So, the final session for today is going to be about Barrett's esophagus, esophageal cancer and complications. We'll lead things off with Prasad Iyer, who's a professor of medicine, director of the esophageal interest group at the Mayo Clinic, talking about risk stratification of screening and surveillance in Barrett's esophagus. Prasad, welcome. We look forward to your talk. Thank you. Again, like everyone else, I'd like to thank both Vani and Pratik for their kind invitation. This looks like a very nice course. So these are my disclosures. And my objectives are, under the screening heading, to help the audience understand current recommendations for Barrett's screening, discuss Barrett's and esophageal adenocarcinoma risk assessment scores that are available, and at least my thoughts on the challenges in the implementation of these risk assessment scores, and maybe a vision for the future. Under surveillance, we'll discuss factors which have been associated with progression in Barrett's, some clinical risk progression scores that have been published, biomarkers that have been used for surveillance and risk stratification, as well as some final thoughts on challenges and implementation. So this slide essentially summarizes our current Barrett's to esophageal adenocarcinoma paradigm, which is that those with risk factors that I have listed here are at risk for developing Barrett's esophagus, which then progresses via low and high-grade dysplasia to esophageal adenocarcinoma. And we really have, perhaps, three tools to influence this progression paradigm, the first of which is screening those who have risk factors to identify the presence of metaplasia. And then once we have identified those with dysplasia, to treat their reflux and place them in endoscopic surveillance to hopefully detect dysplasia. And if we find dysplasia, to implement endoscopic therapy to prevent the progression to esophageal adenocarcinoma. And as many of you know, this has now become standard of care. And we and others have published that if you find cancer early enough, you can treat it effectively with significantly improved survival compared to when the cancer is detected after the onset of symptoms. So this is a quick summary of the updated guidelines. These were just published. And essentially, the guidelines say that in patients with chronic and or frequent reflux with three or more risk factors, which includes age, male sex, Caucasian race, obesity, smoking, or a confirmed family history, you can screen endoscopically or you could consider screening endoscopically. And the guidelines do say that the swallowable esophageal cell collection device combinations with biomarkers, whether it be cytosponge, whether it be the esophagap, whether it be esocheck, are potential alternatives to endoscopic screening. Now this is a busy slide. And essentially, I've summarized the most recent guidelines from not just the ACG, but also the recently published AGA clinical practice update, ASG guidelines, and then some older guidelines from the American College of Physicians and the British Society of Gastroenterology. And really, the common themes are that screening can be considered in those with multiple risk factors who have chronic reflux symptoms. And all of them are saying no population screening, that is, no screening without the presence of risk factors. So in some ways, this is distinct from colon cancer screening. So in an attempt to perhaps simplify this integration of these disparate risk factors, a number of scores have been published in the literature ranging from what is called the M. Barrett, which is the University of Michigan score, the Hunt score, the Kunzman score. The Locke score is actually named after Dr. Rick Locke, who was one of my colleagues, and also the Gerson score named after Lauren Gerson, all of whom have tried to integrate a variety of these risk factors that I've listed on the second column to the right. And as you can see, many of them are the same risk factors that we have talked about. But unlike having to count them, there are online calculators where you can plug in these numbers and these risk factors will spit out a score and then the probability of what the risk of Barrett's or esophageal cancer is. And Joel Rubinstein, in a recent paper that he published in gastroenterology using a prospective cohort study wherein over 1,000 patients underwent an endoscopy at the University of Michigan with a prevalence of about 7%, looked at the accuracy of each of these five scores in predicting the presence or absence of Barrett's esophagus. So the prevalence of Barrett's was about 7%, and this is data that is not necessarily new, but he compared all of these risk scores together and essentially showed us that if you use only reflux, your area under the curve is about 0.58. On the other hand, if you add risk factors, whether it be age, whether it be male sex, whether it be BMI, whether it be smoking, the M. Barrett score actually uses waist-hip ratio, you tend to do better. How much better? Maybe about 10 points better. So you go up from 0.58 to about 0.68. So that was one message that the paper had. He also looked at a unique cohort from the Kaiser Permanente medical system where over 200,000 patients way back in the 60s and 70s had submitted questionnaires and also had anthropometry, and he used them, he used this data to predict how many of them actually would have developed an esophageal adenocarcinoma or G-junction cancer. So you can see this cohort is very unique, and they actually linked that up to the SEER registry and to other cancer registries and found about 319 cancers in these patients who had had this questionnaire data and anthropometry measured about 40 years ago. And essentially, again, he showed the same thing, that just like for Barrett's, if you only use GERD, you don't do very well in prediction. But if you use some of these other scores, like the Hunt score, the Kunzman score, or the M. Barrett score, you actually do somewhat better, with the Kunzman score being the best, about 0.73. He also looked at G-junction cancers, and the rates of the performance is not as good as you can see for esophageal cancer, but again, the same message. The Kunzman score does a little bit better. So I just wanted to briefly mention another approach. So these are the previous scores, our scores that have been sort of put together from clinical observation. We took a different approach, wherein we tried to cast a wider net and come up with an artificial intelligence-powered risk assessment score. Now how did we do this? Our aim was to develop a score using machine learning, which would predict the risk of developing Barrett's or esophageal adenocarcinoma one to five years before its diagnosis, using all risk factors that are available in an electronic medical record. So we have EPIC at Mayo Clinic, and this is something that we used. And we used this data platform called CDAP, which has about 7 million patient records in collaboration with a company called Enference. And software was used to identify first cases with Barrett's and esophageal cancer using both ICD codes as well as natural language processing algorithms from patient notes. And essentially, we used data that was available on these cases and controls between one and five years before the onset of diagnosis, or before the diagnosis was actually made. We call that the observation time. We excluded any clinical features that was within a year of the diagnosis. And this was done essentially to reduce bias from clinical features which are symptomatic of the disease itself. And this is just a funnel here, which tells you that after all the inclusion and exclusion criteria, we ended up with about 8,500 Barrett's patients and about 1,500 esophageal cancer patients. This was paired with a cohort group of 252,000 and an endoscopy negative cohort of over 150,000. And again, this is complex terminology. I am still getting to understand, but this was a convolutional neural network, which used a model called a transformer model, which is really good at understanding sequential data. So as opposed to the best analogy is as opposed to random words in a sentence, it actually uses data in terms of how the sentences are organized in a sentence, how the words are organized in a sentence to make sense of the data. So using this, we came up with this machine learning predictive model for both Barrett's and esophageal cancer with, I would say, somewhat better prediction than what at least has been published in the literature. So this is just another example of what could be done. So how do we see this evolving in the future? So this is actually a figure I borrowed from a recent publication from Dr. Fitzgerald's group. But really, if you were to go into the general population, you would have to identify those who are at risk. Now these methods can be any of these scores I showed you. These could be apps. These could be a tool like I just showed you in terms of interrogating the electronic medical record. And then the next step would be something which would be non-endoscopic, perhaps, with high sensitivity, reasonable specificity. Finally, only those who are positive with this non-endoscopic test will then be referred on to standard endoscopy. And hopefully, we can then identify most of these patients who are at risk for esophageal cancer development. So I'm going to switch gears and move to risk stratification in surveillance. And again, this is a summary of what was recently published in the ACG guidelines. And I think this is for the first time where we have stratified non-displastic Barrett's on the basis of length. So all our prior guidelines had said you could bring back non-displastic Barrett's every three to five years on the basis of very strong evidence that is now available from systematic reviews, from cohort studies. Non-displastic Barrett's short segment can be brought back every five years because there's fairly strong data to suggest that the risk of progression is lower in short segment. And then for long segment Barrett's, surveillance every three years, low-grade dysplasia, either endoscopic eradication therapy or surveillance, high-grade dysplasia or T1A cancers would be endoscopic eradication therapy. So let's move on to risk stratification in Barrett's. And really, what we are talking about here is how do we identify those or predict those who are going to progress to either high-grade dysplasia or esophageal cancer. And unlike the previous slide where we are really using only dysplasia, and more recently the length of the Barrett's as a factor predicting progression, we and others have shown in systematic reviews and meta-analyses that there are several other risk factors that could be integrated into a score to predict the risk of progression. So this was a systematic review that we published about four years ago showing age, sex, smoking, dysplasia grade, particularly low-grade dysplasia, length of the Barrett's segment, and even some of the medications that patients are on are really predictive of progression risk. So Dr. Parasa, who is one of our faculty here, published a very nice paper with Dr. Sharma in gastroenterology about four years ago where using this multicenter best cohort, they used a cohort of about 2,700 patients and came up with the PIB score, which is the progression in Barrett's esophagus score, which basically integrates a point system. So points are assigned on the basis of sex, on the basis of cigarette smoking, on the basis of the length of the Barrett's segment, and the presence or absence of confirmed low-grade dysplasia. As you can see, low-grade dysplasia is the strongest risk factor for progression. And they divided the score into three categories, low, intermediate, and high-risk, and essentially stratified the risk of progression with these point scores into this pyramid. Now why is this impactful? This is impactful because they were able to show that those who were in the low-risk category progressed at a much lower rate compared to those who are in the high-risk category. They also reported something called the C-statistic, which is really a measure of the strength of prediction of the outcome, and it ranges from 0.5 to 1. The closer the C-statistic is to 1, the better the prediction of the model. So again, this is obviously better than 0.5. So I would call this sort of in the moderate prediction range, but not a bad start. And interestingly, again, they basically showed that the rate of progression was stratified by these point scores. And essentially, they made a proposal that perhaps those who are in the low-risk category could perhaps be discharged from surveillance. And you continue surveillance in either the intermediate risk or in the high risk. Now before any model gets implemented, it has to be externally validated. And indeed, I congratulate Pratik and his group for doing this. This has now been looked at in three populations that I have listed out here, including one in the Mayo Clinic population. And not surprisingly, like any other score developed in one set, once it's applied to another set, its performance sort of drifts down because of some unique characteristics, but still sort of in the 0.6 to 0.7 range. Of course, the integration of low-grade dysplasia into this score is something that I think could be problematic at some point, just because of the inherent inter-observer variability of pathologists in the diagnosis of low-grade dysplasia. So let's switch gears and look at biomarkers. And again, this is a lecture on biomarkers, and Barrett's could last maybe a whole hour in and of itself. But I'm just going to quickly take you through some biomarkers that have been used and looked at to predict progression. The data that we have most is on p53 in terms of its aberrant expression. And Dr. Konda and her colleagues have actually looked at the ability of p53 expression using immunohistochemistry or other assays to predict the risk of progression. And as you can see in cohort studies, the relative risk is 14.3. Again, there are panels of biomarkers, retrospective studies, prospective cohort, relative risk of 38. Again, other biomarkers, including abnormal DNA ploidy, retrospective analyses in case-controlled studies with some impressive odds ratios. Other DNA panels, DNA marker panels, retrospective cohort, need to be externally validated. FISH, which is fluorescence in situ hybridization, was actually looked at in a prospective cohort, relatively small number of progressors, reasonable sensitivity, borderline specificity, still needs to be externally validated. So you can see a theme evolving here, that these markers, there are plenty of markers that have been investigated. Unfortunately, many of them still need to be externally validated. And perhaps their performance scores are not just where they need to be. So the last biomarker test I might mention is something called tissue cipher, which is a tissue systems pathology test. And unlike some of the other biomarkers, this uses nine kinds of biomarkers, as well as about five nuclear characteristics, in a multiplex fluorescent labeling platform. And a proprietary software will integrate data quantitatively from all of these biomarkers and nuclear characteristics into a single score, which can then be divided to split patients into low-risk, intermediate-risk, and high-risk. This assay is actually clinically commercially available. So we published recently a pooled analysis of four studies, which were multicenter, multiple centers from the United States, as well as from Europe. One of the cohorts were actually from the SURF trial, which was a low-grade dysplasia trial. And this was a fairly large cohort of over 500 patients, including 150 progressors and 400 non-progressors. And essentially, our goal was twofold. Number one, to see if addition of the biomarker actually improved the progression ability beyond what we can do with just clinical variables. So remember, we talked about the PIB score, which just integrates clinically available variables into a score, and then you can decide what the progression risk might be. So as you can see, very similar to the PIB score, if you use only clinical variables, your C-statistic is somewhere in the 0.68 to 0.7 range. What happens when we add the tissue cipher score? And as you can see, it displaces most of the other risk factors and becomes its own independent risk factor with an odds ratio of about 6. And the only other variable that remains significant in the multivariable analysis is low-grade dysplasia, which, as we know, is fairly commonly reported. Now what you should see is that the C-statistic, remember, that is the prediction score. The score that sort of gives you the accuracy of the model went up from 0.68 to 0.75. In terms of sensitivity and specificity, the model, the test is very specific, so it has very few false positives. But on the other hand, the sensitivity is still modest. It's still missing a lot of the progressors. So I think this is where work needs to be done. As I alluded to before, there are several challenges in this area. Prospective validation is needed. Unfortunately, the low overall progression rate in Barrett's necessitates large sample sizes, long follow-up, and funding is always a challenge. Many of the biomarkers that have been proposed actually can only be assayed with either specialized sample types, like frozen biopsies or brushings, or the analysis platforms are difficult to standardize. And of course, we always need a commercial partner and reimbursement before any kind of test can be implemented. So in summary, for screening, risk assessment remains critical to the implementation of widespread Barrett screening. Clinical scores and potentially AI-based tools may be options. Surveillance. We have now begun to personalize surveillance on the basis of not just dysplasia grade, but also Barrett's length. And clinical and biomarker tools are making progress. However, as I indicated, several challenges remain for implementation. Thank you. ____________________________________________ ______________________________________

Barrett's esophagus

risk stratification

screening

surveillance

risk assessment scores

biomarkers