TY - JOUR
AU - USAR, Ian Black, MC
AB - ABSTRACT Objective: This article addresses the design of a robust autopilot for the delivery of intravenous anesthesia drugs. Methods: A mathematical framework that expresses the pharmacological variability of a patient population into uncertainty bounds is proposed. These bounds can be effectively used to tune the parameters of a controller to ensure its stability, a key design aspect related to the safety of the overall system. Results: The proposed method is applied to the control of propofol, a powerful hypnotic agent used for sedation and anesthesia. Simulations show that the controller remains stable for all patients considered and that performance are clinically acceptable. Conclusion: This methodology can be an important step forward in the design and regulatory approval of such systems. INTRODUCTION Since the pioneering work of Bickford1 in the 1950s, the development of a system for the automatic administration of anesthesia drugs to patients has remained a topic of intense research worldwide.2,–20 Providing anesthesiologists with an anesthesia “autopilot” would allow them to concentrate on higher level tasks while leaving the minutiae of constant hypnotic/analgesic drug adjustment to the autopilot. This would help to limit the effects of individual patient variability related to the quality of anesthesia, maximize the time spent in a more desirable clinical state, optimize the workload of the anesthesia team, and ultimately improve the safety and quality of care.21 A system designed for the automated administration of intravenous (IV) agents would further enable the wider practice of total intravenous anesthesia (TIVA), an anesthesia technique where only IV agents are administered. Several studies have shown TIVA to be associated with an improved patient outcome.22,23 For the military, TIVA provides a smaller logistical footprint as compared to the more traditional and widespread inhaled anesthesia practice. The smaller logistical tail derives from a smaller delivery mechanism, and the ability to use TIVA-related devices (e.g., IV pumps) and sedation drugs in multiple patient settings. This redundancy allows for readily available inventory and access to maintenance. In addition, the traditional delivery mechanism for volatile anesthesia in austere environments, the draw over vaporizer, has an uncertain future. Manufacturing, maintenance, training, and compliance issues have become more troublesome over time, leading to the search for another solution. In 2004, a meeting of tri-Service consultants in anesthesia discussed the future of combat anesthesia and considered TIVA an important technique for the future for combat casualty care.24 Unfortunately, compared to inhaled anesthesia systems, TIVA is more difficult to manage because of the lack of an available method to measure the drug concentration in plasma. Currently, volatile gas measurements are used by providers to guide their titration strategy for tailoring their anesthetic use. TIVA can be especially problematic in hemorrhagic casualties and is also more difficult to manage as some of the IV drugs have very fast onset of action and elimination. Despite the inherent difficulties related to its administration, TIVA is currently used in the United States, Europe, and many other developed countries in healthy patients undergoing elective surgeries. Supporting infusion technologies, such as target-controlled infusion (TCI) pumps, were developed to help anesthesiologists compensate for complex drug pharmacokinetics (PK) and pharmacodynamics (PD). These pumps integrate PK–PD population-normed patient models to calculate the ideal drug infusion profile in order to obtain a desired target blood plasma or effect-site concentration. Based on their experience, the anesthesiologist sets the target concentration and lets the system adjust the infusion rates automatically and continuously. Although these open-loop model-based TCI systems are currently being used in Europe and most developed countries, they are not used in the United States, where the Food and Drug Administration has cited that the true concentration in the patient's blood may be significantly different than the targeted desired concentration, thereby potentially leading to confusion from the care provider side, and, ultimately, harm to the patient. The main pitfall of TCI systems is their inherent open-loop nature. There is no quantifiable endpoint that the system can use to adjust its infusion profile to the specificity of the patient. The use of TCIs is predicated on the expectation that the care provider will adjust the targeted concentration if the observable drug effect does not fit the surgery or patient's need. Although the current practice of anesthesia consists of the provider as the controller, there is an opportunity to improve practice. For instance, the idea of autopilot acting in a closed-loop manner, where a measure (or set of measures) is used to determine the drug effect and automatically adjust the infusion rate, is particularly attractive. In recent years, an increasing body of evidence has shown that such closed-loop systems are technically and clinically feasible, safe, and capable of maintaining an adequate anesthetic depth while requiring minimal to no operator interaction.15,19 The control engineering community has been increasingly tapped by clinical researchers to design control algorithms based on methods borrowed from the robust control theory. This particular field pertains to the control of complex processes with uncertainty in some or all of its components. A robust controller is essentially static. It has a fixed form and fixed parameters that are specifically designed to remain stable and provide adequate performance as long as the process remains within its uncertainty bounds (i.e., the process may evolve between its uncertainty bounds during normal operation without compromising the stability and performance of the closed-loop system). The advantage of such a controller is that its operation is very well defined because of its fixed nature. To illustrate the proposed method, we propose to derive the parameters of a robust controller designed to drive and maintain patients in an adequate hypnotic depth suitable for surgery, using IV propofol and the Wavelet-based Anesthetic Value for Central Nervous System (WAVCNS), a processed electroencephalography (EEG) parameter that quantifies the effect of the drug on the brain using wavelet analysis.25,–27 The proposed method is based on the quantification of the pharmacological variability of a large population of patients (18–60 years old, 50–110 kg) in response to the administration of propofol. The variability is derived based on PK–PD models obtained from 44 anesthesia inductions, which forms a representative sample of the target population. Based on this variability, we show how the parameters of a simple proportional-integral-derivative (PID) controller can be optimally tuned to ensure that the controller remains stable for this particular population. Using simulations, we further show that the robust PID controller provides a clinically acceptable level of performance. METHOD The proposed robust control design is based on a 4-step approach: (1) the identification of the drug pharmacological model from a number of patients representative of the targeted population, (2) the mathematical characterization of the uncertainty observed from the identified models, (3) the tuning of the controller closed-loop transfer function in such a way that stability is guaranteed for any patient fitting the uncertainty profile, and (4) the verification in simulation that the system's performance across the study population meets clinical expectations. This last point will be discussed further in the “Results” section. Step Number 1: Identification of the Drug Pharmacological Model From a Number of Patients Representative of the Targeted Population The anesthesia pharmacology community has invested significant efforts in deriving models for most anesthesia drugs used in clinical practice. Propofol benefits from an extensive body of published research and a large number of PK models can be found that describe the drug uptake, distribution, and elimination based on the infusion profile.28,–32 The propofol PD, relating the effect to the blood plasma concentration, has also been modeled for various clinical endpoints. However, an important feedback variable is to have PD models specifically derived for the quantitative effect observation being used. Here, we developed a proprietary measure of hypnotic depth, the WAVCNS index, based on scalp EEG signals acquired and processed by the NeuroSENSE Monitor (NeuroWave Systems, Cleveland Heights, Ohio). The WAVCNS index is a dimensionless number between 100 and 0, where 100 denotes the awake state (no hypnotic effect) and 0 is deep pharmacological coma (maximal effect that can be observed in the EEG). The WAVCNS is characterized by a fast response time, with minimal computational delay between a change in EEG (reflecting a change in drug effect) and its WAVCNS characterization. In a previous study,33 we collected induction EEG data from 44 adult patients undergoing elective orthopedic surgeries. The WAVCNS index was calculated based on the collected data. In addition, we recorded both the start and end times of the propofol bolus, its dosage, as well as the patients age, weight, height, and gender. These data allowed us to derive the PK–PD pharmacological model as shown in Figure 1. The model comprises a traditional 3-compartment mamillary PK model, a first-order linear PD function, and a nonlinear Hill characteristic to capture the observation saturation at low and high doses. FIGURE 1. View largeDownload slide Pharmacological model for propofol. FIGURE 1. View largeDownload slide Pharmacological model for propofol. In addition, we added a pure time delay Td to capture the arm-to-brain travel time of the drug. We also added the trending dynamics H(s) of the WAVCNS observation variable. This function captures the technological dynamics that exist between a change in the EEG (reflecting a change in drug effect) and the change in the WAVCNS index. In a previous work,34 we have shown that H(s) can be described as a second-order linear function, which comes in contrast to other quantitative EEG indexes. Being able to adequately model the dynamics H(s) of the sensor allows us to effectively compensate for the impact of the sensing technology on the PD model. As such, the PD model reflects only the drug effect dynamics. Since blood samples were not available, the propofol blood plasma concentration was inferred from the administration profile based on the PK model of Schuttler and Ihmsen32 Using standard identification techniques, we derived for each patient data set the drug travel time Td, the PD time constant kd, the 50% effective concentration EC50, and the Hill parameters (E0, E100, and γ). The complete model set has been published.33 A representative sample for the younger age group (18–29 years old) is presented in Table I using our latest WAVCNS algorithm (v.3.0.0.1). TABLE I. Propofol PD Model Parameters (Emax = 0) Patient ID No  Demographics  PD Parameters  Linear Model  Hill Saturation  ASA  Age (years)  Weight(kg)  Height(cm)  Gender  Td (s)  s−1 . 10−3  μg/mL  E0(1)  Y(1)  G1:≥18–<30 Years  007  nc  21  100  178  M  22  133.5  3.2  98.3  4.7  008  nc  28  59  168  F  4  44.4  3.1  99.7  2.5  010  2  26  90  190  M  44  25.0  2.4  96.1  2.0  015  2  21  53  157  F  45  51.5  3.8  100.0  1.2  015  2  19  90  185  M  39  85.7  3.8  98.4  2.3  023  1  28  60  162  F  18  82.5  3.9  100  2.1  030  1  24  78  170  M  32  44.5  2.9  97.5  2.7  035  2  19  68  164  F  12  26.7  1.9  93.8  2.1  038  2  25  70  170  M  7  35.2  3.4  98.0  1.9  046  1  23  81  180  M  9  32.8  2.8  97.5  2.8  048  1  18  83  178  M  17  46.4  2.8  99.8  2.3  053  2  21  67  163  F  4  26.2  2.4  100.0  2.5  053  1  22  88  183  M  9  50.4  2.5  96.1  2.6  066  1  21  59  162  M  18  160.5  3.6  100.0  4.1  071  1  19  72  176  M  20  75.0  2.5  99.6  1.9  Mean    22.3  74.5  172    20.0  61.4  3.0  98.3  2.5  Patient ID No  Demographics  PD Parameters  Linear Model  Hill Saturation  ASA  Age (years)  Weight(kg)  Height(cm)  Gender  Td (s)  s−1 . 10−3  μg/mL  E0(1)  Y(1)  G1:≥18–<30 Years  007  nc  21  100  178  M  22  133.5  3.2  98.3  4.7  008  nc  28  59  168  F  4  44.4  3.1  99.7  2.5  010  2  26  90  190  M  44  25.0  2.4  96.1  2.0  015  2  21  53  157  F  45  51.5  3.8  100.0  1.2  015  2  19  90  185  M  39  85.7  3.8  98.4  2.3  023  1  28  60  162  F  18  82.5  3.9  100  2.1  030  1  24  78  170  M  32  44.5  2.9  97.5  2.7  035  2  19  68  164  F  12  26.7  1.9  93.8  2.1  038  2  25  70  170  M  7  35.2  3.4  98.0  1.9  046  1  23  81  180  M  9  32.8  2.8  97.5  2.8  048  1  18  83  178  M  17  46.4  2.8  99.8  2.3  053  2  21  67  163  F  4  26.2  2.4  100.0  2.5  053  1  22  88  183  M  9  50.4  2.5  96.1  2.6  066  1  21  59  162  M  18  160.5  3.6  100.0  4.1  071  1  19  72  176  M  20  75.0  2.5  99.6  1.9  Mean    22.3  74.5  172    20.0  61.4  3.0  98.3  2.5  ASA, American Society of Anesthesiologists physical status. View Large TABLE I. Propofol PD Model Parameters (Emax = 0) Patient ID No  Demographics  PD Parameters  Linear Model  Hill Saturation  ASA  Age (years)  Weight(kg)  Height(cm)  Gender  Td (s)  s−1 . 10−3  μg/mL  E0(1)  Y(1)  G1:≥18–<30 Years  007  nc  21  100  178  M  22  133.5  3.2  98.3  4.7  008  nc  28  59  168  F  4  44.4  3.1  99.7  2.5  010  2  26  90  190  M  44  25.0  2.4  96.1  2.0  015  2  21  53  157  F  45  51.5  3.8  100.0  1.2  015  2  19  90  185  M  39  85.7  3.8  98.4  2.3  023  1  28  60  162  F  18  82.5  3.9  100  2.1  030  1  24  78  170  M  32  44.5  2.9  97.5  2.7  035  2  19  68  164  F  12  26.7  1.9  93.8  2.1  038  2  25  70  170  M  7  35.2  3.4  98.0  1.9  046  1  23  81  180  M  9  32.8  2.8  97.5  2.8  048  1  18  83  178  M  17  46.4  2.8  99.8  2.3  053  2  21  67  163  F  4  26.2  2.4  100.0  2.5  053  1  22  88  183  M  9  50.4  2.5  96.1  2.6  066  1  21  59  162  M  18  160.5  3.6  100.0  4.1  071  1  19  72  176  M  20  75.0  2.5  99.6  1.9  Mean    22.3  74.5  172    20.0  61.4  3.0  98.3  2.5  Patient ID No  Demographics  PD Parameters  Linear Model  Hill Saturation  ASA  Age (years)  Weight(kg)  Height(cm)  Gender  Td (s)  s−1 . 10−3  μg/mL  E0(1)  Y(1)  G1:≥18–<30 Years  007  nc  21  100  178  M  22  133.5  3.2  98.3  4.7  008  nc  28  59  168  F  4  44.4  3.1  99.7  2.5  010  2  26  90  190  M  44  25.0  2.4  96.1  2.0  015  2  21  53  157  F  45  51.5  3.8  100.0  1.2  015  2  19  90  185  M  39  85.7  3.8  98.4  2.3  023  1  28  60  162  F  18  82.5  3.9  100  2.1  030  1  24  78  170  M  32  44.5  2.9  97.5  2.7  035  2  19  68  164  F  12  26.7  1.9  93.8  2.1  038  2  25  70  170  M  7  35.2  3.4  98.0  1.9  046  1  23  81  180  M  9  32.8  2.8  97.5  2.8  048  1  18  83  178  M  17  46.4  2.8  99.8  2.3  053  2  21  67  163  F  4  26.2  2.4  100.0  2.5  053  1  22  88  183  M  9  50.4  2.5  96.1  2.6  066  1  21  59  162  M  18  160.5  3.6  100.0  4.1  071  1  19  72  176  M  20  75.0  2.5  99.6  1.9  Mean    22.3  74.5  172    20.0  61.4  3.0  98.3  2.5  ASA, American Society of Anesthesiologists physical status. View Large Step Number 2: Mathematical Characterization of the Uncertainty Observed From the Identified Models The 44 PK–PD models obtained during the identification procedure comprises a linear part (the PK, and the first-order PD dynamics) and a nonlinear element that captures the low- and high-dose saturation. The Hill saturation affects the gain of the model depending on the operating point: for instance, the Hill steepness is typically higher around the middle point and smaller as the effect-site concentration reaches values for which the observation saturates. The Hill function can thus be substituted with a variable gain KH, which can theoretically take any value between 0 and a maximum gain KH,M. A gain of 0 essentially indicates that the observable effect fully saturates, at which point closed-loop operation would no longer be possible. As such, we first need to define the desired operating range of the control system. It is reasonable to assume that the anesthesiologists will want to drive and maintain their patients in cortical states spanning light sedation (WAVCNS of 70) to deep hypnosis (WAVCNS of 30). Considering these endpoints, the maximum KH,max and minimum KH,min Hill gains can be calculated, as shown in Figure 2. After the linearization of the Hill function, we obtain a total of 88 patient models: 44 models using KH,min, and 44 models using KH,max. FIGURE 2. View largeDownload slide Linearization of the nonlinear h(x) Hill function. FIGURE 2. View largeDownload slide Linearization of the nonlinear h(x) Hill function. Each of these linearized models captures how a theoretical input sine waveform is transformed by the system, both in terms of its amplitude and phase. As a result, for each frequency ω, the PK–PD model can be expressed in terms of its gain g, and phase shift φ. This information can be further represented in the complex Nyquist plot (Fig. 3). Like most processes in nature, the PK–PD model gain tends to decrease with frequency, although the lag time increases. As a result, a Nyquist representation of the model creates a path that starts on, or close to, the positive x-axis in low frequencies and tends to turn in a clockwise fashion around the origin of the complex plot as the frequency increases. By plotting all 88 models, an uncertainty space containing all the models takes shape (Fig. 3). To mathematically model this space, it is useful to consider the distribution of the models at each frequency ω. In example of Figure 4, we can see that the models are contained within a circle C centered on G0 and of radius r. Any patient model G can thus be defined as:   (1) The parameter Δ is any complex number, whose norm is less or equal to 1, and as such, it describes a unity circle in the complex plane. The model G can, therefore, be expressed as a function of the nominal model G0 and an uncertainty term defined by the weight function w:   (2) The weight function w is also commonly referred to as relative uncertainty. FIGURE 3. View largeDownload slide Nyquist plot of all 88 propofol models and associated uncertainty bounds. Each model is represented by a path that wraps clockwise around the origin of the plot. As the frequency increases, the model gain tends to decrease (i.e., the path gets closer to the origin), whereas the phase lag increases (the path turns around the origin). This is very typical of a natural biologic process. FIGURE 3. View largeDownload slide Nyquist plot of all 88 propofol models and associated uncertainty bounds. Each model is represented by a path that wraps clockwise around the origin of the plot. As the frequency increases, the model gain tends to decrease (i.e., the path gets closer to the origin), whereas the phase lag increases (the path turns around the origin). This is very typical of a natural biologic process. FIGURE 4. View largeDownload slide Unstructured uncertainty: at each frequency ω, all patient models (for both Kh,min and kh,max) are contained within uncertainty bounds defined by circle C. FIGURE 4. View largeDownload slide Unstructured uncertainty: at each frequency ω, all patient models (for both Kh,min and kh,max) are contained within uncertainty bounds defined by circle C. It should be noted that the multiplicative uncertainty framework defined in equation (2) is conservative, as the circle C encompasses a larger space that the one strictly defined by the individual models. A less-conservative approach would be to consider the polygonal shape, such as in Figure 4. However, the mathematical expression of such a shape may be difficult to achieve and parameterize as a function of ω. When designing a negative feedback controller C, we need to ensure that the loop transfer function L = CG0H + wCG0HΔ does not encircle the (−1,0) Nyquist coordinate to guarantee the stability of the system. This translates into a Robust Stability (RS) condition, where:   (3) It can be shown that the RS condition can be further expressed as:   (4) Where T is the complimentary sensitivity function defined as T = CG0 . (1 + CG0H)−1. This is an important result as it shows that the relative uncertainty defines an upper bound that the complimentary sensitivity function cannot exceed to guarantee stability. We invite interested readers to refer to the excellent book from Skogestad and Postlethwaite35 for a more in-depth treatment on this subject. Step Number 3: Tuning of the Controller Closed-Loop Transfer Function in Such a Way That Stability is Guaranteed for Any Patient Fitting the Uncertainty Profile To design the closed-loop controller, one approach is to first select a controller structure that is well suited for the application. In this case, a simple PID controller provides the basic functionalities required: an integrator with time constant τi to minimize steady state error, a derivative action with time constant τd to increase the stability margin in higher frequencies (by estimating future errors), and a proportional gain K to rapidly compensate the current error. Using the nominal model G0 defined above, we tuned the PID parameters to obtain a phase margin of at least 45° and a gain margin of at least 6 dB, which are typical design objectives. Yet, when plotting the complementary sensitivity function and the inverse of the relative uncertainty, we found that this initial controller violates the RS condition (Fig. 5). As a result, we cannot guarantee the stability of the system in view of the system uncertainty. FIGURE 5. View largeDownload slide Robust stability condition check. FIGURE 5. View largeDownload slide Robust stability condition check. An optimization procedure by iterative search is then necessary to find the parameter set {K, τi, τd} that maximizes the area between ‖T‖ and ‖1/w‖ while ensuring that the product ‖w . T‖ remains below 1 at all frequencies. The optimization procedure yielded the parameter set {K = 0.0076, τi = 118, τd = 20} using a 1-s sampling rate. Using this parameter set, the RS condition is respected, which mathematically guarantees the stability of the closed-loop system. RESULTS The controller resulting from the optimization procedure is, by definition, stable. However, this does not imply adequate performance. As such, simulations must be carried out as part of the design procedure to verify that clinical performance can be achieved. Ideally, a closed-loop controller should drive and maintain patients in an adequate anesthetic state as rapidly and precisely as would be accomplished by an experienced anesthesiologist. For instance, the NeuroSENSE manufacturer recommends keeping the WAVCNS index between 60 and 40. The time to induce an awake patient and reach this range can be an important performance parameter. In addition, having low or no steady state error is another desirable endpoint. Finally, minimizing the undershoot (i.e., too low transitory WAVCNS values denoting a temporary overdosing) is key to minimize unwanted response (e.g., bradycardia and hypotension). For the purpose of the simulation, we evaluated the response of the closed-loop system to induce each patient to a target WAVCNS level of 50, an appropriate depth to allow surgical access in most patients. We also added a measurement noise commensurate in amplitude and frequency content with the noise typically seen on the WAVCNS index. Simulating noise in the feedback variable helps in assessing whether this noise is amplified by the controller in a significant way. In our first attempt, we found the system remained well behaved and stable. However, the undershoot (indicative of a temporary overdosing) exceeded 20 units in some patients. This could potentially put patients at risk for undesired cardiorespiratory depression and was not considered to be adequate. The controller “overreaction” resulting in this type of significant undershoot is typically a sign that the controller operates close to its instability region. Hence, one way to minimize the undershoot is to reduce the gain and cutoff frequency of the controller to increase its stability margin. This results in a smoother control action, albeit with a longer settling time. Note that changing the PID parameters always entails verifying that the RS condition still holds true. Another method not involving the retuning of the controller parameters consists in increasing the stability margin of the overall closed-loop system by minimizing the effect of the arm-to-brain time delay through predictive action. A Smith Predictor is an effective and simple method that effectively minimizes the phase shifting effect of pure delays. The Smith Predictor estimates the effect of the time delay by using a model of the patient and removes this effect from the feedback error. The stabilization effect of the Smith Predictor only affects the closed-loop system in the high frequency region. The Smith Predictor has limited to no effect in the low frequency region. Figure 6 shows the simulated results for all 44 patients, using the PID controlled defined in the previous section and augmented with a Smith Predictor. For all patients, the Smith Predictor PK–PD model was defined for a 30-year-old, 70 kg male patient. As expected, the controller initially starts the same way for all patients. As the patient specificity comes into play (e.g., sensitivity to drug and PK), the infusion profile time course is updated by the controller to minimize the feedback error. Once the system reaches steady state, and assuming no change in extraneous inputs (e.g., set point, output disturbance, etc.), the integrator action becomes dominant and results in the minimization of the steady state error. This is one advantage of a controller structure that integrates an integrator: there is no steady state error in the controlled variable, whereas patient variability results in infusion rate time courses that converge toward different values. In addition, the maximum undershoot resulting from an overdosing was under 10 WAVCNS units. The average settling time to reach the 60 to 40 region was 3.6 minutes ([3.1; 4.5] minutes, 1 SD), which is clinically adequate. FIGURE 6. View largeDownload slide Simulated results of a propofol induction (WAVCNS target of 50). FIGURE 6. View largeDownload slide Simulated results of a propofol induction (WAVCNS target of 50). If the settling time of the controller had been unsatisfactory, methods aimed at minimizing the uncertainty should then be investigated (e.g., tightening the target population to a predefined age group or limiting the operating range of the controller to a smaller range of WAVCNS values). By minimizing the uncertainty, additional stability margin can be obtained. Optimizing the PID parameters to meet the RS condition would have then resulted in larger controller gain and higher cutoff frequency, thereby providing improved performance. If adequate performance remains elusive, the robust control framework, where a unique controller is used for a given population of patients, may not be suitable. Adaptive control techniques may then be needed. DISCUSSION The technical and clinical feasibility of anesthesia closed-loop systems has now been established by a number of research groups. However, the question of the regulatory path forward to test and commercialize these systems is becoming more pressing. Medical device manufacturers are more likely to invest in the development of a commercial solution if the regulatory hurdles can be addressed. Although the International Electrotechnical Commission36 has established a standard for the design of Physiological Closed-Loop Control Systems, the question of safety while operating in closed loop is a key aspect for regulatory bodies. Agencies like the Food and Drug Administration require scientific evidence establishing patient safety in view of pharmacological variability and controller stability, i.e., that the closed-loop operation remains well behaved and stable even when the patient's sensitivity to the drug evolves between extremes. In this article, we have proposed a design methodology for an anesthesia close-loop controller that can address such regulatory concerns. Even when considering a large segment of the overall patient population, we were able to tune a simple PID controller to remain mathematically stable despite a large pharmacological variability. This, in turn, ensures that the system will remain well behaved even with the most sensitive patients (i.e., the patient will not alternate between an awake/conscious state and deep pharmacological coma during the course of the surgery). RS for a large patient population is a safety aspect that is key to gaining regulatory approval to conduct human trials and commercialize the device. It is important to note, however, that stability can theoretically always be achieved by derating the controller sufficiently. Having a robust control algorithm does not imply that the controller will provide a level of performance that is clinically acceptable. Verifying the performance level ahead of running clinical trial is another important design aspect. In this work, we have shown that the controller will settle in an acceptable amount of time, with minimal under and overshoot. Based on this result, a validation study in a human patient population can be envisaged. It should also be noted that, as with any other closed-loop system design, the selection of the sensing technology is a critical issue. Using a different depth-of-anesthesia measure for which a linear and time invariant sensor model is unavailable would have translated in adding a component in the closed-loop feedback path whose dynamic behavior is not defined. This technological uncertainty would have in turn compounded the overall system uncertainty, which would have imposed further limitations on the controller performance. Finally, we believe that a closed-loop TCI system, such as described here, can provide U.S. anesthesiologists with a supporting technology to deliver safe and effective TIVA. By leapfrogging over the scientific limitations of the classical TCI technology, a closed-loop TCI version may gain easier regulatory acceptance while further improving the quality of care and pharmacological safety. This is particularly important in patients likely to undergo multiple surgical procedures in a short time span. ACKNOWLEDGMENTS The authors gratefully acknowledge the support of the U.S. Army Medical Research and Materiel Command for funding this work under contract no. W81XWH-11-C-0078. REFERENCES 1. Bickford R Automatic electroencephalographic control of general anesthesia. Electroencephalogr Clin Neurophysiol  1950; 2: 93– 6. Google Scholar CrossRef Search ADS   2. Belville J, Attura G Servo control of general anesthesia. Science  1957; 126: 827– 30. Google Scholar CrossRef Search ADS PubMed  3. Zbinden A, Frei F, Westenskow D, Thomson D Control of end-tidal halothane concentration. Part B: verification in dogs. Br J Anaesth  1986; 58( 5): 563– 71. Google Scholar CrossRef Search ADS PubMed  4. Schwilden H, Stoeckel H, Schuttler J Closed-loop feedback control of propofol anesthesia by quantitative EEG analysis in humans. Br J Anaesth  1989; 62: 290– 6. Google Scholar CrossRef Search ADS PubMed  5. Schwilden H, Schuttler J Model-based adaptive control of volatile anaesthetics by quantitative EEG. In: Control and Automation in Anaesthesia , pp 163– 74. Edited by Schwilden H, Stoeckel H Berlin, Germany, Springer-Verlag, 1995. 6. Linkens D Adaptive and intelligent control in anesthesia. IEEE Control Systems Magazine  1992; 12( 6): 6– 11. Google Scholar CrossRef Search ADS   7. Kenny G, Mantzaridis H Closed-loop control of propofol anaesthesia. Br J Anaesth  1999; 83: 223– 8. Google Scholar CrossRef Search ADS PubMed  8. Leslie K, Absalom A, Kenny G Closed-loop control of sedation for colonoscopy using the bispectral index. Anaesthesia  2002; 57: 690– 709. Google Scholar CrossRef Search ADS PubMed  9. Absalom A, Sutcliffe N, Kenny G Closed-loop control of anesthesia using bispectral index. Anesthesiology  2002; 96: 67– 73. Google Scholar CrossRef Search ADS PubMed  10. Struys M, De Smet T, Versichelen L, et al.   Comparison of closed-loop controller administration of propofol using BIS as the controlled variable versus ‘standard practice’ controlled administration. Anesthesiology  2001; 95: 6– 17. Google Scholar CrossRef Search ADS PubMed  11. De Smet T, Struys M, Neckebroek M, et al.   The accuracy and clinical feasibility of a new bayesian-based closed-loop control system for propofol administration using the bispectral index as a controlled variable. Anesth Analg  2008; 107( 4): 1200– 10. Google Scholar CrossRef Search ADS PubMed  12. Gentilini A, Frei C, Glattfelder A, et al.   Multi-tasked closed-loop control in anesthesia. IEEE Eng Med Biol Soc Mag  2001; 20( 1): 39– 53. Google Scholar CrossRef Search ADS   13. Locher S, Stadler K, Boehlen T, et al.   A new closed-loop control system for isoflurane using bispectral index outperforms manual control. Anesthesiology  2004; 101: 591– 602. Google Scholar CrossRef Search ADS PubMed  14. Liu N, Chazot T, Genty A, et al.   Titration of propofol for anesthetic induction and maintenance guided by the bispectral index: closed-loop versus manual control: a propsective, randomized, multicenter study. Anesthesiology  2006; 104( 4): 686– 95. Google Scholar CrossRef Search ADS PubMed  15. Liu N, Chazot T, Hamada S, et al.   Closed-loop coadministration of propofol and remifentanil guided by bispectral index: a randomized multicenter study. Anesth Analg  2011; 112( 3): 546– 57. Google Scholar CrossRef Search ADS PubMed  16. Le Guen M, Liu N, Bourgeois E, et al.   Automated sedation outperforms manual administration of propofol and remifentanil in critically ill patients with deep sedation: a randomized phase II trial. Intensive Care Med  2013; 39( 3): 454– 62. Google Scholar CrossRef Search ADS PubMed  17. Glass P, Rampil I Automated anesthesia: fact or fantasy. Anesthesiology  2001; 95: 1– 2. Google Scholar CrossRef Search ADS PubMed  18. Huang J, Lu YY, Nayak A, Roy R Depth of anesthesia estimation and control. IEEE Trans Biomed Eng  1999; 46: 71– 81. Google Scholar CrossRef Search ADS PubMed  19. West N, Dumont GA, Van Heusden K, et al.   Robust closed-loop control of induction and maintenance of propofol anesthesia in children. Paediatr Anaesth  2013; 23( 8): 712– 9. Google Scholar CrossRef Search ADS PubMed  20. Biswas I, Mathew P, Singh R, Puri G Evaluation of closed-loop anesthesia delivery for propofol anesthesia in pediatric cardiac surgery. Paediatr Anaesth  2013; 23( 12): 1145– 52. Google Scholar PubMed  21. Dumont GA, Ansermino J Closed-loop control of anesthesia: a primer for anesthesiologists. Anesth Analg  2013; 117( 5): 1130– 8. Google Scholar CrossRef Search ADS PubMed  22. Visser K, Hassink EA, Bonsel GJ, Moen J, Kalkman CJ Randomized controlled trial of total intravenous anesthesia with Propofol versus inhalation anesthesia with isoflurane-nitrous oxide: postoperative nausea and vomiting and economic analysis. Anesthesiology  2001; 95( 3): 616– 26. Google Scholar CrossRef Search ADS PubMed  23. Chandler J, Myers D, Mehta D, et al.   A randomized trial to compare total intravenous anesthesia with propofol and remifentanil to inhalational sevoflurane anesthesia. Pediatr Anaesth  2013; 23( 4): 309– 15. Google Scholar CrossRef Search ADS   24. Barras P, McMaster J, Grathwohl K, Blackbourne L Total intravenous anesthesia on the battlefield. US Army Med Dep J ; January– March 2009: 68– 72. 25. Zikov T, Bibian S, Dumont GA, Huzmezan M, Ries C Quantifying cortical activity during general anesthesia using wavelet analysis. IEEE Trans Biomed Eng  2006; 53( 4): 617– 32. Google Scholar CrossRef Search ADS PubMed  26. Zikov T, Bibian S, Ries C, Jaromczyk A, Voltz D Performance of BIS and WAVcns to predict loss of consciousness during routine propofol induction. Article presented at American Society of Anesthesiologists Annual Meeting, October 16, 2006, Chicago, IL. Available at http://www.asaabstracts.com/strands/asaabstracts/abstract.htm;jsessionid=CCF870CE6C820BF5896BAA1F534E4A3B?year=2006&index=5&absnum=1324; accessed July 14, 2014. 27. Agrawal G, Clement F, Scholtes F, Bibian S, Zikov T Relationship between WAVcns index and suppression ratio. Article presented at American Society of Anesthesiologists Annual Meeting, October 19, 2010, San Diego, CA. Available at http://www.asaabstracts.com/strands/asaabstracts/abstract.htm;jsessionid=FDC63224EB3EA88FAD38F8CEF2EEB0D2?year=2010&index=5&absnum=1838; accessed July 14, 2014. 28. Gepts E, Jonckheer K, Maes V, Sonck W, Camu F Disposition kinetics of propofol during alfentanil anaesthesia. Anaesthesia  1988; 43: 8– 13. Google Scholar CrossRef Search ADS PubMed  29. Marsh B, White W, Morton N, Kenny G Pharmacokinetic model driven infusion of propofol in children. Br J Anaesth  1991; 67: 41– 8. Google Scholar CrossRef Search ADS PubMed  30. Kataria B, Ved S, Nicodemus H, et al.   The pharmacokinetics of propofol in children using three different data analysis approaches. Anesthesiology  1994; 80: 104– 22. Google Scholar CrossRef Search ADS PubMed  31. Schnider T, Minto C, Gambus P, et al.   The influence of method of administration and covariates on the pharmacokinetics of propofol in adult volunteers. Anesthesiology  1998; 88: 1170– 82. Google Scholar CrossRef Search ADS PubMed  32. Schuttler J, Ihmsen H Population pharmacokinetics of propofol: a multicenter study. Anesthesiology  2000; 92: 727– 38. Google Scholar CrossRef Search ADS PubMed  33. Bibian S, Dumont GA, Huzmezan M, Ries CR Patient variability and uncertainty quantification in anesthesia: Part I—PKPD modeling and identification. Article presented at IFAC Symposium on Modeling and Con, 2006, Reims, France. Available at http://www.ifac-papersonline.net/Detailed/29918.htmltrol in Biomedical Systems; accessed July 14, 2014. 34. Bibian S, Dumont GA, Zikov T Dynamic behavior of BIS, M-Entropy and NeuroSENSE brain function monitors. J Clin Monit Comput  2011; 25: 81– 7. Google Scholar CrossRef Search ADS PubMed  35. Skogestad S, Postlethwaite I Multivariable feedback control . West Sussex, England, John Wiley & Sons, 1996. 36. International Electrotechnical Commission Medical Electrical Equipment, Part 1-10: general requirements for basic safety and essential performance: collateral standard: Process requirements for the development of physiologic closed-loop controllers . Geneva, Switzerland, IEC, 2007. Available at http://marketplace.aami.org/eseries/scriptcontent/docs/Preview%20Files/iec60601-1-10_0711_preview.pdf; accessed July 13, 2014. Footnotes 1 This article was selected for plenary talk at the Military Health System Research Symposium, Fort Lauderdale, Florida, August 14, 2013. Reprint & Copyright © Association of Military Surgeons of the U.S.
TI - Closed-Loop Target-Controlled Infusion Systems: Stability and Performance Aspects
JO - Military Medicine
DO - 10.7205/MILMED-D-14-00380
DA - 2015-03-01
UR - https://www.deepdyve.com/lp/oxford-university-press/closed-loop-target-controlled-infusion-systems-stability-and-eLStqTUCU0
SP - 96
EP - 103
VL - 180
IS - suppl_3
DP - DeepDyve
ER -