and insulin injections are
200
S. Taherinasab, S. Soleimaniasl and S. Taherinasab Mathematics and Computers in Simulation 194 (2022) 198–209

Table 1
Daily consumption of food.
First Second Day
54 82 73 62 91 57 Meal Rate (g)
8 12 19 7 12 21 Meal time (hours)

measured subcutaneously. Therefore, the dynamics of the system are delayed. The model structure is selected as a
delayed second-order conversion function. The inputs of the system are the consumed food and the injected insulin,
and the output is the blood glucose level. The system has two inputs and a single output. The following points are
considered when data collection is needed to identify each patient’s specific model:
• Virtual tests should be as close as possible to the patients’ real-life conditions. Insulin levels should be as high
as the patient’s blood glucose output within a safe range (55–280 mg/dl).
• The selected inputs should be sufficiently stimulated and rich. This means that the input signal should vary
in an appropriate range to see all possible modes. If the input signal does not be rich enough, the estimated
output would overlap with the actual output, but the parameters will not be converted to actual convergent
parameters. Sometimes the identified ARX, ARMAX, and Box Jenkins model maybe not properly identified
due to an improper input, as the food and insulin inputs occur simultaneously and proportionally [18]. To
overcome this problem, the amount of food, the time it takes to differentiate between days, and insulin intakes
should be selected from the patient’s closed-loop control.
• The data obtained from the closed-loop control of the patient is richer than the open-loop control data. Because
in the open-loop data, the insulin control signal sums up the effects of a basal insulin profile and some
synchronized snack basals. Therefore, in the closed-loop data, the insulin signal depends on the measured
data from the glucose sensor, in which temporal variations provide the level of required stimulation.
As mentioned above, the system identification protocol is defined as follows. GIM simulation software is used
to collect closed-loop control data [21]. This software simulates the Dalla Man model through MATLAB with a
Windows user interface that allows the user to easily simulate the 24 hour daily life of a type 1 diabetes patient,
a healthy person, or a type 2 diabetes patient. In this software, three meals are considered. An insulin injector
based on the PID controller is designed to be injected into the patient’s body [26]. In Table 1, two datasets are
compiled. Both of them contain insulin input, food input (Meal), and glucose sensor outputs. The first set is used
for identification and the second set is used for validation.
There are three-time delays in the system physiology:

1. The first delay time is when the patient eats a meal. It is the time it takes for the blood glucose level to rise
after eating.
2. The second delay time relates to the method of measuring blood glucose. Releasing the glucose from the blood
into the adipose tissue under the skin takes 5 to 15 min and the glucose sensor measures it subcutaneously.
So, there is a delay between the plasma glucose of the model and the measured one by the sensor.
3. The third delay time is attributed to the delay from the uptake of insulin from the skin to the bloodstream and
its effect on glucose levels. Therefore, the dynamics of the system is delayed. Therefore, the model structure
is chosen as a delayed second-order conversion function. The model is discrete and MISO.

The system has two inputs and a single output. So, there are two subsystems, the first subsystem is from insulin
to glucose and the second one is from carbohydrate to glucose. The equation can be defined as:
b k
Y (s) = 2 e−t0 s U (s) + 2 e−t1 s D(s) (1)
s + α1 s + α2 s + c1 s + c2
in which Y (s), U (s), D(s) are laplace conversions of glucose output, insulin input, and food input respectively. The
overall delay time including sensor delay and insulin uptake delay is significant, knowing an estimate of which
helps the experts with its identification. So, to determine an initial estimation the feed input was set to zero. Then
an insulin step input was applied to the system and its output was measured. The delay time was measured as
40 min. Using the system identification toolbox in MATLAB and the nonlinear least squares with a line search
201
S. Taherinasab, S. Soleimaniasl and S. Taherinasab Mathematics and Computers in Simulation 194 (2022) 198–209

Table 2
The relation between the Z domain and the Laplace second-order
system, the first rows show the different poles and the second row
shows the first real poles.
X (z) X (s)
sin wT
e−at z −1 W
1 − 2e−at z −1 cos wT + e−2at z −2 (s + a)2 + w2
(e−at − e−bt )z −1 b−a
(1 − e−at z −1 )(1 − e−bt z −1 ) (s + a) + (s + b)

Table 3
F I T and M S E values of the identified model.
T0 T1 Fit (%) MSE
5 0 82.71 26.79
5 1 80.6 13.79
5 2 85.6 9.59
5 3 82.71 25.94
5 4 75.15 63.89

Table 4
Values of the identified model parameters.
Amount Parameter
−0.001719 B
0.005871 a1
2.07 ∗ 10−5 a2
42 t0
5.349 ∗ 10−8 K
0.01357 c1
2.034 ∗ 10−5 c2
11 t1

algorithm, the parameters of the two conversion functions are identified. The continuous equations are obtained
from the equations of Table 2.
Bz −T0 K z −T1
Y (z) = U (z) + D(z) (2)
1 + A1 z −1 + A2 z −2 1 + C1 z −1 + C2 z −2
Considering the relations of z-domain to laplace and the sampling period, the T0 value is considered as 5 which
leads to a t0 of 40 min. To confirm the delay time T1 , the identification process starts with a constant T0 and different
values of T1 . Utilizing MSE and goodness of the fit criteria, the best value of T1 is considered such that the goodness
of fit becomes the most and MSE be the least (Table 3).
N
1 ∑
MSE = ( ŷ(i) − ŷ(i))2
N i=0
(   )
 y − ŷ 
The goodness of fit = 100 1 −   (3)
 y − mean( ŷ)

Thus, the magnitude of the delay T1 is 2, which is equivalent to t1 of 10 min. Table 4 represents the parameters
of the identified model.

2.4. MRAS design using the Lyapunov theory

Lyapunov theory of stability is used to obtain the law of regulating the controller parameters. First, a differential
equation for the error e = y − ym is obtained containing some adjustable parameters. Then, by defining a Lyapunov
202
S. Taherinasab, S. Soleimaniasl and S. Taherinasab Mathematics and Computers in Simulation 194 (2022) 198–209

function and satisfying the asymptotic stability conditions, the parameter matching rules are computed in a way that
the error tends to zero. The general equation of the controller is considered as RU = T UC − SY which is obtained
by setting R = 1, T = f , S = q0 s + q1 , and u c = setpoint which changes the control signals to Eq. (4).
U = f u c − q0 ẏ − q1 y (4)
As mentioned in the preceding section, the reference model was firstly designed for the process without a delay
in the controller. Then using the Smith predictor idea, the controller for the main process was obtained. Whereas
the overall form of the process may be expressed as.
y b0
G(s) = = 2 (5)
u s + a1 s + a2
By placing Eq. (4) into Eq. (3), the closed-loop system equation is obtained as:
f b0
y= 2 uc (6)
s + (a1 b0 q0 )s + (a2 + b0 q1 )
In this equation f, q0 and q1 are the control parameters. The reference model is considered as:
ym bm
G m (s) = = 2 (7)
uc s + am1 s + am2
So, comparing Eqs. (6) and (7) yields the parameter errors as:
b̃0 = f b0 − bm
ã1 = a1 + b0 q0 − am1
ã2 = a2 + b0 q1 − am2 (8)
To reduce the error, it is common to define a differential equation. It can also be extracted from Eqs. (5) and
(6):
ÿ + (a1 + b0 q0 )y + (a2 + b0 q1 )y = f b0 u c
ÿm + am1 ẏm + am 2 ym = bm u c (9)
Therefore, for the error dynamics, Eq. (10) is obtained:
ë + am 1 ė + am 2 e = b̃0 uc − ã1 ẏ − ã2 y (10)
Assuming that γ0 , γ1 , γ2 are positive, the Lyapunov function is defined as.
1 1 1
V = am1 e2 + ė2 + b̃02 + ã12 + ã22 (11)
γ0 γ1 γ2
dv
In Eq. (11), V is positive. This function may be defined as a Lyapunov function only when dt
is negative. The
derivative of this equation is:
2 2 2
V̇ = 2am1 eė + 2ėë + b̃0 b̃˙ 0 + ã1 ã˙ 1 + ã2 ã˙ 2 (12)
γ0 γ1 γ2
By placing Eqs. (8) and (10) in Eq. (12), the following relation is obtained:
1 ˙ 1 1
V̇ = −2am1 ė2 + 2b̃0 [ėuc + b̃] ˙ + 2ã2 [−ėy + ã˙ 2 ]
+ 2ã1 [−ė ẏ + ã] (13)
γ0 γ1 γ2
The derivative of the Lyapunov function becomes semi-negative if the expressions in dv dt
= −2am 1 ė2 brackets
dv 2
must be zero in Eq. (13) which leads to) dt = −2am 1 ė . Therefore, the rules for setting parameters are as follows.
γ0 t0
∫
b̃˙ = −γ0 ėuc ⇒ f = − ėucdt
b0 0
γ2
∫ t0
ã˙ 2 = γ1 ėy ⇒ q1 = − ėydt
b0 0
γ1
∫ t0
ã˙ = γ1 ė ẏ ⇒ q0 = − ė ẏdt (14)
b0 0
203
S. Taherinasab, S. Soleimaniasl and S. Taherinasab Mathematics and Computers in Simulation 194 (2022) 198–209

Fig. 2. Diagram of the reference model for patients whose blood glucose is 180 mg/dl at the beginning of the simulation.

Since the derivative of the Lyapunov function is semi-negative, the error and the parameter values are bounded. In
this method, it is not necessary to converge the parameters to nominal values, and the system stability is maintained
even if the parameters are not converged.
The second derivative of the Lyapunov function is also calculated to show that the error, in addition to being
bounded, tends to zero.
V̈ = −4am1 ëė
= −4am1 ė(b̃0 u c − ã1 ẏ − ã2 y − am1 ė − am2 e) (15)
Because u c , y, and e are bounded, ė and ẏ are also bounded. Consequently, V̈ is bounded too, so the error tends
to zero. It should be noted in model choosing that preventing hypoglycemia (blood glucose below 70 mg/dl) is
preferable for the patient over hyperglycemia (blood glucose greater than 180 mg/dl) because the hypoglycemia
risks are serious and occur rapidly, while the effects of hypoglycemia are long term. When large and sudden
changes in the blood concentration occur (e.g., after ingestion), the existence of a rapid reference model may lead to
over-administration. Thus, despite the slow subcutaneous absorption of insulin, the risk of hypoglycemia increases.
Therefore, the reference model should be chosen more slowly than the patient model in order not to reach the final
setpoint value quickly. The reference model poles were 0.25 times that of the patient model additionally when it is
multiplied by setpoint, the initial value of the reference model at t = 0 will be equal to the patient’s baseline blood
glucose and the final value is equal to the setpoint. Fig. 2 shows the graph of the controller reference model.

2.5. Improving perturbation repulsion by modified Smith predictor

In Fig. 3, Smith predictor uses Ĝ P internal model to predict the delay-free response of the process (y p ). Then,
y p is compared to the desired setpoint to determine the control command u. Also, Smith prediction of the outgoing
external perturbations compares y1 to the process output; and the y − y1 the difference value is fed through f and
contributes to the overall system error. Therefore, the design of the f filter can be used to improve the perturbation
repulsion.
According to Fig. 3, Eqs. (16) and (17), can be written as:
1 + C Ĝ P − C Ĝ P e−t0 s F C Ĝ P e−t0 s
y(s) = d+ ysp (16)
1 + C Ĝ P 1 + C Ĝ P
1 + C Ĝ P (1 − e−t0 s ) 1 + C Ĝ P (1 − e−t0 s F)
e(s) = ysp (s) − y(s) = ysp − d (17)
1 + C Ĝ P 1 + C Ĝ P
Therefore, if 1 − e−t0 s F ≈ 0, the disturbance effect in error decreases regardless of the mismatch between G P
and Ĝ P . The Turksoy et. al. [14] have suggested Eq. (18) for et0 s approximation:
1 + B(s)
e t0 s = (18)
1 + B(s)e−t0 s
204
S. Taherinasab, S. Soleimaniasl and S. Taherinasab Mathematics and Computers in Simulation 194 (2022) 198–209

Fig. 3. Improved Smith predictor block diagram.

where B is a low pass filter with the same time constant and with the Ĝ P internal model. The poles of the
internal model in the previous sections are −0.0029 ± 0.0035i, so the time constant of Ĝ P is 34.8276. By choosing
1
B = 1+344.8276s in F(s) equation, the results can be expressed as:
1 + B(s) 2 + 344.82s
F(s) = = (19)
1 + B(s)e−40s 1 + 344.82s + e−40s

2.6. Simulation

All the aforementioned models are average population models. So, a single patient whose parameters are equal
to the population’s average parameters is defined. Therefore, these models can predict the dynamics of the average
population and are not capable of predicting variability between individuals. This method is not suitable for virtual
tests of controllers. Hence, it is necessary to have a group of patients with different parameters to consider the
variability of parameters in a single patient and among different patients. The metabolism variability of real patients
has two forms:
1. Intrapersonal variability means that certain parameters of each patient individually change during a day.
2. Interpersonal variability, which is related to differences in some parameters, such as individual weight, fasting
blood glucose concentration, and so on between different patients.
To create a virtual patient population from the Dalla Man model, the parameters K p1 (internal glucose
production), Vm0 , K m0 , Fcns (glucose consumption) and K e1 (renal glucose excretion rate) are selected for the
change [27]. To represent the intrapersonal and interpersonal metabolic variability simultaneously, a sinusoidal
oscillation is added to the above parameters. The general shape of the variable in time is expressed as Eq. (20).
{ }
2π
p(t) = p0 (1 + α sin t +ϕ ) (20)
60δ
where P0 is the nominal value of the parameter, a is the amplitude of sine with a random value in the interval
[−0.75, 0.5], δ is a random period in the interval [1, 24], and ϕ is a random phase in the interval [0, 2π ]. Therefore,
using a normal distribution, it is possible to create the parameters of virtual patients randomly. Fig. 4 shows the
blood glucose levels of 10 different virtual patients in the open-loop control.
In the open-loop control, insulin is injected by the patient before eating based on the amount of food being
consumed. Also, two additional simulations of the PID controller combination with Smith predictor and open-loop
control are performed to assess the performance improvement of the reference model controller [26].

3. Results
In this part, the results of the PID with Smith predictor (PIDSP controllers), Model Reference with Smith
Predictor (MRSP), and open-loop control for the average Dalla Man patient (the disease whose parameters are
205
S. Taherinasab, S. Soleimaniasl and S. Taherinasab Mathematics and Computers in Simulation 194 (2022) 198–209

Fig. 4. The virtual patient open-loop control.

Fig. 5. The patient’s response, the first graph indicates blood glucose, the second graph indicates injected insulin, and the third graph is the
food consumed. The dashed curve is the open-loop controller, the small dotted curve shows the PID controller, and the other dotted curve
is related to the MRMSP controller.

the average of all patients) simulations are presented during a day. It is assumed that the patient has consumed 47
g of carbohydrates for breakfast at 8 a.m., 77 g for the lunch at midday, and 64 g for the dinner at 8 p.m. As it
is shown in Fig. 5, the nature of the perturbation in this system is traumatic. Also, while hyperglycemic events
have occurred for the PID controller and the open-loop control, the MRSP controller has been able to maintain the
patient’s blood glucose in the safest state (180–70 mg/dl). Furthermore, the performance of PIDSP and MRMSP
controllers is evaluated compared to the open-loop control for 30 virtual patients during single daylight (1440 min)
in terms of qualitative and quantitative indices. Food input protocol (e.g., patient’s food input protocol) is average.
The initial conditions of the system or the initial patient’s glucose concentration are determined randomly with a
normal distribution having an average of 143.4 mg/dl and a standard deviation of 9.3. Also, the setpoint value of
the controllers is set to 120 mg/dl.
On the issue of glycemic control, recent articles (since 2009) have used the following qualitative and quantitative
indicators presented by Kovatchev to evaluate their controller performance [15]. CVGA is a graphical representation
of min/max values in the virtual or real patient population. The CVGA enables the simultaneous evaluation of blood
glucose quality in patients, thus playing an important role in the design of a closed-loop control algorithm and its
performance evaluation. Available for a specified period (e.g., one day), this is how CVGA is obtained. For each
patient, a point is specified where the x coordinate represents the patient’s lowest blood sugar and the y coordinate
represents the highest blood glucose over the defined period. The x range is from 50 to 110. Therefore, the optimal
blood glucose setting is in the right-hand corner. The overall appearance of the map is a cloud of points in different
206
S. Taherinasab, S. Soleimaniasl and S. Taherinasab Mathematics and Computers in Simulation 194 (2022) 198–209

Fig. 6. CVGA diagram for 30 patients under control; the multiplication indication for the PIDSP controller, square points for the MRMSP
controller, and solid circles represent the open-loop control.

Table 5
Comparison of the quantitative indicators for the three methods.
PIDSP MRMSP Open-loop Index
147.14 (8.15) 123.14 (14.9) 186.78 (21.4) Average blood glucose
78.95 (5) 96.85 (5.7) 51.93 (10.7) % Normal range
0.157 (0.848) 0.134 (0.59) 0 % Hypoglycemia
22.87 (4.93) 3.03 (5.7) 48.09 (10.7) % Hyperglycemia

areas. CVGA is a qualitative indicator. The closer the patient points to the area A, the better the performance of
the controller. CVGA diagrams are obtained for three methods [23]. Thus, the blood glucose level of some patients
under open-loop control is greater than 110 mg/dl. On the outside of the CVGA chart x axis, the chart is separated
by a dot.
As it is shown in Fig. 6, most patients under open-loop control suffer from severe hyperglycemia (BG >
280 mg/dl). However, the CVGA diagrams of the reference model control strategy are more congested in area A
than in the PID controller. So, they represent better performance. Combining the CVGA diagrams of Fig. 6, it can be
seen that the MRMSP controller is more effective than the PIDSP controller because of its structural adaptability
to the variability of metabolism between patients. The percentage indices, being a good example of quantitative
indices, is defined as the percentage ratio of the time in which the patient’s blood glucose is in the range of normal
(70–180 mg/dl), hypoglycemic (BG < 70 mg/dl), and hyperglycemic (BG > 180 mg/dl) to the total test time
(one day). From a more complete perspective, the indicators of severe hypoglycemia (BG < 50 mg/dl) and severe
hyperglycemia (BG > 280 mg/dl) can be examined. In Table 5, the mean values of the percentage indices of the
three methods are compared for 30 patients. The numbers in parentheses show the standard deviation of the set.
In all three control methods, a small proportion (nearly zero percent) of the patient’s blood glucose is within the
range of hypoglycemia. The main reason is that in the hypoglycemic conditions, especially severe hypoglycemia, the
patient is at the risk of death and coma. Long-term hyperglycemia may cause serious damage to the heart, kidneys,
and eyes. Also, the average blood glucose index in the MRMSP controller is lower than the other two methods and
is closer to the set point of 120 mg/dl. Also, the values of normal and hyperglycemic range indices in the reference
model strategy have improved. Therefore, according to the hypoglycemia percentage indicators, it can be said that
the three methods are similar in terms of hypoglycemia occurrence, but in terms of preventing hyperglycemia, the
reference model strategy has been much better. Whereas the rate at which the patient population is within the normal
range (70–180 mg/dl), 96% of the entire simulation timeframe indicates that the reference model strategy is capable
of improving the lives of patients with type 1 diabetes.
207
S. Taherinasab, S. Soleimaniasl and S. Taherinasab Mathematics and Computers in Simulation 194 (2022) 198–209

4. Conclusion
In this paper, some strategies of glycemic control are presented. The Dalla man And Hovorka and modified
Bergman are the most effective and widely used models that have been used repeatedly in this research in the articles
of the past few years. The main limitation in the performance of the controller in all these various methods is the
accuracy of the utilized model in representing the insulin/glucose dynamics. The model reference Smith Predictor
adaptive controller is designed for the problem of closed-loop control of type 1 diabetes patients. Performance of
the proposed algorithm has been compared to the PID controller with smith predictor, based on the quantitative and
qualitative indicators, simulation results show that the proposed control scheme is effective in fasting conditions,
meal disturbance rejection, and robustness against inter-patients variability. As the Smith Predictor structure can
withstand the intrinsic delays of the system, it causes no major hypoglycemic or hypoglycemic events during the
simulation. Also, despite the controller’s versatility, it shows a good performance against interpersonal metabolism
variability. Given the importance of preventing hypoglycemic status, in all three discussed methods in this article,
namely open-loop control, reference model control, and PID control, the rate of hypoglycemic events is negligible.
However, the strength of the reference model adaptive structure is in the significant reduction of the hyperglycemic
status compared to the entire patient control period. Therefore, the patient’s blood glucose levels are within the
normal range for a longer time. Furthermore, the longer-term complications of hyperglycemia will decrease, and
the life quality of the patient will improve.

References
[1] I. Ahmad, F. Munir, M.F. Munir, An adaptive backstepping based non-linear controller for artificial pancreas in type 1 diabetes patients,
Biomed. Signal Process. Control 47 (2019) 49–56.
[2] D. Boiroux, A.K. Duun-Henriksen, S. Schmidt, K. Nørgaard, N.K. Poulsen, H. Madsen, J.B. Jørgensen, Adaptive control in an artificial
pancreas for people with type 1 diabetes, Control Eng. Pract. 58 (2017) 332–342.
[3] C. Califano, E. Scharbarg, N. Magdelaine, C.H. Moog, A nonlinear time-delay realization for gastroparesis in patients with diabetes,
Annu. Rev. Control 48 (2019) 233–241.
[4] A. Cinar, Multivariable adaptive artificial pancreas system in type 1 diabetes, Curr. Diabetes Rep. 17 (10) (2017) 88.
[5] G.R. Cocha, C. Amorena, A. Mazzadi, C. D’Attellis, Geometric adaptive control in type 1 diabetes, in: 12th International Symposium
on Medical Information Processing and Analysis (Vol. 10160, 101600R), International Society for Optics and Photonics, 2017.
[6] G. Eigner, J.K. Tar, L. Kovács, Adaptive control solution for T1DM control, in: 2015 IEEE 10th Jubilee International Symposium on
Applied Computational Intelligence and Informatics, IEEE, 2015, pp. 215–220.
[7] G. Eigner, J. Tar, L. Kovács, Fixed point transformation-based adaptive control for type 1 diabetes mellitus, T. Autom. Control Comput.
Sci. 60 (74) (2015) 1.
[8] F.M. El-khatib, E.R. Damiano, S.J. Russell, U.S. Patent Application No. 15/748, 333, 2018.
[9] I. Hajizadeh, M. Rashid, A. Cinar, Plasma-insulin-cognizant adaptive model predictive control for artificial pancreas systems, J. Process
Control 77 (2019) 97–113.
[10] I. Hajizadeh, M. Rashid, S. Samadi, J. Feng, M. Sevil, N. Hobbs, K. Turksoy, et al., Adaptive and personalized plasma insulin
concentration estimation for artificial pancreas systems, J. Diabetes Sci. Technol. 12 (3) (2018) 639–649.
[11] P. Herrero, J. Bondia, O. Adewuyi, P. Pesl, M. El-Sharkawy, M. Reddy, P. Georgiou, et al., Enhancing automatic closed-loop glucose
control in type 1 diabetes with an adaptive meal bolus calculator–in silico evaluation under intra-day variability, Comput. Methods
Programs Biomed. 146 (2017) 125–131.
[12] H. Heydarinejad, H. Delavari, Adaptive fractional order sliding mode controller design for blood glucose regulation-4-3, in: Theory
and Applications of Non-Integer Order Systems, Springer, Cham, 2017, pp. 449–465.
[13] M. Hoskins, Diabeloop artificial pancreas approved in Europe, 2018, https://www.Healthline.com/diabetesmine/diabeloop-pre-artificial-
pancreas-approved#2.
[14] H.P. Huang, C.L. Chen, Y.C. Chao, P.L. Chen, A modified smith predictor with an approximate inverse of dead time, AIChE J. 36
(7) (1990) 1025–1031.
[15] B.P. Kovatchev, E. Otto, D. Cox, L. Gonder-Frederick, W. Clarke, Evaluation of a new measure of blood glucose variability in diabetes,
Diabetes Care 29 (11) (2006) 2433–2438.
[16] N. Magdelaine, L. Chaillous, I. Guilhem, J.-Y. Poirier, M. Krempf, C.H. Moog, E. Le Carpentier, A long-term model of the
glucose-insulin dynamics of type 1 diabetes, IEEE Trans. Biomed. Eng. 62 (2015) 1546–1552.
[17] N. Magdelaine, P.S. Rivadeneira, L. Chaillous, A.L. FournierGuilloux, M. Krempf, T. MohammadRidha, M. Ait-Ahmed, C.H. Moog,
The hypoglycemia-free artificial pancreas project, IET Syst. Biol. 14 (2020) 16–23.
[18] M. Messori, G.P. Incremona, C. Cobelli, L. Magni, Individualized model predictive control for the artificial pancreas: In silico evaluation
of closed-loop glucose control, IEEE Control Syst. Mag. 38 (1) (2018) 86–104.
[19] M. Messori, J. Kropff, S. Del Favero, J. Place, R. Visentin, R. Calore, F. Boscari, et al., Individually adaptive artificial pancreas in
subjects with type 1 diabetes: A one-month proof-of-concept trial in free-living conditions, Diabetes Technol. Ther. 19 (10) (2017)
560–571.
208
S. Taherinasab, S. Soleimaniasl and S. Taherinasab Mathematics and Computers in Simulation 194 (2022) 198–209

[20] A. Nath, D. Deb, R. Dey, S. Das, Blood glucose regulation in type 1 diabetic patients: An adaptive parametric compensation
control-based approach, IET Syst. Biol. 12 (5) (2018) 219–225.
[21] N. Resalat, W. Hilts, J.E. Youssef, N. Tyler, J.R. Castle, P.G. Jacobs, Adaptive control of an artificial pancreas using model identification,
adaptive postprandial insulin delivery, and heart rate and accelerometry as control inputs, J. Diabetes Sci. Technol. 13 (6) (2019)
1044–1053.
[22] D. Romeres, M. Schiavon, A. Basu, C. Cobelli, R. Basu, Chiara Dalla Man, 1862-P: Insulin-Independent Glucose Utilization During
Exercise Is Impaired in Type 1 Diabetes: A New Model Based Analysis, 2019.
[23] J. Smieja, Automatic control and feedback loops in biology and medicine, in: Conference on Automation, Springer, Cham, 2019, pp.
3–12.
[24] C. Toffanin, R. Visentin, M. Messori, F. Di Palma, L. Magni, C. Cobelli, Toward a run-to-run adaptive artificial pancreas: In silico
results, IEEE Trans. Biomed. Eng. 65 (3) (2017) 479–488.
[25] S. Trevitt, S. Simpson, A. Wood, Artificial pancreas device systems for the closed-loop control of type 1 diabetes: What systems are
in development? J. Diabetes Sci. Technol. 10 (3) (2016) 714–723.
[26] R. Visentin, E. Campos-Náñez, M. Schiavon, D. Lv, M. Vettoretti, M. Breton, C. Cobelli, et al., The UVA/Padova type 1 diabetes
simulator goes from single meal to single day, J. Diabetes Sci. Technol. 12 (2) (2018) 273–281.
[27] D.P. Zaharieva, S. McGaugh, R. Pooni, T. Vienneau, T. Ly, M.C. Riddell, Improved open-loop glucose control with basal insulin
reduction 90 min before aerobic exercise in patients with type 1 diabetes on continuous subcutaneous insulin infusion, Diabetes Care
42 (5) (2019) 824–831.

Further reading
[1] Y. Batmani, S. Khodakaramzadeh, Blood glucose concentration control for type 1 diabetic patients: A multiple-model strategy, IET Syst.
Biol. (2019).
[2] K.J. Bell, C.E. Smart, G.M. Steil, J.C. Brand-Miller, B. King, H.A. Wolpert, Impact of fat, protein, and glycemic index on postprandial
glucose control in type 1 diabetes: Implications for intensive diabetes management in the continuous glucose monitoring era, Diabetes
Care 38 (6) (2015) 1008–1015.
[3] A. Beneyto, J. Vehi, Postprandial fuzzy adaptive strategy for a hybrid proportional derivative controller for the artificial pancreas, Med.
Biol. Eng. Comput. 56 (11) (2018) 1973–1986.
[4] W. Liu, G. Zhang, L. Yu, B. Xu, H. Jin, Improved generalized predictive control algorithm for blood glucose control of type 1 diabetes,
Artif. Organs 43 (4) (2019) 386–398.
[5] D. Cai, J. Song, J. Wang, D. Shi, Glucose regulation for subjects with type 1 diabetes using active disturbance rejection control, in:
2019 Chinese Control Conference, CCC, IEEE, 2019, pp. 6970–6975.
[6] Y. Toyoda, S. Saitoh, Adaptive regulation of glucose transport, glycolysis and respiration for cell proliferation, Biomol. Concepts 6 (5–6)
(2015) 423–430.
[7] K. Turksoy, E.S. Bayrak, L. Quinn, E. Littlejohn, A. Cinar, Adaptive multivariable closed-loop control of blood glucose concentration
in patients with type 1 diabetes, in: 2013 American Control Conference, IEEE, 2013, pp. 2905–2910.

209