PRINTER-FRIENDLY VERSION AVAILABLE AT ANESTHESIOLOGYNEWS.

COM

Artificial Intelligence,
Coming to an Airway Near You?
A

KRISTOPHER M. SCHROEDER, MD, FASA

ll
rig

Co

Professor
ht

Vice Chair of Faculty Development

py
s

Department of Anesthesiology
rig ed.
re

University of Wisconsin School of Medicine and Public Health

ht
se

Madison
rv

©
20

RICHARD E. GALGON, MD, MS

24
Re

Staff Anesthesiologist
M
pr

MediTech Solutions, LLC

cM

Clinical Staffing Solutions

od

Waunakee, Wis.
uc

ah in w
tio

on

Schroeder is a paid speaker for Northwest Anesthesia Seminars, and receives royalties from
n

Springer Inc. from publication of The Essential Guide to Healthcare Professional Wellness.
Pu
bl
is
ho

hi
ng
le
or

G
ro
in

D
espite the fear for catastrophe and societal
up
pa

calamity, the incorporation of AI into our

un ou
rt
w

le

daily lives, both in and out of the OR,

ith

ss
ot

continues to accelerate.
he
tp

rw
er

is
m

e
is

no
si
on

te

Rooted in the collective history of our society is a simul- However, despite the fear for catastrophe and societal
d.
is

taneous fascination with and deep-seated mistrust of the calamity, the incorporation of AI into our daily lives, both in and
pr

potential indications and dangers associated with automati- out of the OR, continues to accelerate. In fact, the global AI
oh

zation and artificial intelligence (AI). This interest is manifested healthcare market landscape currently hovers around $22 bil-
ib

across popular culture, and the technology is generally her- lion, but is projected to balloon to $208 billion by 2030.1 For-
ite

alded as either comedic and benevolent (i.e., Iron Man’s AI tunately, AI is in its infancy, and it is in this stage where there
d.

assistant J.A.R.V.I.S, Data from “Star Trek: The Next Genera- are the greatest opportunities for intervention and molding
tion” or C-3PO from “Star Wars”) or malevolent and a threat to to ensure the technology is crafted in a way that unburdens
human existence (i.e., HAL 9000 in Arthur C. Clarke’s “Space healthcare professionals and improves patient care.
Odyssey” series, Ultron from the Marvel Universe or Skynet
from “Terminator”). Short of the elimination of all human exis- Expanding Technology’s Reach
tence, there are less existential, but more commonly encoun- There are several promising areas within the practice of
tered, concerns expressed by more than merely Luddites, anesthesiology that might be augmented or improved with
and might include the potential for intellectual theft, academic the use of AI. Where this technology has already gained
misconduct or elimination of certain lines of employment. substantial traction is in the analysis and integration of

6 ANESTHESIOLOGYNEWS.COM
data that is now created through publication and patient Can AI Improve Airway Management
encounters. This unfathomable amount of information can And Operator Training?
quickly overwhelm the capabilities of humans, but it is in The field of airway management is one that historically has
this data vastness that AI and machine learning platforms been a leader in the area of technology improvement and
can be leveraged for our collective benefit. However, only widespread adoption. Tinkerers and advanced thinkers have
recently has the cost of computation power and data stor- long been fascinated with methodologies that might improve
age decreased to the extent that large-scale data analy- the safety and efficacy of airway management procedures,
sis is now reasonably available to healthcare practitioners. since failed airway management remains associated with seri-
In an example of the power that machine learning mod- ous patient morbidity and mortality. The widespread adoption
els possess, a study by Myszewski et al demonstrated of extraglottic airway devices and video laryngoscopes has
that sentiment analysis of published abstract language occurred at such a rapid pace that they now occupy elevated
A

produced similar conclusions when compared with tradi- status on difficult airway management algorithms, and it is dif-
ll

tional meta-analysis studies, and the time required rep- ficult now to consider the practice of anesthesiology without
rig

Co

resented only a minute fraction of what was required for these devices.10 Therefore, it stands to reason that the adop-
ht

py

more traditional techniques.2 With these approaches, natu- tion and application of AI technologies might experience sim-
s

rig ed.

ral language processing can offer the ability for healthcare ilar early adoption by those interested in improving airway
re

professionals to perform their own meta-analysis evalu- management. Will robotic intubation become routine? Will AI-
ht
se

ations and for these types of evaluations to incorporate augmented airway exams improve difficult airway predictions?
rv

even near real-time published data. Can AI-guided intubations aid novice airway managers while
20

Similarly, electronic health records now gener- improving intubation success and shortening times to profi-
24
Re

ate great volumes of data with every patient encoun- ciency? Could virtual assistants aid the real-time management
M
pr

ter and machine learning techniques offer the ability to of an airway crisis by guiding decision making in accordance
cM
od

sift through patient comorbidities, laboratory and imag- with difficult airway algorithms and dynamically changing clin-
ing results, vital signs, demographics, and caregiver notes ical conditions? Answers to these questions and many similar
uc

ah in w

to identify patterns and generate objective risk assess- ones are likely to be just a short way off.
tio

on

ments that can predict perioperative morbidity and mortal- As a training platform, virtual reality (VR) may play a cru-
n

Pu

ity.3,4 With this risk assessment and perioperative outcome cial role in ensuring that trainees and experienced clinicians
bl

predictive power, AI models can help to more appropri- are equipped to manage difficulties encountered with airway
is

ately allocate finite healthcare resources and assist in management and master skills associated with the perfor-
ho

hi

decisions relevant to patient disposition or discharge. For mance of uncommonly performed procedures. For example,
ng
le

example, AI can help to predict blood loss and the need the use of VR may offer opportunities to develop muscle
or

for transfusion associated with various patient conditions memory and avoid the unnecessary storage and use of lim-
ro
in

and surgical procedures. If substantial blood loss were ited and potentially costly invasive airway equipment. In addi-
up
pa

to occur, AI platforms have also demonstrated an ability tion, the use of VR might avoid or limit the need to manage
un ou

to assist in clinical decision making and provide recom- equipment or schedule time in busy or distant simulation
rt

mendations that accurately adhere to published transfu-

w

le

sion guidelines.5,6 In OR and recovery room environments,

ith

ss

AI has demonstrated an ability to predict adverse events,

ot

guide fluid management, titrate sedation administration

he
tp

and monitor vital signs to detect early harbingers of clini-

rw
er

cal deterioration.
is
m

AI platforms have simultaneously demonstrated an abil-

e
is

ity to assist in the real-time performance of regional anes-

no
si
on

thesia procedures. In this setting, AI-assisted technologies

te

provide real-time labeling of anatomic structures and sub-

d.
is

sequently contribute to image acquisition optimization

pr

and labeling of relevant anatomy.7,8 These technologies

oh

offer the promise of expanding the potential pool of physi-

ib

cians able to safely perform regional anesthesia because

ite

they provide an augmented ability to intervene in circum-

d.

stances where clinical experience or training may be rel-

atively diminished. In a similar fashion, AI offers the ability
to standardize performance and interpretation of echocar-
diography examinations, thereby improving the accuracy
and applicability of diagnostic exam findings. In addition, Figure 1. VR platform demonstrating what
real-time guidance on suggestions for improved image might be visualized during performance of
acquisition could again extend the pool of potential phy- cricothyroidotomy.
sicians able to comfortably apply this mode of diagnostic Image courtesy of Vantari VR.
imaging in a clinical setting.9

A N E S T H E S I O L O G Y N E W S A I R W AY M A N A G E M E N T 2 0 2 4 7
training environments. Multiple studies have already demon- progress over time could be tracked, and skill performance
strated the promise of VR platforms for augmenting training in metrics could be determined that would indicate proficiency
airway management procedures. Flexible fiberoptic intubation, prior to certification (Figure 2). Patients could be spared being
direct laryngoscopy and emergency cricothyrotomy training subjected to the antiquated tradition of “see one, do one,
platforms have been described, but there remains a lack of teach one,” and trainees could be required to demonstrate
consensus regarding measurable training outcome and quan- quantifiable VR proficiency before gaining an opportunity to
tifiable success variables.11 Despite these limitations, imagine engage in airway management in a clinical setting.
the benefits that might be realized if every OR or department Even seasoned airway management professionals could
of anesthesiology lounge was equipped with a VR platform benefit from the application of VR pathways when planning
that allowed for the ability to practice uncommonly performed and training for certain clinical scenarios. Imagine a scenario
airway management procedures, including retrograde or cri- in which an AI platform could interpret medical imaging to
A

cothyrotomy airway securement interventions, under realistic create VR training platforms that would allow anesthesiolo-
ll

time-sensitive simulated clinical conditions (Figure 1). VR also gists to practice managing the exact airway that they would
rig

Co

offers the ability to rapidly toggle between airway pathologies encounter in the OR for a looming case requiring manage-
ht

py

without the need to maintain a vast library of static and dedi- ment of a difficult airway. Consider how this training might
s

rig ed.

cated manikin models, which are unable to be adapted or to improve device or technique selection, improve anesthesiol-
re

offer variations on airway pathology themes as skills improve. ogist familiarity and comfort, and enhance patient outcomes.
ht
se

Augmented reality (AR) offers the ability to overlay vir-

rv

tual data on top of what is occurring in the physical world. In Predicting the Difficult Airway
20

these virtual and augmented platforms, real-time identifica- When it comes to airway assessment and the prediction of
24
Re

tion of airway structures could contribute to generalizable airway management difficulty, it remains well acknowledged
M
pr

airway anatomy knowledge and understanding that could that significant blind spots exist and what we are unable to
cM
od

be leveraged in later situations that require expert-level assess with our external airway evaluation might result in a
anatomic understanding. significant risk for unanticipated airway management difficul-
uc

ah in w

Where AI may ultimately contribute to these virtual and aug- ties. While the evaluation described by Mallampati provided
tio

on

mented platforms is through the evaluation of trainee perfor- airway proceduralists with an enhanced ability to predict air-
n

Pu

mance and leveraging of strengths and weaknesses to design way management difficulties, the failures of this investigation
bl

diverse training programs that continue to advance trainee skill, to eliminate unexpected difficulties has resulted in a prolifer-
is

device familiarity and procedural response. As trainees prog- ation of airway analyses that still fail to completely eliminate
ho

hi

ress and improve their skill with bronchoscopy, virtual lesions airway management misadventures.12 In fact, difficult laryn-
ng
le

could increasingly result in substantial airway distortion, or goscopy (12.3%), difficult intubation (9%) and failed intubation
or

blood/vomitus could be introduced that would require addi- (0.47%) continue to plague our profession and challenge the
ro
in

tional skill to suction and navigate an obscured airway. Trainee day-to-day clinical management of patients.13 The implications
up
pa

of these difficulties continue to be significant morbidity, mortal-

ity and the need for invasive airway interventions.14-15
un ou
rt

Whatever airway evaluation technique is selected, there

w

le
ith

is generally a tradeoff to be made between sensitivity and

ss

specificity that either misses patients at elevated risk for air-

ot

way management difficulties or exposes an unnecessary num-

he
tp

ber of patients to advanced or awake airway management

rw
er

maneuvers. However, AI and machine learning offer the prom-

is
m

ise of improved difficult airway prediction through the incor-

e
is

poration of multiple data points and the benefit of continued

no
si
on

model improvement. In one recent study, a machine learning

te

approach to airway management that evaluated several com-

d.
is

ponents of patients’ medical records found that the AI model

pr

performed substantially better than the traditional modified

oh

LEMON (Look, Evaluate, Mallampati, Obstruction, and Neck

ib

mobility) criteria, and was more likely to accurately predict first-

ite

pass intubation success and airway management difficulty.16

d.

Patient airway assessment images have also been used

by deep learning platforms to evaluate facial and airway fea-
tures associated with difficulty in airway management. Using
this AI platform, a model was created that exceeded 80% for
Figure 2. VR platform demonstrating an accuracy, sensitivity and specificity associated with intuba-
example of trainee performance evaluation tion difficulties.17 Other models have been created that spe-
in the setting of cricothyroidotomy. cifically address the risk for difficulties encountered in the
Image courtesy of Vantari VR. management of the pediatric airway or specific surgical sit-
uations including thyroid surgery.18,19 In fact, multiple models

8 ANESTHESIOLOGYNEWS.COM
have been described that offer the promise of improved widespread implementation can be reasonably considered.
detection and prediction of airway management difficulties, Additionally, the current platforms/devices would require
although none have yet been broadly incorporated into clini- substantial reductions in their physical footprint to be fea-
cal practice.20,21 These AI-based airway evaluations have also sible in the environments (e.g., prehospital settings) where
demonstrated an ability to benefit from significant automatiza- they may be associated with the greatest benefit.
tion, and fully automated systems that capture and evaluate
images without any anesthesiologist input have performed Lessons to Be Heeded
equally well against manual airway examinations.22 The future use of AI in anesthesiology is a virtual certainty
The ability of AI platforms to “learn” offers great promise as this technology increasingly permeates diverse aspects of
for rapid improvement of quantifiable airway management our culture and daily lives. Advances in AI technology offer tre-
difficulty prediction through the incorporation of evolving mendous promise, and it is tantalizing to speculate how they
A

research findings and the potential identification of pre- might impact our collective approach to the delivery of anes-
ll

viously undescribed features associated with increased thesia care and airway management. However, we are now
rig

Co

difficult airway management risk. Through accurate preop- at a crossroads, where we are compelled to collectively stop
ht

py

erative and automated risk assessment, patient scheduling and ponder how this technology might benefit versus harm
s

rig ed.

may be modified to ensure appropriate staffing or to maxi- our specialty and our patients, and how to rationally adopt this
re

mize the ability to attend to the temporal needs associated tool in a fashion that minimizes implementation-associated bur-
ht
se

with management of challenging airways. dens and maximizes patient outcomes. Simultaneously, avoid-
rv

ing undue restrictions to implementation of this technology and

20

Looking to the Future actively engaging with software engineers represent impor-
24
Re

The use of AI platforms in the process of procedural air- tant considerations when thinking about how this technology
M
pr

way management largely remains in its developmental infancy might be most effectively implemented into clinical practice.
cM
od

and substantial work is required to advance early efforts at The Sedasys machine developed by Johnson & Johnson is
real-time airway structure attention and identification.23-25 Real- a perfect example of how lack of support by anesthesiolo-
uc

ah in w

time identification of airway anatomy has been described gists portends certain doom for any proposed technological
tio

on

but has not yet achieved substantial clinical traction.26 How- advancement. This technology was intended to provide titrat-
n

Pu

ever, this AI-based anatomic identification augmentation may able sedation for routine procedures but was panned for a lim-
bl

soon become more routinely available and critically important ited ability to replicate a human’s ability to maintain procedural
is

in the setting of the novice laryngoscopist. Consider how this and situational vigilance. As a result of this lack of professional
ho

hi

anatomic identification assistance may guide safe intubation support, this technology failed to gain a clinical foothold and
ng
le

efforts performed by novice providers, including medics and ultimately was pulled from production in 2016.35 Additionally, it
or

EMS professionals, and/or in severely soiled airways where is important to consider the novelty of AI technology and how
ro
in

structure identification may be challenging. it might be prone to the creation and propagation of errors.
up
pa

In addition to structure identification, neural network−

Conclusion
un ou

based studies have demonstrated an ability to differentiate

rt

between tracheal and esophageal intubation and avoid cata- This article has highlighted several aspects of airway
w

le

strophic hypoxic outcomes.27 While AI-augmented procedures

ith

management where AI is already or will soon be contribut-

ss

are currently limited to video and fiberoptic bronchoscopy ing to the clinical management of patients requiring airway
ot

airway management platforms, this limitation may become interventions. Other AI roles, including voice-assisted guid-
he
tp

largely moot as the profession moves to more frequent use ance during airway management procedures, automated
rw
er

of video-based techniques, with a diminished dependence on recording of procedural details, postoperative monitoring
is
m

direct laryngoscopy techniques. It is also possible that AR plat- and prediction of airway-related deteriorations and/or com-
e
is

forms will eliminate the requirement to use video imaging plat- plications, and telemedicine support to help local health-
no
si
on

forms as real-world images are also subject to evaluation. care professionals navigate complex airway management
te

procedures, may soon change the delivery of airway man-

d.
is

Rise of the Robots agement services. After we consider the “what ifs” of AI tech-
pr

The implementation of robotic intubation systems also nology and the impact it might have one day on our careers
oh

has gained recent attention as a potential mechanism to and patient outcomes, it is more likely that this technology
ib

improve intubation success and reduce airway manage- will serve as a companion—rather a replacement—that aug-
ite

ment−related complications.28 While this intubation platform ments and assists the work done by anesthesiologists in the
d.

remains in an investigational stage, much of the early work OR and other anesthetizing locations. As such, it remains
involved intubation efforts being actively driven and con- critically important that clinicians collectively learn from mis-
trolled by a human driver.29-32 However, early work suggests steps in technology design and implementation, and work
this process of intubation might, at some point, become with AI platform engineers to develop a technological com-
entirely automated as AI structure identification advances panion that contributes to our ability to enhance patient care
are coupled with improved robotic technology.33,34 However, and does not detract from clinical workflows, patient interac-
hurdles remain and must be overcome before intubation can tions, or healthcare professional engagement and wellness.
be handed over to autonomous robots. Efficacy, safety, effi-
ciency, cost and training obstacles must be favorable before All references are available on AnesthesiologyNews.com.

A N E S T H E S I O L O G Y N E W S A I R W AY M A N A G E M E N T 2 0 2 4 9