Closed-loop Control of Mean Arterial Blood Pressure during Surgery with Alfentanil: Clinical Evaluation of a Novel Model-based Predictive Controller

After approval by the ethics committee of the Canton of Bern (Bern, Switzerland) and obtaining written informed consent, 13 patients with American Society of Anesthesiologists physical status I or II who were scheduled to undergo neurosurgery during general anesthesia were enrolled. Patients younger than 18 yr or older than 65 yr and patients with a history of cerebrovascular insufficiency, coronary artery disease (angina or previous infarction, previous coronary artery bypass graft operation or percutaneous coronary intervention), or untreated arterial hypertension were excluded.

At the time of arrival in the operating room 30 min after premedication with 7.5 mg oral midazolam, the patients were monitored with an electrocardiogram, a noninvasive blood pressure cuff, a pulse oximeter (Dräger PM 8060 from a Cicero Anesthesia workstation; Dräger GmbH, Lübeck, Germany), and an Aspect A-2000 monitor (BIS version 3.3; Aspect Medical Systems, Inc., Newton, MA). An infusion of Ringer's lactate at 2 ml · kg⁻¹· h⁻¹was started through a venous catheter at the nondominant forearm. A 20-gauge arterial catheter was placed in the radial artery of the nondominant forearm during local anesthesia after the patency of the ulnar artery had been evaluated with the Allen test. Skin electrodes were placed on the ulnar nerve of the opposite arm and connected to a train-of-four watch (Organon Teknika BV, RM Boxtel, The Netherlands) for later monitoring of neuromuscular blockade. Anesthesia was induced with 2 mg/kg propofol (Disoprivan®; AstraZeneca, Zug, Switzerland). Alfentanil (Rapifen®; Janssen-Cilag, Baar, Switzerland) was administered with a Harvard 22 infusion pump (Harvard Apparatus, South Natick, MA) driven by the controller software. Sixty seconds before tracheal intubation, an alfentanil bolus of 30 μg/kg was given with the controller in the manual mode (see controller description in the next paragraph). Tracheal intubation was facilitated with 0.3 mg/kg mivacurium (Mivacron®; Organon, Pfäffikon, Switzerland). After intubation, sufficient hypnosis was maintained with isoflurane. The end-tidal concentration was manually adjusted by increasing or decreasing the end-tidal concentration in steps of 0.1–0.2 vol% in case the BIS value exceeded the target range of 50 and 60 for longer than 3 min. After intubation, the alfentanil controller was switched from the manual mode to the closed-loop mode. An infusion of mivacurium (0.3–0.5 mg · kg⁻¹· h⁻¹) was started to maintain neuromuscular blockade (train-of-four count 0–2 of four twitches). Patients were ventilated with oxygen in air and a fresh gas flow of 3 l/min to end-tidal carbon dioxide concentration of 35 mmHg. To determine the performance of the selected pharmacokinetic parameters for alfentanil (table 1), arterial blood samples were drawn during stable phases of the study periods (one sample in period 1, one or two samples in period 2). The samples were processed as further detailed in the  appendix.

A model predictive controller was implemented in an X-Oberon real-time system (Institute of Robotics, Swiss Federal Institute of Technology, Zurich, Switzerland) (fig. 1). In previous studies, only one input variable was used to control one output variable (single input–single output). Conversely, this control system had two input variables and was targeted to keep the MAP (primary input variable) as close as possible to the set-point MAP (MAP^ref) while minimizing the predicted Cp Alf (secondary input variable). A detailed description of the controller algorithm and the algorithms for artifact detection was published elsewhere.11Briefly, in subjects under noxious stimulation, a linear relation between MAP and effect site alfentanil concentration was assumed (equation 1) according to data from the literature13and from a pilot study14:

where ΔMAP = decrease in MAP (mmHg) induced by alfentanil, k = 0.088 (mmHg · ml · ng⁻¹), and Ce^Alf= effect site alfentanil concentration (ng/ml). The k^e0relating the plasma and the effect site concentration was originally derived from electroencephalography data15of alfentanil. Because a specific k^e0for the hemodynamic effects was not available, we assumed that for our purpose the electroencephalography-based k^e0would be adequate. Primarily, the controller estimated the effect site alfentanil concentration necessary to decrease the MAP to the MAP^refaccording to equation 1and computed the respective alfentanil infusion rate. If MAP was equal to the MAP^ref, the infusion would stop and the alfentanil concentration would decrease. The second input variable (Cp Alf) with a set-point value of around 100 ng/ml was introduced to maintain a certain minimal Cp Alf if MAP was equal to or below the MAP^refbut above the lower constraint value of 60 mmHg and to avoid unnecessary high opioid concentrations. As a third element, constraints for MAP were defined (60 and 120 mmHg). A MAP exceeding the upper constraint induced a higher alfentanil infusion rate compared with the value derived from equation 1. In case the MAP was below 60 mmHg, the infusion temporarily stopped to allow the Cp Alf to decrease. The constraints for Cp Alf were 0 and 400 ng/ml, thus defining a maximal concentration while allowing a total elimination of alfentanil in case the lower MAP constraint was violated.

The controller was tuned to react with moderate changes of Cp Alf during stable periods where MAP lies within the physiologic range of 60–100 mmHg while minimizing the alfentanil consumption. The tuning was achieved by defining weighting parameters for the different input and output variables in the control function. The weighting parameters were selected such as to make the controller more sensible to deviation of MAP than of Cp Alf. Further, the controller was tuned to react more aggressively if the constraints were hit.11 

The set-point MAP (MAP^ref) was initially set at 70 mmHg, the set-point Cp Alf at 100 ng/ml in every patient. During surgery, the MAP^refand the set-point plasma alfentanil concentration (Cp Alf^ref) could be changed at the discretion of the anesthesiologist to adapt to the individual requirements of the patient. The continuously measured MAP was compared with MAP^refevery 5 s, and the controller computed an alfentanil infusion rate to keep the MAP as close as possible to the MAP^ref. From the history of the recorded alfentanil infusion rates, the controller computed the Cp Alf using a three-compartment body surface area–adjusted pharmacokinetic model used by Shafer et al.  in the STANPUMP program16modified by Gentilini et al.  11(table 1). The Cp Alf^refimplied that a minimal Cp Alf was maintained even when the MAP was below the lower constraint of MAP (60 mmHg). In case of hypotension (MAP < 60 mmHg, longer than 30 s) in the absence of hypovolemia, 5 mg ephedrine was given intravenously.

Three operating modes of the system were defined. In the passive mode, the target parameters were set while the drug administration was stopped. In the manual mode, alfentanil administration was allowed at fixed infusion rates set by the anesthesiologist while the computer ran the pump and recorded the infusion rate (open-loop system). In the control mode, the alfentanil infusion rate was adjusted by closed-loop control. The start and end of the manual and control mode periods were recorded in every patient. During anesthesia, the controller could be switched from manual to control mode by pressing one button on the touch screen, without interrupting data acquisition.

The artifact detection algorithms implemented in the controller have been described in detail previously.11Briefly, a model-based and a signal-based approach were applied in parallel. With the system in the manual mode and the control mode, a real-time prediction of the MAP was computed according to equation 1. In case an artifact was detected in the arterial blood pressure signal, the MAP value sent to the controller was switched from the actual measured to the predicted MAP (with the prediction based on the previous five measured MAP values), and an alert was displayed on the screen. In case the artifact persisted for longer than 60 s, the infusion rate was set to zero and another alert on the screen prompted the anesthesiologist to switch from control to manual mode. Because it was a pilot system, the same anesthesiologist (C.B.) performed the anesthesia in all of the patients.

Heart rate, arterial blood pressure, oxygen, carbon dioxide and isoflurane concentrations, oxygen saturation (pulse oximetry), BIS, the controller operating mode (passive, manual, closed loop), MAP^ref, Cp Alf^ref, alfentanil infusion rate, predicted plasma and effect site alfentanil concentrations, and the validity of the arterial blood pressure signal with artifact flag were recorded every 5 s on a computer hard disk. The data from intubation to 5 min before skin incision (period 1) and the data from 5 min before skin incision to skin closure (period 2) were analyzed separately. Signs of inadequate anesthesia were defined according to Ausems et al.  17The number of episodes with inadequate anesthesia (longer than 60 s) was computed for both study periods from the automatic hemodynamic data records and the manual records of clinical data (sweating, flushing, etc. ). In addition, the area under the curve of mean arterial blood pressure and heart rate beyond ±20% of baseline was computed for the first and second study periods and for each subject. For comparison between subjects, the area under the curve values were normalized to the duration of time period, i.e. , divided by the duration of the related study period.

The performance of the controller was evaluated with similar parameters as were used for evaluation of computer-controlled infusions18and which have been used for evaluation of other closed-loop controllers.19The performance parameters were calculated for MAP and Cp Alf similarly by replacing measured MAP by predicted Cp Alf in the following formulas.

The mean absolute difference (MAD) between measured MAP and MAP^refwas calculated according to equation 2for each subject:

where Nⁱ= total number of measurements in subject i .

The weighted performance error (PE) for the i  th patient and the j  th time window was calculated according to equation 3:

The median prediction error (MDPE), reflecting the bias, the median absolute prediction error (MDAPE), reflecting the inaccuracy, and the wobble, reflecting the variability, were then calculated according to equations 4–6:

where PE^ijrepresents the j  th observation in the i  th subject, Nⁱrepresents the total number of observations in the i  th subject, and t^ijrepresents the j  th time window in the i  th subject. MDPE, MDAPE, and wobble are given as percentages.

To evaluate the effect of the second input variable (reference alfentanil concentration), post hoc  simulations were performed using the controller software fed with the recorded MAP values from the study, the recorded reference MAP and Cp, and the height and weight from each patient. The alfentanil infusion rate, the predicted plasma and effect site alfentanil concentrations, and the alfentanil total dose were computed with the Cp Alf^refconsidered (double input–single output [DISO] mode) and ignored (single input–single output [SISO] mode). The parameters from the two simulations were compared with a paired t  test, with a significance level of 0.05.

Thirteen patients were enrolled in the study; 2 patients were excluded from data analysis because of loss of arterial blood pressure signal and because of protocol violation at induction. The characteristics of the 11 patients (5 women and 6 men) included in the analysis are summarized in table 2.

The characteristics of anesthesia between intubation and skin incision (period 1) and during surgery (period 2) in the 11 patients are summarized in table 3. Short episodes of inadequate anesthesia according to the criteria by Ausems et al.  17(systolic arterial blood pressure 15 mmHg greater than individual normal in 2 or tachycardia greater than 90 min⁻¹longer than 1 min) in period 1 (before surgery) were recorded in 6 patients (one episode in 1 subject due to hypertension; a median [range] of 3 [1–3] episodes due to tachycardia in 5 additional subjects). The median duration (range) of these episodes was 115 (65–195) s. In period 2 (surgery), such episodes were recorded in 6 patients, with a median (range) of 1 (1–4) episodes per subject due to hypertension and 4 (1–5) due to tachycardia. The median (range) duration of these episodes was 110 (80–160) s. No flushing, sweating, or movement was observed.

The mean (SD) normalized areas under the curve of MAP beyond ±20% of baseline were 2.8 (2.2) and 5.1 (3.9) mmHg between induction and incision and between incision and skin closure, respectively. The mean (SD) normalized areas under the curve of heart rate beyond ±20% of baseline in the two study periods were 5.8 (8.3) and 4.3 (9.3) beats/min, respectively.

The median performance error of the predicted Cp Alf compared with the measured value was −30.2%, signaling a bias toward underdosing (fig. 2). The median absolute performance error was 30.2%. These numbers are in the range of previous data.16,20 

The isoflurane concentrations and the BIS values of every subject are presented in figures 3A and B. Only minimal changes of the isoflurane concentrations were necessary to keep the BIS within the target range, even upon skin incision. In the presurgery period, there were some accidental sharp and short-lasting isoflurane increases (spikes in fig. 3A) because of moving of the patient and rapid BIS increase during positioning (prone position).

Ephedrine (5 mg) was given in 9 of 11 subjects during period 1 (presurgery) and in 4 of 11 subjects during period 2 (surgery). In the subjects given ephedrine, the median (range) numbers of doses in periods 1 and 2 were 2 (1–4) and 1 (1–2), respectively.

The data on the controller performance in the two study periods are summarized in tables 4 and 5. Between intubation and incision and between incision and closure, the controller was in the control mode 98.9 (1.5) and 95.5 (4.3)% of the time. The reason to change from control to manual mode was an unreliable MAP due to artifacts. In one patient, an additional bolus of alfentanil was given in the manual mode because of a massive blood pressure response to skin incision. The predicted Cp Alf of every subject is presented in figure 4A, and the MAP values are presented in figure 4B.

The reference MAP, which was first set at 70 mmHg in every patient, was changed between 70 and 90 mmHg by the attending anesthesiologist in 8 of 11 patients (fig. 4C). In these subjects, the median (range) number of changes was 3 (1–5). In the remaining 3 subjects, the Cp Alf^refwas changed instead (fig. 4D), with a median (range) number of changes of 1 (1–2).

The alfentanil infusion rate and the predicted Cp Alf in a typical subject are presented in figure 5to illustrate the controller performance.

The offline simulation results are summarized in table 6and figure 6. The predicted Cp Alf values from the DISO simulation were similar to those recorded from the clinical study (fig. 4A). With the SISO simulation, ignoring Cp Alf ^ref, the cumulative alfentanil dose and the mean (SD) Cp Alf were substantially higher than with the DISO simulation.

Mean arterial blood pressure has only rarely been used as an input variable for closed-loop control of anesthesia and analgesia.21,22Our controller mimicked the dosing rules applied by many anesthesiologists in clinical practice and clinical studies17,23: increasing the opioid concentration by administering an intravenous bolus or increasing the infusion rate if the patient shows a hemodynamic response to surgical stimulation, and decreasing the opioid concentration in case the patient does not respond and to avoid excessive opioid concentrations.

The set-point precision of MAP was significantly different in the two study periods (MDPE, table 4) reflecting hypotension observed in the unstimulated patients, most of them needing ephedrine treatment, even though the predicted alfentanil concentration was only 122 (35) ng/ml. The set-point precision of Cp Alf during the preoperative period (MDPE, table 5) was in the same range as the set-point precision of MAP. During surgery, however, it was significantly worse. This reflects the controller tuning where the priority was set to control MAP. Our simulation illustrates the effect of the Cp Alf^refin reducing the alfentanil dose, especially during surgery (fig. 6).

Although this controller may be considered a DISO, it was designed to control only one of the three components of general anesthesia. An ideal anesthesia control system would integrate closed-loop control of hypnosis, analgesia, and neuromuscular blockade (multiple input–multiple output [MIMO]), but this has never been attempted for multiple reasons: the lack of a specific parameter to measure analgesia, the number of confounding factors influencing MAP together with noxious stimulation, and the effect of opioid–hypnotic–anesthetic drug interactions on blood pressure and electroencephalography, which could provoke controller instability and put the patient at risk of severe hypotension or inadequate anesthesia. In this study, the hypnotic component of anesthesia was represented by the BIS, and the end-tidal isoflurane concentration was kept rather constant (table 2and fig. 3). The mean (SD) BIS values and the mean (SD) isoflurane concentrations were similar to values published by Locher1for their isoflurane BIS controller tested in a similar patient population. In view of the fact that only minimal adjustments of the isoflurane concentration were necessary to keep the BIS in the target range, we presume that in the context of adequate hypnosis, a closed-loop administration of an opioid according to blood pressure results in a similar anesthetic state as closed-loop administration of a hypnotic according to BIS in the context of adequate analgesia. Fixing either the hypnotic or the analgesic component and extending the concept of a SISO controller by using a second input variable to optimize the output variable as in our study may solve the potential problem of controller instability. This controller may thus be a further step toward a future MIMO controller.

There are certain limitations in our study: The mean arterial blood pressure is not specific for the response to noxious stimulation and may be affected by a number of covariates, such as blood loss, heart failure, arrhythmia, manipulation of great blood vessels, and a variety of drugs (e.g. , inhalation anesthetics, propofol, β blockers, vasodilators). Nevertheless, the performance of this controller was similar to the performance of other recently published hypnotic controllers.1,3,19One reason for the good performance may be that we enrolled healthy patients (American Society of Anesthesiologists physical status I or II) undergoing surgery associated with only minimal blood loss. Only one subject had a blood loss more than 500 ml (2,200 ml), which was adequately replaced with colloids. Therefore, we do not have any data on controller stability in situations with larger blood loss; in such cases, continuous and sufficient volume replacement would be essential to preserve stability during closed-loop alfentanil administration. The frequent administration of ephedrine in the presurgery study period reveals that this controller in the context of the isoflurane concentrations necessary to keep the BIS level in the target range is not able to prevent hypotension in periods with limited noxious stimulation. One reason might be the choice of alfentanil instead of remifentanil, which would have allowed a more rapid decrease of the concentration after induction. Alfentanil was selected for better comparability of the results with our previous study.1It was even the purpose of the second control parameter (Cp Alf^ref) to keep a minimal opioid concentration also in case of hypotension, which in general is treated by administration of fluids or vasoactive drugs, to maintain a minimal level of anesthesia. Because a closed-loop controller cannot foresee the time of skin incision, a substantial hemodynamic response may follow this strong stimulus if the opioid concentration had been decreased to zero in the period of low stimulation before.

The anesthesiologist decreased the reference Cp Alf in two subjects to allow rapid emergence. Another potential limitation is the use of the k^e0value of alfentanil for electroencephalography effects. Because a k^e0of alfentanil to suppress hemodynamic response to noxious stimulation has not been determined, we had to assume that it would not substantially differ from the known k^e0of electroencephalography effect.

The opioid requirement to maintain the MAP within the target range was substantially different among the study population (coefficient of variation, 26–32%). The anesthesiologist in charge had two possibilities to respond to an increased or decreased opioid requirement of the individual subject: decreasing the MAP^refor increasing Cp^refand vice versa . The number of changes of the reference values was limited, however, so that the system remained truly closed loop and was not turned into a “physician closed-loop” system.

We conclude that this model predictive controller has a similar set-point precision as previous electroencephalography-based controllers for hypnotic administration and provides adequate alfentanil dosing during surgery. It may be useful to standardize opioid dosing for research purposes and can be considered a further step toward a future MIMO controller.

The authors thank Rolf Lauber, Ph.D. (Head of the Analytical Laboratory, Department of Anesthesiology, University Hospital of Bern, Bern, Switzerland), and his laboratory staff for the analysis of the alfentanil probes and Peter Schumacher, M.Sc., Ph.D. (Research Engineer, Department of Anesthesiology, University Hospital of Bern), for critical review of the manuscript.

Two to three arterial blood samples were withdrawn in each subject, one before and the other(s) during surgery. The samples were centrifuged at 3,500g  for 30 min at 4°C, and the plasma was then stored at −20°C until analysis. Alfentanil plasma concentration was measured by gas chromatography–mass spectrometry24with a detection limit of 0.4 ng/ml. The capillary column was a DB-XLB (J&W Scientific Inc., Folsom, CA) with 12 m length, 0.2 mm ID, and a film thickness of 0.33 μm. ²H⁵-Alfentanil was used as internal standard (Janssen Pharmaceutica, Beerse, Belgium). Quantification occurred in the selected ion monitoring mode, using the m/z 289 and 294 ion fragments for alfentanil and D⁵-alfentanil, respectively with a dwell time of 100 ms each. Plasma samples were extracted with Clean Screen® tubes (United Chemical Technologies Inc., Bristol, PA) with a mean recovery of alfentanil (n = 15) at concentrations 5–240 ng/ml of 89.5% (range, 72–105%). The measurements were made on the same day with coefficients of variation of 2.0, 2.2, and 4.3% at concentrations of 240, 64, and 5 ng/ml, respectively.