IN this issue of Anesthesiology, Struys et al.  1provide a challenge to us as anesthesiologists. Can automated closed-loop anesthesia improve patient outcomes? In this small study, they demonstrated that outcomes (as measured by hemodynamic control and initial recovery) were better in the group to which propofol was administered via  a closed-loop control system than in the group to which propofol was administered by a clinician.

Anesthesiologists often compare themselves to pilots because they have similar work environments and work functions. Automated flight control is standard in professional aviation industry because it reduces workload during busy times and reduces certain types of pilot error. We in anesthesia have been slow or even reluctant to adopt such technology, although the first demonstration of closed-loop anesthesia was described by Bickford in 1950. 2Comparing automated flight to automated anesthesia is glib; however, making the comparison and then determining how they are different allows one to establish what the challenges are for anesthesia and, from this, to determine how these issues are being resolved.

A generic closed-loop system is shown in figure 1. 3The set points for flying (speed, altitude, and so forth) are readily defined and measurable. What constitutes anesthesia is still hotly debated, with no sensor able to measure changes in depth of anesthesia. What has been established is that alteration in consciousness is at least a required component of anesthesia. Although consciousness itself may be difficult to define, recent work has demonstrated that derivatives of the electroencephalogram, i.e. , Bispectral Index, 4electroencephalographic entropy, 5and auditory evoked potentials, 6can correlate with changes in consciousness. Thus, sensors now exist to measure at least one component of the anesthetic, making it amenable to closed-loop control. This has been a major step for developing clinically acceptable closed-loop anesthesia. However, improvements in sensor technology and artifact detection and elimination remain challenges for their routine use in closed-loop anesthesia. 7 

Fig. 1. The components of a typical automated drug delivery system.

Fig. 1. The components of a typical automated drug delivery system.

Close modal

The next objective is to take the step from a drug dose (the control signal for unconsciousness) to the desired set point (e.g. , Bispectral Index value). The algorithm is not as simple as delivering x mg for a given change in Bispectral Index units. The ability to obtain the set point for unconsciousness is dependent on achieving in the brain (biophase) a concentration of the drug that will produce this set point. In turn, the concentration of the drug in the biophase is determined both by the dose administered and by its pharmacokinetics. Innovations in pharmacokinetic modeling have enabled the development of target-controlled delivery systems. 8They use pharmacokinetic models to calculate the required dosing scheme to achieve a desired concentration. These developments have been crucial in providing closed-loop anesthesia.

Unfortunately, biologic systems have significant variance within a population, and the pharmacokinetic parameters used within the models are likely to result in inaccuracies for any given individual. In addition, there may be up to 10-fold differences in the relation between concentration and effect (pharmacodynamics) among individuals. Pharmacodynamics may also vary over time and during surgery according to the intensity of the applied noxious stimulus. This adds further complexities to the algorithm. To accommodate this, an adaptive control signal is added by which the algorithm is altered or adapted to the individual’s unique response. Variability may be caused by pharmacokinetics or pharmacodynamics. One may assume that pharmacodynamics between individuals do not vary and thus any difference in the measured effect is caused by the pharmacokinetic parameters. Alternatively, one may assume that the pharmacokinetic parameters are always correct but that pharmacodynamics vary among individuals. Struys et al.  have taken the latter approach. During induction, propofol was administered using an open-loop mode (i.e. , target concentration rather than target effect was the initial set point). This relation between concentration and effect was used to construct a Hill curve for each individual. In this early stage of development, we need to establish the optimal approach for closed-loop systems used to provide anesthesia. Is the approach of Struys et al.  better than altering the pharmacokinetic parameters? Did the authors achieve an accurate estimation of the individuals’ Hill curves? All subsequent adjustments in drug administration were dependent on the accuracy of this curve. Their system performed well (achieving a median absolute performance error of less than 10%), but this may have been because the system was not fully tested or stressed. Loss of consciousness is only one component of the anesthetic state. Analgesia also plays a significant role. In the study by Struys et al. , the patients were all administered a maintenance dose of 0.25 μg · kg−1· min−1remifentanil. This dose of remifentanil is sufficient to produce a maximal reduction in the minimum alveolar concentration of isoflurane. 9It is likely that during surgery, the required concentration of propofol was only that needed to achieve loss of consciousness. This is confirmed by the small adjustments in propofol concentration that occurred during the study. 1Would the control algorithm have worked as well if the dose of analgesic administered had been much less, thereby requiring much larger adjustments in propofol concentration? Would the accuracy of the original Hill curve under these circumstances have been adequate, and would the methods proposed for making these adjustments have been sufficient and appropriate for these larger changes in concentration? Would overshoot or undershoot have occurred? Would the responses have been rapid enough for the changing stimuli encountered during surgery with the potential for awareness or movement? It is important that at this stage of development of closed-loop anesthesia, we do not extrapolate any results beyond the narrow confines of the specific study design, especially because of the critical role of analgesia in the performance of the system.

The study design chosen by Struys et al.  for comparing the two groups may be criticized by some. In the control group, propofol was titrated manually to achieve set points that are commonly used during anesthesia, i.e. , blood pressure, heart rate, and sympathetic tone. In contrast, the closed-loop group had as their set point a Bispectral Index value. Thus, they compare two groups whose anesthetic was adjusted to two very different measures of adequate anesthesia. Is this acceptable? It depends on the question being answered. If the question is “Can a machine titrate propofol to a desired Bispectral Index better than humans?” then the study design is faulted. A more fundamental question is “Does closed-loop anesthesia titrated to a set point of unconsciousness provide a better outcome than anesthesia titrated to set points that are currently used most commonly?” This question is appropriately answered by the study design chosen by Struys et al.  One may argue that the improved outcomes they noted were caused by titration to a Bispectral Index rather than by the use of a closed-loop system. The study leaves this as an unanswered question that needs to be answered as the natural evolution of good science. Their study has necessitated that we now carefully consider the value of closed-loop anesthesia.

Closed-loop systems for anesthesia are more difficult to design and implement than those for aviation, but considerable progress has been made. It is only with the development of reliable and robust monitors of consciousness that we could even consider developing clinically viable closed-loop systems. The evolution of pharmacokinetic models, including the biophase and the implementation of such models into drug delivery systems, has made closed-loop delivery of anesthesia a viable possibility. The study by Struys et al.  implies that the development of closed-loop delivery systems for anesthesia are no longer esoteric. Rather, the challenge is now to establish fully the safety, efficacy, reliability, and utility of closed-loop anesthesia for its adoption into the clinical setting.

Struys MMRF, De Smet T, Versichelen LFM, Van de Velde S, Van den Broecke R, Mortier EP: Comparison of closed-loop controlled administration of propofol using BIS as the controlled variable versus “standard practice” controlled administration. A nesthesiology 2001; 95: 6–17
Bickford RG: Automatic electroencephalographic control of general anesthesia. EEG Clin Neurophysiol 1950; 2: 93–6
Reves JG, Jacobs JR, Glass PSA: Automated drug delivery in anesthesia. ASA Refresher Course in Anesthesiology 1991; 19: 127–37
Glass PS, Bloom M, Kearse L, Rosow C, Sebel P, Manberg P: Bispectral analysis measures sedation and memory effects of propofol, midazolam, isoflurane and alfentanil in normal volunteers. A nesthesiology 1997; 86: 836–47
Bruhn J, Ropcke H, Hoeft A: Approximate entropy as an electroencephalographic measure of anesthetic drug effect during desflurane anesthesia. A nesthesiology 2000; 92: 715–26
Iselin-Chaves IA, El-Moalem H, Gan TJ, Ginsberg B, Glass PSA: Changes in the auditory evoked potentials and the Bispectral Index following propofol or propofol and alfentanil. A nesthesiology 2000; 92: 1300–10
Schuttler J, Schwilden H: Present state of closed-loop drug delivery in anesthesia and intensive care. Acta Anaesthesiol Belg 1999; 50: 187–91
Schwilden H, Stoeckel H, Schuttler J, Lauven PM: Pharmacological models and their use in clinical anaesthesia. Eur J Anaesth 1986; 3: 175–208
Lang E, Kapila A, Shlugman D, Hoke JF, Sebel PS, Glass PSA: Reduction of the minimum alveolar concentration of isoflurane by remifentanil. A nesthesiology 1996; 85: 721–8