AI in Anesthesiology

REVIEW ARTICLE Deborah J. Culley, M.D.
, Editor
ABSTRACT
Artificial intelligence has been advancing in fields including anesthesiology.
This scoping review of the intersection of artificial intelligence and anesthesia
Artificial Intelligence in research identified and summarized six themes of applications of artificial
intelligence in anesthesiology: (1) depth of anesthesia monitoring, (2) control
Anesthesiology
of anesthesia, (3) event and risk prediction, (4) ultrasound guidance, (5) pain
management, and (6) operating room logistics. Based on papers identified
in the review, several topics within artificial intelligence were described and
Current Techniques, Clinical summarized: (1) machine learning (including supervised, unsupervised, and
reinforcement learning), (2) techniques in artificial intelligence (e.g., classical
Applications, and Limitations machine learning, neural networks and deep learning, Bayesian methods),
Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

and (3) major applied fields in artificial intelligence.
Daniel A. Hashimoto, M.D., M.S., Elan Witkowski, M.D., M.P.H.,
The implications of artificial intelligence for the practicing anesthesiologist are
Lei Gao, M.D., Ozanan Meireles, M.D., Guy Rosman, Ph.D.
discussed as are its limitations and the role of clinicians in further develop-
Anesthesiology 2020; 132:379–94 ing artificial intelligence for use in clinical care. Artificial intelligence has the
potential to impact the practice of anesthesiology in aspects ranging from
perioperative support to critical care delivery to outpatient pain management.
A rtificial intelligence has been defined as the study of

algorithms that give machines the ability to reason and
perform functions such as problem-solving, object and word
(ANESTHESIOLOGY 2020; 132:379–94)
recognition, inference of world states, and decision-mak-

ing.1 Although artificial intelligence is often thought of as Anesthesiology as a field is well positioned to potentially
relating exclusively to computers or robots, its roots are benefit from advances in artificial intelligence as it touches
found across multiple fields, including philosophy, psychol- on multiple elements of clinical care, including perioperative
ogy, linguistics, and statistics. Thus, artificial intelligence can and intensive care, pain management, and drug delivery and
look back to visionaries across those fields, such as Charles discovery.We conducted a scoping review of the literature at
Babbage, Alan Turing, Claude Shannon, Richard Bellman, the intersection of artificial intelligence and anesthesia with
and Marvin Minsky, who helped to provide the foundation the goal of identifying techniques from the field of artificial
for many of the modern elements of artificial intelligence.2 intelligence that are being used in anesthesia research and
Furthermore, major advances in computer science, such as their applications to the clinical practice of anesthesiology.
hardware-based improvements in processing and storage,
have enabled the base technologies required for the advent Materials and Methods
of artificial intelligence. The Medline, Embase, Web of Science, and IEEE Xplore
Artificial intelligence has been applied to various databases were searched using combinations of the follow-
aspects of medicine, ranging from largely diagnostic ing keywords: machine learning, artificial intelligence, neu-
applications in radiology3 and pathology4 to more ther- ral networks, anesthesiology, and anesthesia. To be included,
apeutic and interventional applications in cardiology5 papers had to be focused on the design or application of
and surgery.6 In April 2018, the U.S. Food and Drug artificial intelligence–based algorithms in the practice of
Administration approved the first software system that anesthesia, including preoperative, intraoperative, postop-
uses artificial intelligence—a program that assists in the erative, and surgical critical care as well as pain manage-
diagnosis of diabetic retinopathy through the analysis of ment. All English-language papers from 1946 to September
images of the fundus.7 As the development and applica- 30, 2018 were eligible. Peer-reviewed, published literature,
tion of artificial intelligence technologies in medicine including narrative review papers, were eligible for inclu-
continues to grow, it is important for clinicians in every sion as were peer-reviewed conference proceeding papers.
field to understand what these technologies are and how Editorials, letters to the editor, and abstracts were excluded
they can be leveraged to deliver safer, more efficient, more as were any studies involving animals, fewer than 10 patients,
cost-effective care. or simulated data only. Papers that involved monitoring of
This article is featured in “This Month in Anesthesiology,” page 1A. Supplemental Digital Content is available for this article. Direct URL citations appear in the printed text and are
available in both the HTML and PDF versions of this article. Links to the digital files are provided in the HTML text of this article on the Journal’s Web site (www.anesthesiology.org).
Submitted for publication November 26, 2018. Accepted for publication July 29, 2019. Published online first on September 12, 2019. From the Surgical Artificial Intelligence and
Innovation Laboratory (D.A.H., E.W., O.M., G.R.) and Department of Anesthesia, Critical Care, and Pain Medicine (L.G.), Massachusetts General Hospital, Boston, Massachusetts;
Computer Science and Artificial Intelligence Laboratory, Massachusetts Institute of Technology, Cambridge, Massachusetts (G.R.).
Copyright © 2020, the American Society of Anesthesiologists, Inc. All Rights Reserved. Anesthesiology 2020; 132:379–94. DOI: 10.1097/ALN.0000000000002960
ANESTHESIOLOGY, V 132 • NO 2 February 2020 379

Copyright © 2020, the American Society of Anesthesiologists,
<zdoi;. Inc. Unauthorized reproduction of this article is prohibited.
DOI: 10.1097/ALN.0000000000002960>
REVIEW ARTICLE
patient parameters broadly but were not specific to the review papers that covered multiple categories. Included
practice of anesthesia (e.g., general ward-based monitoring articles and their categorization are listed in Supplemental
and alarms, sleep study analysis) were excluded. Digital Content (http://links.lww.com/ALN/C51). From
Four reviewers screened articles for inclusion or exclu- these categories, we identified the major topics of artificial
sion using Covidence (Melbourne, Australia). Each article intelligence that were captured in this search.
was screened by two independent reviewers. A third reviewer The predominant focus across most of these studies
would mediate any disagreement between two screeners. has been to investigate potential ways that artificial intelli-
Reference lists of included papers were hand-searched by gence can benefit the clinical practice of anesthesiology not
one reviewer and included if the inclusion criteria were met. through the replacement of the clinician but through aug-
Emphasis was placed on extracting themes relating to mentation of the anesthesiologist’s workflow, decision-mak-
applications of artificial intelligence. Although specifics ing, and other elements of clinical care. Thus, although
regarding the numerous algorithms that can be used in arti- artificial intelligence is an expansive field, the results of

ficial intelligence studies and applications were outside the this review demonstrated the literature’s focus on machine
scope of this review, common techniques uncovered in the learning and its applications to clinical care.To structure the
literature were reviewed as examples. findings of the scoping review, we have presented below
the following: machine learning and the types of learning
Results algorithms, major topics in artificial intelligence, artificial
A total of 2,171 titles were identified from the database intelligence techniques, and the six aforementioned catego-
search, and 1,368 abstracts were eligible for screening. Of ries of applications of artificial intelligence to anesthesiol-
these, 394 full-text studies were assessed for inclusion, and ogy that were found in the literature.
173 manuscripts were included in the final analysis (fig. 1).
Based on the 173 included full-text articles, we identified Machine Learning and Learning Algorithms in Artificial
the following categories of studies: (1) depth of anesthesia Intelligence
monitoring, (2) control of anesthesia, (3) event and risk pre- Although different taxonomies of artificial intelligence have
diction, (4) ultrasound guidance, (5) pain management, and been previously described, common to many of these is
(6) operating room logistics. Seven articles (n = 7) were broad the categorization of machine learning as one of the major
Fig. 1. Preferred Reporting Items for Systematic Reviews and Meta-Analyses diagram of screening and evaluation process.
380 Anesthesiology 2020; 132:379–94 Hashimoto et al.

Copyright © 2020, the American Society of Anesthesiologists, Inc. Unauthorized reproduction of this article is prohibited.
Artificial Intelligence in Anesthesiology
subfields of artificial intelligence.Traditional computer pro- automatically classified for hypothesis generation for future
grams are programmed with explicit instructions to elicit research.11 Similarly, unsupervised learning can be used
certain behaviors from a machine based on specific inputs to identify patients who could most benefit from certain
(e.g., the primary function of a word processing program drugs, such as asthmatics who would benefit most from glu-
is to display the text input by the user). Machine learn- cocorticoid therapy, based on genomic analysis.12
ing, on the other hand, allows for programs to learn from Reinforcement Learning refers to the process by which
and react to data without explicit programming. Data that an algorithm is asked to attempt a certain task (e.g., deliver
can be analyzed through machine learning are broad and inhalational anesthesia to a patient, drive a car) and to learn
include, but are not limited to, numerical data, images, text, from its subsequent mistakes and successes.13 A biologic
and speech or sound. A common manner with which to analogy to reinforcement learning is operant conditioning,
conceptualize machine learning is to consider the type where the classical example is of a rat taught to push a lever
of learning algorithm used to solve a problem: supervised through the use of food-based reward. However, reinforce-

learning, unsupervised learning, and reinforcement learn- ment learning problems today are more sophisticated. For
ing.8 Each of these types of learning algorithms has a range example, Padmanabhan et al.14 used reinforcement learning
of techniques that can be applied and will be covered in the to develop an anesthesia controller that used feedback from
next section. a patient’s bispectral index (BIS) and MAP to control the
Supervised learning is a task-driven process by which an infusion rates of Propofol (in a simulated patient model).
algorithm(s) is trained to predict a prespecified output, such In this scenario, achieving BIS and MAP values within a
as identifying a stop sign or recognizing a cat in a photo- set range result in a reward to the algorithm whereas values
graph. Supervised learning requires both a training dataset outside the range result in errors that prompt the algorithm
and a test dataset. The training dataset allows the machine to perform further fine-tuning.
to analyze and learn the associations between an input and Machine learning problems are often divided into
desired output while the test dataset allows for the assess- those that require classification (dividing data into dis-
ment of the performance of the algorithm on new data. In crete groups) and those that require regression (model-
many studies, one large dataset is subdivided into a training ing data to better understand the relationship between
set and a test set (often 70% of the data for training, 30% two or more continuous variables with the potential for
for testing). prediction). A frequent example of classification is image
For example, Kendale et al.9 conducted a supervised recognition (e.g., recognizing a cat vs. a dog), whereas a
learning study on electronic health record data with the frequent example of regression is prediction (e.g., predict
goal of identifying patients who experienced postinduction house prices from preexisting data on square footage).15
hypotension (mean arterial pressure [MAP] less than 55 Supervised, unsupervised, and reinforcement learning
mmHg). The training dataset included 70% of the patients approaches can each be used to tackle problems of clas-
and variables such as American Society of Anesthesiologists sification and regression depending on the nature of the
(ASA) physical status, age, body mass index, comorbidities, question and the type of data available.
and medications as well as the blood pressures of the patients.
The various algorithms used by Kendale et al.9 could then Techniques in Artificial Intelligence
analyze the training dataset to learn which variables were There are various techniques and models within each of
predictive of postinduction hypotension. The test dataset the three approaches to machine learning described above.
was then analyzed to assess how accurately the algorithm Although a detailed description of the specific methods
could predict postinduction hypotension in the remaining and algorithms used in machine learning are outside the
30% of patients.9 Some studies use external validation (i.e., scope of this review, it can be useful to have an introductory
the use of a separate dataset) to assess the generalizability of familiarity with basic concepts of the more popular tech-
the algorithm to other data sources.10 A study using elec- niques used in artificial intelligence research.
tronic health record data from one hospital in their training
and testing datasets may then incorporate electronic health Fuzzy Logic – A Historical Perspective. Fuzzy set theory and
record data from a separate hospital to further validate the fuzzy logic were first described in 1965.16 Although fuzzy
performance of their algorithm on a different data source, logic in and of itself is not necessarily artificial intelligence,
demonstrating generalizability to other populations. it has been used within other frameworks to facilitate
Unsupervised learning refers to algorithms identifying artificial intelligence–based functions. Standard logic allows
patterns or structure within a dataset. This can be useful only for the concepts of true (a numerical value of 1.0) and
to find novel ways of classifying patients, drugs, or other false (a numerical value of 0.0), but fuzzy logic allows for
groups. Bisgin et al.11 used unsupervised learning techniques partial truth (i.e., a numerical value between 0.0 and 1.0). A
to mine data from the Food and Drug Administration drug comparison may be made to probability theory where the
labels to identify the major topics (e.g., specific adverse probability of a statement being true is evaluated (e.g., “a
events, therapeutic application) into which drugs could be pancreaticoduodenectomy will be scheduled tomorrow”)
Hashimoto et al. Anesthesiology 2020; 132:379–94 381

REVIEW ARTICLE
versus the degree to which a statement is true (e.g., “there

is an 80% chance of a pancreaticoduodenectomy being
scheduled tomorrow”). It is meant to resemble human
methods of decision making with vague or imprecise
information.
Fuzzy logic often uses rule-based systems (e.g., if…then
systems) that have been largely used in control systems
where precise mathematical functions do not accurately
model phenomena. For example, an anesthesia monitor
to detect hypovolemia was developed using fuzzy logic to
approximate the presence of mild, moderate, and severe
hypovolemia based on normalized values of heart rate

(HR), blood pressure, and pulse volume that were divided
into categories of mild, moderate, and severe. The monitor
used rules built on fuzzy logic such as “If (electrocardio-
Fig. 2. An illustrative example of a decision node. Several ter-
gram-HR is mild) and (blood pressure is mild) and (pulse minologies can be used to describe decision trees. The root
volume is severe), then (hypovolemia is moderate).”17 node is the start of the tree, and branches connect nodes. A
The development of such rules requires expert human child node is any node that has been split from a previous node,
input to determine an appropriate rule set for the machine whereas a decision node is any node that allows two or more
to follow, and early studies in fuzzy logic and other adap- options to follow it. A chance node is any node that may repre-
tive control mechanisms help set the foundation for explor- sent uncertainty.
ing more modern approaches to imprecise information or
incomplete data. More recent research in the field has used
artificial intelligence approaches to better evaluate and use geometrically and then determine whether additional input
data to trigger the rule functions of fuzzy systems. Thus, data falls into a category based on the training examples
although research in fuzzy logic systems is ongoing (espe- plotted nearest to it (based on Euclidian distance). Based
cially in control systems for applications such as medica- on the specific approach used, this may be based on a single
tion delivery), progress in artificial intelligence research has nearest point (1-nearest neighbor) or the weight of a group
focused on using more data-driven techniques in machine of points (k-nearest neighbor).19 Support vector machines
learning to achieve goals first explored by researchers work- are another type of supervised learning algorithm that can
ing with fuzzy logic. also be used for both classification and regression. Support
vector machines map training data in space and then use
Classical Machine Learning. Machine learning uses features, hyperplanes to optimally separate the data into representa-
or properties within the data, to perform its tasks. To tive categories or clusters (fig. 3). New data are then clas-
analogize to an example in statistical analysis, features sified based on their location in the space relative to the
would be analogous to independent variables in a logistic hyperplane.20
regression. In classical machine learning, the features are
selected (often referred to as hand-crafted) by experts to help Neural Networks and Deep Learning. One of the most popular
guide the algorithms in the analysis of complex data. methods today for performing work in machine learning is
Decision tree learning is a type of supervised learning the use of neural networks. Neural networks are inspired
algorithm that can be used to perform either classification by biologic nervous systems and process signals in layers
(classification trees) or regression tasks (regression trees). As of computational units (neurons).21 Each network is made
its name implies, this set of techniques uses flowchart-like up of an input layer of neurons comprised of features that
tree models with multiple branch points to determine a tar- describe the data, at least one hidden layer of neurons
get value or classification from an input. Each node within that conducts different mathematical transformations on
a tree has an assigned value, and the final node represents the input features, and an output layer that yields a result
an endpoint and the cumulative probability of arriving at (fig. 4). Between each layer are multiple connections
that point based on the preceding decisions (fig. 2). Hu et between neurons that are parameterized to different
al.18 used decision trees to predict total patient-controlled weights depending on the input-output maps. Thus, neural
analgesia (PCA) consumption from features such as patient networks are a framework within which different machine
demographics, vital signs and aspects of their medical history, learning algorithms can work to achieve a particular task
surgery type, and PCA doses delivered with the promise of (e.g., image recognition, data classification).
using such approaches to optimize PCA dosing regimens.18 Modern neural network architecture has expanded to
The k-nearest neighbor algorithms are a group of allow for deep learning, neural networks that use many
supervised learning algorithms that assesses training data layers to learn more complex patterns than those that

(e.g., speech).22 Unlike in classical machine learning where

features are hand-crafted, deep learning self-learns features
based on the data itself. It analyzes all available features within
the training set to determine which features allow for the
optimal achievement of the deep neural network’s given task
(e.g., object recognition from an image). Thus, deep learning
is potentially a powerful tool with which to analyze very
large datasets where hand-crafted feature selection may be
inadequate and/or inefficient in achieving an effective result.
Given their flexibility in analyzing various types of
data, neural networks are a technique that have now been
applied to other subfields of artificial intelligence, including

natural language processing and computer vision. Within
the field of anesthesia, multiple examples of the applica-
Fig. 3. An illustrative example of support vector machines. tions of neural networks exist, including depth of anesthesia
The goal of the support vector machines algorithm is to find the monitoring and control of anesthesia delivery as described
hyperplane that maximizes the separation of features. The solid further below.
black line represents the optimal hyperplane, whereas the dotted
lines represent the planes running through the support vectors.
Bayesian Methods. Bayes’ theorem provides a description of
The empty circle and the solid triangle represent support vec-
tors—the data points from each cluster that represent the clos- probability of an event based on previous knowledge or
est points to the optimal hyperplane. The dashed line represents data about factors that may influence that event. In many
the maximum margin between the support vectors. studies in the medical literature, a frequentist approach to
statistics is applied, wherein hypothesis testing occurs based
on the frequency of events that occur in a given sample
of data as a representation of the study population of
interest. A Bayesian approach differs by using the known
previous probability distribution of an event along with the
probability distribution represented in a given dataset.23
Bayes’ theorem provides the foundation for many tech-
niques in artificial intelligence as it allows for both mod-
eling of uncertainty and updating or learning repeatedly
as new data is made available.24 Bayesian techniques have
been incorporated into many common tasks such as spam
filtering, financial modeling, and evaluation of clinical
tests.25,26 Although delving into the specific methodologies
of Bayesian methods is outside the scope of this review,
Bayesian methods that are being used with increased fre-
quency in the medical literature include Bayesian networks,
hidden Markov models, and Kalman filters.27
Applied Fields in Artificial Intelligence

Fig. 4. An illustrative example of a three-layer neural network. Two other topics identified by the search as relevant to anes-
The input layer provides features such as electroencephalogram thesiology were natural language processing and computer
(EEG) power and entropy, the patient’s mean arterial pressure
vision, and these topics are both applications of and subfields
(MAP), and the patient’s heart rate variability (HRV) to the net-
work. A hidden layer transforms inputs into features usable by within artificial intelligence that use machine learning.
the network. The output layer transforms the hidden layer’s
activations into an interpretable output (e.g., patient awake vs. Natural Language Processing. Natural language processing is
asleep). a subfield of artificial intelligence that focuses on machine
understanding of human language. Before the advent of
natural language processing, computers were limited to
are discernible from simple one- or two-layer networks. reading machine languages or code (e.g., C++, JAVA, Visual
Subtypes of deep learning networks that one may encoun- Basic). Instructions programed in code are compiled by a
ter are convolutional neural networks, which can process computer to process a set of instructions to yield a desired
data composed of multiple arrays, and recurrent neural net- output.With natural language processing, machines can strive
works, which are better designed to analyze sequential data toward the understanding of language that is used naturally by

REVIEW ARTICLE
humans. Natural language processing, however, is not simply anesthetic depth. The attention to these two modalities of
recognizing letters that construct a word and then matching measurement was not surprising given research efforts in
them to a definition. It strives to achieve understanding of reducing the risk of intraoperative awareness and previous
syntax and semantics to approximate meaning from phrases, literature suggesting that low BIS and burst-suppression on
sentences, or paragraphs.28 electroencephalography during anesthesia may be associated
In medicine, the most natural application of natural with poorer outcomes.37,38 In addition, careful monitoring
language processing is to automated analysis of electronic of MAP has also been noted in the literature, likely due to
health record data. Although the move to electronic health the association of low MAP with postoperative mortality.39
records has shifted a considerable amount of documenta- Machine learning approaches are well-suited to analyze
tion to checkboxes, dropdown menus, and prepopulated complex data streams such as electroencephalographies;
fields, free text entry remains a critical component of cli- thus, a range of electroencephalography-based signals was
nician documentation, allowing clinicians to communicate found to have been investigated to measure depth of anes-

with one another and to document not just important thesia. Literature from the early 1990s described the use of
findings but also the thought process behind medical deci- neural networks to evaluate electroencephalography power
sions. Natural language processing is currently being used spectra as a signal in discriminating awake versus anesthe-
to help extract information from free text fields to build tized patients and identified the potential of specific fre-
more structured databases that can be further analyzed to quency bands in assessing the effects of certain drugs.40,41 As
identify surgical candidates, assess for adverse events, or to use of index parameters of depth of anesthesia such as BIS
facilitate billing.29–31 increased in popularity, neural networks and other machine
learning approaches were used to analyze electroencepha-
Computer Vision. With the growing attention on autonomous lography data with the goal of approximating BIS through
vehicles, computer vision has become one of the most widely other electroencephalography parameters.42,43
recognized subfields of artificial intelligence. Computer More recent papers have used artificial intelligence
vision refers to machine understanding of images, videos, and techniques and spectral analysis to more directly analyze
other visual data (e.g., computed tomography). Automated electroencephalography signals to estimate the depth of
acquisition, processing, and analysis of images are part of anesthesia. Mirsadeghi et al.44 studied 25 patients and com-
the computer vision process that leads to understanding pared the accuracy of their machine learning method of
of a scene. Color, shape, texture, contour, and focus are analyzing direct features from electroencephalography sig-
just some of the different elements that can be detected nals (e.g., power in different bands [delta, theta, alpha, beta
and analyzed by computer vision systems. While object and gamma], total power, spindle score, entropy, etc.) in
detection and recognition (e.g., identifying another car on identifying awake versus anesthetized patients against the
the road) have been popularized in the lay press, computer BIS index. Their accuracy in using electroencephalography
vision encompasses the field of work that includes research features was 88.4% while BIS index accuracy was 84.2%.44
on describing the visual world in numerical or symbolic Similarly, Shalbaf et al.45 used multiple features from elec-
form to allow for interpretation of images for subsequent troencephalographies to classify awake versus anesthetized
action (e.g., determining that the car on the road ahead of patients (as four possible states of awake, light, general, or
your will be slowing down at the stop sign, prompting the deep anesthesia) during sevoflurane with 92.91% accuracy
need to slow your vehicle as well). compared with the response entropy index which had an
In medicine, the application of computer vision to accuracy of 77.5%. This same algorithm demonstrated with
pathology and radiology have led to systems capable of generalization to Propofol and volatile anesthesia patients
assisting clinicians in reducing error rates in diagnosis by with 93% accuracy versus the BIS index’s 87% accuracy.45
identifying areas on slides and x-rays that have a high prob- These papers highlight the power of artificial intelligence
ability of demonstrating pathology.32,33 Furthermore, com- techniques in creating models that can efficiently consider
puter vision has been used to automatically identify and linear and non-linear data together to generate predictions
segment steps of laparoscopic surgery, suggesting that con- that maximize the utility of each available variable.
text awareness is possible with computer vision systems.34 Whereas BIS and electroencephalography were the sub-
In anesthesiology, computer vision has largely been applied ject of the majority of the papers identified, other clinical
to the automated analysis of ultrasound images to assist with signals have been investigated as well. Zhang et al.46 recorded
identification of structures during procedures.35,36 mid-latency auditory evoked potentials from patients and
used neural networks to assess the accuracy of these sig-
Applications of Artificial Intelligence in Anesthesiology nals in determining when patients were awake (96.8%
Depth of Anesthesia Monitoring. Use of artificial intelligence accuracy), receiving adequate anesthesia (86% accuracy),
to improve depth of anesthesia monitoring was identified in and emerging from general anesthesia (86.6% accuracy).
42 papers.The majority of these papers focused on use of the In addition, clinical variables such as heart rate variability
BIS (Medtronic, USA) or electroencephalography to assess have been investigated to approximate sedation level as

typically measured by the Richmond Agitation Sedation of stay, awareness, etc.), and these papers were categorized
Scale.46 Related to depth of anesthesia monitoring, Ranta into those pertaining to operative (immediate preoperative
et al.47 conducted a database study of 543 patients who had assessment; n = 26 papers), postoperative (n = 14 papers), or
undergone general anesthesia with 6% reporting intraop- critical care-related events (n = 13 papers).
erative awareness. Neural networks analyzed clinical data For perioperative care risk prediction, various techniques
from these patients such as blood pressure, heart rate, and in machine learning, neural networks, and fuzzy logic have
end-tidal carbon dioxide but did not use any electroen- all been applied. For example, neural networks were used to
cephalography data. The prediction probabilities for these predict the hypnotic effect (as measured by BIS) of an induc-
networks in predicting awareness was, at best, 66% though tion bolus dose of propofol (sensitivity of 82.35%, specificity
it had a high specificity of 98%.47 As described above about of 64.38% and an area under the curve of 0.755) and was
neural networks and deep learning, the strength of such found to exceed the average estimate of practicing anesthe-
approaches over classical regression techniques is a neural siologists (sensitivity: 20.64%, specificity 92.51%, area under

network’s ability to engage in self-learning of the features in the curve of 0.5605).61 At the tail end of an operation, Nunes
the dataset, selecting features that will best predict a target et al. (2005)62 compared neural networks and fuzzy models
value (e.g., awareness) rather being fed the features thought in their prediction of return of consciousness after general
to be most predictive by a human expert. anesthesia (propofol + remifentanil) in a small sample of 20
patients, finding mixed results for all models. Neural networks
Control of Anesthesia Delivery. Thirty-three (n = 33) papers have also been used to predict the rate of recovery from neu-
were identified that pertained to control systems for romuscular blockade63 and hypotensive episodes postinduc-
delivery of anesthesia, neuromuscular blockade, or other tion64 or during spinal anesthesia,65 while other machine
related drugs. Control systems in anesthesia were reviewed learning approaches have been tested to automatically classify
in detail by Dumont and Ansermino48 with explanations pre-operative patient acuity (i.e., ASA status),66 define diffi-
of feedforward and feedback systems as well as multiple cult laryngoscopy findings,67 identify respiratory depression
examples of different closed-loop systems.48 As automated during conscious sedation,68 and to assist in decision-making
delivery of anesthesia also requires a determination of the for the optimal method of anesthesia in pediatric surgery.69
depth of anesthesia by a machine, approaches to control With regard to specific event detection in both the
require the measurement of clinical signs or surrogate operating room and the intensive care unit (ICU), Hatib
markers of anesthetic depth. Thus, the evolution of control et al.70 used two large databases (total of 1,334 patients)
system research in anesthesia is evident in the various targets of waveform data from arterial lines to develop and test
used to approximate depth of anesthesia. a logistic regression model that could predict hypotension
In the 1990s, clinical signs and measurements such as (i.e., mean arterial pressure less than 65 mmHg sustained
blood pressure were used as signals to the control systems for 1 min) up to 15 min before its occurrence on an arte-
to determine how to regulate delivery of anesthetics.49,50 rial line waveform.70 Other ICU database studies have used
As new metrics to measure depth of anesthesia were devel- machine learning models to predict morbidity,71,72 weaning
oped, targets shifted to measuring values such as BIS. Early from ventilation,73 clinical deterioration,74,75 mortality,76 or
work used purely empirical approaches to tuning delivery readmission77 and to detect sepsis.78
of anesthesia to achieve a target BIS.51 As use of BIS became Machine learning approaches to critical care have not
more widespread, researchers began to use more sophisti- been limited to large database studies only. In a single-cen-
cated fuzzy logic systems or reinforcement learning to ter randomized control trial comparing a machine learn-
achieve anesthetic control using BIS as a target measure.52–54 ing alert system (using six vital sign parameters as features)
Our search did not return papers on the use of direct elec- versus an electronic health record–based alert system that
troencephalography measures for anesthesia control. used other criteria for the prediction of sepsis, the machine
Control systems that use machine learning have also been learning alert system outperformed Systemic Inflammatory
used to automate the delivery of neuromuscular block- Response Syndrome criteria, Sequential Organ Failure
ade,55,56 and these systems have also incorporated forecasting Assessment score, and quick Sequential Organ Failure
of drug pharmacokinetics to further improve the control of Assessment score in the detection of sepsis. Its use resulted
infusions of paralytics.57 Three papers were identified that in a 20.6% decrease in average hospital length of stay and,
described the use of artificial intelligence to achieve control more importantly, a 58% reduction in in-hospital mortality.79
of mechanical ventilation58,59 or to automate weaning from
mechanical ventilation.60 Ultrasound Guidance. Eleven (n = 11) papers were identified
that described the use of artificial intelligence techniques
Event Prediction. Our search identified 53 papers that could to assist in the performance of ultrasound-based procedures,
loosely be classified as pertaining to event prediction. We and neural networks were the most commonly employed
refer to event prediction as any studies that engage in method of achieving ultrasound image classification.
prediction of an effect or event (e.g., complication, length Smistad et al.36 used ultrasound images of the groin from 15

REVIEW ARTICLE
patients to train a convolutional neural network to identify to operating room logistics, such as scheduling of
the femoral artery or vein while distinguishing it from other operating room time or tracking movements and actions
potentially similar appearing ultrasound images such as of anesthesiologists. Combes et al.86 used a hospital database
muscle, bone, or even acoustic shadow. Closer investigation containing extensive information on staffing, operating
of the network found that it would analyze horizontal edges room use per procedure and staff, and post anesthesia care
in the ultrasound with greater priority than vertical edges to unit use with the electronic health record to train a neural
identify vessels with an average accuracy of 94.5% ± 2.9%.36 network to predict the duration of an operation based on
In addition to specific structure detection in ultrasound the team, type of operation and a patient’s relevant medical
images, researchers have also used neural networks to assist history; however, prediction accuracy of their models never
in the identification of vertebral level and other anatomi- exceeded 60%.86 In a different example, fuzzy logic and
cal landmarks for epidural placement. Pesteie et al.35 used neural networks were used to optimize bed use for patients
convolutional neural networks to automate identification undergoing ophthalmologic surgery by modeling the type

of the anterior base of the vertebral lamina,35 whereas of case, modeling surgeon experience, staff experience, type
Hetherington et al.80 used convolutional neural networks to of anesthesia and the experience of the anesthesiologist,
automatically identify the sacrum and the L1–L5 vertebrae patient factors, and comorbidities with error rates ranging
and vertebral spaces from ultrasound images in real time from 14% to 19% depending on the type of case.87 In a
with up to 95% accuracy.80 different application of machine learning, Houliston et al.88
analyzed radio frequency identification tags to determine
Pain Management. Nine (n = 9) papers were identified the location, orientation, and stance of anesthesiologists in
that related to the field of pain management, ranging from the operating room. Though their analyses were limited to
prediction of opioid dosing to identification of patients simulated operating rooms with mannequins, the authors
who may benefit from preoperative consultation with a proposed the use of similar tracking applications with real
hospital’s acute pain service.81,82 Brown et al.83 used machine patients to better understand potential impacts on patient
learning to analyze differences in functional magnetic safety based on the interaction of anesthesiologists with the
resonance imaging data collected from human volunteers various equipment in the room.88
who were exposed to painful and nonpainful thermal
stimuli, demonstrating that machine learning analysis of Discussion
whole brain scans could more accurately identify pain than This scoping review identified six main clinical applica-
analysis of individual brain regions traditionally associated tions of artificial intelligence research in anesthesiology: (1)
with nociception.83 depth of anesthesia monitoring, (2) control of anesthesia, (3)
Identification of pain was not limited to imaging tech- event and risk prediction, (4) ultrasound guidance, (5) pain
niques, Ben-Israel et al.84 developed a nociception level management, and (6) operating room logistics. From these
index that was based on machine learning analysis of applications, a summary of the most commonly encoun-
photoplethysmograms and skin conductance waveforms tered artificial intelligence sub-fields (e.g., machine learning,
recorded from 25 patients undergoing elective surgery. computer vision, natural language processing) and methods
However, the nociception level index was based on a target (classical machine learning, neural networks, fuzzy logic)
described as the combined index of stimulus and analge- was presented. Most applications of artificial intelligence for
sia that was defined and validated within the same study, anesthesiology are still in research and development; thus,
including an arbitrary ranking of intraoperative noxious the current focus of artificial intelligence within anesthesi-
stimuli.84 In an attempt to identify more objective bio- ology is not on replacing clinician judgment or skills but on
markers for pain, Gram et al.85 used machine learning to investigating ways to augmenting them.
analyze electroencephalography signals from 81 patients in Some of the reviewed studies demonstrated improve-
an attempt to predict patients who would respond to opioid ments over existing methods and technologies in measur-
therapy for acute pain; the results demonstrated only 65% ing clinical endpoints such as depth of anesthesia, whereas
accuracy of preoperative electroencephalography assess- others showed no difference or even worse performance
ment of patients who would respond to postoperative opi- of artificial intelligence compared with currently available
oid therapy.85 Furthermore, Olesen et al.81 conducted a big and widely used techniques. This dichotomy in the perfor-
data study searching for single nucleotide polymorphisms mance of artificial intelligence is not surprising. Although
in 1,237 cancer patients that could predict opioid dosing artificial intelligence as a field has been in existence for over
for these patients; however, their study did not discover any 50 years, the recent resurgence of and increased attention
single nucleotide polymorphism associations with opioid on artificial intelligence as a potential enabling technology
dosing in this population.81 across industries has only occurred in the last 10 to 15 yr
or so. The rapid advances that have been occurring in deep
Operating Room Logistics. Three (n = 3) papers described learning and reinforcement learning over this time frame
the use of artificial intelligence to analyze factors relating have been attributed to a “big bang” of three factors: (1) the

availability of large datasets, (2) the advancement of hard- As described in the Results section, one strength of
ware to perform large, parallel processing tasks (e.g., use of machine learning is its ability to learn from data; and this
graphical processing units for machine learning), and (3) a learning can happen continuously as more data becomes
new wave of development for artificial intelligence archi- available. This provides a distinct advantage over static algo-
tectures and algorithms. As a result, the lay press has written rithmic approaches (e.g., many of the existing risk predic-
of examples of impressive feats of artificial intelligence in tion models that have relied on a regression model run
fields such as autonomous driving, board games (e.g., chess, once) that analyze data once to determine how variables
Go), and complex strategy-based computer games (e.g., might predict outcomes or other clinical factors. A learn-
Starcraft). However, these applications have the advantage ing algorithm can be used statically or continuously as data
of being constrained environments with very specifically are updated, and the choice of the approach often depends
defined rules (as in gaming) or have had catastrophic fail- on its intended use. As of April 2019, the Food and Drug
ures that have gone along with their impressive successes Administration had approved only static or locked algo-

(as exemplified by fatalities involving autonomous vehicles). rithms for medical use (e.g., detection of retinal pathology)
Deep Blue’s victory over chess grandmaster Garry as these algorithms can be expected to return predictable
Kasparov in 1998 stands as a famous example of a machine results for a given set of inputs. However, the Food and
outperforming a human at a complex task, and even Drug Administration is considering policies for the intro-
non–artificial intelligence–based computer programs have duction of adaptive or continuously learning algorithms
continued to demonstrate their superiority over humans based on pre-market review and clear (as yet to be deter-
in playing chess. These early programs were developed mined) guidelines around data transparency and real-world
through simple optimization and search techniques or clas- performance monitoring of these algorithms.91 It is unclear
sical machine learning approaches that required input from which field of medicine will see the first adaptive learning
expert players to help the machine develop effective strate- algorithm approved for clinical use, but a diagnostic rather
gies to select the best moves in a given game scenario. More than therapeutic application may be a more likely candidate.
recently, AlphaZero is an example of artificial intelligence This review was conducted as a focused scoping
that uses deep neural networks and reinforcement learn- review intended to assess the breadth of artificial intelli-
ing to achieve its success as it learned from each play of gence research that has been conducted in anesthesiology
the game to improve its strategy. Provided just the rules of rather than a systematic review designed to answer specific
chess, Go, and shogi, and a few hours of self-play to learn questions on the effectiveness or utility of artificial intel-
strategy, AlphaZero played at a superhuman level, beating ligence–based technologies in clinical practice. Although
the world experts and the other top computer programs in additional papers that may be relevant to artificial intelli-
each game.89 gence and anesthesiology may not have been captured with
We do not anticipate that artificial intelligence will our search parameters, the breadth demonstrated in this
master all aspects clinical anesthesiology in a matter of review led us to identify areas where artificial intelligence
hours (or even years) as it has done with board games. will have implications for practicing clinicians, how clini-
Although games such as Go and shogi are complex, they cians can influence the future of artificial intelligence adop-
are defined by well-understood rules that can be navi- tion, and the limitations of artificial intelligence.
gated efficiently by artificial intelligence algorithms even
when the possible set of moves is incredibly large (e.g., the Implications of Artificial Intelligence for Practicing
estimated number of possible Go games is a googolplex Clinicians
100
or 10(10 ).90 Clinical medicine, on the other hand, has sig- The practice of modern anesthesiology requires the anes-
nificant uncertainty; for much of the clinical data that we thesiologist to gather, analyze, and interpret multiple data
interpret during daily practice serve as surrogates for an streams for each patient. As the healthcare system has
underlying physiologic or pathophysiologic process rather increasingly moved from analog to digital data, practic-
than a direct measure of the process itself. Reinforcement ing clinicians have been asked to rely on ever-expanding,
learning and other approaches in artificial intelligence are data-intensive workflows to accomplish their daily tasks.
thus left to tackle much more specific problems where the The electronic health record and the anesthesia infor-
rules of clinical medicine are better understood, resulting mation management system are just two of the interfaces
in narrow artificial intelligence that seems constrained to presented to clinicians.92 Fortunately, information manage-
accomplish one task at a time. However, feats such as arti- ment systems that allow for automated extraction of clinical
ficial intelligence’s conquest of complex games and recent variables (e.g., vital signs, drug delivery timestamps, etc.) have
advances in artificial intelligence developed from large, eased the burden of documentation on the anesthesiologist.
well-curated clinical datasets70 raise hope that as we build At the same time, the clinician must now consider how best
better, more complete datasets, we could see more sophis- to interpret the increasing amount of available data for the
ticated and clinically meaningful applications of artificial delivery of anesthetic and critical care. The application of
intelligence in medicine. artificial intelligence technologies should now emphasize

REVIEW ARTICLE
aiding the clinician in maximizing the clinical utility of the accuracy of artificial intelligence’s decisions and predictions
data that is now captured electronically. are based on accepted standards to which we compare the
Intraoperative and ICU monitoring of patients under artificial intelligence. Some accepted standards are objec-
anesthesia has relied on the experience of anesthesiolo- tive and immutable: prediction of mortality can be validated
gists to titrate anesthetics, neuromuscular blockade, and against data that has recorded patient deaths. However,
cardiovascular medications to safely maintain sedation and other accepted standards are subject to interpretation. In
physiologic support. As medical technology has advanced, the studies we reviewed on ultrasound guidance, training of
anesthesiologists are now expected to weigh and consider the selected machine learning method for identification of
multiple sources of data to safely manage a patient’s anes- structures, landmarks, etc., was dependent on human label-
thesia. Although commercially available devices such as the ing of the target in the training set (i.e., supervised learning).
BIS and SEDLINE (Masimo, USA) monitors have offered Thus, assessment of the accuracy of the artificial intelligence
the promise of simplifying the assessment of hypnosis, these method was also based on comparison of the machine label

devices have limited algorithms that can be unreliable at to that of the human, underscoring the importance of hav-
extremes of age,93 are dependent on the type of anesthetic ing reliable, consistent human-generated labels.
agent(s) used,94 and are susceptible to interference (from Identification of structures in ultrasounds can vary
motion, electrocautery, etc.). On the other hand, anal- between novice and experienced sonographers. Thus, for
ysis of electroencephalography data can be complex and a machine learning study or device development, it would
time-consuming for anesthesiologists who are having to be important to ensure that experienced sonographers
analyze multiple data streams. were labeling training data and were the standard against
Shortcomings and barriers to electroencephalography which we compare a machine’s performance. Similarly, as
usage during anesthesia could be mitigated by artificial we demonstrated in our review, much of the work on mon-
intelligence as it excels in analyzing complex, large data sets, itoring depth of anesthesia in the 1990s and early 2000s was
and our review has revealed multiple ongoing efforts to focused on the use of BIS as the target to assess awake versus
identify prediction features for depth of anesthesia or con- anesthetized states in patients. BIS provided a convenient
sciousness states from electroencephalography signals and accepted standard target for artificial intelligence as a single
physiologic parameters.95 Furthermore, because intraopera- numerical value could be used to approximate an appro-
tive awareness is a relatively rare event, big data approaches priate depth of anesthesia; however, as recent clinical litera-
could be leveraged to overcome issues of data scarcity sur- ture has called into question the utility of BIS in measuring
rounding awareness. This further highlights the importance depth of anesthesia, clinicians and researchers will have to
of building and curating robust datasets that can provide an collaborate to identify safe, valid target markers of hypnosis
appropriate source of training material for the development on which to train artificial intelligences.97,98 The continued
of future artificial intelligence systems. development of technologies that use artificial intelligence,
Examples such as the Philips eICU database (The such as control systems for the delivery of medications such
Netherlands) and the Multiparameter Intelligent as anesthetics and paralytics, will be dependent on the iden-
Monitoring in Intensive Care database demonstrate how tification of—and agreement on—accepted standard targets
current monitors and electronic health records or informa- on which an artificial intelligence can tune its performance
tion management systems can be leveraged to help auto- in achieving a given task.
mate the data collection process during episodes of care,96 Big data approaches to personalized medicine could play
and datasets that combine traditional healthcare data (e.g., a large role in the delivery of anesthetic care in the future.
claims data, vital signs, laboratory values) with specific data Though a search of single nucleotide polymorphisms asso-
such as images or video could further unlock the potential ciated with opioid dosing was not fruitful in a study on
of artificial intelligence algorithms that excel at analyzing pain management,81 artificial intelligence has been used to
and integrating vast amounts of data.6 With enough data, great effect in other fields to identify and further investigate
tasks such as real-time event prediction, automated adjust- potential genetic markers of response to specific therapies.12
ment of target-controlled infusions, and computer-assisted The scope of this review may not have captured current
or even robotically autonomous ultrasound-guided proce- or ongoing work on genome wide association studies that
dures could become real possibilities. However, accurate could uncover markers for optimal drug selection, dosing,
assessment of phenomena such as intraoperative awareness or adverse reactions.
will be reliant on agreed upon standards of measurement as At this point in time, it is hard to predict the full potential
well as good annotation of data. of artificial intelligence applications as we continue to make
Good annotation of data is critical for the success of significant strides in hardware and algorithm design as well
artificial intelligence in medicine. As artificial intelligence as database creation, curation, and management—advances
research attempts to approximate human performance in that will undoubtedly catalyze even further advances in
diagnostics and therapeutics or surpass human abilities artificial intelligence. Before pulse oximetry, trainees were
in prediction, we must remember that assessment of the taught to recognize cyanosis and other signs of hypoxia;

before automated sphygmomanometry, adept tactile esti- node could be reviewed and assessed while deep learn-
mation of blood pressure from palpation of the pulse was ing is currently assessed through inductive means. That is,
a skill sought after by clinicians. Innovation led to devices each node within a deep learning model may not provide
that leveled the playing field for clinicians to be able to a clear explanation of why certain predictions were made;
provide care based on reliable clinical metrics of oxygen but the model can be asked to present relevant features or
saturation and blood pressure. Currently, the greatest near- examples from its training data of skeletal x-rays to explain
term potential for artificial intelligence is in its ability to why a particular prediction was made on the bone age of a
offer tools with which to analyze massive amounts of data patient.101 In addition to concerns about transparency and
and offer more digestible statistics about that data that cli- trust in models, artificial intelligence can excel in demon-
nicians can use to render a medical decision. Artificial intel- strating correlations or in identifying patterns, but it can-
ligence could thus provide anesthesiologists at all levels of not yet determine causal relationships—at least not to a
expertise with decision support—whether clinical or pro- degree that would be necessary for clinical implementation.

cedural—to enable all clinicians to provide the best possible As research continues to expand in the field, clinicians will
evidence-based care to their patients. have to critically vet new research findings and product
As Gambus and Shafer99 point out in their recent edi- announcements, and medical education and training will
torial on artificial intelligence, humans have the ability to necessarily have to incorporate elements of literacy in arti-
extrapolate from the known to the unknown based on a ficial intelligence concepts.
more complete understanding of scientific phenomena, but Artificial intelligence algorithms are also susceptible to
artificial intelligence can only draw conclusions from the bias in data. Beyond the basic research biases that clinicians
data it has seen and analyzed.99 As more and more elements are taught (e.g., sampling, blinding, etc.), we must also con-
of clinical practice become digitized and accumulated into sider both implicit and explicit bias in the healthcare system
databases, we may one day see the development of artifi- that can impact the large-scale data that is or will be used
cial intelligence systems that have a more complete under- to train artificial intelligences. Eligibility of specific patient
standing of clinical phenomena and thus greater potential populations for clinical trials, implicit biases in treatment
to deliver elements of anesthesia care autonomously. decisions in real world care, and other forms of bias can
significantly affect the types of predictions that an artificial
Limitations and Ethical Implications of Artificial intelligence may make and influence clinical decisions.102,103
Intelligence Char et al.104 give the example of withdrawal of care from
patients with traumatic brain injury. An artificial intelli-
Artificial intelligence is not without its limitations. The
gence may analyze data from a neuro-ICU and interpret
hype surrounding artificial intelligence has reached a fever
a pattern of fatality after traumatic brain injury as a fore-
pitch in the lay press, and unrealistic expectations can result
gone consequence of the injury rather than as secondary to
in eventual disillusionment with artificial intelligence if
the clinical decision to withdraw life support.104 Therefore,
clinicians, patients, and regulators do not see the expected
it is imperative that practicing clinicians partner or develop
revolution in healthcare that is anticipated with artificial
a dialogue with data scientists to ensure the appropriate
intelligence technologies.100 Therefore, it is critical to under-
interpretation of data analyses.
stand that using artificial intelligence–based techniques will
not necessarily result in classification or prediction that is
superior to current methods. Artificial intelligence is a tool Future Directions
that must be deployed in the right situation to answer an As this review has demonstrated, artificial intelligence algo-
appropriate question or solve an applicable problem. rithms have not yet surpassed human performance; how-
One criticism of artificial intelligence, particularly of ever, artificial intelligence’s ability to quickly and accurately
neural networks, is that the methods can result in black sift through large stores of data and uncover correlations
box results, where an algorithm can notify a clinician or and patterns that are imperceptible to human cognition will
researcher of a prediction but cannot provide further infor- make it a valuable tool for clinicians. In pathology, artificial
mation on why such a prediction was made. Efforts in intelligence has been demonstrated to augment the ability
explainable artificial intelligence are being made to improve of clinicians in making diagnoses, such as in the reduction
the transparency of algorithms. The aim of explainable arti- of error rate in recognizing cancer-positive lymph nodes.32
ficial intelligence is to produce models that can more read- This reduction error occurred not because the artificial
ily explain its findings (e.g., by showing which features it intelligence outperformed humans—on the contrary it was
may have relied on to generate its prediction) with the end worse—but because it could narrow the area of a histo-
goal of improving its level of transparency and therefore pathologic slide that a human pathologist had to review,
increasing human trust and understanding of its predictions. allowing more attention to be placed on a smaller area.
Some techniques in artificial intelligence will more readily Similarly, we anticipate that artificial intelligence technol-
lend themselves to explanation than others. For example, ogy that can contribute to monitoring depth of anesthesia,
decision trees allow for great transparency as each decision maintaining drug infusions, or predicting hypotension in an

REVIEW ARTICLE
operation will allow practicing anesthesiologists to be more Research Support

effective and efficient in the care they provide.
Support for this study was provided solely from institutional
Medicine and the practice of anesthesiology is still, at its
and/or departmental sources.
core, a uniquely human endeavor as both science and art.
Although algorithms may one day exceed human capabilities
Competing Interests
in integrating complex, gigantic, structured datasets, much
of the data that clinicians gather from patients comes from Dr. Hashimoto is a consultant for Verily Life Sciences
the clinician–patient relationship that is established when (Mountain View, California), the Johnson & Johnson
patients bestow trust on their doctor. Although anesthe- Institute (Raynham, Massachusetts), and Gerson Lehrman
siologists can develop the knowledge and training to trust Group (New York, New York). He serves on the clinical
artificial intelligence models, it remains to be seen to what advisory board of Worrell, Inc. (Minneapolis, Minnesota).
extent patients will be willing to trust algorithms and how Dr. Meireles is a consultant for Olympus (Tokyo, Japan)

patients wish to have results from algorithms communicated and Medtronic (Dublin, Ireland). Dr. Rosman is a
to them.Therefore, qualitative research will be needed to bet- research scientist at Toyota Research Institute (Cambridge,
ter understand the ethical, cultural, and societal implications Massachusetts). Drs. Hashimoto, Meireles, and Rosman are
of integrating artificial intelligence into clinical workflows. co-inventors on a patent application submitted in relation
Anesthesiologists should continue to partner with data to computer vision technology for assessment of intraop-
scientists and engineers to provide their valuable clinical erative surgical performance. The other authors declare no
insight into the development of artificial intelligence to competing interests.
ensure that the technology will be clinically applicable, that
the data used to train algorithms are valid and representa- Correspondence
tive of a wide population of patients, and that interpreta- Address correspondence to Dr. Hashimoto: Surgical
tions of that data are clinically meaningful.105 Integration Artificial Intelligence and Innovation Laboratory,
and analysis of data by models does not mean the mod- Massachusetts General Hospital, 15 Parkman Street, WAC
els understand the implications of that data for specific 339, Boston, Massachusetts 02139. dahashimoto@mgh.
patients. Thus, anesthesiologists should partner with other harvard.edu. Information on purchasing reprints may be
providers (e.g., surgeons, interventionalists, intensivists, found at www.anesthesiology.org or on the masthead page
nurses) and patients to help develop the strategy for the at the beginning of this issue. Anesthesiology’s articles are
optimal use of artificial intelligence. Anesthesiology as a made freely accessible to all readers, for personal use only, 6
field has been a leader in the implementation and achieve- months from the cover date of the issue.
ment of patient safety initiatives, and artificial intelligence
can serve as a new tool to continue innovations in the References
delivery of safe anesthesia care.
1. Bellman R: An introduction to artificial intelligence:
Conclusion Can computers think? San Francisco, Boyd & Fraser
The field of anesthesiology has a long history of engaging Pub Co, 1978
in research that incorporates aspects of artificial intelligence. 2. Buchanan BG: A (very) brief history of artificial intel-
Artificial intelligence has the potential to impact the clinical ligence. Ai Magazine 2005; 26: 53
practice of anesthesia in aspects ranging from perioperative 3. Thrall JH, Li X, Li Q, Cruz C, Do S, Dreyer K,
support to critical care delivery to outpatient pain manage- Brink J: Artificial intelligence and machine learning
ment. As research efforts advance and technology develop- in radiology: Opportunities, challenges, pitfalls, and
ment intensifies, it will be crucial for practicing clinicians criteria for success. J Am Coll Radiol 2018; 15(3 Pt
to provide practice-based insights to assist in the clinical B):504–8
4. Salto-Tellez M, Maxwell P, Hamilton P: Artificial
translation of artificial intelligence.
intelligence-the third revolution in pathology.
Histopathology 2019; 74:372–6
Acknowledgments
5. Deo RC: Machine learning in medicine. Circulation
The authors thank Elizabeth Y. Zhou, M.D. (Department 2015; 132:1920–30
of Anesthesiology and Critical Care, Perelman School of 6. Hashimoto DA, Rosman G, Rus D, Meireles OR:
Medicine, Philadelphia, Pennsylvania), for her insights on Artificial intelligence in surgery: Promises and perils.
clinical care and data streams in cardiac anesthesia and Ann Surg 2018; 268:70–6
Caitlin Stafford, B.S., C.C.R.P. (Department of Surgery, 7. FDA: FDA permits marketing of artificial intelli-
Massachusetts General Hospital, Boston, Massachusetts) for gence-based device to detect certain diabetes-related
her insights on implementation and management of large eye problems. Edited by FDA. US Food and Drug
clinical databases. Administration, 2018

8. Russell S, Norvig P: Artificial Intelligence: A Modern 25. Kirkos E, Spathis C, Manolopoulos Y: Data mining
Approach, 3rd Edition. Upper Saddle River, New techniques for the detection of fraudulent financial
Jersey, Prentice Hall, 2009 statements. Expert Syst Appl 2007; 32: 995–1003
9. Kendale S, Kulkarni P, Rosenberg AD, Wang J: 26. van den Berg JP, Eleveld DJ, De Smet T, van den Heerik
Supervised machine-learning predictive analyt- AVM, van Amsterdam K, Lichtenbelt BJ, Scheeren TWL,
ics for prediction of postinduction hypotension. Absalom AR, Struys MMRF: Influence of Bayesian
Anesthesiology 2018; 129:675–88 optimization on the performance of propofol tar-
10. Wanderer JP, Rathmell JP: Machine learning for anes- get-controlled infusion. Br J Anaesth 2017; 119:918–27
thesiologists: A primer. Anesthesiology 2018; 129:A29 27. Kukacka M: Bayesian Methods in Artificial Intelligence,
11. Bisgin H, Liu Z, Fang H, Xu X, Tong W: Mining WDS’10 Proceedings of Contributed Papers, 2010, pp
FDA drug labels using an unsupervised learning tech- 25–30
nique–topic modeling. BMC Bioinformatics 2011; 12 28. Nadkarni PM, Ohno-Machado L, Chapman WW:

Suppl 10:S11 Natural language processing: An introduction. J Am
12. Hakonarson H, Bjornsdottir US, Halapi E, Bradfield Med Inform Assoc 2011; 18:544–51
J, Zink F, Mouy M, Helgadottir H, Gudmundsdottir 29. Murff HJ, FitzHenry F, Matheny ME, Gentry N, Kotter
AS, Andrason H, Adalsteinsdottir AE, Kristjansson KL, Crimin K, Dittus RS, Rosen AK, Elkin PL, Brown
K, Birkisson I, Arnason T, Andresdottir M, Gislason SH, Speroff T: Automated identification of postop-
D, Gislason T, Gulcher JR, Stefansson K: Profiling of erative complications within an electronic medical
genes expressed in peripheral blood mononuclear cells record using natural language processing. JAMA 2011;
predicts glucocorticoid sensitivity in asthma patients. 306:848–55
Proc Natl Acad Sci U S A 2005; 102:14789–94 30. Cohen KB, Glass B, Greiner HM, Holland-Bouley
13. Sutton RS, Barto AG: Reinforcement Learning: An K, Standridge S, Arya R, Faist R, Morita D, Mangano
Introduction, MIT press Cambridge, 1998 F, Connolly B, Glauser T, Pestian J: Methodological
14. Padmanabhan R, Meskin N, Haddad WM: Closed- issues in predicting pediatric epilepsy surgery candi-
loop control of anesthesia and mean arterial pressure dates through natural language processing and machine
using reinforcement learning. Biomed Signal Process learning. Biomed Inform Insights 2016; 8:11–8
Control 2015; 22: 54–64 31. Friedman C, Shagina L, Lussier Y, Hripcsak G:
15. Ng A: Supervised Learning, CS229 Lecture Notes. Automated encoding of clinical documents based on
Edited by Ng A, Stanford University, 2018 natural language processing. J Am Med Inform Assoc
16. Zadeh LA: Fuzzy sets. Information and Control 1965; 2004; 11:392–402
8: 338–353 32. Ehteshami Bejnordi B, Veta M, Johannes van Diest P,
17. Baig MM, Gholamhosseini H, Kouzani A, Harrison van Ginneken B, Karssemeijer N, Litjens G, van der
MJ: Anaesthesia monitoring using fuzzy logic. J Clin Laak JAWM, Hermsen M, Manson QF, Balkenhol M,
Monit Comput 2011; 25:339–47 Geessink O, Stathonikos N, van Dijk MC, Bult P, Beca
18. Hu YJ, Ku TH, Jan RH, Wang K, Tseng YC, Yang SF: F, Beck AH, Wang D, Khosla A, Gargeya R, Irshad H,
Decision tree-based learning to predict patient con- Zhong A, Dou Q, Li Q, Chen H, Lin HJ, Heng PA,
trolled analgesia consumption and readjustment. BMC Haß C, Bruni E, Wong Q, Halici U, Öner MÜ, Cetin-
Med Inform Decis Mak 2012; 12:131 Atalay R, Berseth M, Khvatkov V,Vylegzhanin A, Kraus
19. Hastie T, Tibshirani R, Friedman J: Overview of O, Shaban M, Rajpoot N, Awan R, Sirinukunwattana
Supervised Learning, The Elements of Statistical K, Qaiser T, Tsang YW, Tellez D, Annuscheit J, Hufnagl
Learning: Data Mining, Inference, and Prediction. P, Valkonen M, Kartasalo K, Latonen L, Ruusuvuori
New York, Springer, 2016, pp 9–42 P, Liimatainen K, Albarqouni S, Mungal B, George A,
20. Hastie T, Tibshirani R, Friedman J: Support Vector Demirci S, Navab N, Watanabe S, Seno S, Takenaka Y,
Machines and Flexible Discriminants, The Elements Matsuda H,Ahmady Phoulady H, KovalevV, Kalinovsky
of Statistical Learning: Data Mining, Inference, and A, Liauchuk V, Bueno G, Fernandez-Carrobles MM,
Prediction. New York, Springer, 2016, pp 417–58 Serrano I, Deniz O, Racoceanu D, Venâncio R; the
21. McCulloch WS, Pitts W: A logical calculus of the ideas CAMELYON16 Consortium: Diagnostic assessment
immanent in nervous activity. Bulletin of Mathematical of deep learning algorithms for detection of lymph
Biophysics 1943; 5: 115–33 node metastases in women with breast cancer. JAMA
22. LeCun Y, Bengio Y, Hinton G: Deep learning. Nature 2017; 318:2199–210
2015; 521:436–44 33. Tajmir SH, Lee H, Shailam R, Gale HI, Nguyen JC,
23. Bland JM, Altman DG: Bayesians and frequentists. BMJ Westra SJ, Lim R, Yune S, Gee MS, Do S: Artificial
1998; 317:1151–60 intelligence-assisted interpretation of bone age radio-
24. Ghahramani Z: Probabilistic machine learning and graphs improves accuracy and decreases variability.
artificial intelligence. Nature 2015; 521:452–9 Skeletal Radiol 2019; 48:275–83

REVIEW ARTICLE
34. Volkov M, Hashimoto DA, Rosman G, Meireles OR, 45. Shalbaf A, Saffar M, Sleigh JW, Shalbaf R: Monitoring
Rus D: Machine learning and coresets for automated the depth of anesthesia using a new adaptive neuro-
real-time video segmentation of laparoscopic and fuzzy system. IEEE J Biomed Health Inform 2018;
robot-assisted surgery, IEEE International Conference 22:671–7
on Robotics and Automation (ICRA), 2017, pp 754–9 46. Nagaraj SB, Biswal S, Boyle EJ, Zhou DW, McClain
35. Pesteie M, Lessoway V, Abolmaesumi P, Rohling RN: LM, Bajwa EK, Quraishi SA, Akeju O, Barbieri R,
Automatic localization of the needle target for ultra- Purdon PL, Westover MB: Patient-specific classifica-
sound-guided epidural injections. IEEE Trans Med tion of ICU sedation levels from heart rate variability.
Imaging 2018; 37:81–92 Crit Care Med 2017; 45:e683–90
36. Smistad E, Lovstakken L, Carneiro G, Mateus D, 47. Ranta SO, Hynynen M, Räsänen J: Application of arti-
Peter L, Bradley A, Tavares J, Belagiannis V, Papa JP, ficial neural networks as an indicator of awareness with
Nascimento JC, Loog M, Lu Z, Cardoso JS, Cornebise recall during general anaesthesia. J Clin Monit Comput

J: Vessel Detection in Ultrasound Images Using Deep 2002; 17:53–60
Convolutional Neural Networks, Med Image Comput 48. Dumont GA, Ansermino JM: Closed-loop control of
Comput Assist Inter, Springer, 2016, pp 30–8 anesthesia: A primer for anesthesiologists. Anesth Analg
37. Fritz BA, Maybrier HR, Avidan MS: Intraoperative 2013; 117:1130–8
electroencephalogram suppression at lower volatile 49. Tsutsui T, Arita S: Fuzzy-logic control of blood pres-
anaesthetic concentrations predicts postoperative delir- sure through enflurane anesthesia. J Clin Monit 1994;
ium occurring in the intensive care unit. Br J Anaesth 10:110–7
2018; 121:241–8 50. Zbinden AM, Feigenwinter P, Petersen-Felix S,
38. Kertai MD, Pal N, Palanca BJ, Lin N, Searleman Hacisalihzade S: Arterial pressure control with isoflu-
SA, Zhang L, Burnside BA, Finkel KJ, Avidan MS; rane using fuzzy logic. Br J Anaesth 1995; 74:66–72
B-Unaware Study Group: Association of perioperative 51. Absalom AR, Sutcliffe N, Kenny GN: Closed-loop con-
risk factors and cumulative duration of low bispectral trol of anesthesia using Bispectral index: Performance
index with intermediate-term mortality after cardiac assessment in patients undergoing major orthopedic
surgery in the B-Unaware Trial. Anesthesiology 2010; surgery under combined general and regional anesthe-
112:1116–27 sia. Anesthesiology 2002; 96:67–73
39. Sessler DI, Sigl JC, Kelley SD, Chamoun NG, Manberg 52. Shieh JS, Kao MH, Liu CC: Genetic fuzzy model-
PJ, Saager L, Kurz A, Greenwald S: Hospital stay and ling and control of bispectral index (BIS) for gen-
mortality are increased in patients having a “triple low” eral intravenous anaesthesia. Med Eng Phys 2006;
of low blood pressure, low bispectral index, and low 28:134–48
minimum alveolar concentration of volatile anesthesia. 53. Lowery C, Faisal AA: Towards efficient, personalized
Anesthesiology 2012; 116:1195–203 anesthesia using continuous reinforcement learning
40. Veselis RA, Reinsel R, Sommer S, Carlon G: Use of for propofol infusion control. 2013 6th International
neural network analysis to classify electroencephalo- IEEE/EMBS Conference on Neural Engineering
graphic patterns against depth of midazolam sedation in (NER) 2013: 1414–7
intensive care unit patients. J Clin Monit 1991; 7:259–67 54. Zaouter C, Hemmerling TM, Lanchon R, Valoti E,
41. Veselis RA, Reinsel R,Wronski M: Analytical methods Remy A, Leuillet S, Ouattara A: The feasibility of a
to differentiate similar electroencephalographic spectra: completely automated total IV anesthesia drug deliv-
neural network and discriminant analysis. J Clin Monit ery system for cardiac surgery. Anesth Analg 2016;
1993; 9:257–67 123:885–93
42. Ortolani O, Conti A, Di Filippo A, Adembri C, Moraldi 55. Lendl M, Schwarz UH, Romeiser HJ, Unbehauen R,
E, Evangelisti A, Maggini M, Roberts SJ: EEG signal Georgieff M, Geldner GF: Nonlinear model-based
processing in anaesthesia: Use of a neural network predictive control of non-depolarizing muscle relax-
technique for monitoring depth of anaesthesia. Br J ants using neural networks. J Clin Monit Comput
Anaesth 2002; 88:644–8 1999; 15:271–8
43. Benzy VK, Jasmin EA, Koshy RC, Amal F: Wavelet 56. Shieh JS, Fan SZ, Chang LW, Liu CC: Hierarchical
Entropy based classification of depth of anesthesia. rule-based monitoring and fuzzy logic control for
2016 International Conference on Computational neuromuscular block. J Clin Monit Comput 2000;
Techniques in Information and Communication 16:583–92
Technologies (ICCTICT) 2016: 521–4 57. Motamed C, Devys JM, Debaene B, Billard V: Influence
44. Mirsadeghi M, Behnam H, Shalbaf R, Jelveh of real-time Bayesian forecasting of pharmacokinetic
Moghadam H: Characterizing awake and anesthetized parameters on the precision of a rocuronium tar-
states using a dimensionality reduction method. J Med get-controlled infusion. Eur J Clin Pharmacol 2012;
Syst 2016; 40:13 68:1025–31

58. Martinoni EP, Pfister ChA, Stadler KS, Schumacher 70. Hatib F, Jian Z, Buddi S, Lee C, Settels J, Sibert K,
PM, Leibundgut D, Bouillon T, Böhlen T, Zbinden Rinehart J, Cannesson M: Machine-learning algorithm
AM: Model-based control of mechanical ventilation: to predict hypotension based on high-fidelity arterial
Design and clinical validation. Br J Anaesth 2004; pressure waveform analysis. Anesthesiology 2018;
92:800–7 129:663–74
59. Schäublin J, Derighetti M, Feigenwinter P, Petersen- 71. Gao L, Smielewski P, Czosnyka M, Ercole A:
Felix S, Zbinden AM: Fuzzy logic control of mechan- Cerebrovascular signal complexity six hours after
ical ventilation during anaesthesia. Br J Anaesth 1996; intensive care unit admission correlates with outcome
77:636–41 after severe traumatic brain injury. J Neurotrauma
60. Schädler D, Mersmann S, Frerichs I, Elke G, Semmel- 2016; 33:2011–8
Griebeler T, Noll O, Pulletz S, Zick G, David M, 72. Gao L, Smielewski P, Czosnyka M, Ercole A: Early
Heinrichs W, Scholz J, Weiler N: A knowledge- and asymmetric cardio-cerebral causality and outcome

model-based system for automated weaning from after severe traumatic brain injury. J Neurotrauma
mechanical ventilation: Technical description and 2017; 34:2743–52
first clinical application. J Clin Monit Comput 2014; 73. Gottschalk A, Hyzer MC, Geer RT: A comparison
28:487–98 of human and machine-based predictions of success-
61. Lin CS, Li YC, Mok MS, Wu CC, Chiu HW, Lin YH: ful weaning from mechanical ventilation. Med Decis
Neural network modeling to predict the hypnotic Making 2000; 20:160–9
effect of propofol bolus induction. Proc AMIA Symp 74. Bonds BW, Yang S, Hu PF, Kalpakis K, Stansbury LG,
2002: 450–3 Scalea TM, Stein DM: Predicting secondary insults
62. Nunes CS, Mendonca TF, Amorim P, Ferreira DA, after severe traumatic brain injury. J Trauma Acute Care
Antunes L: Comparison of neural networks, fuzzy and Surg 2015; 79:85–90; discussion 90
stochastic prediction models for return of consciousness 75. Jalali A, Bender D, Rehman M, Nadkanri V, Nataraj C:
after general anesthesia. Proceedings of the 44th IEEE Advanced analytics for outcome prediction in intensive
Conference on Decision and Control 2005: 4827–32 care units. 2016 38th Annual International Conference
63. Santanen OA, Svartling N, Haasio J, Paloheimo of the IEEE Engineering in Medicine and Biology
MP: Neural nets and prediction of the recovery rate Society (EMBC) 2016: 2520–4
from neuromuscular block. Eur J Anaesthesiol 2003; 76. Clermont G, Angus DC, DiRusso SM, Griffin M,
20:87–92 Linde-Zwirble WT: Predicting hospital mortality for
64. Lin CS, Chang CC, Chiu JS, Lee YW, Lin JA, Mok patients in the intensive care unit: A comparison of
MS, Chiu HW, Li YC: Application of an artificial neu- artificial neural networks with logistic regression mod-
ral network to predict postinduction hypotension els. Crit Care Med 2001; 29:291–6
during general anesthesia. Med Decis Making 2011; 77. Desautels T, Das R, Calvert J, Trivedi M, Summers
31:308–14 C, Wales DJ, Ercole A: Prediction of early unplanned
65. Lin CS, Chiu JS, Hsieh MH, Mok MS, Li YC, Chiu intensive care unit readmission in a UK tertiary care
HW: Predicting hypotensive episodes during spi- hospital: A cross-sectional machine learning approach.
nal anesthesia with the application of artificial neural BMJ Open 2017; 7:e017199
networks. Comput Methods Programs Biomed 2008; 78. Desautels T, Calvert J, Hoffman J, Jay M, Kerem Y,
92:193–7 Shieh L, Shimabukuro D, Chettipally U, Feldman MD,
66. Zhang L, Fabbri D, Lasko TA, Ehrenfeld JM, Wanderer Barton C, Wales DJ, Das R: Prediction of sepsis in the
JP: A system for automated determination of perioper- intensive care unit with minimal electronic health
ative patient acuity. J Med Syst 2018; 42:123 record data: A machine learning approach. JMIR Med
67. Moustafa MA, El-Metainy S, Mahar K, Mahmoud Inform 2016; 4:e28
Abdel-magied E: Defining difficult laryngoscopy find- 79. Shimabukuro DW, Barton CW, Feldman MD, Mataraso
ings by using multiple parameters: A machine learning SJ, Das R: Effect of a machine learning-based severe
approach. Egypt J Anaesth 2017; 33: 153–8 sepsis prediction algorithm on patient survival and
68. Berkenstadt H, Ben-Menachem E, Herman A, Dach hospital length of stay: A randomised clinical trial. BMJ
R: An evaluation of the Integrated Pulmonary Index Open Respir Res 2017; 4:e000234
(IPI) for the detection of respiratory events in sedated 80. Hetherington J, Lessoway V, Gunka V, Abolmaesumi
patients undergoing colonoscopy. J Clin Monit P, Rohling R: SLIDE: Automatic spine level iden-
Comput 2012; 26:177–81 tification system using a deep convolutional neu-
69. Hancerliogullari G, Hancerliogullari KO, Koksalmis E: ral network. Int J Comput Assist Radiol Surg 2017;
The use of multi-criteria decision making models in 12:1189–98
evaluating anesthesia method options in circumcision 81. Olesen AE, Grønlund D, Gram M, Skorpen F, Drewes
surgery. BMC Med Inform Decis Mak 2017; 17:14 AM, Klepstad P: Prediction of opioid dose in cancer

REVIEW ARTICLE
pain patients using genetic profiling: Not yet an option 93. Sciusco A, Standing JF, Sheng Y, Raimondo P, Cinnella
with support vector machine learning. BMC Res G, Dambrosio M: Effect of age on the performance
Notes 2018; 11:78 of bispectral and entropy indices during sevoflurane
82. Tighe PJ, Lucas SD, Edwards DA, Boezaart AP, Aytug pediatric anesthesia: A pharmacometric study. Paediatr
H, Bihorac A: Use of machine-learning classifiers to Anaesth 2017; 27:399–408
predict requests for preoperative acute pain service 94. Olofsen E, Sleigh JW, Dahan A: Permutation entropy
consultation. Pain Med 2012; 13:1347–57 of the electroencephalogram: A measure of anaesthetic
83. Gonzalez-Cava JM, Arnay R, Perez JAM, Leon A, drug effect. Br J Anaesth 2008; 101:810–21
Martin M, Jove-Perez E, Calvo-Rolle JL, Casteleiro- 95. Eagleman SL, Drover DR: Calculations of con-
Roca JL, Juez FJD, Garcia HP, AlfonsoCendon J, sciousness: electroencephalography analyses to deter-
Gonzalez LS, Quintian H, Corchado E: A Machine mine anesthetic depth. Curr Opin Anaesthesiol 2018;
learning based system for analgesic drug delivery. 31:431–8

International joint Conference SOCO’17-CISIS’17- 96. Ghassemi M, Celi LA, Stone DJ: State of the art review:
ICEUTE’17. 2017; 649: 461–70 The data revolution in critical care. Crit Care 2015;
84. Ben-Israel N, Kliger M, Zuckerman G, Katz Y, Edry R: 19:118
Monitoring the nociception level: A multi-parameter 97. Avidan MS, Jacobsohn E, Glick D, Burnside BA, Zhang
approach. J Clin Monit Comput 2013; 27:659–68 L, Villafranca A, Karl L, Kamal S, Torres B, O’Connor
85. Gram M, Erlenwein J, Petzke F, Falla D, Przemeck M, Evers AS, Gradwohl S, Lin N, Palanca BJ, Mashour
M, Emons MI, Reuster M, Olesen SS, Drewes AM: GA; BAG-RECALL Research Group: Prevention of
Prediction of postoperative opioid analgesia using clin- intraoperative awareness in a high-risk surgical popu-
ical-experimental parameters and electroencephalog- lation. N Engl J Med 2011; 365:591–600
raphy. Eur J Pain 2017; 21:264–77 98. Avidan MS, Zhang L, Burnside BA, Finkel KJ,
86. Combes C, Meskens N, Rivat C,Vandamme JP: Using Searleman AC, Selvidge JA, Saager L, Turner MS,
a KDD process to forecast the duration of surgery. Int Rao S, Bottros M, Hantler C, Jacobsohn E, Evers AS:
J Prod Econ 2008; 112: 279–93 Anesthesia awareness and the bispectral index. N Engl
87. Devi SP, Rao KS, Sangeetha SS: Prediction of surgery J Med 2008; 358:1097–108
times and scheduling of operation theaters in ophthal- 99. Gambus P, Shafer SL: Artificial intelligence for every-
mology department. J Med Syst 2012; 36:415–30 one. Anesthesiology 2018; 128:431–3
88. Houliston BR, Parry DT, Merry AF: TADAA: Towards 100. Sicular S, Brant K: Hype Cycle for Artificial
automated detection of anaesthetic activity. Methods Intelligence, 2018. Edited by Gartner I. Gartner, 2018,
Inf Med 2011; 50:464–71 pp 1–73
89. Silver D, Hubert T, Schrittwieser J, Antonoglou 101. Lee H,Tajmir S, Lee J, Zissen M,Yeshiwas BA, Alkasab
I, Lai M, Guez A, Lanctot M, Sifre L, Kumaran D, TK, Choy G, Do S: Fully automated deep learning
Graepel T, Lillicrap T, Simonyan K, Hassabis D: A gen- system for bone age assessment. J Digit Imaging 2017;
eral reinforcement learning algorithm that masters 30:427–41
chess, shogi, and Go through self-play. Science 2018; 102. Murthy VH, Krumholz HM, Gross CP: Participation
362:1140–4 in cancer clinical trials: Race-, sex-, and age-based
90. Walraet M, Tromp J: A Googolplex of Go Games. disparities. JAMA 2004; 291:2720–6
Cham, Springer International Publishing, 2016, pp 103. Schulman KA, Berlin JA, Harless W, Kerner JF,
191–201 Sistrunk S, Gersh BJ, Dubé R, Taleghani CK, Burke
91. Proposed Regulatory Framework for Modifications JE, Williams S, Eisenberg JM, Escarce JJ: The effect of
to Artificial Intelligence/Machine Learning (AI/ race and sex on physicians’ recommendations for car-
ML)-Based Software as a Medical Device (SaMD) - diac catheterization. N Engl J Med 1999; 340:618–26
Discussion Paper and Request for Feedback, Food and 104. Char DS, Shah NH, Magnus D: Implementing machine
Drug Administration, 2019, pp 1–20 learning in health care - Addressing ETHICAL
92. Liem VGB, Hoeks SE, van Lier F, de Graaff JC: What CHALLENGES. N Engl J Med 2018; 378:981–3
we can learn from Big Data about factors influencing 105. Weber GM, Mandl KD, Kohane IS: Finding the
perioperative outcome. Curr Opin Anaesthesiol 2018; missing link for big biomedical data. JAMA 2014;
31:723–31 311:2479–800


AI in Anesthesiology

Uploaded by

Copyright:

Available Formats

AI in Anesthesiology

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

AI in Anesthesiology

Uploaded by

Copyright:

Available Formats

REVIEW ARTICLE Deborah J. Culley, M.D.

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

A rtificial intelligence has been defined as the study of

recognition, inference of world states, and decision-mak-

ANESTHESIOLOGY, V 132 • NO 2 February 2020 379

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

380 Anesthesiology 2020; 132:379–94 Hashimoto et al.

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

Hashimoto et al. Anesthesiology 2020; 132:379–94 381

versus the degree to which a statement is true (e.g., “there

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

382 Anesthesiology 2020; 132:379–94 Hashimoto et al.

(e.g., speech).22 Unlike in classical machine learning where

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

Applied Fields in Artificial Intelligence

Hashimoto et al. Anesthesiology 2020; 132:379–94 383

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

384 Anesthesiology 2020; 132:379–94 Hashimoto et al.

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

Hashimoto et al. Anesthesiology 2020; 132:379–94 385

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

386 Anesthesiology 2020; 132:379–94 Hashimoto et al.

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

Hashimoto et al. Anesthesiology 2020; 132:379–94 387

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

388 Anesthesiology 2020; 132:379–94 Hashimoto et al.

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

Hashimoto et al. Anesthesiology 2020; 132:379–94 389

operation will allow practicing anesthesiologists to be more Research Support

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

390 Anesthesiology 2020; 132:379–94 Hashimoto et al.

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

Hashimoto et al. Anesthesiology 2020; 132:379–94 391

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

392 Anesthesiology 2020; 132:379–94 Hashimoto et al.

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

Hashimoto et al. Anesthesiology 2020; 132:379–94 393

Downloaded from http://pubs.asahq.org/anesthesiology/article-pdf/132/2/379/516477/20200200_0-00027.pdf by guest on 19 March 2023

394 Anesthesiology 2020; 132:379–94 Hashimoto et al.

You might also like