Professional Documents
Culture Documents
IJFuzzy
IJFuzzy
In this work we present a data-driven modeling of the insulin dynamics in different in silico patients
using a recurrent neural network with output feedback. The inputs for the identification is the rate
of insulin (µU/dl/min) applied to the patient, and blood glucose concentration. The output is insulin
concentration (µU/ml) present in the blood stream. Once completed the off-line modeling, this model
could be used for on-line monitoring of the insulin concentration for a better treatment. The learning
law of the recurrent neural network is inspired by adaptive observer theory, and proven to be convergent
in the parameters and stable in the Lyapunov sense, even with only 13 samples available. Simulation
results are shown to validate the presented modeling.
‡
Corresponding author.
149
April 13, 2010 15:13 00232
or controllers,6 they were used as static function the insulin dynamics is discussed. For this sys-
identifiers,7, 8 provide that the functions to be iden- tem, a black-box structure based on Recurrent Neu-
tified were smooth. The interpretation of a fuzzy ral Networks is presented in Sec. 3, jointly with
system as a neural network, called neurofuzzy net- a training algorithm, where also the boundness
work, has been used to combine the best properties and convergence of the training algorithm are dis-
of both: having the space partition via membership cussed. Examples for the validation of the perfor-
functions, and take advantage of the training algo- mance of the neural identifier are shown in Sec. 4.
rithms designed for neural networks.9 Finally, conclusions and future work are drawn in
When dealing with identification of dynamic Sec. 5.
models, neurofuzzy networks incorporate dynamics
into the structure. This kind of neurofuzzy net-
works are then called recurrent neurofuzzy net- 2. Problem Statement
works (RNFN). One of the very first approaches Glucose is the principal energy substrate of the
given10 considered one-bit memory in each neu- human body, but at least 80% of body cells can not
ron. In this context, several models have been absorb it by their own, so they need insulin to help
proposed, using different structures.9, 11–15, 37, 38 In them to do it. Insulin is a hormone produced in pan-
training of recurrent networks, the usual train- creas by Beta cells (β cells), that also regulates the
ing schemes (as backpropagation or gradient tech- glucose storage by liver, muscles, and adipose tissue,
niques16 ) are not longer useful and schemes that take in order to have glucose available for fasting periods.
into account the dynamics of the network ought to When insulin is higher than its basal level these tis-
be used, such as Backpropagation-Through-Time,17 sues absorb glucose to store it, and the lack of insulin
or Recurrent-Learning-Through-Time.18 However, activates the mechanisms to release it.22
these techniques tend to be sluggish and unstable In Diabetes Type-1 β cells are destroyed by
in the presence of disturbances. autoimmune mechanisms, so the pancreas can not
Regarding models to be used in control or esti- produce insulin anymore. In consequence, the glu-
mation problems, structures in state-space repre- cose concentration is always higher than its basal
sentation are preferred in some applications.19, 20 level. Current medical treatment suggests three or
However, they require the actual measurement of the four daily glucose measurement, with the same num-
physical states of the system, which is not always ber of subcutaneous insulin injections.23 Insulin can
the case. Previously,21 a network with a stable train- also be delivered by an insulin pump, and there are
ing algorithm is proposed, where it generates its own real time glucose monitors with implantable sensors.
states in order to have the same input-ouput dynam- This represents the sensor and actuator for a closed
ics as a real system, alleviating the need of measuring loop control for glucose. There are several research
those of the system. works about this topic.24–27
In this paper we propose a data-driven modeling Insulin therapy can have some complications. In
in state-space representation of the insulin dynam- the first place, the accumulation of insulin injections
ics in diabetic Type-1 patients, by analyzing the can lead to a excessive concentration of insulin in
response of the blood stream insulin concentration blood (also known as hyperinsulinemia) that pro-
when insulin is directly fed into the body and when duces hypoglycemia, a life-threatening condition.
the glucose concentration is measured in real-time. If hyperinsulinemia persists, but within euglycemia
The training of the network is done using lab tests range, it can produce insulin resistance, and then
that can measure both glucose and insulin concen- impairing the effect of the therapy. Otherwise, pro-
trations in the bloodstream. Once the training is longed lack of insulin, can lead to excessive lipol-
completed, the objective is to have the network ysis and unrestrain fatty acid oxidation producing
to reproduce the insulin dynamics when only glu- metabolic acidosis and dehydration.28 So, it is desir-
cose measurements and the injected insulin rate are able to design a system that can estimate the con-
available. centration of insulin level in the bloodstream to help
This work is structured in the following man- that those limit conditions are not neither surpassed
ner: In Sec. 2 the basic modeling problem of nor fall below a certain level.
April 13, 2010 15:13 00232
A Discrete-Time Recurrent Neurofuzzy Network for Black-Box Modeling of Insulin Dynamics 151
with (A, C) observable, and f¯(y, u) smooth. In the θu2,i = [ θu2,i,1 θu2,i,2 . . . θu2,i,n ]T ,
same way, it is assumed that (1) fulfills the sufficient θy,i = [ θy,i,1 θy,i,2 . . . θy,i,n ]T .
April 13, 2010 15:13 00232
Remark 1. The main function that the fuzzy be obtained as a function of the input and output sig-
structure has in the neurofuzzy network is to pro- nals from (1). The consequent part in the fuzzy rules
vide with a space partition i, and then each section is designed in such a way to meet with the persistence
be modeled by a nonlinear system in the form of the of excitation condition (see below). However, other
fuzzy rule Ri . type of fuzzy rules may be considered, as long as they
satisfy the persistence of excitation condition.
If the defuzzification is defined as a weighted aver-
r r
age ηk+1 = i=1 Fi ηk+1i
/ i Fi , then the regressor Remark 4. Similarly to some previous work,34 the
ϕ can be defined by regressor in this proposed scheme can be proven to
comply with the persistence of excitation condition.
ϕ(yk , uk ) = [ϕu1 (yk , uk )ϕu2 (yk , uk )ϕy (yk , uk )]
∈ n×3nr , (4) Remark 5. By construction, it follows from (5)
that the regressor ϕ is always bounded if u is
and ϕu1 , ϕu2 , ϕy ∈ 4 can be expressed as bounded.
[ F1 In×n . . . Fr In×n ]u1,k
ϕu1 (yk , uk ) = r , 3.1. Observation-training algorithm
i Fi
A Discrete-Time Recurrent Neurofuzzy Network for Black-Box Modeling of Insulin Dynamics 153
A Discrete-Time Recurrent Neurofuzzy Network for Black-Box Modeling of Insulin Dynamics 155
blood can be obtained from the patient in any desired Fig. 9. Training and identification: Real case — Patient
time length. For this case, a total of 400 min and a #1. yd is the response of the in silico patient, y is the
sampling time Ts = 0.2[min] are taken. Three dif- output identified by the network.
ferent patient models are identified, using the input
signals (insulin and glucose) depicted in Figs. 4–6 for
each patient. As indicated before, the Fuzzy C-Means shown in Figs. 3–5 for each patient. It can be seen
algorithm is run in order to obtain the initial mem- that the parameters remain stable and convergent.
bership functions, which are depicted in Fig. 2. The After training, the network is compared in parallel
convergence of the training algorithms, as well as the with the output obtained from the in silico patient.
progression of the RMS error in normalized data are These results are depicted in Figs. 6–8.
April 13, 2010 15:13 00232
A Discrete-Time Recurrent Neurofuzzy Network for Black-Box Modeling of Insulin Dynamics 157
C-Means algorithm using the input and output sig- 3. J. T. Sorensen, A physiologic model of glucose
nals depicted in Figs. 5–7 (corresponding to patient metabolism in man and its use to design and assess
#1, #2 and #3 respectively), and then the training improved insulin therapies for diabetes, Ph.D. disser-
tation, Massachusetts Institute of Technology. Dept.
algorithm is run and the system and the network are of Chemical Engineering (1985).
compared in parallel. In this example, the dynamics 4. M. Wilinska, L. J. Chassin, H. Schaller, L. Schaupp,
are compared using a validation set where more data T. Pieber, and R. Hovorka, Insulin kinetics in type-
is available. It can be seen that the parameters are 1 diabetes: continuous and bolus deliveryinsulin
still convergent, as can be seen in Figs. 9–11, and that kinetics in type-1 diabetes: continuous and bolus
delivery, Biomedical Engineering, IEEE Transac-
the dynamics is captured by the network, although
tions 52-1 (2005) 3–12.
some high-frequency information is lost because of 5. C. D. Man, R. Rizza and C. Cobelli, Meal simulation
the increased sampling time. model of the glucose-insulin system, IEEE Trans-
action on Biomedial Engineering 54 (2007) 1740–
1749.
5. Conclusions 6. K. Narendra and K. Parthasarathy, Identification
In this work we presented a data-driven modeling of and control of dynamical systems using neural net-
works, IEEE Tras. Neural Networks 1 (1990) 4–27.
the insulin dynamics in different in silico patients
7. G. Cybenko, Approximation by superpositions of a
using a discrete-time recurrent neural network with sigmoidal function, Math. Con. Signal Syst. 2 (1989)
output feedback, where the input signals are consid- 303–314.
ered as the rate of insulin (in µU/dl/min) applied to 8. J.-S. R. Jang, Anfis: Adaptive-network-based fuzzy
the patient, and the blood glucose concentration, and inference system, IEEE Transactions on Systems,
the output is insulin concentration (µU/ml) present Man and Cybernetics 23 (1993) 665–685.
9. C.-F. Juang, A tsk-type recurrent fuzzy network for
in the blood stream. With a membership initializa- dynamic systems processing by neural network and
tion based on the Fuzzy C-Means algorithm, and a genetic algorithms, IEEE Transactions of Fuzzy Sys-
learning law inspired from adaptive observer theory, tems 10(2) (2002) 155–170.
it was proven to be convergent in the parameters 10. J. Hopfield, Neural networks and physical systems
and stable in the Lyapunov sense even under model with emergent collective computational abilities,
Proceedings of the National Academy of Sciences
mismatch and low sampling rates, and when just
79(8) (1982) 2554–2558.
few data is available. Simulation and experimental 11. V. Gorrini and H. Bersini, Recurrent fuzzy sys-
results were shown in order to validate the presented tems, Proc. IEEE Int. Conference on Fuzzy Systems
modeling, using only input-output data and prac- (1994), 193–198.
tically no knowledge of the models of subsystems 12. A. D. Back and A. C. Tsoi, A time series modeling
involved in the system. Future work includes con- methodology using fir and iir synapses, Proc. Work-
shop on Neural Networks for Statistical and Eco-
trol schemes based in the identified model, as well as nomic Data (1990), 187–194.
analysis of the obtained model. 13. P. A. Mastorocostas and J. B. Theocharis,
A recurrent fuzzy-neural model for dynamic system
identification, IEEE Trans. onf Systems, Man and
Acknowledgments
Cybernetics — Part B: Cybernetics 32(2) (2002)
This work is partly supported by UNAM-PAPIIT 176–190.
IN120009. The work of Marcos A. González-Olvera 14. A. Poznyak, E. Sanchez, and W. Yu, Differen-
tial Neural Networks for Robust Nonlinear Con-
and Ana G. Gallardo-Hernández is supported by
trol: Identification, State Estimation and Trajectory
Conacyt. Tracking (World Scientific, 2001).
15. M. A. Gonzalez-Olvera and Y. Tang, Identification
and control of nonlinear dynamic systems via a con-
References strained input-output neurofuzzy network, Advances
1. R. N. Bergman, Y. Z. Ider, C. R. Bowden, in Control Systems Theory and Applications (eds.)
and C. Cobelli, Quantitative estimation of insulin G. Tao and J. Sun (2008).
sensitivity, Am. J. Physiology, Endocrinology and 16. O. Nelles, Nonlinear System Identification (Springer,
Metabolism 235 (1979) E667–E677. Berlin, 2001).
2. V. Bolie, Coefficients of normal blood glucose regu- 17. P. Werbos, Backpropagation through time: what it
lation, Journal of Applied Physiology 16 (1961) 783– does and how to do it, Proceedings of the IEEE
788. 78(10) (1990) 1550–1560.
April 13, 2010 15:13 00232
18. R. J. Williams and D. Zipser, A learning algorithm 28. Larsen, Kronenberg, Melmed and Polonsky,
for continually running fully recurrent neural net- Williams Textbook of Endocrinology (Sanuders,
works, Neural Computation 2(1) (1989) 270–280. USA, 2003).
19. A. Poznyak, T. Poznyak and I. Chairez, Dynamic 29. M. Spong, Robot Dynamics and Control (John Wiley
neural observers and their application for identifica- & Sons, Inc. New York, 1989).
tion and purification of water by ozone, Automation 30. R. Marino, Adaptive observers for single output non-
and Remote Control 67(6) (2006) 887–899. linear systems, Automatic Control, IEEE Transac-
20. W. Yu and X. Li, Fuzzy identification using fuzzy tions on 35(9) (1990) 1054–1058.
neural network, IEEE Trans. on Fuzzy Systems 31. G. Zhang, B. E. Patwo and M. Y. Hu, Forecast-
12(3) (2004) 411–420. ing with artificial neural networks: The state of the
21. M. A. Gonzalez-Olvera and Y. Tang, Identification art, International Journal of Forecasting 14 (1998)
of a class of nonlinear systems by a continuous- 35–62.
time recurrent neurofuzzy network, American Con- 32. Q. Zhang, Adaptive observer for multiple-input-
trol Conference, to be published. multiple-output (MIMO) linear time-varying sys-
22. A. Guyton and J. Hall, Textbook of Medical Physiol- tems, IEEE Transactions on Automatic Control
ogy (W. B. Saunders, 2000). 47(3) (2002) 525–529.
23. A. D. Asociation, Standards of Medical Care in Dia- 33. A. Guyader and Q. Zhang, Adaptive observer for
betes (ADA, 2009). discrete time linear time varying systems, in Proc. of
24. G. Gallardo-Hernandez, L. Fridman, S. Islas- 3th IFAC/IFORS Symposium on Identification and
Andrade and Y. Shtessel, Quasi-continuous high System Parameter Estimation, SYSID 2003.
order sliding modes controllers applied to glucose- 34. G. Bastin and M. R. Gevers, Stable adaptive
insulin regulatory system models, in Proceedings of observers for nonlinear time-varying systems, IEEE
Decision and Control Conference, 9–11 December Transactions on, Automatic Control 33(7) (1988)
2008. 650–658.
25. P. Kaveh and Y. Shtessel, Blood glucose regula- 35. J. Dunn, A fuzzy relative of the ISODATA process
tion using higher-order sliding mode control, Int. and its use in detecting compact well-separated clus-
J. of Robust and Nonlinear Control 18 (2007) 557– ters, Journal of Cybernetics 3(3) (1973) 32–57.
569. 36. J. Bezdek, Pattern Recognition with Fuzzy Objective
26. R. Hovorka, V. Canonico, L. Chassin, U. Haueter, Function Algorithms (Kluwer Academic Publishers
M. Massi-Benedetti, M. Federici, T. Pieber, Norwell, USA, 1981).
H. Schaller, L. Schaupp, T. Vering and M. Wilinska, 37. H. Adeli and X. Jiang, Dynamic fuzzy wavelet neu-
Nonlinear model predictive control of glucose con- ral network model for structural system identifi-
centration in subjects with type 1 diabetes, Physio- cation, Journal of Structural Engineering 132(1)
logical Measurements 25 (2004) 905–920. (2006) 102–111.
27. R. Parker, F. Doyle III, J. Ward and N. Pep- 38. G. Puscasu and B. Codres, Nonlinear system iden-
pas, Robust h∞ glucose control in diabetes using a tification based on internal recurrent neural net-
physiological model, American Institute of Chemical works, International Journal of Neural Systems
Engineers (AIChE) 46(12) (2004) 2537–2549. 19(2) (2009) 115–125.
Copyright of International Journal of Neural Systems is the property of World Scientific Publishing Company
and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright
holder's express written permission. However, users may print, download, or email articles for individual use.