Hi, my name is Seth Gross, and I'm a member of the ASGE AI Task Force. I'm going to talk with you, why machine learning in gastroenterology? These are my disclosures. The objectives of this talk will go over some key background around artificial intelligence, discuss the challenges the gastroenterologist faces in practice, take a disease-based approach to highlight why artificial intelligence in GI has tremendous value, looking at colon polyp detection, barotosophagus, gastric cancer, and evaluating small bowel disease. Artificial intelligence continues to grow with one key goal, to improve our overall quality in care in clinical practice. Specialties like ours, which are image-based, like endoscopy, potentially have the most to gain. Once an endoscopist maximizes their technique, AI may help address some of the clinical pain points amongst endoscopists. So what is artificial intelligence? Artificial intelligence has lots of terms. In a global sense, any technique that enables a computer to mimic human intelligence using logic, if then rules, decision trees, leading to machine learning, which is a subset of AI, including statistical techniques, enabling machines to improve tasks with experience, deep learning, a broader family of machine learning methods based on artificial neural networks and representing learning, followed by deep learning, a deep neural network in artificial intelligence with multiple layers of input and output, and lastly, the ground truth. This is a term used to check machine learning against the real world in terms of accuracy. So there's training used to look at extensive amounts of inputted data, allowing AI to potentially act like humans. So where are we? I think the current state is quite narrow. Right now, there's a lot of recognition or image recognition, but I think the future has a much broader impact of AI, where it exceeds human performance, reasoning, and complex tasks, such as writing a bestselling novel, for instance. So over time, we're going to see the continued growth of artificial intelligence. When I think of artificial intelligence, there's augmentation versus automation. Augmentation is what we're seeing now in colonoscopy, where a bounding box hovers over a polyp, and some systems even have polyp interpretation and characterization, which is mainly available outside the United States. And then there's automation, doing those mundane tasks before the procedure in terms of documentation, and after the procedure, making your report. Imagine machine learning doing this for you. So when we think about the clinical applications of AI in healthcare, there are a lot of areas where we're going to see its value. There's big data analysis, natural language processing, voice recognition. That could certainly help with that procedure report development. Robotics and endoscopy, which we've seen in the past, but I think we're going to see more in the future. There's a lot with image analysis. When we think about colonoscopy and polyp detection, the programs and software that's available today is really about identifying an abnormal area, such as a colon polyp, and then interpreting it. Clinical pathways, when patients present to the hospital or even in the outside practice, being able to better risk stratify, looking at the whole patient and looking at all their data to allow for statistical analysis and give some modeling. And these are things that we're starting to see a little bit in artificial intelligence in gastroenterology. So what are some of the goals of AI in gastroenterology? It's the goals that we always try to work on, but maybe AI can enhance that, increase our quality, improve our diagnostic accuracy, decrease variance of healthcare delivery, enhance our outcomes, and hopefully decrease medical costs. One of the key goals in healthcare today is to offer the highest quality, but balancing the costs of doing this by decreasing the overall healthcare costs. So what are some of the challenges with endoscopy? There's visualization, finding the lesion, such as a colon polyp, and then interpretation. Is this benign? Is this precancerous or is this malignant? And if it is malignant, what's the depth of invasion? Is this something that's resectable by endoscopy or should this go for surgery? So what are some of the potential applications in GI endoscopy? When we think of upper endoscopy, Barrett's esophagus, and there's been some work in this area to really identify those areas of dysplasia, which is going to impact how we further manage this patient. Esophageal cancer, squamous cell cancer is a bit more challenging sometimes to identify, or H. pylori detection in the stomach, and of course, gastric cancer, differentiating cancer versus non-cancerous tissue, since they could be right next to one another. When we think about capsule endoscopy, trying to identify the source of bleeding to potentially decrease the interpretation time that we have to do when we watch these capsule videos, and then celiac disease, being able to detect the subtleties of villus atrophy. There's a lot of work in colonoscopy, especially around polyp detection, but looking at bowel prep assessment, live endoscopy assistance, and for those patients with ulcerative colitis to determine severity of disease, get a sense of where they are, if they're having a flare, and then of course, picking up dysplastic lesions in patients with longstanding ulcerative colitis. So as you could see, the opportunities are quite great, and those are just a small fraction. But why did AI start in gastroenterology? And I think it's due to colonoscopy. Colonoscopy is a procedure that's performed quite often, not just in the United States, but around the world. Colon cancer screening and prevention is something that we can do to disrupt that pathway from a precancerous polyp to colon cancer. Colonoscopy involves both lesion detection, detecting polyps, and then interpretation. And the colon itself is not an easy organ to evaluate. There are deep folds and blind spots. And sometimes also the polyps are very subtle, like a sessile serrated lesion, that it's right in front of you, and you just don't appreciate it. So what are the clinical challenges we face in colonoscopy? We know that there's a lot of risk to colon polyps, being able to make an accurate optical diagnosis all the time, and then, of course, polyp size estimations. We know that depending on the size will determine patient recall. So there's definitely work being done in that area. What is the current workflow of devices around colonoscopy? So we have the endoscopy system, and the image is captured. And this is a plug-and-play device, where the AI system is a box that goes on your tower. The image is captured. It's processed by the computer. And then there's an output, and you can see that green circle, which has a polyp inside of it. And this real-time feedback is quite quick, 0.03 seconds per frame. So there's really minimal lag. This is what a typical layout looks like. You have your tower with your monitor. You have your scope processor. And then underneath is the AI unit. This is a bounding box with a colon polyp. And what you're seeing here in this video is, as the colonoscope is being withdrawn, you see the bounding box go over the area of the polyp. And that's for the endoscopist to pay attention, to look at that area and confirm it's a polyp, and then remove it. So there are several FDA-approved systems. Two of them are commercially available, the Endoscreener and GI Genius. Both of them do computer-aided detection. We don't have computer-aided diagnosis currently here in the United States. And then there's another system called Scout, also doing computer-aided detection. That should probably be commercially available later this year. So how effective is real-time computer-aided detection for colorectal neoplasia in a randomized trial? The purpose of this study looked at three sites in Italy where a lot of AI work is being done. And they saw that ADR increase was 14%, adenomas per colonoscopy, 46%. Benefits with lesion morphology, flat, 42%, polypoid, 36%. And then when you look at location, proximal, 26%, and distal, 53%. And also benefits on size, six to nine millimeters, 78% more likely to detect, and less than five millimeters, 26% more likely to protect. So we are detecting more polyps. This was another study looking at misrate. It was a tandem trial that I was a part of. And you could see here for the total polyp misrate when the AI system was used first, 20%, versus when it was not used after, 33%. The adenoma total misrate was also lower, 20.12% versus 31.25%. Sessile serrated lesions, 7.14 versus 42.1. And then advanced adenomas, 11.11% versus zero. So there was a little discordance with that result. But overall, the misrate using computer-aided detection was less when that went first. There were some meta-analyses looking at prospective AI trials. There were five randomized trials. In this meta-analysis, the ADR with AI was 29.6% versus 19.3%. No difference in detection of advanced adenomas. Mean APC was higher, especially for small lesions less than five millimeters versus non-AI. But we are seeing that the larger the lesion, we're not seeing as much as a difference, especially for polyps greater than 10 millimeters. So we're seeing a benefit for lesions less than 10 millimeters, especially for those less than five millimeters. And of course, there will be debate around the significance of finding those types of polyps. This, again, was another meta-analysis, again, showing benefits of computer-aided detection versus just standard white light colonoscopy. What about impact of AI based on lesion size? I referenced this a moment ago. What you could see here is when you look at adenomas less than five millimeters versus adenomas six to nine millimeters and adenomas greater than 10 millimeters, we see a really nice benefit for polyps less than 10 millimeters. But as we get to larger lesions, those are also easier to see with just white light endoscopy alone. What about colon-polyp characterization? This is the nice classification, type 1 versus type 2 lesions, hyperplastic versus adenomatous. So lesion characterization is something that is being used in Europe and in Asia. And ultimately, I suspect we will see this here in the United States as well once approved. These are the studies looking at polyp interpretation. And the take-home looking at all these different trials is really just to highlight the sensitivity and specificity, all being quite high. Most of them, if not all of them, except for two being over 90%, many of them over 95%, showing that it is very accurate for interpreting these lesions for characterization purposes. Benign versus adenomas, which are precancerous. Looking at different imaging modalities, whether it's narrow band imaging or white light. Those are the two that we mainly do. This is another area where work is being done. Again, not currently available for daily clinical practice, but bowel prep quality is very important and having an objective way to give feedback to the endoscopist to see how they're doing, to make sure that they're irrigating and washing. You see here we're in the SECUM, and it's a 0, 1, 2, and 3 to go with the Boston bowel prep score. And what you're going to see here in this clip is that as the scope is being withdrawn, we see that the Boston bowel prep is being graded. And you can see that because there's some residual debris, it's between a 1 and a 2. And the whole idea here is to clue the endoscopist to continue to do more washing of this area and suctioning to ultimately bring this to a higher level of cleaning. You see now we're getting to Boston bowel prep of 3, but with continued work. And as the instrument is getting withdrawn and we're seeing a better bowel prep, you can see that the grading is changing. But one thing that I think would be quite helpful is to give this real-time feedback to the endoscopist. Polyp measuring, another key thing that I think is very important, not available here in the United States. This was recently launched in Europe. You see there's this virtual ruler and giving a sense of how big this polyp might be. And historically, endoscopists are not great at giving accurate polyp size measurements. Some endoscopists are very good, but you could see here that it's a ruler or you could do a circle. And the whole idea is to give a sense of, is this 5 millimeters, is this 10 millimeters? As we know that, especially between that 9 to 10 millimeter range, that will significantly impact the surveillance interval going from, if it's 9 millimeters, safely do 5 years. And if it's more than that, suggesting advanced adenoma, you're going to bring that person back in a shorter interval at 3 years. And this is the first step. And I think we're going to see more of these measuring systems. The other goal of artificial intelligence is to take our workflow, which you could see here is quite busy. You have your colonoscopy. You're waiting for the pathology. You have to generate your report. You want to make sure that your quality measures are captured, time to seek and withdrawal time, prep quality, ADR. You have to notify your patient. You have to notify the referring physician. Wouldn't it be nice if we could do this a little bit easier, more streamlined and artificial intelligence, again, could help with some of that automation. And the whole goal here is to potentially allow us to spend more quality time with our patients and not so much worry about that electronic paperwork or regular paperwork if individuals are still doing that. And it could lead to higher ADRs and lower interval cancers. When we think of upper endoscopy, Barrett's esophagus is a challenge to identify dysplasia and cancer. We know that the longer the Barrett segment is, the less likely the physician may adhere to the Seattle protocol. And many studies have shown low adherence, 16%, 24%. And then the other thing is there's a meta-analysis that suggests a modest benefit when we do the Seattle protocol surveillance. This is a video clip of a segment of Barrett's esophagus. You could see the normal squamous versus the Barrett's. And now you're going to see this bounding box go over an area right where the squamous and Barrett's transition. And this is a suggestive area of dysplasia. And that's really important, being able to identify the dysplasia, the highest degree of dysplasia, if present, because that will certainly impact the clinical management of this patient. Is this someone that we could just do surveillance on? But if they have dysplasia, like in this example, you would then consider endoscopic therapy. And this could be really valuable and maybe move us away from the Seattle protocol. This is just data showing the accuracy of AI with white light and narrowband imaging. You see some nice sensitivity and specificity, 98.6% and 92.4%. Specificity, 88.8% and 99.2%. And then when you look at the different focuses, they're standard. And for those that use Zoom, again, quite accurate. Comprehensive AI diagnosis, 96.4% versus 94.2% for specificity. Gastric cancer, again, another challenge, differentiating cancer and non-cancerous lesions. Optical detection. And this was a study looking at 11 experts versus the computer-aided detection system. And you could see here the accuracy was 85.1% using AI, which was in line with most of the experts, but actually doing better than three of them that were hovering in the 70 percentile range or even lower. Again, the sensitivities for the computer-aided detection, 87.4%, in line with a lot of the experts, but at the same time outperforming several of them, suggesting the benefits of artificial intelligence for gastric cancer differentiation. And this is why. This is why it's so difficult. I really just want to highlight this picture right here where you have the blue box and the red box. And the blue square represents a cancerous lesion right next to the red square, which is non-cancerous. This is really focal. This is really subtle. So being able to differentiate this is really critical, and AI could potentially help us with that. Capsule endoscopy has been around for a long time. It revolutionized how we evaluate the small bowel, being able to identify bleeding, causes of bleeding in the small bowel. There's something called the suspected blood indicator, had a sensitivity of about 60% for active bleeding. This, of course, improved with the adaptive frame rate to improve resolution. But again, reading is challenging, looking at 50,000 to 60,000 frames and being able to differentiate polyps versus nodules versus epithelial tumors versus submucosal tumors and venous structures. You see a bunch of bounding boxes highlighting areas of abnormality. I think probably one of the most challenging is differentiating a submucosal tumor versus just a prominent fold in the small bowel. Since sometimes it's difficult because the capsule just continues to go with the small bowel motility. So can AI impact reading speed? This was a study looking at a training set of over 150,000 images, validating it with even more images. And you see that the sensitivity and specificity was quite high, over 99%. And this is what's really exciting, being able to take the conventional reading time, which of course there's a range, but going down to 5.9 minutes, but maintaining sensitivity and maintaining accuracy, actually doing better when you use the AI system when you compare conventional reading versus AI. But imagine being able to shrink the average reading time down, which probably around 45 minutes, down to under 10 minutes. And I think that this is something that is coming. Another thing that we talked about is the role of looking at medical record data to better risk stratify patients. This was a nice study published in 2019, looking at patients presenting with upper GI bleeding. They looked at clinical assessment and laboratory testing, and they compared that to traditional bleeding scores. You see the area of under the curve was a 0.9, and it showed that machine learning algorithm specificity was a 26% compared to the bleeding scores, which was 12%. So imagine being able to better risk stratify your patient, who needs to get admitted to get an endoscopy and ongoing care and who could safely get discharged. And this is the power of what our artificial intelligence could do. So to conclude why AI and GI endoscopies in immunospaced area, AI could potentially shrink the skill gap amongst endoscopists. GI conditions such as colon cancer prevention and GI bleeding impact a significant portion of the healthcare population, and AI will continue to grow and be interwoven to improve the GI patient workflow. So we're gonna see benefits with quality and outcomes and making our workflow a bit better. So the patient and the physician could come out ahead as artificial intelligence matures in gastroenterology. Thanks so much for your time.

machine learning

gastroenterology

artificial intelligence

endoscopy

computer-aided detection systems

adenoma detection rates