Sie sind auf Seite 1von 4


1, JANUARY 2011

Closed-Loop Anesthetic Drug Concentration Estimation Using Clinical-Effect Feedback

Jin-Oh Hahn , Member, IEEE, Guy A. Dumont, Fellow, IEEE, and J. Mark Ansermino
AbstractThis letter presents a novel closed-loop approach to anesthetic drug concentration estimation using clinical-effect measurement feedback. Compared with the open-loop prediction used in current target-controlled infusion systems, closed-loop estimation exploits the discrepancy between the measured and predicted clinical effects to make corrections to the drug-concentration estimate, achieving improved robustness against variability in the patient pharmacokinetics and pharmacodynamics. A robust estimator, which processes drug administration and clinical-effect measurements to estimate the plasma- and effect-site drug concentrations, is designed using -synthesis theory. Initial proof of principle of the closed-loop estimation is demonstrated using the Monte Carlo simulation of surgical procedures with a wide range of patient models. Closed-loop estimation results in statistically signicant reductions in median percentage, median absolute percentage, and maximum absolute percentage drug-concentration errors compared to open-loop prediction. Index TermsAnesthetic drug concentration, clinical-effect feedback, closed-loop estimation.
Fig. 1. Open-loop prediction versus closed-loop estimation of anesthetic drug concentrations. (a) Open-loop prediction. (b) Closed-loop estimation.

I. INTRODUCTION NOWLEDGE of the anesthetic drug concentration in a patient during a surgical procedure is benecial for both patient monitoring and control of drug administration. Highdelity estimates of anesthetic drug concentration allow anesthesiologists to make relevant dose adjustments as well as appropriate higher-level clinical decisions. It may also benet computer-controlled anesthesia, including, but not limited to, targeting and limiting the drug concentration itself [1], [2], as well as monitoring and maintaining homodynamic stability [3]. The current practice of anesthetic drug concentration estimation is primarily based on open-loop prediction [see Fig. 1(a)], in which the drug concentration is estimated by solving the patient pharmacokinetics (PK) described in terms of a multicom-

Fig. 2.

PK/PD and effect monitor models of anesthesia process.

Manuscript received June 17, 2010; revised July 17, 2010 and August 24, 2010; accepted September 6, 2010. Date of publication September 16, 2010; date of current version December 17, 2010. This work was supported in part by the National Sciences and Engineering Research Council of Canada and by the Canadian Institute of Health Research under their Collaborative Health Research Project program. Asterisk indicates corresponding author. J.-O. Hahn is with the Department of Mechanical Engineering, University of Alberta, Edmonton, AB T6G2R3, Canada (e-mail: G. A. Dumont is with the Department of Electrical and Computer Engineering, University of British Columbia, Vancouver, BC V6T1Z4, Canada (e-mail: J. M. Ansermino is with the Department of Anesthesiology, Pharmacology and Therapeutics, University of British Columbia, Vancouver, BC V6T1Z4, Canada (e-mail: Color versions of one or more of the gures in this paper are available online at Digital Object Identier 10.1109/TBME.2010.2076811

partmental model. In the absence of feedback, this approach inevitably suffers from interpatient variability. Any discrepancy between the real PK of an individual patient and its model counterpart results in a mismatch between the true versus estimated drug concentrations [4]. The goal of this letter is to determine how much a closedloop approach, exploiting the clinical-effect measurement [see Fig. 1(b)], can improve the accuracy of anesthetic drug concentration estimation. A depth of anesthesia index called WAVCNS [5] is used in the estimation of intravenously administered propofol concentrations at the plasma and effect sites. If successful, the closed-loop estimation is expected to outperform its open-loop counterpart in suppressing the adverse inuence of patient variability by feedback correction of the drugconcentration estimate (see Fig. 1). On the other hand, it must also effectively deal with unknown surgical stimuli combined with analgesic suppression (shown as d in Fig. 2), which acts as a disturbance and distorts the clinical-effect measurement; the measured clinical effect represents the combined consequence of anesthetic drug administration and surgical stimuli, which counteract the drugs anesthetic effect. Therefore, closed-loop estimation is required to optimize the tradeoff between the robustness against patient variability

0018-9294/$26.00 2011 IEEE


and the rejection of unknown surgical-disturbances. This letter presents a robust closed-loop estimator that processes drugadministration and clinical-effect measurements to estimate the anesthetic drug concentrations. The initial proof of principle of the closed-loop estimation is demonstrated using Monte Carlo simulation of surgical procedures, which uses randomly generated patient models. The results suggest that closed-loop estimation results in statistically signicant reductions in median percentage, median absolute percentage, and maximum absolute percentage errors (PEs) over the open-loop prediction. This letter is organized as follows. Section II describes the design of the closed-loop estimator using the -synthesis theory. Section III details the Monte Carlo simulation and statistical analysis conducted in this letter to compare closed-loop estimation and open-loop prediction. The results are presented and discussed in Section IV, followed by the conclusion in Section V. II. CLOSED-LOOP DRUG CONCENTRATION ESTIMATION A. Patient Model A patient under anesthesia can be described by the cascade connection of a PK model, a pharmacodynamics (PD) model, and a monitor model shown in Fig. 2. The input to this process is the drug infusion rate (IH ), while the output is the clinicaleffect measurement, here the depth of hypnosis (WAVCNS ). The surgical stimulation is modeled as an output disturbance because it counteracts the effect of anesthetic drug in an unpredictable manner, and thereby distorts the WAVCNS measurement. 1) PK and PD Models: The three-compartment model of propofol developed by Sch ttler and Ihmsen [6] is used to deu scribe the PK. Its state-space representation is given by the following: xPK = APK xPK + BPK u yPK = Cp (1)

model derived by Zikov et al. [5] is used y(s) = WAVCNS (s) = 100 1 [E(s) + d(s)] (8s + 1)2 (4)

where d is the disturbance caused by surgical stimuli. Note that maximum (=1) and minimum (=0) clinical effects correspond to WAVCNS = 0 and WAVCNS = 100, respectively. 3) Closed-Loop Estimator Design Model: The closed-loop estimator design model is obtained by local linearization of the PD model around the operating regime of E = 0.5 during the maintenance phase of anesthesia. Combining (1)(4) results in the following closed-loop estimator design model: 0 0 0 APK kd 0 0 k 00 x= d x 0 0 1 013 1 013 2561 EC50 41 641 + BPK 031 051 641 u(t Td ) d

= Ax + Bu(t Td ) + Ed y = [ 0 0 0 0 1 0 ] = Cx. B. Closed-Loop Estimator Design The anesthetic drug delivery process model (5) is subject to signicant amount of uncertainty arising from the patient variability and unknown surgical disturbances. The populationbased PK and PD models [6] have limited predictive accuracy. Moreover, the PK and PD model parameters may change with uctuations in the individual patients physiologic condition. Furthermore, the clinical-effect measurement is persistently corrupted by unpredictable surgical stimuli. Therefore, the closedloop estimator must be robust against these multiple confounding factors. The tasks of achieving robustness against PK/PD variability and rejecting surgical disturbances conict with each other. The clinical-effect measurement is required to be accurate for highdelity feedback correction of drug concentration estimation in the presence of PK/PD variability, whereas the PK/PD model is required to be accurate in order to recognize and reject the surgical-disturbance component in the measured clinical effect. To cope with these conicting requirements, the closed-loop estimator is designed based on the -synthesis theory [7], which offers a systematic framework to optimize the tradeoff between the robustness against patient variability and the rejection of surgical disturbances. Rearranging (5) in the following, where the matrices A and B are decomposed into nominal (bar) and uncertain (tilde) components: x = Ax + Bu + Ed = (A + A)x + (B + B)u + Ed = Ax + Bu(t Td ) + Fw (6) (5)

where xPK contains drug concentrations in the plasma, fast peripheral, and slow peripheral compartments, u = IH , and Cp is the plasma concentration. To account for the PD lag associated with the distribution of drug into the effect site (i.e., the brain), the following delay plus rst-order model is used [5]: Ce (s) = eT d s kd Cp (s) s + kd (2)

where Td and kd are the transport delay and the rate of drug distribution, respectively, from the plasma to the brain. Finally, the saturating Hill equation is used to describe the relation between Ce and the anesthetic effect E(Ce ) =
EC50 Ce + Ce


where EC50 is the 50% effect concentration, and is the cooperativity coefcient. 2) WAV C N S Monitor Model: The WAVCNS index reects the combined effect of the anesthetic drug and the unknown surgical disturbance. It takes values ranging from 100 to 0, where WAVCNS = 100 corresponds to fully awake state and WAVCNS = 50 corresponds to adequate anesthesia. The WAVCNS monitor

where Fw = Ax + Bu(t Td ) + Ed is the lumped uncertainty term. The elements of F are determined such that w 2 1 is satised. The robust -estimator for the process


(6) designed in this letter assumes the following form: x = A + Bu(t Td ) + x (7)

where the feedback is the output of a dynamic system with the clinical-effect estimation error as its excitation signal x (s) = L(s)(y C) = [C (sI A ) B + D ] (y C). x (8) Our objective is to design a -estimator (8) that is robust against both the PK/PD variability and unknown surgical stimuli. Based on (5), (7), and (8), the following frequency-domain expression of drug-concentration error dynamics is obtained: Fw(s) (9) where e = x x. The robust -estimator can be designed by minimizing the H norm of the frequency-weighted closedloop disturbance-to-error transfer function We (s)Tew (s) [8], where We (s) is a weighting function specifying the desired error bounds. This letter employs the following rst-order lter as We (s) to obtain a low-order closed-loop estimator: We (s) = k 1 2 s + 1 [I44 1 s + 1 2EC50 042 ] (10) e(s) = Tew (s)w(s) = {I + sI A

} sI A

where i , i = 1, . . . , 6 were chosen to encompass the resultant effect of up to 30% random PK/PD parametric perturbations as well as surgical stimuli of up to 20 WAVCNS units. The parameters of the weighting function (10) were chosen as fol1 1 lows: k = 5.0, 1 = 2.0 105 Hz, and 2 = 1.0 109 Hz, which requires the error bounded by 20% up to 2.0 105 Hz, beyond which increase in the size of the error is permitted. Two representative anesthetic drug-administration scenarios were simulated: target-controlled infusion (TCI) and closedloop administrations. In the TCI scenario, the drug was administered so as to target an effect-site concentration of 3 g/mL.In the closed-loop-control scenario, the drug was administered to maintain the WAVCNS value at 50, which is known to correspond to adequate depth of hypnosis [5]. For both scenarios, a realistic surgical stimulation prole, adopted from Dumont et al. [9], was applied to the patient models. B. Statistical Analysis For TCI and closed-loop scenarios, respectively, the plasma and effect-site PEs associated with each of the 880 simulations were calculated for both closed-loop estimation and open-loop prediction
s P Ej = 100

where k species the error bound at steady state, 1 species the observer bandwidth, and 2 > 1 . This choice of We (s) empha1 sizes small low-frequency estimation errors ( < 1 ), but allows relatively large errors in higher frequencies. For the amount of frequency-dependent patient-model uncertainty and surgical disturbance specied in its design, -synthesis yields a closedloop estimator that minimizes worst case drug-concentration estimation error by seeking optimized tradeoff between uncertainty robustness and disturbance rejection [7]. Based on (9) and (10), the robust -estimator can be designed using the D-K iteration procedure [7]. III. METHODS A. Monte Carlo Simulation The demographic and PD data of 44 patients aged 1860 (3712) listed in Dumont et al. [9] were used to obtain 44 nominal PK and PD models (1)(3), as described in Sch ttler and u Ihmsen [6] and Dumont et al. [9], respectively. Then, 20 random Monte Carlo models were created for each of the nominal models by applying up to 30% random perturbations to all the PK and PD parameters, yielding 880 models in total that were subsequently used for the Monte Carlo simulation. For each patient, the open-loop prediction and the closedloop estimation were designed based on the nominal models and tested against the aforementioned 20 random Monte Carlo models. The open-loop prediction estimates the drug concentration using (1) and (2). For the closed-loop estimation, the robust -estimator (7) was designed by applying the D-K iteration procedure in the MATLAB robust control toolbox [10]. The PK/PD variability was specied as follows: Fw = Ax + Bu(t Td ) + Ed = diag {i } (11)

Cs (j) Cs (j) Cs (j)


s where P Ej denotes the PE for the sample j at the site s (p for plasma and e for effect site), and Cs (j) and Cs (j) are the true versus open-loop-predicted or closed-loop-estimated drug concentrations. The median PE (MDPE), median absolute PE (MDAPE), and maximum absolute PE (MXAPE) were calculated as described in Dumont et al. [9] and used to compare the closed-loop estimation and the open-loop prediction. The closed-loop-estimations reduction in errors over the open-loop prediction was assessed by calculating the median value of the 880 closed-loop-estimated MDPE, MDAPE, and MXAPE normalized by their open-loop counterparts. The statistical significance of the error reduction was determined by applying the two-sample z-test to the 880 pairs of closed-loop-estimated and open-loop-predicted MDPE, MDAPE, and MXAPE.

IV. RESULTS AND DISCUSSION The distribution of the closed-loop-estimated compared to the open-loop-predicted plasma and effect-site MDPE, MDAPE, and MXAPE over 880 random Monte Carlo patient simulations are summarized in Table I for TCI and closed-loop-control scenarios. The resulting reduction in error, together with the p-values, accomplished by the closed-loop estimation is summarized in Table II. The closed-loop estimation with the use of clinical-effect measurement as feedback correction can signicantly improve anesthetic drug concentration estimation. The closed-loop estimation outperformed the open-loop prediction in terms of both mean and SD of the error metrics assessed (see Table I), except for the MDPE of Cp and Ce under the TCI scenario and the MXAPE of Cp under the closed-loop-control scenario. Furthermore, the closed-loop estimation resulted in statistically




signicant reductions in both plasma- and effect-site drugconcentration errors (see Table II). The reduction of SD achieved by the robust -estimator indicates that, in contrast to the openloop prediction (in which the PK/PD variability directly deteriorates the drug-concentration estimates), the clinical-effect feedback is indeed helpful in suppressing the propagation of the adverse inuence of the PK/PD variability into the drugconcentration errors. In essence, closed-loop estimation can potentially improve the performance of drug-delivery control in real clinical settings. In addition to the clinical-effect feedback, the strength of the robust -estimator originates from its explicit specication of the amount of the PK/PD variability and unknown surgical disturbances it must tolerate [see (6) and (11)], and its subsequent use for the optimization of the closed-loop estimator design. Compared with the open-loop prediction, which is passively exposed to these confounding factors, the closed-loop estimation with the robust -estimator is tuned to optimally exploit the clinical-effect measurement (which is distorted by unknown surgical stimuli) to yield more accurate drug-concentration estimates in spite of the PK/PD variability. This robust -estimator-based closed-loop estimation may have signicant benets over open-loop prediction, as suggested by the proof-of-principle results presented in this letter. However, the benet of the closed-loop approach appears to be limited by the conict between the robustness against the PK/PD variability and the surgical-disturbance rejection. Indeed, the conservative specication of the PK/PD variability (i.e., 30%) and the surgical disturbance (i.e., 20 WAVCNS units) that the closed-loop estimator is required to tolerate does restrict performance of the estimator, especially in the absence of any a priori knowledge on the amount of PK/PD variability and the degree of surgical stimuli for an individual patient undergoing a specic procedure. Specifying a reduced amount of PK/PD variability in (11) may result in a robust -estimator with improved surgicaldisturbance rejection (since the PK/PD model used in the design of the robust -estimator becomes more accurate so as to discern the effect of anesthetic drug from the effect of surgical stimuli).

Conversely, specifying a reduced degree of surgical stimulation in (11) may allow the robust -estimator to further suppress the variability in the drug-concentration error (since the clinicaleffect measurement becomes more reliable with less distortion, which enables a more rigorous feedback correction). In this regard, there appears to be additional opportunities to improve the closed-loop estimation by developing on-line patient and surgical stimuli characterization methods that can contribute to reduce the variability of PK/PD and surgical disturbances in the design of closed-loop estimators. V. CONCLUSION A novel closed-loop approach to the estimation of anesthetic drug concentration was described. The closed-loop estimation with the clinical-effect measurement feedback can improve the drug-concentration error compared to the traditional open-loop prediction, in the presence of the PK/PD variability and the unknown surgical stimuli distorting the clinical-effect measurement. We are currently investigating opportunities for improving the closed-loop estimation by incorporating on-line characterization of the individual patient PK/PD. REFERENCES
[1] G. E. Van Poucke, L. J. Brandon Bravo, and S. L. Daly, Target controlled infusions: Targeting the effect site while limiting peak plasma concentration, IEEE Trans. Biomed. Eng., vol. 51, no. 11, pp. 18691875, 2004. [2] A. Gentilini, C. Schaniel, M. Morari, C. Bieniok, R. Wymann, and T. Schnider, A new paradigm for the closed-loop intraoperative administration of analgesics in humans, IEEE Trans. Biomed. Eng., vol. 49, no. 4, pp. 289299, Apr. 2002. [3] R. R. Rao, B. W. Bequette, and R. J. Roy, Simultaneous regulation of hemodynamic and anesthetic states: A simulation study, Ann. Biomed. Eng., vol. 28, no. 1, pp. 7184, 2000. [4] V. Sartori, P. M. Schumacher, T. Bouillon, M. Luginbuehl, and M. Morari, On-Line Estimation of Propofol Pharmacodynamic Parameters, in Proc. IEEE EMBC, Shanghai, China, 2005, pp. 7477. [5] T. Zikov, S. Bibian, G. A. Dumont, M. Huzmezan, and C. R. Ries, Quantifying cortical activity during general anesthesia using wavelet analysis, IEEE Trans. Biomed. Eng., vol. 53, no. 4, pp. 617632, Apr. 2006. [6] J. Sch ttler and H. Ihmsen, Population pharmacokinetics of propofol: A u multicenter study, Anesthesiology, vol. 92, no. 3, pp. 727738, 2000. [7] S. Skogestad and I. Postlethwaite, Multivariable Feedback Control Analysis and Design. Chichester, UK: Wiley, 1996. [8] A. M. Pertew, H. J. Marquez, and Q. Zhao, H observer design for lipschitz nonlinear systems, IEEE Trans. Automat. Control, vol. 51, no. 7, pp. 12111216, Jul. 2006. [9] G. A. Dumont, A. Martinez, and J. M. Ansermino, Robust control of depth of anesthesia, Int. J. Adaptive Control Signal Process., vol. 23, no. 5, pp. 435454, 2009. [10] G. Balas, R. Chiang, A. Packard, and M. Safonov, Robust Control Toolbox Users Guide. Natick, MA: MathWorks, 2010.