Artificial Intelligence in Gastroenterology (AI for GI)
Summary
TLDRDr. Brennan Spiegel, a gastroenterologist at Cedars-Sinai, discusses the impact of artificial intelligence (AI) in gastroenterology. He explores how AI, particularly natural language processing (NLP) and image recognition, is transforming diagnosis and treatment in the field. Topics include automated polyp detection in colonoscopies, AI-assisted diagnosis of conditions like Barrett's esophagus and ulcerative colitis, and the potential for personalized medicine. Spiegel also touches on the limitations of AI, emphasizing the importance of human judgment and the collaboration between machines and doctors for better patient outcomes.
Takeaways
- 🤖 AI in gastroenterology is rapidly advancing, focusing on how human intelligence and AI can complement each other, termed as 'augmented intelligence'.
- 📚 Artificial Intelligence (AI) includes components like machine learning, natural language processing (NLP), and image recognition, which are relevant to medical applications like gastroenterology.
- 💡 AI systems are being developed to assist doctors by automating tasks such as polyp detection in colonoscopies, improving accuracy and efficiency in diagnosis.
- 🏥 AI can augment doctors' abilities rather than replace them, allowing more time for patient consultation, decision-making, and empathy, as discussed in the book 'Deep Medicine'.
- 🔬 AI has proven useful in areas like image analysis in gastroenterology, where it can accurately identify cancerous or precancerous conditions like Barrett’s esophagus and gastric cancer.
- 📈 In clinical trials, AI-assisted reviews increased the diagnosis rate of high-grade dysplasia and cancer by over 14%, showing the potential impact on patient outcomes.
- 👁️ AI systems in colonoscopy can increase the adenoma detection rate, improving overall outcomes in screening and diagnosis of colorectal conditions.
- 🔍 AI's application in capsule endoscopy helps detect ulcers and erosions with high accuracy, reducing the burden on clinicians sifting through large image sets.
- 🧪 AI can analyze blood work (CBC) to predict colon cancer risk, offering a non-invasive way to screen for potential issues before using more invasive procedures like colonoscopies.
- 🧠 The adoption of AI in healthcare is in its early stages, and its future success will depend on more prospective validation studies, cost-effectiveness evaluations, and integration into daily clinical workflows.
Q & A
What is the main focus of Dr. Brennan Spiegel's presentation?
-The main focus of Dr. Brennan Spiegel's presentation is the exciting and rapidly evolving role of artificial intelligence (AI) in gastroenterology, with an emphasis on how AI can augment human abilities in clinical settings.
What is the difference between artificial intelligence (AI) and augmented intelligence as described by Dr. Spiegel?
-According to Dr. Spiegel, artificial intelligence (AI) refers to computer systems simulating human behavior, while augmented intelligence emphasizes the collaboration between human intelligence and machines to enhance natural human capabilities.
What are some components of AI that Dr. Spiegel highlights in the presentation?
-Dr. Spiegel highlights several components of AI, including machine learning, natural language processing (NLP), and vision systems, with a focus on how they apply to gastroenterology.
How has AI shown its capabilities in gastroenterology according to the presentation?
-AI has demonstrated its capabilities in gastroenterology through natural language processing (NLP) for diagnostic support, machine learning for risk prediction, and vision systems for detecting abnormalities such as polyps during colonoscopy.
What is an example of AI being used in real-time during colonoscopy as mentioned by Dr. Spiegel?
-An example is the use of AI to detect polyps in real-time during colonoscopy by placing boxes around potential polyps, improving adenoma detection rates and assisting gastroenterologists in identifying abnormalities that might otherwise be missed.
What concerns are raised regarding the use of AI in gastroenterology?
-Dr. Spiegel mentions concerns such as selection bias in the images used for training AI models, the varying quality of real-world images, and the need for prospective validation studies to ensure the efficacy and reliability of AI tools in clinical practice.
How has AI improved the detection of gastric cancer and Barrett's esophagus?
-AI has shown promise in detecting early gastric cancer and Barrett's esophagus through convolutional neural networks that analyze endoscopic images, achieving high accuracy in identifying dysplasia and cancerous changes with strong area under the curve (AUC) scores.
What potential role does AI have in diagnosing conditions like ulcerative colitis and IBS?
-AI can aid in diagnosing ulcerative colitis by distinguishing active from healing inflammation using endocytoscope images, and it may also be able to identify imperceptible abnormalities in IBS, potentially leading to faster and more accurate diagnosis.
What is Dr. Spiegel's perspective on using AI in routine gastroenterology practice?
-Dr. Spiegel believes AI has the potential to significantly enhance gastroenterology by assisting with polyp detection, reducing radiation exposure, and identifying microscopic abnormalities. However, he stresses that more research and real-world validation are necessary before widespread adoption.
What are some future challenges for AI in gastroenterology, as discussed by Dr. Spiegel?
-Future challenges include conducting more prospective validation studies, proving the impact of AI on patient outcomes, assessing cost-effectiveness, and figuring out how to integrate AI tools into routine workflow without disrupting the clinical process.
Outlines
well hi I'm Brennan Spiegel I'm a,gastroenterologist...
This paragraph discusses well hi I'm Brennan Spiegel I'm a, gastroenterologist here at Cedars-Sinai, and director of Health Services Research, for our health system and today I'm, going to give you a brief overview on, some of the exciting and fascinating, updates in the use of artificial, intelligence in gastroenterology this is,a big topic why don't we get started, here I'm going to pull up my slides...
good example of the difference between,humans and computers...
This paragraph discusses good example of the difference between,humans and computers here's this unusual,situation all the humans know, instinctively what to do the computer,thinks it's just going to cut everybody,off and that's just not going to fly now, this is an even more concerning,situation that occurred a few years ago,with this woman who just struck, um this passenger I'm sorry this,this woman who is in the passenger seat,I should say was not...
Mindmap
Keywords
💡Artificial Intelligence (AI)
💡Natural Language Processing (NLP)
💡Machine Learning
💡Image Recognition
💡Endoscopy
💡Convolutional Neural Networks (CNN)
💡Barrett's Esophagus
💡Automated Polyp Detection
💡Area Under the Curve (AUC)
💡Deep Medicine
Highlights
Introduction to AI in gastroenterology: AI is not replacing human intelligence but augmenting it with technologies like natural language processing (NLP) and image recognition.
Definition of AI: AI is a branch of computer science simulating intelligent behavior in machines, like passing the Turing test to imitate human intelligence.
AI's potential in gastroenterology: AI can help evaluate images in real time during procedures, like colonoscopy, and aid in diagnostic processes.
Natural language processing (NLP): AI has advanced in understanding and processing human language, highlighted by systems like IBM Watson, which outperformed Jeopardy champions.
AI in vision systems: AI in dermatology and gastroenterology helps classify conditions, like skin cancer and Barrett's esophagus, through image recognition.
The importance of image quality: Research shows that AI's accuracy may decrease with low-quality or real-world images, pointing out the need for more real-world validation.
AI for capsule endoscopy: Deep learning systems can effectively analyze thousands of images in capsule endoscopy to detect erosions, ulcers, and other abnormalities.
AI-assisted polyp detection: AI has shown 96% accuracy in identifying polyps during colonoscopies, leading to higher adenoma detection rates in randomized trials.
Machine learning for CBC-based predictions: AI systems analyzing complete blood counts (CBC) can estimate colon cancer risk without relying solely on colonoscopy.
AI for inflammation in ulcerative colitis: AI can distinguish active from healing ulcerative colitis using an endocytoscope, showing high sensitivity and specificity.
AI reducing radiation exposure: AI in interventional endoscopy can reduce radiation exposure for patients and providers by deciding when radiation should be applied.
AI and IBS diagnosis: AI has shown promise in identifying subtle abnormalities in IBS patients, sparking debate about its potential to revolutionize IBS diagnostics.
Potential overuse concerns: Some experts argue against using colonoscopy for IBS diagnosis, but AI's ability to detect imperceptible differences offers new biological insights.
Adoption challenges: AI still faces validation, cost-effectiveness studies, and workflow integration issues before being widely adopted in clinical practice.
AI and medical board exams: The speaker discusses using AI (ChatGPT) for medical board exam preparation and compares its performance with humans, showing its current limitations.
Transcripts
well hi I'm Brennan Spiegel I'm a
gastroenterologist here at Cedars-Sinai
and director of Health Services Research
for our health system and today I'm
going to give you a brief overview on
some of the exciting and fascinating
updates in the use of artificial
intelligence in gastroenterology this is
a big topic why don't we get started
here I'm going to pull up my slides
right now and start off with my cover
slide
I have no specific conflicts of interest
as it pertains to artificial
intelligence
now we now know that
social intelligence is crashing Upon Our
Shores you've probably seen chat GPT
I'll come back to that and
we're figuring out how do we combine the
human brain and all of its innate
capabilities with computers
and in some cases computers can easily
out thank us
in other instances computers are
virtually useless right now and so a
term that's starting to evolve rather
than thinking of all of this as
artificial intelligence is augmented
intelligence that's the other AI can we
augment our natural abilities as humans
and the Consciousness that we bring to
the world with machines and where is
that interface in medicine and in
Gastroenterology in particular and
that's what I'm going to talk about
today
so let's start with some basic
definitions you know what is artificial
intelligence well here's sort of the
Webster textbook definition it's simply
a branch of computer science dealing
with the simulation of intelligent
behavior in computers
it's also the capability of the image of
a machine to imitate intelligent human
behavior you may have heard of a touring
test which was put forward by Alan
Turing which was a test to see if a
human can tell the difference between
another human and a computer when faced
with the same sort of set of
conversations or interactions and we're
starting to see computers that can
pretty easily pass the touring test and
imitate intelligent human behavior
now artificial intelligence has many
different components we're not going to
go into a lot of detail today I'm going
to spend most of our time talking about
natural language processing and and
image recognition which is particularly
important in gastroenterology but I just
want to emphasize that there are many
different branches machine learning NLP
vision and speech systems and within
those there are even finer distinctions
that we're not going to go into today
which is more of a largely clinical talk
rather than an overview of AI but I just
want to be complete in showing you the
extent of what AI is now considered to
be the different methods involved
so let's talk a little bit about natural
language processing which is one area
where AI has really shown incredible uh
capabilities and you know we've heard
about this years ago when
um the Jeopardy champion here went up
against
at the time the IBM Watson computer
which has now really been eclipsed and
this was an amazing moment for
artificial total legal ease for 1200.
Watson what is Executor right
same category 1600 answer
there are the developers very excited to
see Watson the Double Jeopardy question
and it was at that moment that
um he knew he was going to lose to this
computer so that was a big moment sort
of in the public
um perception of AI uh other examples
are you know chess Grand Masters like
Gary Kasparov you know being defeated by
deep blue and computers being able to do
amazing work on a chessboard for example
now machine learning is another really
important area and we've been hearing a
lot about that with regards to
um driverless cars and automated driving
and of course Tesla's made a lot of
inroads although continues to struggle
with obtaining you know human-like
driving capabilities this is um waymo
that's been working on a form of
automated driving for many years I've
routinely as recently as yesterday see
this vehicle driving around Los Angeles
they are constantly testing out their
systems extensively but this is a video
years ago that's smooth though this car
is approaching an accident look at all
the human drivers pulling to the left
using their turn signals they can see
the lights from a long ways away the
emergency lights in that intersection
they know they can't go straight through
the intersection they're all merging
left cutting into the line probably
looking each other in the eye the waymo
decides now it's going to cut in the
line way ahead of about 50 people it
appears that the driver took over and
moved off to the right so just a really
robotic looking decision so a really
good example of the difference between
humans and computers here's this unusual
situation all the humans know
instinctively what to do the computer
thinks it's just going to cut everybody
off and that's just not going to fly now
this is an even more concerning
situation that occurred a few years ago
with this woman who just struck
um this passenger I'm sorry this
this woman who is in the passenger seat
I should say was not paying attention in
the driverless vehicle then all of a
sudden strikes a uh pedestrian and you
can see the moment that she looks up and
sees that she's actually striking his
pedestrian who who tragically died so
you know AI is here but it's not
completely here yet now Vision systems
have been particularly Advanced
particularly in medicine and this really
got its start with Dermatology which
makes sense looking at the surface of
skin can send photographs to your
dermatologist they can make a diagnosis
often just by looking at the image but
what if we can use a computer to
classify skin cancer using in this case
deep neural networks what you're looking
at here is a series of receiver operator
characteristic curves or Roc curves
you'll recall that the area under the
curve of 1.0 is a perfect perfectly
accurate test and here what they did is
they classified images by expert
clinicians and then train the computer
and they found that the AAC for cancer
melanoma and so on was extremely high
with an AUC and the 0.94 plus range
extremely accurate
now of course computers can get fooled
this is a classic image of puppy dogs
and fried chicken they look suspiciously
similar and computer can sometimes be
challenged in distinguishing these kinds
of things or here are dogs and blueberry
muffins which also look shockingly
similar but you know humans can
immediately tell them apart computers
can struggle so we still have a lot of
work to do there this book by Eric topel
called Deep medicine how artificial
intelligence can make Healthcare human
again really talks about the limitations
of computers but also the opportunities
ironically to use computers to allow us
to actually increase the humanity in
healthcare possibly by giving doctors
more time to do the things that humans
do best like Consulting with patients
and conducting shared decision making
and empathizing and understanding the
emotional components of Health Care and
so in this there's a table here of the
so-called outlandish expectations for
artificial intelligence and Healthcare a
partial list you know it's going to
outperform all doctors it's going to
diagnose the undiagnosable treat the
untreatable we're hearing a little bit
more of this these days and this book's
several years old now
but we are not going to lose our jobs to
computers in general uh if we figure out
how to continue to evolve as doctors the
doctors that don't use artificial
intelligence might end up getting
replaced faster than those who do and
that's a theme that comes up frequently
and why we're talking about this in this
conference now Chachi BT came out in uh
just recently in
um November and you may have seen this
it's pretty incredible so I just
literally typed it into it I'm giving a
talk
about the use of AI in gastroenterology
what should I discuss that's a question
I just posed to chat GPT and you can see
how quickly
and starts to answer the question the
use of AI and GI has been rapidly
evolving and here are some ideas you
might want to talk about automated polyp
detection and AI systems being trained
to detect polits and colonoscopy images
diagnostic support to help diagnose IBD
or Celiac or IBS as I'll show you today
risk prediction for various conditions
like predicting colon cancer risk or
Barrett's esophagus risk personalized
medicine using AI to predict how
patients will respond to different
treatments image analysis to look at
endoscopy images or capsule endoscopy
images okay so the computer has already
written my talk for me there are many
exciting developments in the use of a
ingi your talk could focus on any of
these topics so there you go so let's
get started here with that introduction
talking about endodermatology you know
we think of GI in a way as like
Dermatology but we're looking at the
inside so we know that ai's been very
effective in uh x extrodermatology what
about endodermatology so we always spend
time looking at these lumps and bumps
and figuring out which ones are
concerning and which ones maybe not so
and just think about all the
opportunities for computers to help
assist gastroenterologists in evaluating
in real time those images this is an
article published in the red Journal a
couple years ago looking at the promises
and the pitfalls of AI and endoscopy and
so there's been a lot of efforts to use
AI to characterize the surface features
of polyps and estimate a probability of
it being a tubular adenoma or anything
else to help us decide whether to even
remove it although I think we're going
to for the most part remove these
whether maybe we can discard them
without sending them to pathology if the
computer are so sure that something is
let's say a hyperplastic polyp these are
open questions so there are many
opportunities for GI with
endodermatology Barrett's esophagus
gastric cancers small bowel ulcers colon
pulse IBD even IBS as I'll show you in a
second so I'm going to talk about a few
of these today starting with the top
looking for Barrett's esophagus
associate neoplasia using wide area
transepithelial sampling and AI program
so a computer can actually look at the
histopathology and quickly identify
suspicious suspicious histopathology and
in this randomized trial in which there
is AI assisted review of the histopath
it increased the diagnosis of high grade
dysplasia and cancer by over 14 percent
that's really quite remarkable and
something we need to be thinking about
to support our Pathologists this is a
study using convolutional neural
networks to diagnose The Invasion depth
of gastric cancer again using images
based these are conventional Endoscopy
in this case quite sophisticated Ai and
once again looking at the area under the
curve comparing to let's say experience
endoscopy or Junior endoscopus computer
did pretty much just as well as the
experience endoscopus and estimating the
depth of penetration based upon the
surface features so again really
important opportunities for clinical
gastroenterology with a very high AUC
what about for detecting early gastric
cancer this is a study that evaluated uh
AI for early gastric cancer and the
study looks at a number of outcomes but
I just want to highlight some of the key
features in this text here as it says
here there remains considerable
selection bias amongst the images that
have been used so this is an opportunity
to improve outcomes and in addition it
says that only high quality images were
selected performance may suffer when the
software is confronted with the varying
image quality and distractors an
endoscopus routinely encounter in real
life so this is just pointing out that
we need to think about these caveats
before we go Hook Line and Sinker with
all of these studies there are really
some issues that need to be addressed
and they Pro they propose some steps to
really help improve some of the research
here the first is to use real world
quality images rather than highly
selected images and also to do
prospective evaluation so we need to do
more of that
now another area that has achieved some
interest is using automated detection of
erosions and ulcerations with wireless
capsule endoscopy again using this deep
convolutional neural network approach
again using surface Imaging in this
study the capsule can look at all sorts
of interesting features and once again
has a very high AUC in terms of
predicting evidence of ulcers and
erosions and that can be a big deal
somebody's on NSAIDs or has
surreptitious bleeding and we've got you
know 50 000 images to go through
wouldn't it be nice if the computer can
really do the work for us and so this is
an example showing that that can be
achieved with a very high AUC another
study here showing the same thing
now what about using AI to identify
histologic inflammation in the setting
of ulcerative colitis this is another
study where they used an endocytoscope
to help look for evidence of ulcerative
colitis using artificial intelligence to
distinguish active from healing
ulcerative colitis and so the
endocytoscope here has this a camera
that presses right into the mucosa and
creates these beautiful images you can
see on the right it almost looks like an
actual histopathologic specimen and
using this approach which has been
around all for a long time but
augmenting it with artificial
intelligence these investigators found a
pretty high sensitivity and specificity
and very high positive negative
predictive value and accuracy for
diagnosing or distinguishing active
inflammation versus non-active
now here we're going to turn our
attention to the use of artificial
intelligence in colonoscopy
um and so there's been a lot of very
fascinating work being done here some of
you may have used this system before in
your own Practice A system that will put
a box around polyps in real time now as
I look at these polyps I feel like I
would have seen them but some of the
smaller ones you can sort of Miss
obviously there can be false negatives
false positives true negatives true
positives two by two table but the idea
is this is pretty unobtrusive and can
help quickly identify lesions as we're
doing the procedure and so in this study
this was a study finding real-time 96
accuracy during screening colonoscopy in
identifying polyps and this was a
randomized prospective trial to see
whether use of the system can improve
the yield of endoscopy or colonoscopy
and importantly what they did find is a
higher adenoma detection rate of 29 in
the the AI supported colonoscopy versus
only 20 now both these are pretty low so
one questions if more experienced
endoscopists perhaps would find quite as
big of a Delta but nonetheless in this
randomized trial they did find a
difference including the mean number of
adenomas detected overall
now uh this is another study looking at
predictions of findings at screening
colonoscopy using a machine learning
algorithm based on complete blood counts
so this is a machine learning algorithm
out of Israel that simply looks at the
CBC through the electronic health record
is pulls up the CBC and all of its
components including you know the
differential and from that can put an
estimate on whether there might be colon
cancer this is a pretty incredible
potentially high yield way to screen for
colon cancer you can imagine there might
be false negatives false positives but
they did find some accuracy compared to
colonoscopy so we're don't even need to
just rely on colonoscopy colonoscopy
colonoscopic image analysis but also on
other types of data available to us
uh and this is a sort of review looking
at AI for the determination and
management strategy for diminutive
colorectal polyps hope hype hype hope or
help and this is a study that suggests
that we could use this to support a risk
resect and discard strategy in other
words if the Imaging at the time of
polypectomy can reliably distinguish
whether a lesion needs uh
histopathologic review or not that would
be really helpful if you look down here
the 97 not negative predictive value
meaning that if this if the AI was said
there was no problem you could pretty
much assume there's no problem and
potentially does it discard and save a
lot of money
now this was an interesting study
looking at Interventional endoscopy and
seeing if AI could be used to reduce
radiation exposure so they use an AI
system to decide when radiation should
be used or not based upon the images
that are on the screen and they're able
to demonstrate a significant reduction
in radiation exposure to both patients
and providers in the use of AI for this
interesting application
now I'm just about done but I want to
point out just over the weekend saw this
tweet that came from kenwen SIA who's on
Twitter and he just pointed out this new
study that is using I colonoscopy with
AI to diagnose IBS and that's quite
interesting because we usually think IBS
as a normal colonoscopy well that's not
necessarily true there's been some
research going back several years that
some patients with IBS may have
erythematous patches maybe they have an
increased risk of
diverticulosis for example but they used
a computer to distinguish and with a
high degree of accuracy with an area of
the curve of 0.95 could distinguish IBS
versus no IBS and so that seems really
interesting but then there's certain
sort of you know key opinion leaders
that sort of weigh in this is Alex Ford
from the UK who says well well you know
given that we should not even be
performing colonoscopies and people with
IBS uh this seems like a pointless
exercise
well I don't know if it's pointless I I
honestly feel that's short-sighted
because yes we're not suggesting no
one's suggesting right now that we
should be doing colonoscopies in people
with IBS this is a research study but
isn't it interesting if it's true that
there might actually be
abnormalities that are otherwise
imperceptible to human eyes and if so
what is the biological consequences of
that what can we learn about the biology
of IBS for example if it turns out there
are interesting differences that a
computer can identify and maybe we don't
need to do a colonoscopy in everyone but
maybe we can simply do a flex Sig in the
office and quickly diagnose IBS and
avoid a whole bunch of future tests the
point is that no one's going to change
How We Do you know IBS work up right now
but that shouldn't hold us back from
thinking creatively about using AI to
hopefully rigorously distinguish among
different conditions so I think whenever
you hear this is pointless or this is
useless you know you got to think about
about the technology adoption life cycle
we're very early in Ai and there are
certain innovators that will put forward
ideas and then there's sort of a Chasm
you need to have a certain number of
innovators using something before early
adopters get on board then the early
majority of the late majority and the
so-called laggards who really kind of
ossified want to keep the status quo
honestly I think this is about how open
are you to new ideas if you're not real
open to new ideas you'll say something
like that last study was just quote
pointless and of no use at all I think
that's really short-sighted to be honest
all right well we have many challenges
so we need more prospective validation
studies we need to demonstrate the
impact on patient outcomes we need to
evaluate whether it's cost effective
study how to implement this into routine
workflow and these are all things that
are not yet a clear some use cases are
more advanced than others I will say
that I recently posed a series of
questions from the the GI board exam
based upon our books that we've written
these acing the GI board books to chat
GPT and um well it definitely struggled
with some of the questions but it did
really well on others and so we're doing
a study now comparing it to humans so if
you guys are willing to take an
anonymous quiz
um I'm going to actually distribute that
on my Twitter account and you can go in
and take the quiz we're going to see how
you do compared to the computer it's
Anonymous so it doesn't No One's Gonna
know how you did just don't look up the
answers we want to see how uh computer
the computer compares to the humans I
actually subjected
um chat GPT to that quiz it really
didn't do great so I asked it to draw an
oil painting of itself struggling
through a medical board exam and it
actually Drew this depiction on the
right of itself taking the test which I
find absolutely fascinating that it can
draw a picture like that all right I'm
going to end there thank you very much I
hope this was a useful albeit brief
overview of artificial intelligence in
general and for gastroenterology a lot
more to come we'll be hearing incredible
updates over the next year and Beyond
and stay tuned to this space thanks so
much
Ver Más Videos Relacionados
4 Ways Artificial Intelligence is Transforming Healthcare
The Rise of Artificial Intelligence through Deep Learning | Yoshua Bengio | TEDxMontreal
Barbara Gallavotti | Che cosa pensa l'Intelligenza artificiale
IMT - AI (1) - Apa itu AI?, Sejarah, dan Penerapan AI
China's AI Hospitals Are More Advanced Than You Think
Smart Nation 2.0: Professor Joseph Sung on artificial intelligence in science
5.0 / 5 (0 votes)