Hello everyone, and welcome back from our first break. Now let's get right into session number two, which is about how to use AI in your daily practice. And again, we'll follow the same format of two brief lectures followed by a panel discussion. And now it's my pleasure to introduce the moderators of session number two, which is Seth Gross and Philip Chu. Seth is based in New York and is a very active member of the ASGE AI task force. And Philip is based in Hong Kong at the Chinese university and has a great expertise in both robotic surgery and how to apply that into the field of gastrointestinal endoscopy and AI. So welcome Seth and Philip, and I'll turn it over to you. Thank you very much, Patrick, and good morning and good afternoon, good evening. So I'm really honored to chair this session with Professor Steve Gross, and this session is about applying AI to your daily practice. And I'd like to introduce our first speaker, who is Dr. Sarvanti Parisa. She is a practicing gastroenterologist at the Swedish Medical Center. She has specialized training in epidemiology and biostatistics with a passion to end and advance clinical care through minimal application of AI. As a practicing clinician with significant research experience in large database analysis and prior collaborative work with national and international research institutions in various area of AI, Dr. Parisa is uniquely positioned to bridge the gap between AI concepts development and practical implication for patient care. Welcome. Hi there, I am Sarvanti. Thanks for the kind introduction. And you've all heard about the basic concepts of artificial intelligence and the technologies that are already existing. And now we'll briefly segue into how we can use AI to redefine quality metrics in endoscopy. So that will be the goal of my talk today. And these are my disclosures. So a brief outline of my topic would be first defining or understanding why we need quality metrics, what's the purpose, and how are we measuring quality in endoscopy currently, and how we might be doing that in the future. We already heard about the current state of technology as it pertains to AI. And I will kind of sum this up with some recent developments in gastroenterology in the space and with some, and I'll conclude. So why are quality metrics in healthcare important, right? So when you're doing a procedure or providing care to patients, we need to understand where we stand in terms of how we are performing. So we need to quantify quality. That's the first step. And then once you are able to quantify that quality, you are able to identify which practices are leading up to higher quality care and higher efficiency, and then recognize those practices and understand the opportunity to learn about how care is delivered. And at the end, develop a framework using these different components to improve quality. So that's exactly why we need to measure quality, have some quality metrics. Now, I'm just going to give a brief overview of how quality measures are in general in healthcare. And these are not very specific to any specific area, and it kind of easily translates to endoscopy as well. In general, the way we quantify quality in healthcare is by three specific types of measures. One is the structural measure. The second is the process measure. And the third, which is a lot of importance on, is outcome measures. Now, what exactly are structural measures? So a lot of times we are, as gastroenterologists, are not directly involved in working around structural measures. These are like, what's your hospital bed size? Do you have an electronic medical record system? What kind of electronic medical record system do you have? What documentation software you have? And so forth. So those are your structural measures. And the second and most important one, which we heavily focus within endoscopy space, is process measures. Now, process measures are something that we do on a daily basis, where you can use these measures as proxies for outcome measure. A classic example would be our adenoma detection rate or withdrawal time in colonoscopy, where we think that if we are, based on data, that as our adenoma detection rate increases, your risk of interval cancer decreases. So the hard outcome there is interval cancer, but the process measure could be withdrawal time or sequel intubation or your ADR. So in general, these are how the quality measures in endoscopy are kind of classified. Now, now that we know that we do have a lot of guidance from multiple GI societies in terms of different quality metrics and some ways to how to calculate them, now I would like to kind of pinpoint some of the pain points in measuring quality as it pertains to endoscopy currently. One important aspect is the feasibility. As Dr. Berzin mentioned in his talk, the math of calculating these metrics is not hard. For example, ADR is a simple calculation, but retrieving the data from the enormous pool of your electronic health record, your documentation software, your pathology reports, and so forth becomes a hassle that involves a lot of human man hours, which leads to a lot of expenses that we spend. So just to give an idea about how much that costs for actually measuring, the cost of calculating the quality measures in healthcare is about $15 billion a year. And if you translate it to some kind of human service, it equates to having one personal assistant for each physician across the United States. So that's how much we are actually investing in these quality measures across the board. So feasibility is an important pain point here. The second point is data burden. So you have a lot of information. You might be doing 10 or 15 colonoscopies a day, and you have to collate all that information to develop your ADR metric. And then the third point I would like to also emphasize is when you're making it work from a feasibility standpoint and a data burden standpoint, you have employed nurses or yourself to do this pulling of the quality metrics, it can lead up to unintended consequences. This time is coming from patient care somewhere else that's being compromised. So these are the current challenges in quality metrics as it pertains to endoscopy. And I will briefly segue into how AI can be used or has been shown to be used to develop quality metrics or develop automatic quality assist programs. So this is a study from Jin Rai Su where they actually used AI-based automatic quality assist program for colonoscopy. It was a randomized control trial. So in the neural network model, what they incorporated was four important aspects that we normally track during colonoscopy. One is the time of the withdrawal phase. We all know that the minimum is at least six minutes for us to meet the metric. So that is automatically calculated. The second one is stabilizing the withdrawal, supervising the withdrawal stability. Now, what does this metric mean? For example, let's say I have a loop in the sigmoid colon and I'm getting to the cecum, and then I'm withdrawing and my ascending colon is fine, but as I get through the hepatic flexure, the scope all of a sudden falls back into mid-transverse. The scope probably did take a look at the hepatic flexure and the transverse colon, but it wasn't enough for a decent examination. So supervised withdrawal stability tells us that at x frames per second is the speed at which you should withdraw so that you can actually see and examine the entire mucosa of the colon. Now, the third part that we also document routinely in colonoscopy and is a quality metric is evaluating bowel prep. It'll be great if it can be done automatically, right? And of course, they did incorporate detection of colorectal polyps in real time into this composite model, which they use for a randomized control trial. Now, I'm not going to go into the ATBT details of the entire, some of these trials, but the concept is to show that there are different ways as to how we can automate the current existing quality measures and how AI can also help with development of new quality measures. So in this randomized control trial, what you can see is the adenoma detection rate for the AQCS algorithm or the system was much higher compared to controls. It's 28 versus 16. And those metrics kind of improve across all metrics, like polyp detection rate, mean number of adenomas detected, and mean number of polyps detected. Now, the second piece that a lot of researchers, including gastroenterologists and data scientists and computer scientists have worked around is the question of the percentage of the mucosa observed. Now, withdrawal time is a proxy for how much mucosa has been observed. So in this picture on the left side, what you see is a tool or a proof of concept developed by Google Health, which is a navigation system to understand what percentage of the mucosa has been observed. There's several such algorithms. I'm just going to give it as a proof of concept to understand that the percentage mucosa observed could be directly measured instead of an indirect measure like withdrawal time. Now, in this study or the algorithm that Google Health calls C2D2, what they do is as the scope is being withdrawn, it captures the image and provides a navigation map of the mucosa seen. And at the end, what you see is blue and red areas showing the mucosa actually visualized. And the red areas are areas of the colonic mucosa that hasn't been examined by the endoscopist. And this can come out as a score, meaning saying that 80% of the mucosa has been observed or only 20% has not been observed and so forth. Now, how does this work? So these are the pictures of the colon, static images of the colon. And then what you see here is different areas showing what part of the colon has been observed by the algorithm. And in this picture, what you see is the algorithm performed really well, meaning it attained 93% accuracy. But the image down in the middle row shows 42, like 0.42, right? The reason this is not as good as 93 is basically the technique. So as an endoscopist, if you are not having the scope in the lumen centered, then you're probably, in this picture, you're not seeing the upper end of the mucosa. And hence, the accuracy of the vision here is only 0.42. Now, in the last slide or this row, what you see is the mucosal inspection is only 0.22, because even though you probably have aligned your scope in the center of the lumen, which I can't find here, is the image quality, right? If you are not able to see the mucosa because of bowel prep or bubbles or blurry vision, then you can't say that you have actually examined the mucosa. So these are other parameters that, as gastroenterologists, we can't change, but there is software to help us navigate these problems. Now, this is another paper published by Zizo, and what they introduce is an interesting concept, what they call eBoston bowel preparation score. So what this does is, instead of having endoscopists determine the BPPS score, it automatically, and in a cumulative fashion, depending on the segment of the colon, determines the bowel preparation score as a cumulative ratio, and then spits out the final result. So it does update or looks at all the image frames every 30 seconds and then provides a cumulative score. So this would be the future where it's not just us taking a subjective stand on how bowel preparation has been, but it's more standardized and objective. Now, Tyler Berzin briefly covered this colonoscopy quality metrics using optical character recognition as well as NLP. So this was a software developed, I think, by Cleveland Clinic, and this study basically shows, as you can see, if you go back in time and use the software for four parameters, which is polyp detection rate, adenoma detection rate, inadequate bowel preparation rate, and failed secal intubation, you see that the manual review versus NLP alone versus NLP with optical character recognition is almost the same across all these four parameters. So it's a ray of promise to say that we can automate these kind of existing quality measures using different kinds of software. Now, I'm going to just segue again into the EGD part of quality metrics. And in this slide, what you're seeing is a study done by Lianlian Wu that was published in GUT in 2019, and the software is called WiseSense, and right in the manuscript, it's called WiseSense. And the goal is, as you insert the scope in the upper GI tract, the scope starts taking pictures automatically and then filters the images to find areas of significant interest. In this case, they were looking for blind spots and the common documentation, picture documentation that we provide to our patients, and then automatically identifies all the different areas of the upper GI tract, including blind spots, and provides an evaluation report that gets there. So as an endoscopist, you're not clicking the images anymore. The software is doing it for you. Now, how does that happen? So on the left side here, what you see is three rows. One is in vitro, in vivo, and unqualified. Again, this brings up two important points, is recognition of what is GI mucosa versus not. That's what this classification here is, in vitro versus in vivo. And then once the scope is in the upper GI tract, you also have to distinguish between images that have no significance in terms of recognizing lesions. So if it is a blurred image or if you have bubbles in the image, then that's not an image that you want to use for filtering these images up here. So that's how the software works. And what they have found is when the primary endpoint was blind spot rate, which was not great in the WISENs, the controls where WISENs wasn't used had better blind spot rate, but the inspection time was higher in WISENs. And in photo documentation completeness, the report that you finally give to the patient, WISENs plus endoscopies versus just endoscopies was performing the best. And the second best was WISENs versus endoscopies, which is 90 versus 79%. So what this proof of concept tells us is that we can use AI or AI algorithms to kind of automate report generation, including some kind of detection for blind spots as well. So in summary, artificial intelligence in the future will enable standardization of quality metrics and improve outcomes by removing subjectivity, as we see heavily currently in measuring quality metrics as it pertains to endoscopy. It is meant to augment the intelligence of the clinician and remove the non-value-added work from our practice. So I always say imagination is more important than knowledge, and I hope you gain some concepts as to how you can use AI to some of your pain points in endoscopy. Thank you very much for your attention.

artificial intelligence

medical practice

quality metrics

gastroenterology

endoscopy