Good morning to the participants. Thank you to Dr. Kim and the ASG for the invitation to speak today. We're going to go over artificial intelligence and gastroenterology. And really, I want to give you sort of a broad overview of this field as it currently stands, and then to talk about essentially any future hurdles that we might think about and how we should possibly be thinking about this as we move forward. And so let's start here. We just went over the outline with what is artificial intelligence. So artificial intelligence is really the ability of a computer to perform tasks similar to a human being. And what this will involve is everything that a human being takes in to process an output. So we take in, recognize speech, we recognize images, we process that, and then we make a decision. And so that is what artificial intelligence is aiming to emulate. So artificial intelligence has a lot of terms. And essentially, you think about it sort of as a large umbrella under which there are different kinds of computer systems. So machine learning is a system where a machine can learn, so to speak, from data, but it's not actually programmed. So it's inputting a lot of data elements, and it's going to give out certain results. Deep learning is a form of machine learning where the algorithm is now developing what we call neural networks. And it's trying to use these layers of nonlinear processing. So it's not one to the other, but it's trying to take in different parts of an image or data. And then it extracts, the computer will extract relevant features, predict an outcome. And computer vision sort of overlaps all of these because you're introducing image data or video data, and that is going to require a different level of computing. And so these terms are often overlapping. And an important thing to recognize is that you can think of AI terminology in terms of what is that AI aiming to accomplish? Are you trying to understand an image? Are you trying to recognize what the provider is trying to say? So are you recognizing their speech? And how are you doing it? Are you doing it just with plain machine learning? Are you doing it with deep learning algorithms, with convoluted neural networks? And so really, they're all part of a large bucket of artificial intelligence. But oftentimes, a particular task requires multiple inputs and the contribution of different kinds of computer learning to be involved. And I think this has represented the burdening of this field in the publications. And so this is just all of medical AI publication. And you can see that in the last few years, there's been an exponential growth in the number of publications related to medicine in the field of AI. But GI very closely mirrors this. And so we've had a huge uptick in people interested in the field, developments in the field. And this is why it's a really exciting time to be part of the application of AI to GI. So this is a short list of the approved devices. Most of these have been approved in the European Union. In the United States for imaging, both arteries and Poseidon ultrasound or butterfly network have been approved for liver lesion detection. And as many of you are aware, there's devices that are approved for colon polyp detection by the FDA. And these include GI Genius from Medtronic, Endoscreener from Wision, and most recently, the Scout system from Iterative Health. So today, what I'd like to do is kind of go over with you sort of the major publications to give you a sense and understanding of where AI is being applied within the GI tract. And I refer you to this review article, which is quite helpful to give you an overall understanding of AI. But it's being used all the way up in the pharynx to detect pharyngeal cancer when patients are having in the upper endoscopies, both white light and narrowband imaging. You can use it to detect cancer in the esophagus as we'll go over both adenocarcinoma and squamous cell cancer for detection of GI cancers, H. pylori atrophic gastritis, and increasingly to understand the extent of examination performed and quality control is going to be a big issue going forward, both for upper and lower endoscopy. It's also being used more in a natural language processing form for predicting the risk of upper GI bleeding and what maneuvers need to be performed. You can use it now for endoscopic ultrasound. So there's diagnosis of gist in the stomach, and then in the liver, we're using it to detect liver lesions, both on endoscopic imaging as well, that's primarily by EUS, and by cross-sectional imaging, including ultrasound, CT, and MRI. And then there's a lot of work being done in natural language processing of large data to predict the risk and the outcome for patients with fatty liver disease, viral hepatitis, PSC, and to predict which patients might benefit from liver transplant, which may not. Similarly, in the pancreas, we're using it for detection of disease activity, localization of a tumor, prediction of outcomes of surgery. And then moving forward into the small bowel, there's been a lot of advances. Really, they're quite pioneering in the use of artificial intelligence in video capsule endoscopy. And this is for detection of lesions, detection of bleeding sources in patients with obscure GI bleeding. It's also being used to stage Crohn's disease increasingly in the small bowel. And in the colon, of course, CAD-E is what most of us know, that's computer-aided detection of polyps and cancer in the colon. But really, what's more exciting to me is CAD-X, which is the differentiation of lesions. Can you detect which of these are going to be benign versus malignant? And we'll talk about this a little bit more. In inflammatory bowel disease, there's a lot of work being done in scoring of disease activity, because this is something that is very nebulous. It's a lot of inter-provider differences. And so, even for clinical trials right now, there is a need to have a central reader where someone reads all studies so that there's no difference in how it's read. But if AI can take that over and make that standardized, that's going to be really exciting. And there is a lot of work and progress done in that field. My own work is in the detection of dysplasia in IBD. And so, this is a very difficult population because these lesions are often varied morphology. Some of them can be very flat. Some of them can stick out like regular polyps, but they're not actually polyps. They're just benign lesions. And so, it's going to be challenging, but it's something that we're very excited to do here at Mayo and Rochester. And then, again, looking at patient-reported outcomes, patient app data to check for what might be causing triggers for irritable bowel syndrome and other such patient-related symptoms. So, we're going to go through this by the organ. I'm going to give you sort of highlights of the most recent or relevant outcomes. And so, in Barrett's esophagus, computer-aided detection systems have been shown to detect Barrett's esophagus with higher accuracy than expert endoscopists. And so, Barrett's is an extremely common condition, but it's very hard if you're not trained to find these abnormalities. And you can see here that the AI system was able to delineate with nearly perfect accuracy the area that was abnormal, and this allows you to target your biopsies. But I want you to look on the left side of the screen here to how many data sets and how many images were required to train the system. And this sort of gives you a sense of the burden of annotation that AI systems for image being used for image analysis currently require. You can see here five data sets, large volume of images. These have to be manually marked because this is what creates the ground truth to train that AI algorithm. Similarly, we found that you can combine different AI algorithms. And so, this was a paper published, which in the first step just detects your Barrett's and shows you in white light endoscopy, so your regular light that you do during a procedure, where the Barrett's might be, what areas to maybe focus on. And then you can switch to a different form of light imaging. So, for example, this is the Olympus scope, which comes with narrow band imaging. And you can switch to that and focus only on the areas that the initial algorithm tells you are high risk. And you could really identify now lesions that are high risk versus low risk. It allows you to target your biopsies. And this then allows you to have a much more tailored approach to the patient. Similarly, for esophageal squamous cell cancer, something we don't see very much in the West, but very common in the East is you can use computer aided diagnosis or CAD systems to really delineate the lesion. So, you can see in panel A, up and down, that's what the endoscopic image looks like. But now you can diagnose the site of the cancer by the overlay of the AI. Similarly, it predicts high grade dysplasia as well as low grade dysplasia. And these are very difficult, very subtle lesions. If you look at the endoscopic view, a physician that's, you know, doing a quick EGD may miss this. And therefore, that's where the importance of these AI systems really comes into play. Similarly, when you have in any of these cancers now, as we advance all of our endoscopic techniques, our goal is to try and take out as many of these with the endoscope. And so, there are techniques including endoscopic mucosal resection or EMR, or endoscopic submucosal dissection or ESD, which goes to an even deeper layer. But it's very important to recognize that sometimes even ESD is not adequate to take out a lesion. And you do require to have, the patient has to have a surgical resection. So, if you can use an AI algorithm to predict for you what the depth of invasion of that lesion is in real time, this is going to be really helpful to decide whether you should attempt an endoscopic resection or just send the patient to surgery. And AI systems are being used for this as well. And on the left panel, you can see it's being used for patients with Barrett's cancer, where T1B lesions are really ideally referred for surgery. And T1A lesions ideally should be removed with ESD. And so, you can see here how this might significantly change a patient's outcome. And similarly, for squamous cell cancer in the right panel. As just to give you an introduction of what else is happening outside of endoscopy, digital pathology is really gaining traction. And so, whole slide histopathology is being used to differentiate between dysplastic Barrett's, which might progress onto cancer and non-dysplastic Barrett's. And this is an area that requires a lot of, there's a lot of uncertainty. So, even now in real time, when a patient is diagnosed with Barrett's esophagus in the mucosa, this requires confirmation by a second pathologist. It really has significant implications for the patient. And so, but non-dysplastic Barrett's is really quite safe. And so, it's really important to differentiate between them. And that's why the application of deep learning here is really exciting. Because now, if once you have longitudinal databases, you can then follow up and make sure even on a long-term basis that the outcomes and the predictions of the AI were in fact correct. Another area that is being used now is this concept of wide area sampling. So, rather than just biopsying a specific spot, if you do diagnose Barrett's, there's a company called CDX, which currently does this 3D sort of computer analysis of the brushing. So, you just brush an entire large area, and then you're not worried that you might miss something. And this is going to find high-grade dysplasia adenocarcinoma. Then you go in with the second exam and try and identify the exact spot. But for large-scale screening, this could have a significant impact. Similarly, as you're trying to decide now you have a patient who has cancer, what are you going to do? Well, your patient may require radiation. And so, this is just to give you a sense of what's going on outside of the field of endoscopy. Automated segmentation of the esophagus from CT imaging helps radiologists to decide what they need to do. There's also 3D reconstructions that can help surgeons plan their surgical approach. So, switching gears from cancer to more benign conditions. Eosinophilic esophagitis is increasingly detected, and we believe the incidence is also increasing across the United States. So, this is a condition which is benign. It's where you have infiltration of inflammatory cells into the esophageal wall. This causes patients to have inflammation and edema in the wall, causing them to have strictures. They might present oftentimes with food getting stuck in their throat. But it's a very subtle disease. It can be, you know, the features are not classic like a big polyp or something sticking out. And so, you want to have an ability to identify this perhaps before the patient presents with an obstruction or an acute setting. And so, the AI is now being used to train to identify these subtle features of EOE. And the gold standard here is biopsy. So, you biopsy and you know what's going on. And you can see here that the AI seen in the dark blue line here performed better than even consultants and, of course, fellows and beginner trainees. And so, this is doing really…this is exciting because you can, again, screen large populations for a condition that is easily treated. Similarly, there are motility disorders now being studied. So, oftentimes, the esophagus may appear normal, but patients still have trouble with the way the esophagus moves. And so, we do a test called a high-resolution manometry. And there are certain classic features, and then there are a lot of undifferentiated features. And so, this is an area…this is published earlier this year, which is really exciting because if you can identify at least the known swallow types, which is what it was…the study was aiming to achieve, then the next step is maybe you can use more of a unsupervised learning and see if combining this high-resolution manometry data with patients' clinical data, you can now try to sort out all of these undifferentiated motility disorders that patients do present with. And we can maybe try and sort these out with the help of large machine learning datasets. So, let's move forward into the stomach here. We're going down the GI tract. It has…AI has been used to diagnose H. pylori infection in the stomach. So, again, this is not as common in the West as it is in the East, but where H. pylori is ubiquitous. And here, you can diagnose H. pylori gastritis, so inflammation with high accuracy. And what this is helpful for, then, is that this could obviate the need for biopsies to assess whether there is H. pylori or not, if you can just do it with imaging. Similarly, a consequence of H. pylori infection is atrophic gastritis, which then leads to other complications, such as neuroendocrine tumor formation. And so, if you can diagnose areas of atrophic gastritis, then you're going to be able to focus your examination on these areas. And this study, again from China, showed that there was…even compared to experts, the AI algorithm was able to diagnose areas with higher accuracy. And the way they did this was once the expert or the AI algorithm marked these areas, biopsies were taken to confirm, and that's really the gold standard there. Looking at gastric cancer, oftentimes, this can be very subtle, again, hard to find. And so, this is an interesting study because they used entire videos to diagnose the early gastric cancer. And that is really hard to do. It's easier to do on still images, but the stomach has a larger area, surface area, than the esophagus. And so, identifying that exact location would be hard in images, but this study was able to show a significant area under the curve. It's not above 0.9, which is really what you want to see, but it's a good first step considering they used video data, which is moving and requires a lot more annotation. Similarly, once you identify a cancer, the AI has been shown now in multiple studies to be able to predict the depth of invasion. Again, very important because depending on how deep that cancer is, it's going to make a difference for the patient about whether they have a noninvasive endoscopic resection or require surgical resection of the entire segment of the stomach. And you can see here that for depth of invasion of gastric cancer, accuracy was close to 95% for all types of light. So, the first row shows you white light, the second one shows you narrowband imaging, and the third one is with staining with indigo carmine, which is often used to delineate lesions. And you can see here how good the system was, and not just one, but multiple studies have been shown to predict invasion. So, these were a couple of meta-analyses that I'll show you that show that overall, AI has extremely high accuracy for all neoplastic lesions of the upper GI tract. The esophagus and the stomach. Duodenal cancer is extremely rare, but there is work being done, including at Mayo and Rochester. We're looking at duodenal adenomas and identification of these lesions and potentially then prediction of the malignant potential, especially in patients with a condition called familial adenomatous polyposis, where they develop lesions around the ampulla. And these require either an extensive procedure or a resection. And so, identifying these and predicting their malignant potential will be very important. Similarly, I'll show you another meta-analysis showing that no matter where you looked at neoplastic lesions, the endoscopists in the solid line here, the predictions of the endoscopists were not as good as prediction of the AI. And this includes barrett's esophagus, stomach cancer, as well as H. pylori status. And so, this is now solid evidence that AI can help with imaging studies. It can be an adjunct to the physician's procedure and an important one of that. So, now we're looking at quality. And this was an interesting study. This is work we're also doing here at Mayo looking at adequacy of an EGD exam. But an upper endoscopy really can be, you know, quite variable in its quality only because there are certain folds and areas of the stomach that have to be examined. And so, this was a publication which looked at a real-time AI system that showed participants in the active group whether they were seeing enough of the stomach or not. And you can see that there was a significant reduction in the number of blind spots in the upper GI tract in patients that were given real-time feedback versus the controls. And it only increased the inspection time by one minute. But just having that feedback loop to say, oops, you haven't seen this segment or you need to look behind this fold really improved the quality of the exam. And I think going forward as we implement more and more of these AI systems in our practice, at some point reimbursement is going to be linked to whether you've done a good enough quality exam or not. So, moving now to machine learning to give you a sense of developments in that space. Upper GI bleeding is one of the, or really any GI bleeding is one of the commonest indications for ER visits. And as we use more anticoagulants is a higher risk for patients. And so, this study looked at a machine learning algorithm that took the patient's symptoms, included all their current vital exams when they presented to the ER, as well as their lab testing. And then they, clinically, we have clinical risk scores that we can check to sort of using the same data to predict the risk of the bleeding so that we can decide if the patient gets discharged home without patient follow-up or requires hospitalization or requires urgent endoscopy. And looking at this, the machine learning algorithm was able to identify with more specificity the number of patients that were safe to be discharged, as opposed to even the risk scores that we use. And these risk scores are not easy to calculate, they're cumbersome, they don't get used often. So, if you had a system that just spat out a risk of bleed at the end of their patient's summary visit in the emergency room, that would be really meaningful and would change practice. So, moving down to the small bowel now, like I told you, there's been a lot of development in the video capsule endoscopy arena. You can see here that CNN-based algorithms have been used to diagnose small bowel protruding lesions, as shown here. And you can see that the expert boxes here by expert endoscopist or capsule readers almost completely overlap with what the computer is showing you. And so, as I'll show you later, this is really helpful because you don't have to spend as much time then looking through images if the CNN has such high accuracy. Similarly, you can detect the bleeding causes in procedures in patients who have obscure GI bleeding, that's the most common indication for a capsule endoscopy. And you can see that the CNN here in the upper left panel detected even subtle lesions, such as a small erosion or a type 1a angioictasia. These could be missed by the human eye if you were scrolling through those images fast. But at least here, the AI can be used as an adjunct to draw your attention to that spot, and then you can identify that and clarify whether it's true pathology or not. Again, as mentioned, small bowel Crohn's disease. Crohn's disease is a rapidly growing area where capsule endoscopy is being employed. Initially, we thought it would be too high risk because it might get stuck, but more and more we're using small bowel capsule to assess for small bowel Crohn's disease. And you can see here that this AI system was able to classify this into either mild, moderate, or severe disease. And this is really helpful in terms of monitoring of patients as they go on medication and they come back for future procedures. And as shown here, this is really helpful data because you can see here that a full length capsule endoscopy video, even in expert hands, the endoscopist reading the procedure alone took 12 minutes, but after the first screening by AI, it was only 3.1 minutes and was with no missed lesions that the expert, that the endoscopist found without the AI missed. And so this is really helpful data. And I think moving forward will become standard of care. So let's move down to inflammatory bowel disease. This is an interesting study where data from IBD patients that's been posted in online forums was sort of scrubbed. It was examined and then cleaned out. And the annotation here was done by actual IBD patients. And the idea is to try to use natural language processing to predict when patients might be in a flare. This apply, you know, as more and more patients are involved in online forums or different apps to monitor their symptoms. If we could predict a flare prior to them having a significant flare where they require steroids or hospitalization, that would really be helpful. We would be able to prevent a lot of these by adjusting their medications, checking for infection. And so this is very early in the stage of development, but I wanted to introduce this idea of using patient level data or patient reported outcomes from social media or other apps to try and tailor their medical care. This is well known now. There's been a lot of publications in AI-aided scoring system for inflammatory bowel disease. We briefly talked about how difficult this is to standardize because different endoscopists based on their level of experience or exposure to IBD might rate an endoscopic inflammation as different degrees. And you can see here that the best performance was either healing or almost healing, which is Mayo zero or one compared to moderate to severe disease, which is Mayo two or three. To take that one step further, now AI is able to predict even histologic remission. So sometimes when the bowel looks entirely normal, you can take biopsies and there might still be inflammation under the surface. But now the AI imaging, and we don't know what it's looking at because of these black box algorithms, but it is able to predict just even a normal appearing colitis bowel, whether there is underlying remission or not. And this has really big implications once this data is well-validated because now you may not need to biopsy as much as you do currently. This would result in both patient safety savings as well as cost savings. And so switching gears now about quality of colonoscopy, and we'll go through a few quality parameters, but one of the parameters of a good quality exam is the patient's bowel prep. And so there are various scores used. The most common and most validated one is the Boston bowel prep score. And there've been multiple algorithms now that have been developed that can give you your Boston bowel prep score automatically at the end of your procedure. One of these was Shyam Thakkar's work, which was really pioneering a few years ago. And there've been other data that have come up over time where you can, at the end of your exam, will get a Boston bowel prep score or sometimes even real-time. Shyam's data also looked at the percentage of colon visualized. So are you seeing, you know, every five centimeters, he has an algorithm that will tell you every five centimeters, you visualize 80% of the last five centimeters of bowel. And perhaps if you set the threshold at 40% or 60%, you probably should go back, wash, and look behind that fold or look a little bit more to improve the quality of your exam. Extent of sacral intubation is a quality metric defined by the ASGE for a good quality colonoscopy exam, for screening colonoscopy. And so if you have a CNN that can identify that appendiceal orifice on its own, now you can identify extent of exam. This helps you to calculate then withdrawal time automatically, which is also another marker of quality and can be linked again to create. All of these variables are helping you to create an automated report, which is really, is going to improve physician efficiency as well as quality. Again, like we talked about, this is another study showing you where you can qualify your bowel prep as inadequate with high sensitivity. We've, I think all of us are familiar with the CAD-E algorithms, which help to increase adenoma detection rate. This has been shown in multiple studies. This is really the field, the area where I'd say in GI endoscopy, we've had the most advances. This is where the FDA has approved the devices for GI endoscopy. And really multiple now meta-analyses have shown that the use of CAD-E systems increases adenoma detection rate. It does improve quality of the exam. And more exciting to me is really CAD-X. So can you use CAD-X algorithms to really differentiate now whether the polyps that the detection algorithm has identified are really benign or malignant. And there've been meta-analyses on these done. A lot of the initial studies required advanced imaging. So either switching to narrowband imaging, blue light imaging, endocytoscopy. And they had a really good sensitivity and specificity area under the curve of 0.96. There's new data. There's one recent publication by Hassan this year, which shows that you do have, even with white light imaging, can achieve a significant sensitivity for CAD-X. And this, the system that they tried met the PV thresholds for both leave in situ, which means small hypoplastic lesions in the rectum, which you can just leave in place. You don't have to worry about it. Or the resected discard strategy, which is more proximal polyps because you know that they're benign. You can, you still remove them, but you don't have to send them to pathology. And since this met all of these different criteria, this is something that can really change practice, but would improve not only the quality of the exam, but also improve costs because you're not sending a whole bunch of unnecessary polyps to pathology. And so looking now at reporting, can you use large data to create, again, your automated reports? And so this was an interesting study where they did a manual review of a large section of colonoscopy report that do natural language processing alone or natural language processing combined with an optical character recognition in case someone had typed out something in particular, or there was a different finding. And you can see here that the, both machine learning approaches did almost as good as manual review. And most physicians are required to maintain reviews of all of these quality metrics in order to maintain certification. This requires, usually it's not a physician, but it's an office manager or someone manually doing a lot of this work. And so this is gonna be a really powerful technology once it's developed and validated so that you can now have an automated report. You can create your own personalized report that you can pull up on your cell phone or your tablet to see how you've done over the last 100 colonoscopies, for example. And so now moving really briefly into the liver and pancreas because that's not my area of interest necessarily, but there have been now prediction tools, for example, to predict the risk of hepatocellular carcinoma, which is a high risk cancer in patients who've already been treated for hepatitis B. And it really uses just 10 lab parameters, very commonly used. This is all chart review, and this is performing better than other risk scores that are out there. We've already talked about how there's multiple, even FDA approved technologies here to predict liver masses in patients and try to differentiate benign from malignant. We've talked about the prediction, but this is interesting. There's an increasing population of obesity in the United States leading to hepatic steatosis or fat in the liver and non-alcoholic fatty liver disease is predicted to be the highest reason for liver transplant in the future. And so if we can identify these patients to predict with simple, we're using, there is AI being applied to both lab tests as well as their basic physical exam characteristics, as well as to imaging techniques to see, well, you have fat in the liver, but this is actually advanced fibrosis. And if you can predict fibrosis in these patients, then you can obviate the need for a liver biopsy, which is a really invasive exam. And so again, being used in all these different aspects of gastroenterology. In patients who are admitted with cirrhosis, natural language processing tools are being used to predict which of these patients might develop hepatic encephalopathy, which is a common complication. Here at Mayo, we developed a tool because we have a very large population of primary sclerosing cholangitis or PSC patients. And this tool will accurately predict whether they're going to have liver failure and therefore the need for liver transplant. And this is really a very helpful tool going forward. Again, looking at pancreatic cancer, both from patient characteristics and natural language processing, and then also work done here by Mike Levy includes the use of EUS to improve finding the lesion, segmenting the lesion, planning operation of that lesion and whether it's operable or not. Looking at GI bleeds again, this was a very interesting study. They looked at patients, over 300,000 patients and applied three different machine learning models to the data that was already in the electronic medical record and they compared this to a validated bleeding prediction score called HASBLED. And you can see here in the image, all three machine learning models did better than the score that we otherwise have to manually correct. And so this would be as more and more patients receive blood thinning medicine, this would be a great tool to be able to predict which of these might develop GI bleeds in the future. So I hope I've kind of given you, this is not even completely comprehensive and we've spanned from one end to the other, but the future of AI as I see it is gonna be to transform all of these guideline-based practices that we currently have to look at a guideline and read up and then apply to more automated and patient-based practices. There is machine learning already being applied to predict the sedation need for endoscopic procedures. So for example, does a patient, would they be okay with conscious sedation or do they need a deeper level of sedation? And this is again, resource utilization, it's gonna help areas that don't have as many resources plan what they need to do. Like I mentioned, we're doing CADI and CADx algorithms in IBD where here where we are, there's other work being done on adequacy of endoscopic resection. Have you taken off all of the lesion? The goal is eventually one of my own goal, we have a large integrated Mayo GI network where we're gonna have universal recording of all of our procedures. And the goal is to kind of create a personalized, individualized endoscopic history record so that if patients come back, you now know exactly what they had the last time of in a video format, not requiring review of images. Just like you would carry your EMR on your phone, you would now have a little video record of your GI procedures. And like we talked about different aspects of quality, which I think will eventually affect reimbursement. So a few pitfalls as you think about AI, remember that machines at the end of the day can sometimes be binary, they might only go disease versus healthy. And therefore I think you need to really think about the question you're asking the machine to answer, the data you're giving it to answer with, and maybe it's just trying to refine differential diagnoses rather than these conditions, especially when it's a widespread, when it's a spectrum of disease. And so far, many of these studies have been single center. Some of them are multicenter, but most have been single center, mostly image-based. And so I think developing larger datasets would be important. As you think about large datasets, things to think about are data privacy issues. Do you, how are you going to protect the patient data when you're sharing across large sites? Is a solution federated learning buckets or cloud environments where you can perform this? Or do you need to think about ways that you can encrypt data so that they're not traceable? We talked about the large burden of annotation that currently is performed by medical professionals or at least supervised by medical professionals. This adds a significant amount of cost. And so do we consider bootstrapping annotation? Well, it decreases the quality of the data, but maybe that's something we have to think about. And so auto annotation is another thing, but again, has to be supervised at some level, which is costly. We have to think about who has legal liability for algorithms that are employed. Is it physicians? Is it the people, the companies that are making these systems? What are the ethics about using AI? Is there something, do we think about, are we doing patients any harm? Are we, we have to always think about negative predictive values and make sure we're not missing important findings that we otherwise may have acted upon. And eventually I think there's gonna have to be legislation to think about how you can employ these, what the reimbursement will be so that this helps to fund future research. And one of the other things that I think is important is that all of these algorithms are currently black box algorithms. It's hard to assess what that system is looking at, and therefore a little bit harder to assess whether that outcome is important or not, unless you have long-term data. And I think that's where large volume, good quality data is gonna be very important going forward. Just to give you a guide, since many of you are industry thinking about how you might apply these to trials, the FDA does require randomized controlled trials be performed. And in 2020, the consort guidelines that are created for randomized controlled trials added these following multiple criteria, actually for AI studies, but important ones were that you must be able to describe in detail what you were putting into the system and what you were expecting to receive out of it so that you could really understand what that AI is trying to do, obviously what the AI itself is, but then what skills did you recover? Did you apply this to expert gastroenterologists? Did you apply to trainees? Did you apply it to EMR that was acquired from a particular center or a particular scope? And that's where you really have to figure out describe what you're doing, describe the study setting. How did you deal with errors either on the human aspect or the ground truth aspect or the AI aspect? And then how did you interact with, how did the human have to interact with the AI? Was it a bounding box that showed up? Was it something that was done post hoc? And all of these are gonna have to be more and more transparent as we go forward. I think this is one of the reasons because there's so much regulation that really there've been only four devices or so approved in gastroenterology. And you can see here that the bulk 70% approved in 2021 were radiology devices and cardiology is another leader in this space. And so I think because our field is so large and we're so nascent, that's why it's so exciting to be here at this time. I also want you to think about ethics as we develop larger datasets and these cohorts in AI. How are we gonna protect the privacy and the confidentiality of patients? What does patient consent look like? Do we consent upfront for the usage of this data for development of artificial intelligence or not? We don't consent when, currently we do for apps on our phones that have to be applied to medical data. How are you gonna link radiology to pathology, to endoscopic imaging and the EMR? How are you going to ensure that there's equity in datasets? Just because you applied it to a large population, does that apply then in a different country? Does it apply in areas where there may not be as much of computing power? If you need a really high powered computer to run your AI, is that gonna apply in a rural setting? I think these are things that we need to think about to ensure that there's equity as we develop AI. We briefly talked about liabilities at the provider or the commercial entity, because at the end of the day, our AI should aim, the goal of all AI is to democratize the highest standards of care, no matter where you practice in what setting. We talked about giving, very important, think about diverse datasets. Think about, is the coding and the nomenclature the same across different EMRs? It may not be. If you're applying it to different countries or different areas, are there differences in cultural practices in that area? You know, is the same disease in different places called by a different name? And these are important if you're just teaching a machine what to do, because if you don't think about that, the machine definitely won't. I encourage all of you to read this paper by Dr. Folisadme published in Gut this year. It really focuses on ways in which AI has the potential to worsen health inequities. And I think if you go into this with your eyes wide open, it will really allow you to prevent these occurrences or these biases going forward. And finally, on AI acceptability, I will tell you that there is no published data on patient acceptability, definitely in GI, maybe in radiology, but I didn't really delve into that literature. There was a survey done for medical professionals that was published earlier this year, and what medical professionals would consider acceptable for AI. And you can see here on the right what these were. There's a lot of issues with trust of the AI, and then how much are you expected to do? How much liability is it gonna take away from you or add to you going forward? And so I challenge you to think about the future state in terms of, are you going to have, is AI in the future, endoscopy suite, for example, gonna be a bunch of different hardware boxes from different manufacturers that's gonna look like a Home Depot with different boxes and then you plug something in and et cetera, or are the algorithms with a master box that you can choose? I'm doing an EGD on a patient who's a high risk for Barrett's I think I'm gonna choose this one today and apply it. But today I'm doing an IBD colonoscopy. So I think I want to know how good or bad the disease is, but maybe this is not the best algorithm for screening. And so having a system where you could pick and choose and algorithms run with minimal space utilization, that would be the ideal future for AI in the endoscopy suite. And with that, I thank you. I'm happy to take any questions. I'm sorry, I think I ran over a couple of minutes, but it's a very vast topic to cover. Thank you very much, Dr. Kualaprabhu. This is a very important topic, of course, I think it is, and certainly you've shown a lot of our patient care. We have a question from Stacy. The question is, what will be the biggest driver for physician adaptation or adoption of AI as the technologies are rolled out? We've seen very slow adoption of AI for colonoscopy, even though the clinical data is clearly positive. Yeah, it's a great question. Again, I think we need more surveys of data. There was a survey of physicians that was performed for just looking at what you think might be, and this was from Tyler Berzin's group a few years ago, for colonoscopy AI, and there was a dichotomy. So academic providers felt that AI, CAD-E for example, would improve the quality of the exam, but private practice providers were really concerned about the time and the reimbursement. And I think this is where, I think if there is legislation that supports reimbursement for the application of AI, it's going to help, because right now there is no reimbursement. For example, if you put in a CAD-E system into your endoscopy suite, you are not getting paid any extra to apply that CAD-E. Now, you might take out more polyps and you might have more pathology costs if you're getting paid for that, but that also increases your procedure duration, which decreases your efficiency. And so that's why I think CAD-E has been really, the systems have been so wonderful to detect more polyps, but I think the ideology eventually is CAD-X, which will tell you which ones are worrisome versus not, and that will really change practice. But physician adoption right now is, I think cost is a big issue. Okay, and I'll ask this question because I know some of the people that are on this call are gonna be maybe interested in this, but do you foresee this being, a lot of the AI metrics you showed us were endoscopic images, which makes sense because it has to do with pattern recognition and being able to identify pathology versus not. Do you foresee this being more of a, it's part of the scope, you press a button, like we press for say MBI and all of a sudden we have the AI turned on or off, or do you foresee this being a extra box on the cart that allows us to do this and any device company, any scope company can utilize their technology in order to implement the AI? Yeah, that's a great question. And that was kind of my last slide there. I think right now as it stands, the FDA is approving boxes with each algorithm and I don't think that's gonna be feasible in the long run. I anticipate there'll be some sort of high-powered computing interface attached to your scope, in which case you'd be able to turn on and off whatever you wanted and it would just be APIs that would, your algorithm would be inserted into this high-powered computer. So for example, if you're a practice that's doing a lot of mobile carting in the hospital endoscopy unit, for example, that may not be the place that you need to have screening colonoscopy necessarily. These are more acute patients, a totally different algorithm. Whereas I don't see, I think the whole paradigm is gonna have to change because right now, like I said, the FDA approves boxes with different algorithms and it's not gonna be feasible to have 15 of those in your room, depending on as different companies produce different algorithms. So I think you should be able to sort of pay as you go for what you use and that should be, that would be the ideal reimbursement model. That will allow, you know, what that increases though is the risk of, you know, your algorithm is out there, it's not in a box. And so that does increase your risk, but I think that's how the future will be once these, once multiple different ones are available. Great, thank you very much for your perspective. Dr. Coelho probably will also be part of the panel discussion later in the day or life in the day of the gastroenterologist. So if there any additional questions that you have about AI or anything else that comes up throughout the day, please write them down and then we'll come back to her as well as the other speakers this afternoon.

artificial intelligence

gastroenterology

machine learning

deep learning

Barrett's esophagus

esophageal cancer

gastric cancer

eosinophilic esophagitis

motility disorders