And to moderate that, it's my pleasure to introduce Brett Peterson. Brett is professor of medicine and the president-elect of the ASGE. And along with him is John Cohen, who is from New York and also a governing board member of the ASGE. So Brett and John, over to you to lead this session. Thank you very much, Prateek. We'll begin with a discussion by Dr. Rubin, Utilizing AI and ML Machine Learning in IBD Diagnosis and Prognostication. Dr. Rubin is the Joseph B. Kirsner Professor of Medicine and Chief of the Section of Gastroenterology, Hepatology, and Nutrition, and co-director of the Digestive Disease Center at the University of Chicago Medicine. He also is an associate faculty member at the McLean Center for Clinical Medical Ethics and an associate investigator in the University of Chicago Comprehensive Cancer Center. In 2020, Dr. Rubin received the Sherman Prize for Excellence in Crohn's and Colitis. He's an associate editor of Gastroenterology, editor of a best-selling book, Curbside Consultations in IBD. And his current research is in the area of biosensor monitoring of IBD, prevention of progressive complications from uncontrolled inflammation, and a variety of collaborative studies related to causes of IBD and complications. Welcome, Dr. Rubin. So I'm gonna cover what I consider to be more of a show and tell about inflammatory bowel disease and the application of artificial intelligence and machine learning. There's a lot going on in our field, some of which is directly relevant to what other presenters today have mentioned, and others are a little bit more far afield. So let me start by acknowledging a couple of disclosures that are relevant to this presentation. I am consulting for companies working in this space. This presentation is my own, and when I do share any information from them, I've been given permission. So there are a number of challenges to IBD. One is featured on this slide, which is that when you see a patient, there's often a long delay between symptom onset, diagnosis, assessment, and then getting them started on therapy. There's also the big challenge in inflammatory bowel disease that we don't have predictive therapeutic biomarkers to tell us which therapies to use. So you can look at this timeline and appreciate how many delays there are in the way we manage IBD. So one area that's desperate for an improved approach to management has to do with diagnosis, prognosis, choices of therapies, and disease monitoring. We've also moved our field forward and are firmly now in what I've called the disease modification era of IBD, and that recognizes the need to control inflammation. And controlling inflammation is an interesting concept to discuss, but in practical reality, becomes a bit more challenging to achieve. This is a consensus report from the International Organization of IBD that has acknowledged and identified a variety of targets that we're supposed to reach in order to control IBD more objectively. And you can notice across the bottom that there is a time course for achieving these different targets. Targets like CRP or a fecal calprotectin, which is an inflammatory protein, are targets that we believe and have now validated to be associated with more stable disease control, response to therapy, and monitoring when patients are in the maintenance phase of their management. This also lends itself very nicely to the potential for interventions of AI or machine learning. And the third point I wanted to add that sets up this entire conversation is that the FDA has continued to partner with industry and to move our field forward in important ways. This is a white paper published now five years ago that identified what the definition of mucosal healing was supposed to be in ulcerative colitis as an acknowledged endpoint of management. And the definition of mucosal healing moved from just endoscopy to a combination of endoscopy with histologic improvement. Now, as is common when you're negotiating and moving along with regulatory bodies, the definition of histological healing is evolving, and the incorporation into practice with pathologists and endoscopists and trialists has been a bit rocky, and we don't yet have uniform diagnoses or definitions to incorporate all of this. This is another area that easily allows for us to imagine some more objective, reliable, and valid ways to assess the bowel. So all of these different examples I've given you demonstrate that the field of IBD is ripe for the application of AI and machine learning. So there are many different places where this can easily apply, whether it be an earlier or more accurate diagnosis in the pathogenesis of the disease, in care delivery, in predicting, and then monitoring for treatment response, very importantly for prognostication. And I would argue as well to improve clinical trial efficiency in terms of standardization of measurements and moving patients through trials in a more effective manner. There are lots of examples, and I've only picked a number to give you some flavor for what's going on in our field. Of course, this has been demonstrated and discussed extensively by others already today. There are a variety of different variables and datasets that inform our approach to machine, deep learning, and artificial intelligence in IBD. We have incredible datasets in our field. They just haven't all been mined sufficiently or harvested for the types of information we need, but there's a lot of good work going on. Let me start with this example, which is a model to predict individualized risk of Crohn's disease complications. In Crohn's, we could define complications as the need for surgery or recurrent surgery or hospitalization or steroids. Now, this model used the variables you see in the table on the left, including the serologic immune markers, as well as some clinical markers, and even a genetic marker, the NOD2 frameshift mutation, and combined them using additional, I would say, machine learning, or at least deep analysis with algorithmic approaches to develop something that was initially called PROSPECT as a way to put a patient with Crohn's disease at the time of diagnosis into one of three categories of low, medium, or high risk that they were gonna have poor outcomes or complications. This is a very important model that has actually moved forward commercially. It's available for free now and online called CDPATH, where you can quite literally have your patient get this result, and you can use it to communicate whether they need a more advanced therapy, a more aggressive monitoring strategy. And you can see examples of the outputs on the right here. So this is an example of application of a variety of complex variables into a model that has been validated in adults and pediatrics for prognosis in Crohn's disease. I've also said to the folks who developed this that this can be used to encourage providers to be more comfortable embracing certain treatment strategies. I also wanna emphasize the next point, which is the endoscopic healing index. The endoscopic healing index was initially developed commercially by Prometheus Laboratories, and they had a number of markers that they then, after doing a variety of deep and algorithmic assessments and multivariable regression, were able to distill down to 13 different molecular markers of mucosal integrity and healing. And you can see the ones that were in the ultimate model on the right there on this slide that included growth factors, matrix remodeling, angiogenesis, and the standard inflammatory marker like CRP. The endoscopic healing index was then validated using a variety of different sample sets from clinical trials and IBD centers. And after doing so, demonstrated that you could correlate this particular serum marker of inflammation with endoscopy with a high degree of predictive validity, as well as demonstrating change over time. So this has been incorporated as a more accurate predictor in Crohn's, at least for now, as a way to monitor disease, inflammation, and response to therapy. This is another example of applying machine learning and a neural network to assess multiple complex variables and come up with a model that we can apply to practice. There's also some additional work that's been done looking at just laboratory values. And this work from the University of Michigan led by Peter Higgins used a variety of complex variables, but were all laboratory values that you can see highlighted in the first figure in the middle of the screen. And they were able to actually predict the likelihood of response to thiopurine therapy better than using thiopurine metabolites. So standard labs can also be incorporated into a machine learning algorithm. There are other efforts going on to do this with other therapies as well. And we now have a variety of different efforts to use this in capsule endoscopy. In this particular analysis, this is from Denmark, they use the capsule endoscopy videos and all the images that go along with those data collection methods to demonstrate that you could actually predict alterations that were more likely Crohn's disease than other causes for small bowel findings and shorten the reading time, obviously, as well as provide more predictive value for capsule endoscopy in suspected or known Crohn's disease. And in work from Israel, also with Crohn's and also in capsule endoscopy, this group was able to demonstrate that they could predict the presence of strictures using a variety of different approaches. You can see here a demonstration of their methodology using a class activation map, and you can appreciate the very nice ROC curves from their predictive models. Now you could argue, what is the good of knowing that there's a stricture in the small bowel when you drop a capsule, because it's gonna potentially be retained in that setting, but it's a very good example of using an incredibly rich data set to then develop some predictive values that would enable the human reader to do an assessment more effectively and efficiently. Moving on, I wanna emphasize how we can apply this to ulcerative colitis, which might've been what you expected me to spend the whole time talking about, which is how can we use our current scoping technology combined with other assessments to assess the level of endoscopic inflammation and even histologic inflammation. In general, in the traditional approach, endoscopy and histology don't correlate very well, surprisingly, in ulcerative colitis, but the better we get at reading this with the additional impact and assistance of AI, the more we find that endoscopy and histology can correlate very nicely. This work by Takenaka and colleagues in Japan and some additional great work they're doing has demonstrated that this deep neural network has been able to predict with very good degrees of certainty, the degree of inflammation of the mucosa in ulcerative colitis and even demonstrate that it correlates to biopsies of the same areas. Very nice work and ongoing work there. One of the efforts that I'm involved with, as well as some of the other folks who are here today, is in the approach to this through a different company called Iterative Scopes using what you see here, which is the videos collected during clinical trials and in the real world. It developed deep learning network with a MayoScore system, which is how we grade endoscopy and severity of inflammation and a feature classifier that is then able to provide some very nice results where it breaks it down by pixel analysis and by their deep learning, machine learning model to the different degrees of MayoScores. Now, the challenge to this that I wanna bring up is that this may be better than the gold standard of what we consider centralized reading by humans. And so how do you advance AI that might be better than human reading when the human reading is considered the gold standard? It's one of the challenges in AI that others have mentioned today and one that we're also struggling with in IBD. But you can imagine applying this technology in a way to eliminate central reading, which is a delay and a very expensive part of clinical trials and provide more accurate reading that eliminates some of the extra steps and costs in clinical trials. But also, even if we were using it in regular practice to identify patients who are eligible for clinical trials, where we might not have even thought of them for the trial, but the computer helps us think about it ahead of time. And you can see what a video segment of someone with colitis might look like. And also the real-time analysis that's required and the interpretation of how many types of images and scanning that is required across the entire surface area of the bowel. Most recently, I'm gonna share with you the very first patient that we've applied the wide-area sampling from the CDX Diagnostics Company. You're familiar with this for Barrett's esophagus, which is what that image there shows you with the brush. We are now piloting this in ulcerative colitis. And you can see on the left there a patient with a lesion that was not resectable. And our pilot assessment to see if the same brush, cytology, and optical imaging assessment with AI to detect abnormal cells can actually find dysplasia in the colon in colitis. And we're excited to see this work move forward. I'll end with this, which is the work we're doing here at the University of Chicago, to try to apply biosensors as a way to predict inflammation and lead relapses in patients with IBD. You can see here the variety of datasets that we need to combine in order to try to do this. And we're using standard Fitbit devices, although we've branched out now to using Apple Watches and Garmin devices. But using the standard data collection, including sleep, physical activity measured as steps, and heart rate variability data, we have in fact been able to recruit a large number of patients longitudinally and demonstrate that we can actually predict relapse up to a month before it happens clinically, at least in our initial analyses. And so you could imagine biosensors that are monitoring patients in real time and can tell maybe before the patient feels anything that they're starting to have activation of their disease. This would then obviously be paired with digital therapeutics and communication methods that might change the complete flow of how we manage patients in the future. So the practice pearl in IBD is that the movement towards objective disease control is the primary method that we've recognized how AI and deep machine learning approaches might be able to aid us in our future. But we're not quite there yet. I wouldn't say that any of the things I just shared with you are standard of practice. So I'll leave you with just staying tuned to this exciting field. And I wanna thank again the organizers for inviting me to share all of this with you.

artificial intelligence

machine learning

inflammatory bowel disease

diagnosis

therapeutic biomarkers