05164699

Upload: naderjsa

Post on 04-Apr-2018

215 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/30/2019 05164699

    1/6

    Neuromuscular Blockade Nonlinear Model Identification

    B. Andrade Costa, M. Silva, T. Mendonca and J. M. Lemos

    Abstract This paper presents a methodology for parameterestimation of a nonlinear neuromuscular blockade dynamicmodel to be used as a predictive model for automated control,in general anesthesia. The neuromuscular blockade dynamicmodel comprises two blocks connected in series, a pharma-cokinetic model and the pharmacodynamic model. The phar-macokinetic model is a second order linear dynamic modeland describes the redistribution of the drug in the body. Thepharmacodynamic model is a nonlinear function, named asthe Hill equation, and it describes the interaction between theconcentration of the drug in the effect site and the measuredpatients muscle paralysis state. The identification methodologyuses four data points taken from the neuromuscular blockaderesponse obtained with the administration of the first bolus.The four data points are chosen to avoid the identification

    difficulties caused by the presence of the nonlinear behaviorof the Hill equation. This approach enables the identificationof the pharmacokinetic dynamics, that is, the two poles of thesecond order linear dynamic model followed by the estimationof the normalized parameters of the Hill equation.

    Computer simulations show that the proposed identificationmethodology is able to provide good results even when thepharmacokinetic dynamics has an order higher that two. Thissuggests that the methodology may be employed in neuromus-cular blockade automated control as a predictive model, to helpthe initial tuning of the controller parameters or in adaptivecontrol to get a first model that can be improved with onlineidentification using some recursive minimization techniques toadjust the adaptive controller or as an advising mechanism tohelp the anesthesiologist during the anesthesia.

    Keywords: Biomedical engineering, Neuromuscular Blockade,Model Identification, Nonlinear Model

    I. INTRODUCTION

    Computer controlled systems are being considered as a

    promising technology to improve the practice of anesthesia

    [1] [4]. In principle, it is possible to adjust the amount of

    drugs to the patients characteristics, that is, to keep the drug

    concentration in the patients body at a constant, safe and

    adequate level, and by that to have a quicker recovering time

    from anesthesia [3]. Another important field is the utilization

    of automated systems to control anesthesia in animals to

    reduce the costs and work of the veterinary.

    At the core of any computer controlled system for anes-

    thesia automation is a model describing the transport of the

    drug in the human body (pharmacokinetic model) and the

    This work was developed in the project IDeA framework - Integrated De-sign for Automation of Anesthesia, contract PTDC/EEA-ACR/69288/2006

    B. Andrade Costa, is with INESC-ID/DEEC/IST/TU Lisbon, R. AlvesRedol 9, 1000-029 Lisboa Portugal, [email protected]

    M. Silva, and T. Mendonca, are with Dep. Matematica Apli-cada FCUP, Rua do Campo Alegre, 687 4169-007 Porto, [email protected]

    J. M. Lemos, is with INESC-ID/DEEC/IST/TU Lisbon, R. Alves Redol9, 1000-029 Lisboa Portugal, [email protected]

    effect of the drug in the patients state (pharmacodynamic

    model). These are the cases of TCI and TIVA systems

    [2]. These systems use compartmental models obtained by

    processing data, collected from a large sample to characterize

    the properties of a population. However the TCI and

    TIVA systems can be described as operating based on one

    mean model, there is one model able to describe the

    dynamics of all patients! But the current practice shows that

    for neuromuscular blockade control there is a huge variability

    between patients [5] [6]. This suggests that the administration

    of drugs during anesthesia must be based on methods able

    to estimate patients characteristics.

    With compartmental models, each compartment is as-sumed to have homogenous properties, that is, at each time

    instant the drug distribution inside a compartment is uniform.

    This approach is similar as modeling a lumped parameter

    system. Those models may have from 2 compartments to

    12 compartments [7][8][9][10]. Other modeling techniques

    use more deep knowledge of physiology to build complex

    models [11], but this approach needs data and experimental

    procedures that are not available in a standard operating

    room.

    According to [7][8][9][12] a second order linear dynamic

    model is used to describe the pharmacokinetics of drugs

    belonging to the atracuriums family. This motivated the use

    of the second order linear dynamic model in the identificationmethodology presented in this paper, the model has a defined

    structure but the parameters must be estimated. An important

    issue comes from the constraint imposed by ethical and

    practical reasons that constrains the design of identification

    signals to be used in model identification, this makes low or-

    der dynamic models more suitable to describe local dynamics

    than higher order dynamic models where overparametrization

    cause identification problems. The problem is that low order

    models do not fully describe the all dynamics and may

    impose additional constrains in the control design. To tackle

    the problem online identification may be used to adjust the

    low order models.

    The aim of this work is to model the neuromuscu-lar blockade dynamics using the information that a anes-

    thetist/anesthesiologist has when performing her/his job in

    the operating room, by using the index evolution provided

    by the neuromuscular blockade monitor and using the infor-

    mation obtained from evolution of the infusion or from the

    sequence of bolus.

    This paper is structured as follows: Section II describes the

    neuromuscular blockade model based on a 3 compartmental

    model and gives a possible cue for the assumption of a phar-

    macokinetic second order linear model. Section III presents

    17th Mediterranean Conference on Control & Automation

    Makedonia Palace, Thessaloniki, Greece

    June 24 - 26, 2009

    978-1-4244-4685-8/09/$25.00 2009 IEEE 1137

  • 7/30/2019 05164699

    2/6

    Fig. 1. The neuromuscular blockade model with a three compartmentalmodel for the pharmacokinetics, where C1 represents the central compart-ment and C2 represents the effect compartment. The r(t) signal representsthe neuromuscular blockade response.

    the identification methodology. Section IV shows the results

    obtained with computer simulation using a database of 100neuromuscular blockade models that have a pharmacokinet-

    ics of order 4 with a zero. Conclusions are presented in

    the last section.

    II. NEUROMUSCULAR BLOCKADE MODEL

    In this section a neuromuscular blockade model based on

    a three compartmental model is presented [13]. Fig. 1 shows

    the pharmacokinetics diagram with three compartments. The

    central compartment (C1) includes organs such as the cardio-vascular and pulmonary system, brain and other organs that

    are highly irrigated. The second compartment (C2) represents

    the effect site in this case the muscle. The third compartment(C3) represents organs/tissues that have a low rate of drugabsorption/diffusion. Considering a compartment i, Vi repre-sents the total volume ([l]) of the fluid (plasma/blood) in thecompartment. The drugs mass is represented by mi ([mg]),and the volumetric flow rate of fluid leaving the compartment

    i towards compartment j is denoted by qji ([ml.min1]),this quantity in directly correlated with the hearts volumetric

    flow rate, and ki represents the drugs elimination flow rateor the rate at which the drug loses its power.

    By performing a mass balance, the drugs concentrations

    in the three compartments are described by the following set

    of differential equations, where ci(t)= mi(t)/Vi.

    c.1(t) = 11c1(t) + 12c2(t) + 13c3(t) + u(t)

    c.2(t) = 22c2(t) + 21c1(t) (1)

    c.3(t) = 33c3(t) + 31c1(t)

    y(t) = c2(t)

    The ji parameters are shown in table I, q represens the totalflow rate leaving the central compartment and f representsthe fraction of q entering/leaving compartment C2, ki isassumed to be identical in all compartments, this is justified

    by [12], where the neuromuscular blockade drug such as

    TABLE I

    PARAMETERS OF THE THREE COMPARTMENT MODEL 1.

    11 =q+kV1

    12 =fq

    V213 =

    (1f)qV3

    22 =fq+kV2

    21 =fq

    V1

    33 =(1f)q+k

    V331 =

    (1f)q+kV1

    atracurium is transformed by chemical reactions that depend

    on the bodys temperature and on the bloods acidity which is

    assumed identical in all compartments. Note that the model

    (1) may be considered a linear varying parameter model

    because the ji parameters depend on the hearts volumetricflow rate and on the bodys temperature and bloods acidity

    that may change over time.

    The pharmacodynamic model representing the effect of

    the drug is described by the Hill function, eq. (2),

    r(t) =C50

    C50 + (y(t))

    (2)

    where C50 represents the drugs concentration to obtain halfof the full effect, and is a non-dimensional parameter thatcharacterizes the slope of the Hill function.

    A. Analysis of the three compartment model

    The three compartment model given by (1) is represented

    in the state-space form and it is necessary to estimate seven

    parameters, that may be difficult. In order to simplify the

    identification procedure, an input-output representation using

    a transfer functiony(s)u(s) = P(s) will be obtained from (1).

    The transfer function P(s) has three poles and one zero and

    it can be written as shown in (3), with Q(s)= s+22

    s+33.

    P(s) =21

    (s + 11)(s + 22) 1221 1331Q(s)(3)

    It is easy to conclude that P(s) can be simplified to asecond order transfer function if Q(s) 1, that is, 22 33. Now the issue is to demonstrate that there is somephysical/physiological evidence to support that 22 33.Taking the definition of both parameters from table I and

    assuming that (1f)q >> k and fq >> k, this is acceptableotherwise the drug is quickly removed and anesthesia is not

    possible, then(1f)q

    V3 fq

    V2this implies that

    f V2V2 + V3(4)

    Defining now T2 = V2/(f q) and T3 = V3/((1 f)q) asthe time needed to renew the fluids inside compartment 2and 3, and using eq. (4) then T2 = (V2 + V3)/f q andT3 = (V2 + V3)/q. That is, the renewal times for bothcompartments are identical. This can be justified if one

    considers that blood must be renewed in such a way that

    each cell must be supplied with a constant rate of oxygen,

    that is, the flow rate must be proportional to the number of

    cells (volume of the compartment). Assuming that the above

    1138

  • 7/30/2019 05164699

    3/6

    Fig. 2. Description of the time instantes t1 < t2 < t3 < t4 to be usedby the identification methodology.

    argument is true then Q(s) 1 and the transfer function (3)is written as

    P(s) =21

    (s + a)(s + b)(5)

    witha

    andb

    being the transfer function poles computed from

    the parameters 11, 22, 12, 21, 13 and 31.

    III. IDENTIFICATION METHODOLOGY

    In order to present the identification methodology, the

    second order linear model is normalized such that its static

    gain is set to one, and the Hill function parameters are

    modified in such a way that the input-output properties of

    {r(t), u(t)} are maintained. The new model is represented byz(s)u(s) = M(s), with

    M(s) =ab

    (s + a)(s + b)(6)

    and

    r(t) = (C50)

    (C50) + (z(t))

    (7)

    where C50= (a.b.C50)/21. This enables the separation of

    the parameters in two classes, {a, b} are estimated first, and{, C50} are estimated later using a, and b.

    The development of the identification methodology is

    based on the fact that the first neuromuscular blockade bolus

    usually forces the patient to a state of total paralysis, the

    index monitor r(.) goes from the value 1.0 to 0.0 or to avalue near 0.0, as in fig. (2). After some time and dependingon the pharmacokinetics, r(.) starts to increase. As a standardpractice the anesthesiologist manipulates the drug infusion

    rate or the bolus sequence to maintain r(.) near the value of0.1 (10%). This means that there is a time window definedby {t0 < t < tf : r(t) < 0.1} where it is possible to select4 data points with time instantes such that

    r(t1) = r(t4)r(t2) = r(t3)

    (8)

    with t1 < t2 < t3 < t4, fig.(2).The constraints imposed by (8) imply that

    z(t1) = z(t4)z(t2) = z(t3)

    (9)

    Because the pharmacokinetic model response to the first

    bolus is an approximation of the impulsive response of a

    dynamic system, the impulse response equation of a second

    order model hm(t), (10), is used

    hm(t) =ab

    b a(expat expbt) (10)

    to estimate the parameters a and b from the measured values

    t1 < t2 < t3 < t4. These time instants must be selected insuch a way to avoid numeric problems.

    A. Estimating the dynamics ofM(s)

    Using the constraint z(t1) = z(t4) and eq. (10) then a isgiven by

    a =1

    t4 t1ln[

    1 exp0t4

    1 exp0t1] (11)

    with 0 = b a. A similar equation can now be written withparameters t2 and t3

    a =1

    t3 t2ln[

    1 exp0t3

    1 exp0t2] (12)

    The eq.s (12) and (11) are now used to define a function,

    f(), such that

    f() =1

    t3 t2ln[

    1 expt3

    1 expt2]

    1

    t4 t1ln[

    1 expt4

    1 expt1] (13)

    The 0 = b a is obtained by solving the nonlinearequation f() = 0. Knowing the estimate 0 then a andb can be computed from

    a =1

    t4 t1ln[

    1 exp0t4

    1 exp0t1]

    b = 0 + a (14)

    B. Estimating the parameters of the Hill function

    Knowing the estimates a and b and using eq.(10) it is pos-sible to compute z(t1) and z(t2) corresponding respectivelyto the measured values r(t1) and r(t2). Rearranging the Hillfunction it can be written as z = (C50)

    (1/r 1). Thisyields (z(t1)/z(t2)) = (1/r(t1) 1)/(1/r(t2) 1). Byfurther algebraic manipulation then and C50 are obtainedfrom

    = [ln(z(t1)

    z(t2))]1[ln(

    1

    r(t1) 1) ln(

    1

    r(t2) 1)] (15)

    C50 = z(t1)r(t1)

    r(t1) 1(16)

    C. Evaluating the parameter estimation error

    In practice the measurements of the time instants t1 0.2 and a smallerror for r(t) 0.2, as shown in fig.(5). This estimationresult is the worst case obtained with the 100 model database.

    C. Simulation results from the model n.69

    The model number 69 is characterized by the following

    parameters: Static gain Sg = 0.2108, z1 = 0.1230, p1 =0.0360 , p2 = 0.0723, p3 = 0.0996, p4 = 0.2999,C50 = 0.6163, = 4.2189. Note that in this model z1 isnear p3 and it can be considered a zero-pole cancelation. Thenormalized C50 is C

    50 = C50/Sg = 2.9236. The measuredtime instants (in [min]) obtained by linear interpolation the

    1140

  • 7/30/2019 05164699

    5/6

    0 10 20 30 40 50 60 70 80 90 1000

    0.1

    0.2

    0.3

    0.4

    0.5

    0.6

    0.7

    0.8

    0.9

    1

    r(t)

    t [min]

    + Model of order 2

    o Output of the true model

    Fig. 5. Neuromuscular blockade response for the first atracurium bolusof 500g/kg, using model n.18 (symbol - o). Response of the estimatedmodel, symbol +.

    0 10 20 30 40 50 60 70 80 90 1000

    0.1

    0.2

    0.3

    0.4

    0.5

    0.6

    0.7

    0.8

    0.9

    1

    r(t)

    t [min]

    + Model of order 2

    o Output of the true model

    Fig. 6. Neuromuscular blockade response for the first atracurium bolusof 500g/kg, using model n.69 (symbol - o). Response of the estimatedmodel, symbol +.

    bolus sampled data response are t1 = 9.9364, t2 = 11.3413,

    t3 = 49.7039, t4 = 54.4795.Using the estimation methodology, the estimated non-

    linear second order model has p1 = 0.0185, p2 = 0.0725,C50 = 3.1387, = 7.3714.

    In this case, the C50 and C

    50 are similar, and are verydifferent. But if one compares the neuromuscular blockade

    responses of both models they are very similar as shown in

    fig.(6).

    The main conclusion from the above results is that, to

    obtain a good modeling of the bolus response it is necessary

    to obtain a good estimation of the pharmacokinetic model.

    The estimate C50 is near the true value, but has a hugeerror, however this huge error causes a degradation in the

    bolus response estimation.

    V. CONCLUSIONS

    This paper presents an identification methodology to

    obtain a second order non-linear model to describe the

    patient/drug interaction for the neuromuscular blockade in

    general anesthesia. The model is identified using four data

    points taken from the first bolus response. The method

    estimates in first place the pharmacokinetics followed by the

    pharmacodynamics which is described by the Hill equation

    with normalized parameters. The method was evaluated by

    computer simulation using a neuromuscular blockade model

    database with 100 models. The computer simulation results

    show that the proposed identification methodology has a

    good performance when tested with the models of the

    database. The worst case was obtained with models that

    cannot be approximated by a second order system, that is,

    they do not have a near zero-pole cancelation. Despite

    this fact, the estimated models were able to provide a good

    approximation of the neuromuscular blockade response for

    r(.) < 0.2 (20%).

    The main conclusion from the above results is that, to

    obtain a good modeling of the bolus response it is necessary

    to obtain a good estimation of the pharmacokinetics model,

    and this raises the problem of selecting the best structure to

    model the real cases.

    At the present stage there are several unanswered ques-

    tions, that is:

    What is the noise effect in the estimation?

    There is a significant number of real cases with a flat

    response at r(t) = 0.05. How to handle it? What is the closed-loop performance of a controller that

    is designed using above identification methodology?

    As a future work the identification methodology will be

    evaluated with realistic data taken from real cases and if

    possible the methodology will be extended to higher order

    models.

    REFERENCES

    [1] S. Schraag , Theoretical basis of target controlled anaesthesia: history,concept and clinical perspectives, Best Practice & Research ClinicalAnaesthesiology, Vol. 15, No. 1, pp. 1-17, 2001

    [2] DK. Sreevastava, KK. Upadhyaya, Automated Target Controlled Infu-sion Systems: The Future of Total Intravenous Anaesthesia, MJAFI,

    Vol. 64, No. 3; 2008[3] P. Gorce, Economical aspects of concentration-oriented anaesthesia:

    intravenous agents, Best Practice & Research Clinical Anaesthesiol-ogy, Vol. 15, No. 1, pp. 137-142, 2001

    [4] JM. Bailey, WM. Hadda Drug Dosing Control in Clinical Pharma-cology, IEEE Control Systems Magazine, Vol. April, No. April, pp.35-51, 2005

    [5] T. Mendonca, P. Lago PID control strategies for the automationcontrol of neuromuscular blockade, Control Engineering Pratice, Vol.6, pp. 1225-1231, 1998

    [6] H. Alonso, J.M. Lemos, T. Mendonca A Target Control Infusionmethod for neuromuscular blockade based hybrid parameter estima-

    tion, 30th Annual International IEEE EMBS Conference Proceedings,August 20-24, pp. 707-1231, 2008

    [7] ST. Young, KN. Hsiao, A Pharmacokinetic Model To Study Admin-istration of Intravenous Anaesthetic Agents, IEEE Engineering inMedice and Biology, Vol. April/May, pp. 263-268, 1994

    [8] Dorene A. OHara, John G. Hexem, et al., The Use of a PID Con-troller to Model Vecuronium Pharmacokinetics and Pharmacodynam-ics During Liver Transplantation, IEEE Transactions On BiomedicalEngineering, Vol. 44: No. 7, pp. 610619, July 1997

    [9] R. R. Jaklitsch, D. R. Westenskow, A Model-Based Self-AdjustingTwo-Phase Controller for Vecuronium-Induced Muscle Relaxation

    During Anesthesia, IEEE Transactions On Biomedical Engineering,Vol. BME-34, No. 8, pp. 583-594, August 1987

    [10] P. M. Schumacher, K. S. Stadler, et al., Model-based control of neu-romuscular block using mivacurium: design and clinical verification,European Journal of Anaesthesiology, Vol. 23: pp. 691699, 2006

    [11] K.S. Pang, M. Weiss, P. Macheras, Advanced Pharmacokinetic ModelsBased on Organ Clearance, Circulatory and Fractal Concepts, TheAAPS Journal, Vol 9, No. 2, pp-E268-E283, 2007

    1141

  • 7/30/2019 05164699

    6/6

    [12] D. M. Fisher, P. C. Canfell, et al, Elimination of Atracurium inHumans: Contribution of Hofmann Elimination and Ester Hydrolysisversus Organ-based Elimination, Anesthesiology, Vol. 65, pp. 6-12,1986

    [13] Colins A. Shanks Pharmacokinetics of the Nondepolarizing Neuro-muscular Relaxants Applied to Calculation of Bolus and Infusion

    Dosage Regimens, Anesthesiology, Vol. 64, pp. 72-86, 1986

    1142