Journal of Physics: Conference Series

PAPER • OPEN ACCESS

Problems and methods of a closed-loop blood glucose control system

construction
To cite this article: E L Litinskaia et al 2021 J. Phys.: Conf. Ser. 2091 012020

View the article online for updates and enhancements.

This content was downloaded from IP address 198.240.90.91 on 17/12/2021 at 13:26

ICCT 2021 IOP Publishing
Journal of Physics: Conference Series 2091 (2021) 012020 doi:10.1088/1742-6596/2091/1/012020

Problems and methods of a closed-loop blood glucose control

system construction

E L Litinskaia, K V Pozhar and N M Zhilo

Institute of biomedical systems, National Research University of Electronic Technology,
1 Shokin Square, Zelenograd, Moscow 124498, Russia

E-mail: zheka101993@yandex.ru

Abstract. The automation of the insulin therapy in the closed-loop blood glucose control
system is accompanied by such problems as delays of the feedback and the controlling action
effect on blood glucose dynamics, difficulties in low and high glucose levels advance detection
and timely compensation, a human factor, complexity of a mathematical model of control
object building, evaluation of blood glucose control efficiency. In the paper, these automation
problems and methods for their compensation are considered. The closed-loop system based on
the model predictive control, adaptive adjustment of the controlling action and decision-
making support was developed. Experimental testing with UVA/Padova Type 1 Diabetes
Mellitus Simulator showed that the application of suggested methods in the closed-loop allows
increasing the efficiency of the blood glucose maintaining in the target range.

1. Introduction
Type 1 diabetes mellitus is the endocrine chronic disease associated with high blood glucose (BG)
caused by violation of insulin production in the pancreas. To control BG and maintain it in the target
normal range, all over the world personalized closed-loop systems are being developed with different
automation levels, control methods and strategies [1]. The main goal of researches and developments
in the field is to build an automatic control system that will allow minimizing a patient participation in
BG regulation.
Closed-loop control systems consist of an insulin pump as an effector, a continuous glucose
monitor (CGM) as a feedback sensor, and a regulator that provides regulation of controlling action
(subcutaneous insulin infusion) according to the feedback and built-in control algorithms [2]. Existing
and developing systems show promising results in BG maintaining during clinical trials [3–5],
however, along with the high cost of devices, the efficiency and safety of most systems today is not
sufficient for their widespread use by patients with type 1 diabetes.
A number of difficulties related to the system structure, choosing the control method, physiological
features of hormone regulation, mathematical modelling of the system and the efficiency of BG
control evaluation accompanies development of an automated closed-loop system. This work is
devoted to considering these problems and compensating methods including commonly applied as
well as proposed ones.

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
ICCT 2021 IOP Publishing
Journal of Physics: Conference Series 2091 (2021) 012020 doi:10.1088/1742-6596/2091/1/012020

2. Problems & Methods

2.1. Delays in the closed loop

Accuracy of the feedback sensor is calculated from its error and the feedback delay. Nowadays,
invasive CGMs remain the glucose meters with the lowest error (12-25%) and the most effective
feedback sensors [6]. However, the monitors measure glucose concentration in the subcutaneous fat
and therefore the data is received with 15 min delay in comparison with actual BG. The controlling
action also affects BG dynamics with a delay of more than 1.5 hours, depending on the type of insulin.
These two problems lead to the lagging in the closed-loop, which decreases the automated control
efficiency and causes health risks.
The lagging can be compensated by control methods based on advance regulation of the controlling
action. For continuous low-rate insulin infusion (basal delivery), adaptive control algorithms adjusting
the rate of infusion by feedback data are being developed [7, 8]. These algorithms are based on usage
of optimized thresholds inside the target normal range for advance acceleration, braking or stopping of
the controlling action, depending on deviation of measured values from the thresholds.
The proposed method of adaptive basal delivery (figure 1) is based on the use of a nested
supernormal range, within which insulin is injected at the standard rate prescribed by the clinician. If
the measured BG is in zones 1 or 2, the rate of insulin delivery is being proportionally increased until a
certain maximum value reached at the Gcrit level, and is being decreasing within zone 4. Zones 5 and 6
correspond to stopping of the insulin delivery in order to limit an overshoot.

Figure 1. Thresholds and zones of adaptive basal delivery.

Food intake causes a sharp BG increase and requires an impulse controlling action to compensate it
– a high-rate insulin infusion (bolus). Taking into account the delay in the insulin effect, to increase
the efficiency of BG normalization, prediction algorithms are used to calculate the optimal bolus form
and dose according to meal data and to inject insulin in advance. Modern closed-loop systems
commonly use such control methods as PID-regulation, model predictive control (MPC), fuzzy logic
or their combinations.
PID algorithms [9] respond to measured glucose levels and do not perform direct prediction with
taking into account a number of physiological parameters but analysis of the future BG dynamic by its
differential component. Therefore, even with well-optimized constants, PID-regulation is usually
accompanied by significant overshoot and has a fairly long regulation time. When impulse controlling
action, this can lead to a significant sharp decrease of BG and a risk to the patient's health.
A more physiological control method is MPC [10], which predicts the BG dynamics based on the
mathematical modeling of the control object. The method takes into account not only current
measured BG levels but also meal data and physiological parameters of a particular patient when
formatting a controlling action. The disadvantage of the method is the complexity of constructing a
sufficiently reliable mathematical model of the object. However, the method implies the possibility of
the model parameters optimization during the system operation to improve the prediction accuracy.

2
ICCT 2021 IOP Publishing
Journal of Physics: Conference Series 2091 (2021) 012020 doi:10.1088/1742-6596/2091/1/012020

Fuzzy logic methods [11] are based on modeling of physician’s decisions on real-time adjustments
of insulin doses based on CGM data. Compared to MPC and PID-regulation, fuzzy logic control uses
more information from experts and relies less on mathematical modeling of the physiological system.

2.2. High and low BG values

Insulin therapy has a unidirectional effect on BG dynamics, allowing only reducing its value. At the
same time, control and prevention of excessive overshoots leading to low BG is an important task of
the closed-loop control. They usually occur after a meal and are caused by bolus overdosing.
One of the approaches to compensate overshoots is to use a dual-hormone closed-loop system
based on additional infusion of glucagon, which increases BG. However, this method adds
“complexity, possible risks and expense” [12] to the system and is not widely spread. The second
approach to achieve bidirectional effect on BG dynamic is a decision-making support on additional
carbohydrates intake [13]. The method also uses prediction to detect the risk of the low BG and to
calculate the optimal amount of carbohydrates to intake.
In addition, various factors, such as variability in insulin sensitivity, stress, physical activity, can
cause both an unforeseen BG decrease, and its unforeseen increase. High BG can also be observed as a
result of insufficient after-meal insulin infusion. In these cases, when only basal delivery is insufficient
for BG increase compensation, model predictive control allows calculation of a corrective bolus of a
small dose.

2.3. Human factor

Despite the almost automatic BG control, most closed-loop systems require patient to input data about
meals, additional insulin injections, current physical state and activities, which are needed to predict
the BG dynamics and to calculate recommended controlling action values or corrective carbohydrates
amount. Moreover, possible failures of control algorithms and as a result the risk of insufficient or
excessive insulin infusion also limits the ability to build the fully automatic system and requires
patient confirmation or correction of some controlling actions. Patient’s wrong decisions, incorrect or
missing data can lead to risks to the patient’s health. That is why decision support on control and
compensation of the system functioning and operation violations is required.
The proposed concept of such decision support algorithm is based on a comparative analysis of
measured and predicted BG [14]. Measured values that deviate from the prognosis by more than the
certain value are collected in a buffer. Then, by analyzing the trends in the buffers of predicted and
measured values, the most likely events are assumed (figure 2) and advance recommendations to
compensate BG dynamics are offered to the patient.

Figure 2. Events detection: one type of trends correlation can correspond to several possible events.

3
ICCT 2021 IOP Publishing
Journal of Physics: Conference Series 2091 (2021) 012020 doi:10.1088/1742-6596/2091/1/012020

As a result of the trend analysis, such events are detected as the absence or falsity of data about
meal or manual insulin injection, the absence of the meal fact, an insulin pump malfunction, a glucose
sensor error, as well as stress or physical activity.

2.4. Mathematical modeling of the control object

An important step in a control system synthesis is the construction of mathematical models of system
elements, including the control object. These models are used both to optimize system parameters and
to form the basis of model predictive control. The mathematical model of the object should describe
the controlled parameter dynamic in dependence on the controlling action, external disturbances and
other effects arising in the system. In the considering system, the control object is the patient glucose
regulation system. The BG dynamics is determined by a variety of physiological processes, the
mathematical description of which, due to their multifactorial nature, is a non-trivial task.
Mathematical models aimed at a rigorous description of these processes [15] are usually high-order
ODE systems. Some models quite accurately describe particular processes, for example, the delay in
the absorption of insulin into the plasma [16], leading to a delay in the control action mentioned in 2.1,
and food intake [17]. The work uses a mathematical model based on the integration of these models
modified for the case of type 1 diabetes mellitus.
Based on the model, both short-term BG dynamics prediction and the calculation of the optimal
values of the recommended corrective boluses and carbohydrates are carried out. Calculations are
performed by the Dormand-Prince method in Matlab, while the edges of the food intake and insulin
infusion pulses are approximated with steep logistic functions.

2.5. Evaluation of BG control efficiency

When evaluating the efficiency of closed-loop BG control systems in clinical trials, generally accepted
criteria are the percentage of measured BG values within the target range (70–180 mg / dL), the
glycated hemoglobin level, which correlates with the average BG for 3 months observation period and
the number of low and high BG events. However, there is no method to quantify the efficiency of
closed-loop BG control at the stage of a system development and its parameters optimization, taking
into account not only clinical criteria, but also technical ones related to control efficiency and quality.
An integrated method is proposed for efficiency evaluation of BG control based on criteria of
overshoot P (gmin), maximum BG М (gmax), hyperglycemia integral index Jg (S), regulation time Jt (τ)
and safety of the advance bolus infusion B ( g min ), illustrated by the figure 3 and calculated as follows:
g  Gn
P  min , P [0;1] (1)
G fasting  Gn
Gcrit  g max
M , M   0;1 (2)
Gcrit  GHb
1
Jg  , J g   0;1 (3)
1
G (t )  GHb dt
J h TG
Hb


Jt  1  , J t   0;1 (4)
 crit
g min  Gn
BP tm  tm 15  , B  [0;1] (5)
G fasting  Gn

where gmin is a minimum BG, Gfasting is a target BG, Gn is the lower limit of the target range, gmax is a
maximum BG, Gcrit is the upper threshold value, GHb is the average BG value for a normal glycated
hemoglobin index, Jh is the integral value under similar conditions for a healthy person, τcrit is the
maximum regulation time, g min is a minimum BG for B criterion, tm is the meal time.

4
ICCT 2021 IOP Publishing
Journal of Physics: Conference Series 2091 (2021) 012020 doi:10.1088/1742-6596/2091/1/012020

The criteria combination using to calculate the overall efficiency function depends on the purpose
of the method usage. For example, the criteria Jg and Jt are correlated and it is sufficient to use only
one of them. At the same time, the hyperglycemia integral index is more complex in calculation and
requires processing a large amount of data and is applicable rather at the stage of optimization of the
control system parameters, while the regulation time can be used when testing closed-loop systems
during preclinical trials. Criterion B is related to modeling the possibility of meal delay and the
predicted risk to the patient's health in this case. As such, it cannot be evaluated when analyzing the
results of research trials, but it can be used to determine the optimal and safe bolus form.

Figure 3. Illustration of efficiency criteria.

3. Results
Taking into account the considered problems and on the basis of proposed methods, the closed-loop
BG control system was developed [18]. The system comprises of an automatic adaptive adjustment of
basal delivery according to CGM data, and decision support algorithms based on model predictive
control for calculation of optimal after-meal boluses, corrective boluses and carbohydrates as well as
detecting and compensation of the system functioning and operation violations (figure 4).

Figure 4. The structure of the closed-loop insulin therapy system with decision support.

5
ICCT 2021 IOP Publishing
Journal of Physics: Conference Series 2091 (2021) 012020 doi:10.1088/1742-6596/2091/1/012020

In silico testing of the system was performed via UVA/Padova Type 1 Diabetes Mellitus Simulator.
The proposed system was compared with a simulated open-loop one that provides instant bolus
injection and constant basal delivery. A comparative analysis of the BG control efficiency was carried
out according to the method proposed in 2.5. The overall efficiency function was compiled on the
basis of overshoot, maximum BG and regulation time criteria. For each BG fluctuation, the efficiency
increase for the proposed closed-loop system compared to the open-loop one had been calculated.
The control algorithm showed promising results in BG maintaining in adults and adolescents [19].
The 91 % of measures were in the target range during 72-hours simulation with regular meals from 20
to 75 g (figure 5). The adaptive adjustment of the controlling action allowed decreasing the overshoot
during nights in comparison with insulin infusion with constant rate. The average increase in BG
control efficiency was 27.3% with a decrease in the maximum value of 10.6%, a decrease in the
regulation time of 22.5% and an insignificant increase in overshoot of 4.3%.

Figure 5. Results of 72-hours simulation with regular meals from 20 to 75 g.

The ability of the system to detect the risk of low and high BG levels and the quality of proposed
optimized corrective actions was tested. The system allows advance risks detection and
recommendations making, while the simulations shows better BG dynamics when a patient makes a
decision to inject a recommended corrective bolus or to intake the calculated corrective amount of
carbohydrates than without the decision (figure 6).

(a) (b)
Figure 6. BG dynamics with and without corrective action in cases of low (a) and high (b) BG.

6
ICCT 2021 IOP Publishing
Journal of Physics: Conference Series 2091 (2021) 012020 doi:10.1088/1742-6596/2091/1/012020

The developed decision support algorithm detects the violations with high sensitivity (96–100 %,
CI 95 %, p < 0.001) and high specificity (91–100 %, CI 95 %, p < 0.001), that allows preventing BG
deviations from the normal range caused by the human factor or system functioning violations. The
results of evaluating the ability of the algorithm to detect the events of absent meal data and additional
insulin injection data, absent meal fact and pump malfunction for a given threshold time Tthr are
presented in the table 1.

Table 1. Results of developed decision support algorithm testing.

Event Tthr, Time of Sensitivity, % Specificity, %
min detection, min

Meal data absence 90 40.0 ± 6.7 100,0 (CI 95 %, 95.8 – 100.0) 91,5 (CI 95 %, 90.1 – 91.5)

Additional insulin
80 42.4 ± 3.0 100,0 (CI 95 %, 96.9 – 100.0) 99.7 (CI 95 %, 98.6 – 99.7)
data absence
Meal fact absence 120 74.7 ± 19.5 96,0 (CI 95 %, 90.8 – 98.6) 95.0 (CI 95 %, 93.3 – 95.9)

Pump malfunction 90 71.4 ± 16.5 97,0 (CI 95 %, 93.6 – 97.0) 100.0 (CI 95 %, 98.9 – 100.0)

4. Discussion & Conclusion

The construction of an automated closed-loop BG control system is accompanied by a number of
difficulties. This paper considers such problems as delays in a closed-loop associated with the
peculiarities of insulin and glucose absorption; problems of control and timely compensation of high
and low BG; the human factor; problems of mathematical modeling of the control object and
evaluation of BG control efficiency
Control and decision-making support methods, which allow taking into account or compensation
for these problems are considered and proposed. On the basis of the proposed methods a closed-loop
BG control system was developed and tested via the UVA/Padova Type 1 Diabetes Mellitus
Simulator.
Methods of adaptive basal delivery by feedback sensor data based on nested threshold values and
methods of BG dynamics prediction make it possible to compensate for the lagging in the closed-loop
by advance controlling action. Moreover, a comparative analysis of the predicted and measured
profiles, as well as control of their approach to the target range boundaries, provide timely decision-
making support on BG normalization and compensation of system functioning and operation
violations. Despite the fact that these methods allow minimizing the patient participation in the BG
control, they still require additional data input and making corrective and managerial decisions.

Acknowledgments
The reported study was funded by RFBR, project number 19-37-90028.

References
[1] Trevitt S, Simpson S and Wood A, 2016 J. Diabetes Sci Technol 10 714-723
[2] Litinskaia E L and Pozhar K V, 2019 Proceedings of the 2019 III International Conference on
Control in Technical Systems 152-155
[3] Bergenstal R et al., 2016 JAMA 316 1407-1408
[4] Brown S A et al., 2019 N. Eng. J. Med. 381 1701-1717
[5] Buckingham B A et al., 2018 Diabetes Technology & Therapeutics 20 257-262
[6] Kovatchev B P, Patek S D, Ortiz E A and Breton M D, 2015 Diabetes Technology &
Therapeutics. 17 177-86
[7] Forlenza G P et al., 2018 Diabetes Care 41 2155-2161
[8] Litinskaia E L, Pozhar K V and Polyakova E A, 2021 Biomedical Engineering 54 366-370

7
ICCT 2021 IOP Publishing
Journal of Physics: Conference Series 2091 (2021) 012020 doi:10.1088/1742-6596/2091/1/012020

[9] Wolpert H, Kavanagh M, Atakov-Castillo A and Steil G M, 2016 Diabetic Med. 33 235-242
[10] Hovorka R et al., 2004 Physiol. Meas. 25 905–920
[11] Miller S, Nimri R, Atlas E, Grunberg E A and Phillip M, 2011 Diabetes Technology &
Therapeutics 13 983-990
[12] Fabris C and Kovatchev B P, 2020 Artificial Organs 44 671-679
[13] Pérez-Gandía C et al., 2018 Journal of diabetes science and technology 12 243-250
[14] Litinskaia E L and Kozlov M Yu, 2020 Perspektivy nauki 12 54-58
[15] Bazaev N A, Pozhar К V and Rudenko P А, 2014 Meditsinskaia tekhnika 6 8-11
[16] Kraegen E W and. Chisholm D J, 1984 Diabetologia 26 208-213
[17] Dalla Man C, Rizza R.A, and Cobelli C, 2007 IEEE Trans. Biomed. Eng. 54
[18] Litinskaia E L, 2021 Izvestiya vuzov. Elektronika 26 162-171
[19] Litinskaia E L, Pozhar K C and Bazaev N A, 2021 Proceedings of the 2021 IEEE Conference of
Russian Young Researchers in Electrical and Electronic Engineering 2842-2846