Artificial Intelligence-Supported Systems in Anesthesiology and Its Standpoint to Date—A Review ()
1. Introduction
Since the first mortality incident in anesthesiology was reported in 1848 [1] , many further deaths in anesthesiology have occurred [2] - [7] and have been reported up until today [8] [9] . Human error has been identified as the main cause of morbidity and mortality in anesthesiology [2] [3] [7] [10] - [15] . The anesthetist has to work simultaneously with many different aspects, such as multiple manual actions, equipment, choice of procedures, choice of drugs, hospital policies and clinical uncertainty. It is challenging under such conditions due to cognitive overload for anesthesiologists to weigh the outcome probabilities and choose the best management plan for each case [15] . For these circumstances, artificial intelligence (AI) may be a powerful computer technique to improve patient safety in anesthesiology.
Over the past decade, computer technology has improved significantly. Today, computer hardware is more cost-effective and powerful than ever before and is able to support high-functioning software. Furthermore, a large number of AI software have been developed with proven effectiveness in multiple different daily applications, such as Google Maps, search engines, banking, image recognition, natural speech processing, language translation, textual analysis and self-learning [16] . An example of the effective use of computer software is the Anesthesiology Information Management System (AIMS), which allows for the automatic and reliable collection, storage, and presentation of patient data during the perioperative period [17] . Computers have overcome the simple tasks of data recording and calculation following mathematic formulae to now being able to learn, think and work in a human-like fashion. For example, computer AI is able to interpret X-ray investigations and generate a report or another example is a computer large language model (LLM), such as ChatGPT is able to read, think, understand and answer questions [18] - [23] .
Forty years ago, applications of computer technology in anesthesiology were reviewed in the literature [24] . Twenty-five years later, automatic intravenous anesthesiology was discussed [25] . Ten years after that, robotic anesthesiology topic was often discussed in the literature about how it affects daily practice in anesthesiology [26] [27] [28] . Recently, Thomas M. Hemmerling published the article “Robots Will Perform Anesthesiology in the Near Future” in anesthesiology [29] . Currently, artificial intelligence-supported tools in anesthesiology are being investigated, evaluated and developed intensively by medical practitioners, engineers and academics in the literature [30] - [39] . The earlier iterations of AI-assisted systems in anesthesiology were simplistic, using single closed-loop control and single control parameters, such as electroencephalogram indices [40] . Later developments of AI systems use multiple closed-loop controls incorporating multiple control parameters, as well as real-time data for machine learning to improve response [41] [42] .
Nowadays, AI in anesthesiology is a very common topic discussed not only in the literature, but also in the media. Many medical technology companies showcase AI-supported anesthesiology equipment on their website [43] [44] [45] . AI uses the advantages of the availability of large amounts of medical data from electronic medical records and powerful modern microcomputers [46] [47] . AI is the technique that enables computers to mimic human intelligence to provide a solution for a problem [21] . AI has proven its applicability in many different areas of medicine, improving patient outcomes such as drug discovery, diagnostic applications in radiology and pathology, as well as therapeutic and interventional applications in cardiology and surgery [46] [47] [48] [49] [50] .
Until today, although AI technology has proven its ability to help different areas, such as drug development, medical data recording, and pathology identification, AI is scarcely used in the current practice of anesthesiology [21] [46] [48] [49] . In the current literature, AI-based studies are focused on theory development and proposals of AI systems [51] . Further to this, it can be noted that few of these have demonstrated themselves practical in the daily use of anesthesiology [51] .
This fact opens the question of what is the availability of commercial AI-supported systems and what are the barriers to the implementation of AI technology in anesthesiology, as well as how the application of AI-supported systems in anesthesiology can be improved in the near future.
2. Aims and Objectives
The aim of this article is to give readers a quick up to date picture of the abilities of AI technology and its potential implementation in anesthesiology, the availability of commercial AI-supported systems for anesthesiology and those in the late stages of development. There will also be a discussion about the barriers to AI use in practice and the potential future AI applications to improve patient safety in anesthesiology.
3. Method
The databases of PubMed, ScienceDirect, Library of Western Australian Health Service, and arXiv—Artificial Intelligence databases were searched using combinations of the following keywords: artificial intelligence, anesthesiology, machine learning, pharmacological robot, mechanical robot, mortality, and morbidity. Worldwide Web search through Google, Bing and Yahoo search engines using commercial names of AI-supported systems. Included in this review are publications with a focus on AI algorithm learning techniques, and the application of AI algorithms in anesthesiology in the perioperative period. Peer-reviewed published articles, and official commercial worldwide websites, published between 2006 and 2023 were eligible for inclusion, including older papers referenced in historical review. Excluded from the review are studies of AI algorithms for use in anesthesiology in theoretical or early stages of development and AI application in non-anaesthetic areas of medicine as they are outside the scope of the aims and objectives of this article. As a literature review article using published information, ethics approval is not required.
4. What Is Artificial Intelligence Technology?
In order to help the reader to answer the question: “should an anaesthetic be AI robot assisted or entirely human controlled?” AI technology should be briefly summarized.
Artificial intelligence is a complex concept as AI which enables computers to mimic human intelligence to provide a solution for a problem is the transformation of mathematic performance to medical application [21] [52] . AI is used to make a robot perform different tasks such as learning, understanding, reasoning, problem solving, word recognition and decision making like humans. Traditionally, a robot is a device which is able to perform physical actions including walking or arm movement. However, nowadays, the meaning of the word “robot” has become wider and is often thought as a system, which can perform physical and/or non-physical activities like reading documents, answering questions, data recalling and summarizing, and object recognition. Artificial intelligence has been defined as the study of algorithms that give machines the ability to reason and perform functions [47] .
4.1. Key Parts in AI Technology
4.1.1. Big Data
Following the widespread use of electronic medical records (EMRs) and increasing modern medical investigations, huge amounts of health and medical data are available [53] . Big data can be used as inputs for training AI algorithms, in this case, it is called “training data” [46] . The data can be numerical, text, image, audio, and video. In order for an AI tool to work accurately and reliably, the used data must be reliable with correct information and in the correct format [40] [47] [53] . The computer rule “garbage in-garbage out” is valid in AI technology [31] [46] .
4.1.2. Machine Learning (ML)
Machine learning is a practical field of AI with the aim to develop software that enables a computer to learn from training data, gain experience and gradually improve its ability to make predictions [54] . ML can be applied to identify patterns and make decisions with minimal human intervention. It is important for automatic anesthesiology delivery as it is necessary for the predictive analytics required for clinical decision-making [55] .
Traditional computer programs are programmed with explicit instructions to elicit certain behaviors from a machine based on specific inputs. Machine learning, on the other hand, allows for programs to learn from and react to data without explicit programming. Most applications of AI for anesthesiology are still in research and development [47] . The current focus of artificial intelligence within anesthesiology is not on replacing clinician judgment or skills but on investigating ways to assist them [47] . Machine learning is driven by an algorithm.
4.1.3. Algorithm
Algorithm is a computer program, which sets the data analysis rules for machine learning to follow [21] [56] . In other words, an algorithm is a mathematical model responsible for machine learning and performing the requested activities. For machine learning, big data is used as inputs for data analysis following an algorithm. The algorithm is trained by training data to improve its performance [21] [56] . The U.S. Food and Drug Administration (FDA) approved the first AI algorithm for use in medicine in April 2018 which analysed the images of the fundus to diagnose diabetic retinopathy [47] . Until today, 70 algorithms have been approved by the FDA for different areas of medicine such as, radiology, cardiology, internal medicine and general practice, 22 of which have perioperative utility [56] . Recently the transparency of these algorithms are discussed and reviewed. Unfortunately, there is no FDA approved algorithm with anesthesiology utility found [57] .
4.2. Types of Machine Learning
4.2.1. Supervised Learning
An algorithm is trained following the rules, which are set by a human expert to predict a pre-specified output. Supervised learning is recognized as “knowledge based” or “expert systems based” machine learning. Supervised learning requires training data and their associated outcomes to be inputted by a human expert in the algorithm’s learning process. For example, a radiologist is required to input many chest X-ray images and its corresponding findings into the AI algorithm to “teach” it, so that when asked to evaluate a new X-ray, the AI is able draw from previous learning to suggest the findings. The data used to assess the performance of the algorithm’s accuracy and reliability is named “test data”. After supervised learning process, the trained algorithm is able to provide the asked output for new data inputs [47] [51] .
4.2.2. Unsupervised Learning
Unsupervised learning is where algorithms identify patterns within an inputted clinical dataset (big data) which is then applied to analyze new data inputs [47] . Unsupervised learning is recognized as a “non-knowledge based” system [58] .
4.2.3. Reinforcement Learning
Reinforcement learning refers to the process, by which an algorithm is used to attempt a certain task via a device, and the result of the task (success or failure) is reported and the algorithm learns from trial and error and improves its next action [47] . This training method is based on rewarding desired outcomes and/or punishing undesired ones [59] .
4.3. Techniques in Machine Learning
There are different techniques for ML, such as, fuzzy logic, standard logic, neural networks, classical learning, deep learning, and Bayesian methods.
4.3.1. Classical Machine Learning
In the classical machine learning process, the training data is selected from big data by experts to guide the algorithms in the learning process [47] .
4.3.2. Neural Networks and Deep Learning
Neural networks, like biological nervous systems, process signals in layers of computational units (neurons) and is the most important technique for machine learning [40] [52] . A neural network consists of an input layer that describes the input data and an output layer that describes the result, with a hidden layer in between which is a data processing step using at least one algorithm [52] . That means many machine learning algorithms can run in a neural network simultaneously to give the outcome. For very big data sets and complicated task requests, many layers of neural networks are required and thus expands to become “deep neural networks” or “deep learning techniques” [47] . Most commonly, supervised learning type is used to train deep neural networks [54] .
4.3.3. Fuzzy Logic
Unlike standard logic or digital logic using binary numbering for positive and negative outcomes, fuzzy logic allows for probability, i.e. processing numerical values between 0.0 and 1.0 [60] . Fuzzy logic allows the algorithm to guess a result then later prove it in the next step. The rule-based systems (i.e. if… then systems) are used in Fuzzy logic [60] . Usually, fuzzy logic technique is used for supervised machine learning type [58] .
4.3.4. Bayesian Method
Bayesian method use algorithms which can model uncertainty and update probability from new data. From the previous results or factors that may influence an event, Bayesian method can provide a description of probability of the event [47] .
4.3.5. Natural Language Processing Model
Natural language processing models are the algorithms which focus on machine understanding of human language [47] . Natural language processing can achieve understanding of syntax and semantics to approximate meaning from phrases, sentences, or paragraphs [47] . The natural language processing model can be used to automatically analyze EMR [40] .
4.3.6. Computer Vision Model
Computer vision model are the algorithms which enable the AI to understand images and videos. In medicine, computer vision has proven its ability to help clinicians in interpreting radiological investigations [47] . For example, computer vision can be applied in anesthesiology to analyze ultrasound images to identify anatomical structures during ultrasound guided procedures [40] [47] .
5. AI-Supported Systems in Anesthesiology
Nowadays, it is easy to find ever increasing daily applications of AI technology such as in the car industry for example, Tesla’s autopilot or voice recognition devices such as Amazon’s Alexa [61] [62] . Further to this, AI has proven its human-like processing abilities in the case of the IBM created system Deep Blue’s victory in chess over grandmaster Garry Kasparov in 1998 [47] .
More topically, in November 2022 the AI robot “ChatGPT” was launched by Open AI. The early testing shows that Chat GPT can read and understand text correctly, write text summaries, reason, answer questions and provide suggestions [19] [63] [64] [65] . This commercial device has performed imperfectly but with acceptable accuracy and reliability for its practical purpose [18] . Fortunately, in these cases errors can be tolerated as they do not directly cause mortality or morbidity.
The application of AI in anesthesiology is different as it requires absolute accuracy and reliability due to its direct effect on patient outcomes [16] . Hence, the AI tools and algorithms used for medical purposes must be rigorously tested, trialed and approved by a governing health body before they can be used in the clinical setting.
For applications of Artificial Intelligence in Anesthesiology, a number of early-stage research and tests have been reported in the literature and many suggestions for AI system evaluation guidelines have been proposed but there are no current internationally accepted guidelines for AI tool performance benchmarking and evaluation [66] .
In the earlier stages of AI in anesthesiology, Liu et al. introduced a dual closed-loop anesthesiology control system which used electroencephalograms (EEGs) as a single monitoring data as the input to control propofol and remifentanil infusion [67] . Today, more complex big data such as electrocardiogram, variability of heart rate, blood pressure, pulse oximetry, EEG based monitoring (e.g. bispectral index (BIS)), mean arterial pressure (MAP) are proposed as control parameters to maintain stable anesthesiology [30] [32] [40] [47] . The anesthesiology control systems usually use sophisticated machine learning types such as reinforcement learning and learning techniques such as deep learning or fuzzy logic to achieve the desired standard [47] .
The current proposed AI robots can perform a number of activities in anesthesiology, such as: clinical decision support, control of anesthetic medication delivery, adverse event prediction, ultrasound interpretation for regional anesthesiology, pain response monitoring and prediction, depth of anesthesiology monitoring and management of operating theatre logistics [47] . Robots in anesthesiology are proposed to reduce the repetitive actions of the workload [68] . This review focuses on the recently proposed robots at the late stages of their development as early developments are far from implementation in clinical anesthesiology and many remain in theoretical stages, which is outside of the scope of our review.
5.1. Clinical Decision Support Systems
Artificial intelligence can provide guidance or suggestions from data it has analyzed to support clinical decision making such as alerting the clinician to possible drug-drug interactions for patient safety [52] . Anesthetists can use the abilities of AI-supported tools to provide safe care and improve patient outcomes from surgery [15] [31] [47] [66] [69] .
Topically, ChatGPT (Generative Pre-trained Transformers) is a natural language model that was launched in November 2022 and trained by OpenAI. ChatGPT has been tested in different areas including medicine, research, education and software engineering [20] . There is a concern of accuracy, as ChatGPT does not provide citations used to generate information for its answer, and when references are asked to be provided, ChatGPT in some cases provided unrelated or non-existent citations [65] [70] . The algorithms in ChatGPT use reinforcement type learning to improve their function by continuous feedback from complex real-world situations [71] . The drawback of this learning type is that it is high cost and time consuming. ChatGPT has the potential to revolutionize the healthcare industry by improving patient outcomes, reducing costs, and facilitating more efficient and accurate diagnosis and treatment [18] .
ChatGPT can use the natural language processing model to automate analysis of electronic medical record data. ChatGPT is currently being tested to extract information from free text such as EMR data or journal articles to build databases and learn [19] . It can then use inputted patient information to identify surgical candidates, assess for risk of adverse events, or facilitate billing [19] . Although ChatGPT provides promising outcomes in areas such as writing software codes and answering general questions, the gap between ChatGPT and applications in anesthesiology remains wide. Recently, ChatGPT was tested on different medical exams such as The United States Medical Licensing Exam (USMLE) and a Korean parasitology examination [72] [73] . ChatGPT performed poorly in the USMLE at a just pass level despite many correct answers to the questions currently published on the internet [70] [72] . Furthermore, ChatGPT’s performance on the parasitology examination was worse than that of medical students who sat the same exam in Korea [73] . ChatGPT has not been trained for medicine specifically, because of this, there are a number of limitations that still remain such as not being able to access medical research articles and an inability to provide a critique on a presented medical article [20] .
AlertWatch-OR by AlertWatch™ is a commercial clinical decision support (CDS) system, which is run by AI technology and designed specifically for use in anesthesiology [44] . It can provide case-specific advice after training using the available big data in combination with physiologic and laboratory data [58] [74] . This CDS system can detect adverse respiratory and haemodynamic events during anesthesiology and provide guidance for management of these intraoperative events to maintain stable anesthesiology and increase compliance to anesthesiology guidelines [74] [75] . Furthermore, the algorithm can be trained by supervised or non-supervised learning types and its ability is improved over time based on the availability of new data sets and user feedback [58] . The AIMS can be used to develop the CDS system to provide real-time advice to deliver safe patient care in anesthesiology as a requested corrective measure can be made immediately [31] [58] .
AlertWatch has a notable strength in its user-friendly visual display utilizing pictured organ systems, colour codes and text descriptions of alerts [76] .
SEDLINE® by Masimo (USA) monitors are commercial AI-supported tool in current clinical use [77] . They assess of depth of sedation using algorithms to process electroencephalogram waveforms to generate a Patient State Index between 0 (no activity) and 100 (alert) [78] . These devices can be unreliable as they heavily depend on the patient age, anesthetic agents used and interference from motion or electrocautery [47] .
5.2. Pharmacological Robots
The most important application of AI and machine learning in anesthetic practice is intraoperative management using closed-loop control algorithms. The typical feedback closed-loop control of intraoperative management is illustrated in Figure 1.
The closed-loop control algorithm of the AI robot collects and analyses a large amount of physiological data from many patients and draws patterns to actuate and maintain patient parameters at setpoints [51] . A closed-loop anesthesiology delivery system which intervenes only on the hypnotic component has successfully been used in the past for total intravenous anesthesiology induction and maintenance [79] . This is important for machine learning to enable prediction of the changes in control parameters (vital signs) of the patient during the operation [79] . During the anaesthetic, the algorithm is asked to detect if the patient parameters start to drift out the desired ranges, which are set by the anesthetist [79] . The algorithm will make the actuator adjust the drug injection rates to maintain a stable anaesthetic and alarm should the parameters move out of the expected control conditions [16] [79] . The difference between the conventional closed-loop control and AI-supported control is that the AI-supported tools take into account more control parameters and are able to learn from new training datasets to improve its performance with experience [51] .
Figure 1. AI-supported feedback closed-loop control in intraoperative management.
Pharmacological robots are AI-supported systems which operated using a closed-loop control algorithm. These robots can inject the required doses of anaesthetic drugs into the patient to reach preset values to maintain a stable anesthetic [74] [79] . The possible advantage of pharmacological robots lies in rural and remote medicine to make anesthesiology accessible at anytime and anywhere without the need of the presence of a trained anesthetist [76] .
An AI-supported robot McSleepy was introduced by Hemmerling et al. in 2013 as a pharmacological robot which can autonomously control sedation, analgesia and neuromuscular blockade simultaneously during induction, maintenance and emergence [41] [74] . McSleepy is an automated closed-loop system delivering a hypnotic agent, analgesic agent and muscle relaxant, remifentanil and rocuronium targeting the parameters of BIS, analgoscore (nociception score based on blood pressure and heart rate) and muscle relaxation using train of four ratio [41] . Hemmerling et al. reported pilot studies of the efficacy of this system have shown good performance indices in a randomized control trial of 186 patients twenty patients undergoing elective thyroidectomies in Pisa, with their anesthesiology remotely controlled in Montreal, as well as its application in complex cardiac surgery [41] [80] [81] .
In an earlier iteration of McSleepy in 2008 as reviewed by Naaz et al., a closed-loop feedback control system maintaining anesthesiology depth using a hand-crafted rule and response algorithm was unsuccessful in management of complex cases [40] . To overcome this weakness, a neural network learning technique was applied [40] .
Recently, a closed-loop system for intravenous anesthesiology delivery has been developed named iControl-RP. This robot is able to make decisions and control the administration of the intravenous agents Propofol and Remifentanil [47] . iControl-RP takes into account neuromonitoring using NeuroSENSE (EEG based quantification of cortical activity including burst suppression ratio) monitoring as well as haemodynamic monitoring data into its decision making [42] . West et al. has found iControl-RP to perform at an acceptable level in clinical practice and is a feasible solution for propofol and remifentanil anesthesiology closed-loop control [42] .
Sedasys® by Ethicon (USA) was an AI-supported robot which can be managed by non-anaesthetic medical staff [82] [83] . However, this device has been withdrawn from the market in 2016, three years after its FDA approval in 2013 along with approval for use in Australia the European Union and Canada for gastrointestinal endoscopy [74] [84] . This robot performed a propofol sedation in healthy adults (patients of ASA 1 and 2). This robot records patient blood pressure, oxygen saturation, respiratory rate and automatically adjusted propofol infusion rate, oxygen flow and gave the operator cues to optimize patient care.
CLADS™ is a manual or automatic closed-loop system for propofol intravenous anesthesiology delivery. This system was developed at the Postgraduate Institute of Medical Education and Research, Chandigarh, India [85] . It uses BIS as the control variable and electroencephalographic indices as the input variable [85] . The CLADS™ system is able to use 5-secondly BIS feedback, taking into account BIS error, to adjust propofol infusion rates following a pharmacokinetic model to maintain preset BIS values [86] .
A clinical trial using CLADS™ was carried out in different groups of patients using the automatic mode of propofol infusion (non-cardiac surgery, cardiac surgery, postoperative sedation, and high-altitude surgery). The trial results showed CLADS™ was effective and efficient as compared to manual control and can be a useful tool to decrease cognitive load for anesthesiologists during anesthesiology [85] [86] .
RUGLOOP-IITM (Department of Anesthesiology, University Medical Center, Groningen, Netherlands) is a robust, patient-individualized closed-loop control system for propofol administration using BIS as control variable. This system uses Bayesian methodology for machine learning [87] . This system uses Bayesian variances to distinguish the specific patient model from the population model [87] . It was found that the RUGLOOP™ system could achieve induction within set time limits and less overshooting of BIS above pre-set values than human control [87] . The closed-loop control of anesthesiology maintenance had similar quality and movement scores to human controlled anesthesiology and was feasible in terms of hemodynamic and respiratory stability [87] .
The CONCERT-CL™ (VERYARK Technology Co., Ltd. (Guangxi, China)) is a closed-loop propofol and remifentanil infusion system to control intravenous anesthesiology [88] . This system, similar to others, is a target-controlled infusion (TCI) system and uses BIS as a control variable [88] . A clinical trial was performed in 2013 on 89 patients, who were between 18- and 65-years-old and have ASA classifications I or II and required general anesthesiology [88] . The CONCERT-CL was used to automatically regulate the TCI of propofol to maintain the BIS value in preset range [88] . The trial results show the CONCERT-CL system can regulate the TCI of propofol successfully as the BIS value is maintained in an adequate range [88] . While using this system reduces the workload of anesthetists’ significantly, it was only a valuable assistant which required the constant supervision of an anesthesiologist [88] .
Many other closed-loop control systems have been tested in the practice of anesthesiology, unfortunately their performances were not feasible and have not been further improved or tested to date [67] [89] [90] . The main challenges faced by these systems were patient safety and efficacy [83] .
Closed-loop anesthesiology delivery systems control the depth of anesthesiology with greater accuracy and using less sedative agents than human controlled anesthesiology as the automated systems change the medication rate more often than human anesthesiologists [29] [79] . This has a sedative agent sparing effect and may improve patient outcomes in the post operative period by reducing the incidence of delirium [42] .
5.3. Mechanical Robots
Mechanical robots are AI-supported tools that also operate by a closed-loop control algorithm. These robots perform the manual tasks of an anesthesiologist automatically. The use of mechanical robots in anesthesiology can reduce some of the anesthesiologists’ repetitive actions. This gives the anesthetist more time to focus on patients and related perioperative issues, hence improving patient safety and service quality [51] [74] .
Endotracheal intubation and regional anesthesiology are the main proposed fields of application of mechanical robots.
5.3.1. Endotracheal Intubation
For the robotic endotracheal intubation development, the AI-supported multipurpose Da Vinci® Surgical System (Intuitive Surgical, Inc., Sunnyvale, CA, USA) was tested successfully in the performance of two anesthesiologist controlled semi-automated simulated fibreoptic intubations [74] [91] . The Da Vinci robot design has multiple robotic arms so it can handle complicated cases with different technical difficulty levels [91] . There have been successful simulated oral and nasal fiberoptic intubations on an airway simulation manikin, while direct laryngoscope intubations are challenging and not feasible [91] . It can be noted that these intubations were technically difficult due in part to the multiple arms and the robotic interface [91] .
The Kepler Intubation System (KIS) is designed to perform intubation automatically or semi automatically [92] . The system consists of a remote control centre and control joystick linked to a robotic arm with a standard videolaryngoscope [92] . In the semi-automatic mode, the robotic arm is controlled by the joystick, which is operated using a specific software and interface [92] . Hemmerling et al. concluded that using the KIS, a human expert is able to achieve remote intubations between 40 - 60 seconds after a trial of 90 simulated manikin intubations [92] .
5.3.2. Regional Anesthesiology
In terms of development of robotic regional anesthesiology, tracking nerves is an ideal task for the computer vision artificial intelligence technique. However, it should be noted that due to the variation in ultrasound images and the similar acoustic impedance between nerves and their surroundings, makes it a more difficult task than AI’s usual daily application in facial recognition technology [33] .
The DaVinci system has been used to perform trials of nerve block and placement of a perineural catheter on a non-human model using ultrasound-guided technique [40] .
The Magellan system (Oceanic Medical Products, Inc., Atchison, KS, USA) is designed to perform semi-automatic regional anesthesiology using ultrasound guided technique for peripheral nerve blocks [40] . The robot has an arm with a nerve block needle at the end and can be controlled by a joystick [33] [74] .
A recent training study of five anesthesiologists used the Magellan robotic arm to assess learning curves on a nerve phantom [93] . The study found steep improvements on learning curves over 10 needle insertions using the robot compared with manual insertion [93] . Notably, this study was limited by its small sample size, lack of performance criteria and low repetitions [93] . Given that earlier attempts had considerably longer performance times than later attempts, the steep learning curves are likely due to the novelty of the technology [93] . The Magellan system uses a custom algorithm which assists in recognition of nerves in ultrasound images [93] . Notably, the robotic assistance system allowed operators to learn needle positioning faster compared to manually and there is a reduction in the variability in time to position needle between subjects [93] .
In the US, AI-supported mechanical robots remain in the development stage, no truly automated closed-loop anesthesiology delivery systems have been approved for commercial clinical use [74] .
5.4. Event Prediction
Various types and techniques of machine learning have been tested for application of perioperative care risk prediction such as supervised learning, reinforcement learning, neural networks, deep learning and fuzzy logic. Preoperatively, risk assessment is an invaluable tool for both elective and emergency surgical cases where prediction of mortality and/or morbidity has the potential to change perioperative management [94] . An intraoperative example is neural networks that can be applied for prediction of the hypnotic effect of an induction bolus dose of propofol with greater accuracy than the average estimate of anesthesiologists in practice [95] .
5.5. Ultrasound Guidance
Ultrasound guided technique is commonly used in mechanical robots as described above such as the Magellan system used to perform regional anesthesiology [96] . The most common machine learning technique employed to achieve ultrasound image classification is neural networks and deep learning [47] . Smistad and Løvstakken used 15 ultrasound images of patient’s groins to train a convolutional neural network in identifying femoral artery and vein, distinguishing it from other femoral anatomical structures such as muscle, bone or acoustic shadow with an accuracy of 94.5% ± 2.9% on average [97] .
5.6. Pain Management
There are currently no commercial AI products available for widespread use in the area of pain management in the perioperative setting. Attempted AI application in pain management has a wide range from the processing of brain imaging for identifying pain to prediction of opioid dose response using biomarkers [47] .
Images were collected from human volunteers who had functional magnetic resonance imaging while exposed to painful and non-painful thermal stimulation and used machine learning to analyse their differences and similarities. They found that machine learning analysis the brain images as a whole more accurately identifies pain than the analysis of individual brain regions associated with nociception [47] .
PainChek™ is an AI tool used to identify pain in patients with dementia and is designed for use in Geriatric medicine and therefore is outside the scope of this review on AI-supported systems in anesthesiology [98] .
Gram et al. analysed electroencephalopathy signals from 81 patients using machine learning techniques and predicted with 65% accuracy which patients would respond to post-operative opioid therapy [99] .
5.7. Depth of Anesthesiology Monitoring
Algorithms have been developed to analyze complex EEG waveforms using machine learning techniques to create an index directly related to measurement of depth of anesthesiology [47] . The strength of using sophisticated neural networks and deep learning techniques is the ability of the algorithm to engage in unsupervised learning to identify patterns that will best predict target values such as awareness based on EEG as well as haemodynamic data [100] . The AI algorithm is able to predict this more reliably than a human expert [47] .
5.8. Operating Room Logistics
Combes et al. trained a neural network using data from an EMR database containing information of staffing, operating room flow and time in post anesthesiology care unit to predict duration of one episode of surgical care [101] . The prediction is based on operating team, type of operation and relevant patient medical history. It is notable that the accuracy of their models never exceeded 60% [101] .
Another example of combined fuzzy logic and neural network techniques to optimize bed use is by Devi et al. who modelled type of surgery, surgical team experience, type of anesthesiology, experience of the anaesthetic team and relevant patient factors in ophthalmologic surgery [102] . They achieved accuracy between 81% - 86% in prediction of duration of surgical care episode depending on case type [102] .
In these models, the potential of AI technology is not to replace the anesthesiologist, but to augment and improve the workflow and decision making of the surgical and anesthesiology teams [47] .
Summarized below (Table 1) are the AI-supported tools in anesthesiology at late stages of development reviewed in this article.
6. Limitations and Discussion of AI-Supported Systems in Anesthesiology
AI algorithms have many times proven their capability in complex human-like thinking for example the chess algorithm “Deep Blue” winning against grandmaster Garry Kasparov twenty-six years ago [47] . Although a number of research on AI application for anesthesiology are published in the literature, but fully developed systems or those ready for clinical trials remain limited [31] [54] [94] [95] [97] [100] [101] [102] [104] . This gap in the literature represents an
Table 1. AI-supported systems for anesthesiology.
opportunity for the development and progression of clinical anesthesiology in line with multiple other industries and areas of medicine successfully implementing AI technology.
Therefore, despite the advancements and strengths of AI in anesthesiology and the promise it offers, the limitations of the application of AI technology in anesthesiology are important to consider. This will enable the capitalization of AI technology to improve patient outcomes by reducing cognitive load for anesthesiologists, improve efficiency, and open the possibility of remote anesthesiology.
6.1. Big Data Processing
AI machine performs prediction or answers a question by using medical training data, which must be reliable and connected to a specific case in a strict format. Unfortunately, a standard data format for ML is not available at present and there exists much incompatibility between data formats from different EMR systems [3] [46] [51] . Furthermore, the quantity and quality of the training data is very important as if small datasets or irrelevant data is used for the algorithm training process, then the output will be similarly unreliable when used to aid decision making [105] .
Current AI developments are based on differing data formats that is not conducive to international adoption. To accelerate and encourage the development and subsequent application of AI technology in anesthesiology, the author believes a common data format between international EMRs is a vital first step to the further development of AI systems for use in anesthesiology. The aforementioned AIMS is a good example of a widely adopted software enabling perioperative patient data collection [17] .
In order for the AI-supported systems to work effectively, they should process big data and give accurate outputs in real time. Currently, acquiring and processing high velocity and large data volumes in an operating room setting is a challenge due to the multiple and complex algorithms used in AI-supported robots for use in anesthesiology [58] . The author believes due to the limited resources in healthcare, the plausibility of upgrading expensive computer systems to support each new generation of algorithms is unrealistic. Therefore, the algorithms should be designed to be simplified and efficient.
6.2. Individual Case Variability
AI robots answer questions based on the information from many factors, but are unable to take into account the variability of significance of each factor for individual patients as there may be wide variation. For example, ML works based on general models such as the pharmacokinetic model, which adequately calculates plasma concentration of drugs, but is unable to predict the effect of such plasma concentration on patient sedation [88] . Furthermore, the monitoring on which measurement of adequate sedation and unawareness is based, predominantly EEG, remains notoriously unreliable [106] . Due to this inadequacy, AI algorithms are still not competent enough in practice to surpass human performance [40] .
The beginnings of development to overcome the wide variability between individual patients can be seen in the novel robots using Bayesian method for machine learning. The continuous updating of data from experience theoretically enables the robots to adapt its decision making from real time data updates during anesthesiology. The author suggests that improvement of user-friendly graphic interfaces and algorithms to enable increased individual patient data inputs is important for the personalization of AI anesthesiology and hence improve patient outcomes.
6.3. Lack of Knowledge and Ability of the Current AI-Supported Systems
Anesthetists make complex considerations for each case involving collecting patient information, review of medical data, designing anaesthetic plans, and performing individualized clinical procedures. The anesthetist works in a multidisciplinary team including the surgeon, pharmacists, medical specialties, pre-habilitation allied health staff etc. to negotiate the surgical plan and perform the optimal anesthesiology. Human intelligence is used to deal with many complex medical and interpersonal situations simultaneously, which mathematical algorithms are unable to replicate [54] . Further to this, as summarized in Table 1, the current pharmacological robots use electroencephalographic indices as a main or only control parameter as a measure of sedation. However, these values are notoriously unreliable and are affected by much interference [21] . Current developed ML algorithms are neither sophisticated enough nor have adequate ability to integrate all aspects of an anesthesiologist’s work; thus, it is impossible for AI technology to replace an anesthesiologist in the near future [40] .
Human intelligence can be used to deal with abnormal cases; thus skill, experience and knowledge of an anesthetist are combined to make life saving decisions. The AI machine actions or gives predictions following mathematical rules, the predicting algorithms used in AI systems do not support human intervention in a clinical setting. It is difficult for the AI user to know the appropriate actions to take when the robot gives a warning or makes an unexpected decision during anesthesiology [54] . Because of this potential need for rescuing interventions, it is unlikely that anesthetic delivery systems would ever function fully autonomously without clinician supervision [55] .
For anesthesiologists to confidently use AI robots, clear visualization features rich in information and explanation needs to be an integral part of the machines. The author suggests that these graphic interfaces should be designed by anesthesiologists to appropriately rationalize information displayed. Computational or engineering related errors which are out of the scope of practicing anesthesiologists should be minimized to avoid unnecessary apprehension.
The author believes, to improve safety and reliability, AI needs time to prove its ability and improvements implemented before it is applied in daily practice safely as patients cannot be treated as test subjects for novel technology with little evidence base. This is a lesson learned from the development of McSleepy, which in its early version used a hand-crafted rule and response algorithm and did not work well to handle complex cases. Later improvements were made by applying neural network machine learning [40] .
6.4. Deskilling of Anesthesiologists
The deskilling phenomenon could be seen when the regional block needle tracking system was utilized [21] . This potential dependence during the training process on AI technology may decrease overall competence among trainees despite reduction in variation [21] . Skills are maintained and improved through experience. The fear then becomes if anesthesiologists come to rely heavily on automated systems, it would lead to the decrease in skills such as intubation or titration of sedatives [33] . Anesthesiologists need to maintain their specialized skill through practice in the cases of technological failures or emergency situations. The author firmly believes the development of AI systems for use in anesthesiology therefore should be designed with the aim of assisting, not replacing anaesthesists.
Further to clinical skills, anaesthesists have a very involved role in teaching trainees through sharing of hands-on experience and valuable insights during cases which cannot be replaced by AI technology such as natural language models [20] .
6.5. AI Technology Error
As we are all familiar with, sophisticated modern computers are still subject to error or failures without warning which can result from software defects, data artifact, hardware failures or logic errors [58] . Potentially significant morbidity or mortality can occur to a patient if such errors transpire in anesthesiology to trigger inappropriate decision making or administering incorrect drugs [58] .
While error in technology remains an inevitable occurrence, there are some improvements which can decrease the rate of error and failure [107] . For this, we can learn and consider strategies from the aviation industry, one with established low failure and error rates despite the complex engineering of AI-supported aircrafts [107] . The first thing to consider is the balance between quality of build and cost-effectiveness of the systems [107] . While the author is aware a compromise will always need to be made, rationalization of the areas of cost-cutting is important as to not compromise patient safety. There should be strict and clear maintenance services if the widespread use of AI-supported systems were to become widely used with clear documentation in logbooks. The machines should be made from high quality components by specialized manufacturers. There should also be an automatic and neutral recording system of errors which cannot be tampered with similar to the black boxes (flight recorders) on aircrafts.
6.6. AI Systems Can Be Black Boxes
Due to the complexities of processing by AI algorithms applying deep learning, how the decisions are made can be difficult to understand, even by the engineers that understand the used mathematical algorithms [46] [47] [54] . Furthermore, commercialized AI systems often have proprietary algorithms which usually remain confidential, much like the method of BIS calculation [108] . This opacity means AI systems operate by “black box” functioning model which understandably creates apprehension among anesthesiologists, discouraging them to rely on AI systems in practice.
Explainability of the decisions made is vital to increase the transparency of AI systems and encourage its trust and integration in anaesthetic practice [109] . This is particularly important in cases where there is a differing opinion between anesthesiologists and AI systems as knowing how the respective decisions are made will enable comparison, rationalization and thus enactment of the optimal decision.
The author proposes, to improve the transparency, the algorithms should include algorithm explanation windows for how decisions are reached as well as for reasoning of error messages that are written in plain language. This would increase the trust anesthesiologists have for AI technology in anesthesiology and therefore increase its’ implementation and acceptance in daily practice.
6.7. Regulation
There exist many obstacles in the strict guidelines and regulations for the implementation of AI technology in clinical practice, and for good reason due to the high stakes nature of anesthesiology [56] . However, this causes a delay in the development and adoption of AI technology in hospitals around the world as only limited AI technologies have been approved for use. For example, only static or locked algorithms that return expected and predictable outcomes have been approved by the FDA [47] .
This is well illustrated in the case of the pharmacological robot Sedasys™. Its failure on the market was in part due to the highly restrictive and conservative function that was programmed by the developers in order to satisfy the criteria for FDA approval [84] . This particular robot was restricted to only being able to titrate propofol infusions down, and required an anesthesiologist to manually increase the dose if the patient’s sedation lightened [84] . Therefore, not only was this device have limited usefulness in clinical practice, it may also have decreased the interest of using AI systems in anesthesiology.
Another challenge exists in patient privacy and confidentiality when accessing patient big data for inputs in AI systems [31] . For this reason, there needs to be regulatory changes to enable assumed consent and an opt-out system to access patient data [94] .
Health regulators need to make strict laws to regulate AI application to prevent harm to patients. It is difficult for regulators to approve complex, “black box” AI-supported robots, as they need a long time to prove their safety and reliability, therefore delaying their development and implementation. The author suggests a potential solution lays in a modular design strategy where simple AI algorithms should be designed and tested first, then slow integration of more algorithm units to build a more complex system.
6.8. Lack of Knowledge and Experience of Practicing Anesthesiologists in Using AI
At present, few anesthesiologists are familiar with the way AI robots work or are involved in the development of these machines [40] . As a consequence of this opacity, there will be less confidence and acceptance in using AI robots in practice hence slowing the development and implementation of AI robots in the routine practice of anesthesiology [54] .
Furthermore, given the current general lack of knowledge and experience of AI in anesthesiology, the anesthesiologist will be met with difficult situations when AI robots make unexpected decisions, thus creating a barrier to using these machines in regular practice.
The author thinks the key to the success of AI applications in anesthesiology is to begin at the beginning. Anesthetists should be heavily involved in the design of AI-supported systems for use in anesthesiology as this enables the rationalization of important features in clinical use. Furthermore, there exists an ongoing need for further clinical studies evaluating the relevance of AI-supported systems and their effectiveness in real-world applications through rigorous outcome assessments.
All anesthesiology personnel should be trained to understand AI technology as without knowledge and experience, and the user will likely achieve poor outcomes using AI-supported systems and feel mistrust for these systems [40] . The author suggests training should include medical staff at all levels, from medical students to senior clinicians, if the introduction of these systems in anesthesiology is to succeed. To achieve this, AI training courses for medical students, junior doctors and anesthetists should be provided frequently to normalize the presence of AI in medicine.
Earlier in this article, there has been a discussion of multiple areas of anesthesiology where AI has potential applications, including pain medicine, event prediction and operating room logistics. However, most robots at later developmental stages are focused on pharmacological and mechanical robots with a very direct potential impact on patient morbidity and mortality. To improve the reception and trust of AI systems in anesthesiology, development of AI could begin with diagnostic and logistical supports first, which can improve patient outcomes without direct potential for harm.
7. Conclusion
In conclusion, artificial intelligence technologies have the potential to reduce the cognitive load on anesthesiologists to enable concentration on other tasks, which improves patient safety and outcomes. The identified limitations above should be addressed to accelerate the development of artificial intelligence for application in anesthesiology. Strict quality control of the systems and algorithms should be undertaken before application on patients, as adverse events have the potential to decrease trust and become a barrier to future endeavours. The cooperation of anesthesiologists and engineers in the development of AI is required in system design to improve usability. Lastly, the design of AI systems should use a modular design concept to simplify regulatory approval processes. If we begin at the beginning and develop simple and explainable AI algorithms, we may be able to see AI integrated into daily anaesthetic practice in the near future.