The last 2 decades have brought important developments in anesthetic technology, including robotic anesthesia. Anesthesiologists titrate the administration of pharmacological agents to the patients’ physiology and the needs of surgery, using a variety of sophisticated equipment (we use the term “pilots of the human biosphere”). In anesthesia, increased safety seems coupled with increased technology and innovation. This article gives an overview of the technological developments over the past decades, both in terms of pharmacological and mechanical robots, which have laid the groundwork for robotic anesthesia: target-controlled drug infusion systems, closed-loop administration of anesthesia and sedation, mechanical robots for intubation, and the latest development in the world of communication with the arrival of artificial intelligence (AI)–derived chatbots are presented.
We view anesthesiologists as “pilots of the human biosphere.” Administering drugs that act very quickly, anesthesiologists can be viewed as interacting with 3 different biospheres: consciousness, pain, and muscle function, albeit with imperfectly predictable interactions. Physiological and pharmacological knowledge combined with experience are the basis of what anesthesiologists can deliver as state-of-the-art care. In addition, there are physical interventions from anatomically guided insertion of peripheral venous cannulas or epidural catheters to ultrasound (US)–guided insertion of central venous catheters, and insertion of an endotracheal tube–aided sophisticated video-guided laryngoscopes.
Over the past decades, robotic devices have started to become available, some clinically, most for research purposes, with the purpose of helping anesthesiologists perform even better and deal with ever more complex environments.
What Is a Robot?
Karel Čapek probably first used the word “robot” in his play R.U.R., in 1920; the word derives from the Czech word “robota” for “forced labor” or “serf.”1 The robots were abused as slaves until they started a revolution and destroyed mankind. There are several definitions for what a robot is, but for this article, we will use, “any automatically operated machine that replaces human effort, though it may not resemble human beings in appearance or perform functions in a humanlike manner.”1
In anesthesiology, we distinguished originally 2 forms of robots, pharmacological (aiding drug administration) and mechanical (task performance). Most recently, a third form of robot or “bot” is based on artificial intelligence (AI; chatbots or other forms of automated speaking or writing devices for aiding communication between physicians, or with machines or patients).
All these robots present potential advantages and risks. In this article, we offer an overview of robots in anesthesia, from their history, to a prediction of how they might change practice in 2050.
Pharmacological Robots
Total intravenous anesthesia (TIVA) is very popular in Europe, with approximately 50% of general anesthesia cases done using this form of anesthesia, a perceived advantage being rapid turnover of.2 Purported advantages over volatile anesthesia include better environmental profile, early discharge, rapid cognitive recovery, and cost.3–10 There is greater effort needed with repetitive refilling and resetting of syringes, and adjustment of infusion rates by algorithms (since direct plasma concentrations of administered drugs are immeasurable). In addition, TIVA techniques facilitate sedation, and for prolonged surgeries ,even 3 infusion pumps may need managing (eg, propofol, remifentanil and cisatracurium), adding to cognitive strain for the anesthesiologists. The requirement for additional depth of anesthesia monitoring with TIVA (whereas for volatiles, end-tidal monitoring arguably suffices) adds to this cognitive load.
One form of TIVA is target-controlled infusion (TCI), which is available worldwide, except the United States, and by reducing the anesthesiologists’ workload can be said to fulfill our definition of a “robot.”11,12
A Brief History of TCI
The details of TCI systems have been discussed in detail elsewhere. In brief, the system delivers a drug by infusion to a specific target concentration, either in plasma or in a specified modeled tissue compartment of the body. Instead of programming a certain infusion rate, the anesthesiologist chooses a certain target concentration. Pharmacokinetic models in the TCI system are used to then calculate the infusion rate based on the chosen target concentration, and a variety of patient parameters, such as age, weight, or sex. A computer microprocessor calculates the infusion rate and other complex parameters, for example, the “wake-up” time, at any given time. The models are derived from population studies, depending on volunteers or patients. All this is necessary because real plasma drug concentrations cannot be measured.
Infusion models have evolved from a basic 2-compartment model13 to more complex models using a computer-assisted TIVA system.14 These more sophisticated 3-compartment models accounted for bolus dosing, drug elimination, and transfer of a given drug entering the body. The concept of TCI was later applied to other drugs, in order of importance: remifentanil, sufentanil, ketamine, and dexmedetomidine.
One issue is that the pharmacokinetic models originated in the 1980s, and since then, the population profile has changed with increasing age, obesity, and illicit or prescribed drug use. It is therefore important to update the models to account for these changes. Moreover, the original sample group for most models was derived from an area with a predominance of White volunteers: updated models need to be more inclusive. Interestingly, the only updated model for propofol is that of Eleveld et al,15 which extended the earlier Marsh et al16 and Schnider et al17 models based on 20 and 24 volunteers, respectively, to 1033 patients. The model uses age, postmenstrual age (PMA), weight, height, sex, and the utilization—or not—of concomitant anesthetic drugs as covariates. As interesting as this model sounds, it needed to be evaluated from independent researchers. Comparison with the Marsh et al and Schnider et al models using the median absolute performance error (MDAPE) and the median performance error (MDPE) for bias showed that the Eleveld et al model performed slightly better (MDAPE, 22% [range, 4%–50%] vs 25% [6%–58%] for Marsh et al and 26% [2%–54%] for Schnider et al). The MDPEs were −18%, 9%, and −20%, respectively—that is, the Eleveld et al and Schnider et al models tended to overestimate, and the Marsh et al model underestimate the actual plasma concentration.16-18 These results are not surprising since propofol metabolism is influenced by individual factors as discussed above and not accounted for within all models. Moreover, when tested in different populations (eg, Japan), accuracy is not always reproducible.19
In parallel with these advances in TCI propofol, models for TCI remifentanil are developed but not yet commercially available. Kim et al20 recognized the absence of obese volunteers in the original Minto model for remifentanil TCI, and the resulting Kim-Obara-Egan model included a wider patient population. Despite this, in simulation, all 3 remifentanil models (Minto, Eleveld, and Kim) resulted in similar drug administration.21 Supplemental Digital Content S1, https://links.lww.com/AA/E661, indicates the numerous models now available.
A study from 2017 compared no less than 11 pharmacokinetic models for propofol infusions for children and found that the Short model showed the best performance over a longer time course (steady state); however, there are study flaws as induction was done either with propofol or sevoflurane—thus creating a possible bias—and no depth of consciousness monitoring was used.22 The sample size was also low at only 21 children.22
A key practice development is the use of TCI in sedation using dexmedetodimine or propofol and even in nonhospital settings, and depending on the country, even in the hands of nonanesthesiologist care providers. In a very recent study of 101 patients, TCI propofol and remifentanil were used for moderately deep sedation.23 No patient experienced any adverse events, as in loss of responsiveness or desaturation (<94% saturation). TCI was provided by 2 TCI-trained registered nurses with 1 nurse responsible for monitoring the patient in a dental clinic.23 This shows the blessing and the curse of robots: a logical consequence of reducing anesthesiologists’ workload is that they could be replaced entirely!
There is therefore unlikely a single pharmacokinetic model that can be used for all patient populations, throughout the world. The accuracy and bias of the various models will depend on specific patient populations. It is the combination of depth of consciousness monitoring and TCI that is the key to success, the key to an individual patient–orientated anesthesia. Even if for a given patient a model produces a high level of accuracy and low bias in relationship to a given blood concentration, this does not mean that an adequate level of anesthesia is reached. Depth of anesthesia monitoring can be disturbed or falsified by artifacts; however, if the artifact level is very low and analgesia is adequate, depth of consciousness monitoring seems the best chance to titrate TCI. In the future, a model using multiple covariates with sophisticated models may be able to encompass every factor needed, but the likelihood of this reality will lie heavily on the advent and future development of AI programs.
The Robots of OpenAI: ChatGPT and Others
ChatGPT (OpenAI) is a product of the OpenAI initiative: after human language inputs, it can produce human-like responses benefiting from a huge database and advanced deep learning techniques.24
For 1 of the authors (T.M.H.), the arrival of intelligent bots reminds him of the arrival of the Internet many years ago. It makes perfect sense: in a society where writing is considered archaic, knowledge of grammar is rudimentary, and good scientific practice—in general life meaning using objective information—has become almost obsolete, a tool is hailed that reacts to verbal commands only. Any question can be answered by a robot, without the specific need or possibility to “investigate” if the response is valid or not.
Generative pretraining transformer (GPT) models are AI-created language models. Radford et al24 showed that unsupervised pretraining methods using huge datasets (in the millions, they called it “Webtext”) can create language processing systems that can perform human-like tasks—in true robotic fashion. Their original GPT-2 called language system was trained on a very large, diverse dataset underlining a basic machine learning concept that the larger the dataset, the less supervision needs to be involved to create artificially intelligent systems. ChatGPT is a next-generation language model that can produce structured text with well-advanced vocabulary; it can communicate with a user—or machine—and can, in the medical field, be used for patient monitoring, receiving health care–related information from the patient or user.25 Several research papers looked into the use of these bots in the research setting or educational setting, both for health care providers and patients.26–28 There is already at this point the possibility of using these machines in a way similar to decision support systems from the past.29 In fact, such decision support systems could soon be replaced by ChatGPT/AI-based language and database engines. However, these innovative tools need to be used carefully and within the standard ethical boundaries.30 In the present state of development, and this development moves forward at lightning speed, chatbots can be used for the creation of medical notes as well as initiating management—for example, creating a transcript of a patient-anesthesiologist interaction during a preanesthetic visit.31 The chatbot can then execute a note, or the digital file can be charted. Based on the transcript, current chatbots can answer questions, analyze laboratory results, or conduct any other anesthesia-related requests in compliance with any guidelines chosen. Chatbots can be used to “acquire” medical knowledge, or any other knowledge deemed useful by the user. This knowledge not only can be presented in any form chosen by the user—for example, voice, text, or images—but also can be discussed by the bot.31
There are presently several “intelligent” chatbots available. Although most focus is on the OpenAI-created GPT-2, -4, or the latest ChatGPT (launched in November 2022, and trained using supervised and reinforcement machine learning), this is simply because they were the first to be designed. However, there are several other intelligent chatbots, large language models, including but not limited to Amazon’s AlexaAI, Google’s LaMDA, DeepMind’s “Sparrow,” and Microsoft’s BingAI.32
Possible applications and consequences of intelligent bots in anesthesia care were recently described in Dr Richard Novak’s blog.33 He outlines using ChatGPT in a preoperative setting as described above. Having such communication with a chatbot is like having immediately available textbooks or guidelines. Novak poses a scenario: if in an examination situation, a resident is asked to diagnose a 5-year-old child after tonsillectomy who coughs blood, has difficulty breathing, and becomes unconscious, the treatment options offered by the ChatGPT should be exactly the same. Thus, communication with ChatGPT by any practitioner in that real situation could be very helpful. As with any robot, the language model could ease the practitioner’s workload, in this case the cognitive process. These language models basically provide diagnostic or therapeutic advice based on textbook knowledge and standard guidelines. Their limitation at this point is the speed of information transmission: whereas any experienced anesthesiologist can process this information in seconds, the bot would take time to transmit this information to the user. However, speeds are likely to improve. In fact, even while the authors are writing this review—July 2023—there are almost weekly new findings in the world of chatbots. The latest is the finding that Chat-Gpt-3, not even the latest rendition of OpenAI’s ChatGPT, possesses the ability to reason by analogy—something that hitherto was only found in humans.34
Robots Closing the Loop
The typical closed-loop system consists of an output (effect on the patient) in the form of monitoring. The input is the variable that is controlled, in general a drug infusion rate. Actuators, in most cases infusion pumps, manipulate the input, and a control center uses feedback from sensors. A certain target is set by the user, and the system aims to establish or maintain the measured variable as close as possible to the target value, and this over any given time. Schwilden et al35 published the first work of closed-loop propofol control guided by electroencephalogram (EEG) median frequency. Based on adaptive feedback the control mechanism consisted of the following: if the median EEG was in the range of 2 to 3 Hz, propofol was administered after a BET (bolus, elimination, transfer) infusion scheme; if the EEG frequency was outside the range, the difference between the measured and predicted values was used to adapt the propofol infusion rate. Schwilden et al35 formulated the basic principle of closed-loop anesthesia: “Using established pharmacokinetic-dynamic models, one can calculate programmed schemes of drug administration as a first approximation to the individual drug requirement. The controller has to subsequently identify only individual corrections to the suggested administration scheme.” Schwilden et al35 also stressed the need to be able to manually override the closed loop system as a safety measure.
Singh and Nath36 described a multiple-input multiple-output closed-loop system, meaning that not 1 but multiple components of general anesthesia (loss of consciousness, analgesia, or muscle relaxation) were controlled. Several groups have designed dual controller systems for propofol and remifentanil. Our McSleepy system (McGill University) is a tricontrol system for propofol, remifentani,l and rocuronium by closed-loop control.37–40 There are subtle differences in the performance of these systems.
Liu and Fischler’s dual controller system37 uses a proportional-integral-derivative algorithm, with propofol and remifentanil both infused using bispectral index (BIS) monitoring. The controller uses 3 signals from the BIS monitoring: the measured BIS value, electromyography (EMG) activity, and the percentage of burst ratio. The algorithms use a rule-base system to decide if remifentanil infusion is changed, or remifentanil and propofol infusions. Casas-Arroyave et al38 presented a dual controller incorporating BIS, heart rate, and blood pressure—the latter parameters combined using the “Analgoscore.”41,42 They used the patient state index as a clinical control feedback variable for propofol infusion, and the “Analgoscore” for control feedback of remifentanil infusion. Moreover, they extracted several parameters out of the patient state index, such as EMG signals, suppression rates, and spectral edge frequencies, as part of the control feedback for the propofol infusion. To avoid a “cliff-edge” phenomenon––for example, where a score of 60 is “awake” and 59 is “asleep”––fuzzy logic was used.38 Similar to the system developed by Hemmerling and his research group, this group uses something called the Bayesian principle described by De Smet et al43 for closed-loop control. The system uses the “sensitivity” of the patient’s reaction during induction and modifies hemodynamic target values accordingly.40 Using the system for induction—in contrast to the system by Liu et al described above—opens the opportunity to take into account the patient’s reaction to a given dose during the induction and adjust the model algorithms individually. In addition, the system allows the user to indicate “surgical stimulus” with a push button, thus feeding valuable input into the system that can influence infusion rates. Pushed by the user in anticipation of a painful stimulus, the system applies an opioid bolus without changing the infusion rate.
Another dual system described by Zikov et al44 uses a so-called wavelet-based anesthetic value for central nervous system (WAVcns) value, similar to the BIS or patient state index. Remifentanil was controlled using WAVcns values: propofol was infused as a slow actuator to maintain a certain baseline WAVcns target. As in the study by Casas-Arroyave et al,38 induction and maintenance were performed using the so-called icontrol-RP system. The authors noted that the add-on of remifentanil into the closed-loop system improved the accuracy of the system. This could be related to the fact that controlling 2 inputs—propofol and remifentanil infusion—takes better account of the reality that both consciousness and pain need to be managed intraoperatively. McSleepy offers closed-loop control of all 3 components of anesthesia, hypnosis, analgesia, and neuromuscular blockade.45–47
It uses an expert rule-based system, with a focus on artifact control in the hypnosis component. Several predefined push buttons allow the user to inform McSleepy about the progress of surgery at key time points. Once the last button is pushed, the system will automatically stop the infusion to trigger the wakeup period. The communication is also enhanced using live video feeds and voice clips indicating these various stages of anesthesia. The system can also be used to provide anesthesia at distant location.48 There are important safety features implemented—for example, the inability to further administer muscle relaxants when the time point for 20 minutes to the end of surgery has passed (Figure 1). McSleepy was successfully used in conjunction with the DaVinci robot (Intuitive Surgical, Inc), showing the feasibility of performing a completely robotic surgery and anesthesia.49
Closed-Loop Sedation Systems
Sedasys (Johnson & Johnson) is an automated sedation system in the endoscopic suite (Figure 2). The dosing loop was closed using an auditory or vibratory command module that asked the patient to squeeze a handheld switch attached to their hand (Figure 2). There were several design and conceptual flaws in the overall very appealing and necessary device.50
There is, however, a significant pharmacological range of dose to achieve a given sedation target.51–54 Another limitation was that Sedasys was only able to decrease the propofol dose and not to increase it automatically; any increase needed to be effectuated by the physician, rather defeating the object of a closed loop and increasing reliance on human experience or error and introducing a long delay before reaching a sufficiently deep sedation level. Many gastrointestinal specialists working with a dedicated anesthesiologist service were used to deeper sedation levels than Sedasys was designed to provide.55 Developing a closed-loop system for sedation while maintaining spontaneous breathing without using a mechanical airway aid is not an easy task. Whereas a too profound depth of consciousness in a closed-loop general anesthesia system does not affect the airway, a more profound level of sedation in a closed-loop sedation system can necessitate immediate airway intervention or intubation to secure the airway. Such system can therefore not rely solely on monitoring depth of consciousness monitoring alone since it is not known for an individual patient at what level of sedation the airway and breathing might be compromised. There appears to be only 1 closed-loop sedation system that monitors spontaneous breathing.45,56,57 Despite the complexity of the control model, the Varvel parameters yielded MDPE values of approximately 2% and MDAPE levels of approximately 11%, far better than corresponding values for closed-loop anesthesia, which is promising (Figure 3).58,59
Extending Closed-Loop Control
Rinehart and Cannesson have designed a closed-loop system for automated administration of fluids and blood products using the now established concept of goal-directed fluid management in the perioperative period.60–62 The system monitors stroke volume, heart rate, mean arterial pressure, and a dynamic predictor of fluid responsiveness.60,63
Joosten et al64 have broadened the concept to a combination of several closed-loop systems of Liu et al’s dual controller for remifentanil and propofol, closed-loop application of ventilation using the commercially available Zeus ventilator (which allows the adaptation of respiratory frequency and tidal volume to maintain a predefined Etco2 of 32 to 38 mm Hg), and Rinehart’s fluid management system (described above). Compared to control, 44 patients with the multiple closed-loop controller had significantly less time with BIS <40, had less end-tidal hypocapnia, and had a lower fluid balance. There was also quicker neurocognitive recovery in the multiple closed-loop system64 (Figure 4).
Mechanical Robots
Robots for surgery such as the DaVinci robot are now well-established. The precision of the robotic arms cannot be achieved by humans alone, although ultimately the surgeon moves the robotic arm. This removes tremor, allows finer cutting margins, and even allows remote surgery.
The Kepler Intubation System (KIS) was used to intubate a series of patients.65 The system consists of a joystick, a computer as the “electronic brain,” a carbon fiber robotic arm, and a bespoke attachment piece that allows the attachment of the videolaryngoscope (Figure 5). The human controller operates the joystick and to place the videolaryngoscope in the correct position, so simply pushing the joystick allows insertion of the tracheal tube. The theoretical advantages are remote use (distance from patient advantageous, for example, in coronavirus disease 2019 [COVID-19]), steadying of human hand actions, and if necessary, the system can be put on hold, freeing the user’s hands to allow any other movement. For example, when it comes to adjustments of the mouth, however, it is a very cumbersome system to use, large, and clearly not at a commercial stage yet. The small pilot study, however, showed an acceptable success rate, and a simulation study of its regional anesthesia robotic brother, the Magellan system, indicates advantages and steeper learning curves for novice users in comparison to conventional devices.66,67
Other robotic intubation prototypes use a similar concept and have been tested successfully in pigs and simulators (Figure 6).68,69
Identifying the tracheal inlet relies on AI for image recognition.70 Carlson et al70 collected images from users intubating an airway mannequin using a standard videolaryngoscope and applied 4 different machine learning algorithms to yield an acceptable sensitivity of 70% and specificity of 90% for detecting glottis opening. Other research groups have extended this AI concept to pediatric intubations.71
It is only a matter of time until the automated recognition of glottis/vocal cords will be integrated into commercial videolaryngoscopes.
More widely in health care, robots have been introduced to Korean hospitals to address nurse and other workforce shortages, performing routine low-level tasks, and even interacting with patients, especially children or patients with dementia (Figure 7).72,73
Predicting Anesthesia in 2050
In 2050, there will no longer be manual input or recording in any operating room or anesthesia workstation. We will no longer communicate with our machines in writing but with verbal command. Indeed, command may not be accurate as we will in fact exchange opinions, ask for advice, and listen to suggestions, especially as specific decision options will be offered. One could imagine a central command robot, covering the whole hospital, and spanning the entire patient pathway. For film enthusiasts, parallels can be drawn with the computer/machine HAL in Stanley Kubrik’s 2001: A Space Odyssey. This will start with the first contact with the patient, in the preassessment clinic, or before where the robot obtains all the necessary information from the patient without human intervention.74–77 Actions such as planning the details related to the hospital visit, answering questions related to a patients procedure, and even consenting for anesthesia are all possible.78 A final plan will be communicated with the anesthesiologist—human-in-charge (HIC)—and approved.79
Once the patient arrives in the anesthesia room, there will be 1 main anesthesia robot incorporating all necessary tools for induction, maintenance, and emergence of anesthesia, from ventilation, through US and videolaryngoscopic tools including a McSleepy-like anesthesia robot (Figure 8A,B). Communication with the main robot will be a confirmation of the previous anesthesia plan, suggested by the robot and confirmed by the anesthesiologist. Lines were already installed before in the “line robot room,” an intravenous (IV) technician or anesthesiologist (“IV anesth”).
Monitoring will all be wireless and already in place.80 Once the patient is in the anesthesia docking station, the monitoring of vital signs and anesthesia-related—noninvasive—monitoring parameters will be effectuated.
McSleepy would start and induce anesthesia and then prompt the anesthesiologist about all necessary actions. Mask ventilation will be undertaken by a mechanical robot and will be followed by a semiautomated, AI-guided tracheal tube insertion operated via joystick. After intubation, the robot will ventilate the lungs and advise the anesthesiologist about the next steps—for example, performing a preoperative US-guided block. This will be done using something similar to a bolt-snap US blocking robot. A probe will be placed over the designated area, where an acoustic signal indicates correct positioning, allowing the anesthesiologist to trigger the local anesthetic when ready.
In parallel, robotic surgery will take place, and the surgical robot automatically communicates with the anesthesia robot, especially important towards the end of surgery to achieve timely extubation and recovery.
The transport robots move the patient to the recovery room, where a team of recovery robots will take over.
As in our everyday life, robots will slowly take over our work: what will be left for us to do?
REFERENCES