Closed Loop Anesthesia

It can be said that when inducing and maintaining anesthesia, anesthesiologists apply a reactive approach. They select initial doses on the basis of a variety of considerations, observe the results, and then make adjustments on the basis of several factors, at irregularly varying intervals. In control engineering terminology, this constitutes a closed-loop control system, because of the feedback present in the observations and interventions of the anesthesiologist. This closed-loop control system is, however, of a special nature: (i) it has a human controller in the loop, and (ii) as a consequence, the control actions are intermittent and irregular in time. Control theory is increasingly being applied in the development of computer-controlled drug delivery systems, among others in anesthetic closed-loop control applications. The aim of computer-controlled closed-loop systems is to formalize this process of observation and intervention to provide finer and more accurate control. Such systems use a near-continuous signal of drug effect, calculate the error between the observed value and the setpoint value (selected by the user), and use this error in an algorithm to make frequent and regular adjustments to drug administration rates. Even better, some computer-control systems try to predict the future drug effect to make appropriate adjustments well in advance.1 Several conditions are necessary for accurate feedback control. The process being controlled should be defined, and 1 or more real-time representative measures of the system state should be available. In engineering terms the latter signals are designated as the processoutput/s, which have to be controlled according to certain agreed specifications. Ideally also, the control actuators or process input/s should, with minimal or known delay, cause predictable, linear changes in the process. We agree with Glass and Rampil, who posited in 2001 that “closed loop systems for anesthesia are more difficult to design and implement than those for aviation.”2 In aviation, the process of flying is governed by well-known physical laws. The outputs (such as velocity, pitch, and altitude) are clearly defined parameters, and can be measured accurately in real time, and the relationship between the inputs (such as fuel flow, flap, and tail angles) and the outputs is predictable, well-defined, and linear. What is it that makes anesthesia such a challenging control problem? Anesthesia is not a simple, well-understood process. Our understanding of consciousness and the mechanisms of anesthetic-induced loss of consciousness is far from complete. Consciousness is so ethereal that it is difficult to model. At present the best we have are models, such as mean field models of drug action,3 which describe phenomena in the electroencephalogram associated with different brain states. However, the anesthetic literature is replete with references to “depth of anesthesia,” which implies that anesthesia is a continuous function of effect-site concentration. Unfortunately, reality is more complex, because systems such as the brain are commonly nonlinear, bistable systems. The 2006 ASA practice advisory on awareness and brain function monitoring defines anesthesia as a “drug induced loss of consciousness.”4 Because consciousness is either absent or present, this advisory reinforced the assertion by Prys-Roberts in 1987 that there “cannot be degrees of anesthesia nor for that matter can there be variable depths of anesthesia.”5 This assertion is reasonable. Abrupt transitions occur between natural sleep stages, possibly explained by flip-flop states generated by brainstem and hypothalamic nuclei.6–8 Because anesthetic agents may act on key elements of these sleep pathways,9–13 it is likely that threshold events also typify anesthesia. Clearly, there is some way to go before we will completely understand the underlying process we seek to control. The input for an anesthetic control system, anesthetic agent administration, also has less than ideal properties. Drug administration is an asymmetrical process: we can actively infuse but cannot actively remove the drug. Because the relationship between dose and plasma concentration is so complex, target-controlled infusion (TCI) systems are a logical choice of control actuator, so that the control input is a target concentration rather than an infusion rate. Many assumptions underpin the pharmacokinetic models used in TCI systems, and some are obviously incorrect (such as that of instant mixing of an administered drug within the central compartment). Not surprisingly, the predictive accuracy of current models is imperfect, and the choice of model for propofol is controversial.14 There is a significant hysteresis between plasma drug concentration and clinical effect. Effect-site targeted TCI systems may limit phase delays by manipulating plasma concentration to achieve rapid changes in effect-site concentration, but also rely on several layers of unproven assumptions. Those underlying the choice and application of a keo value are critical. After a target concentration alteration, the plasma concentration profiles implemented by the system are critically dependent on the keo. The Schnider15,16 and Minto17,18 models are commonly used for effect-site targeted propofol and remifentanil administration. They incorporate keo values calculated on the assumption that all hysteresis between plasma concentration and clinical effect is explainable by a delay in drug transfer between plasma and effect site, and thus that anesthesia is a smooth, path- and state-independent, symmetrical process. This assumption may also be false. A study using sevoflurane and spectral entropy showed that the parameters of sigmoidal EMax models (EC50 and slope [γ]) were different for induction and recovery.19 The hypothesis that the neural processes and paths involved in induction and recovery are different was supported by an animal study showing that orexinergic pathways were involved in emergence from anesthesia but not induction of anesthesia.20 In an animal study, measured brain drug concentrations at loss of consciousness and at return of consciousness were different.21 If these data are confirmed, then current modeling concepts underpinning effect-site targeting may require significant revisions to more closely approximate and describe pharmacological realities. Is anesthesia such a challenging control problem that it is not worth attempting? Certainly not! In this issue of Anesthesia & Analgesia, Liu et al. report that the computer-controlled system that they have developed is able to provide more stable control of the BIS (bispectral index; Covidien, Mansfied, Massachusetts) than can human anesthesiologists.22 Although this is not the first anesthetic computer control system, it is a brave and pioneering effort because it is among the first to control both hypnotic and analgesic administration. Commendable, too, are the methods used to assess their system. Their study was a randomized controlled trial, which compared machine control with human control. It took place in multiple hospitals, involving multiple anesthesiologists of whom few were experts or enthusiasts for automated anesthesia. Finally, it involved patients with a wide range of pathologies requiring a range of surgical interventions, and included a large number with significant comorbidities. The authors describe their system as a dual-loop, proportional–integral–derivative (PID) controller. On the basis of the assumption that small fluctuations in depth of anesthesia are caused by fluctuation in the severity of noxious stimuli, it is designed to implement changes in the remifentanil target concentration more frequently than for propofol when the absolute value of the BIS error is above 4. When the absolute value is between 2 and 4, changes are made only to the remifentanil target concentration. Although these rules resulted in dual-loop control being associated with improved control stability in comparison with manual control, this came at the cost of larger remifentanil doses. In clinical practice this is often, but not always, acceptable. In the dual-loop group, target remifentanil concentrations were commonly >10 ng/mL, which is outside the comfort range of many clinicians and may be associated with tolerance and hyperalgesia.23–25 The higher remifentanil dose is a result of the heuristic approach to the design of the control algorithm: it includes several rules of thumb and numerical parameters, which have been tuned empirically rather than being designed. Different heuristic rules with other tuning parameters will result in different drug doses. In fact, because of the interpatient variability in drug sensitivity, a fixed set of controller parameters can never be optimal for the whole population. An intelligent computer-controlled system should adapt its control parameters in real time, to optimize them for the specific individual patient.26 Such an adaptive approach is indeed part of the daily practice of the anesthesiologist! From a control engineering point of view, Liu et al.'s is not a dual-loop system. It sets the target concentrations of propofol and remifentanil on the basis of a single output (BIS) and will produce target concentrations roughly in the same proportion, always weighted towards remifentanil. The algorithm that generates these doses is unable to incorporate emerging knowledge of drug interactions. In particular, it cannot incorporate the information gained from studies that have used nonlinear mixed-effect modeling and response surface methodology to study the nature and degree of pharmacodynamic interactions among propofol, remifentanil, and the BIS.27–30 A true dual-loop system should control propofol administration on the basis of a hypnotic end point (e.g., BIS), and remifentanil administration on the basis of a different signal that is a validated signal representing the balance between the level of noxious stimulation and the level of analgesia. Signals that might be suitable include the surgical state index and cortical index.31,32 As is shown in Liu et al.'s study, by making frequent small changes, the system is eventually able to find combinations of propofol and remifentanil effect-site concentrations that limit the error in the controlled signal (BIS). When there is an abrupt change in the dose–response relationship (e.g., change in painful stimulus), the system must again seek a new combination, but can only do so in a fixed way, without taking into account individual sensitivity of the patient to the drugs. It is able to find the “EC50,” but has no facility to learn the γ of the dose–response relationship. The potential consequence is that responses to perturbations may be suboptimal (too small and slow for some, and too large and rapid for other patients), causing oscillations in the target concentrations and in the BIS. It is not surprising then that in the current study, a degree of oscillation appears to have occurred in some patients in the dual-loop group (Liu et al.'s Figure 321). We have previously used simulations to compare the control stability of single input/single output control systems, and believe that adaptive model-based control algorithms offer more stable control, and less oscillation, than does a PID algorithm.33 Additionally, a model-based controller of the hypnotic component, which is able to adapt to the individual patient pharmacodynamics, has been developed.34 New predictive model-based control theories are also under study.1 However, although model-based adaptive and predictive systems are intuitively pleasing and have proven their efficiency in technical systems, further studies are required to evaluate them in the challenging conditions of anesthesia control. Our understanding of anesthesia and anesthetic pharmacology is so limited that some may argue that a simpler PID-type controller is a more appropriate option until we have a stronger evidence base to inform anesthetic dose– response model design and structure. The performance of the computer-controlled system designed by Liu et al. is impressive. Although it performed well in a statistical sense and has been used by a large number of clinicians, it has a lack of theoretical insight. Time, further investigations, and further experience are required to determine how relevant and significant the limitations are. In the search for the “Holy Grail,” anesthesiologists will still have to determine whether or not adaptive, intelligent computer algorithms with dual, interacting, closed-loop systems will facilitate better control and improved outcomes.