I would like to introduce my co-moderators, joining us from Europe, Dr. Evelyn Dekker, who's professor of GI oncology at the University of Amsterdam, and Professor Pradeep Bhandari from Queen Alexander Hospital. He's a professor of GI at Portsmouth University in the UK. Great being with you. As you can see, it's already dark here in Amsterdam, probably the same for Professor Bhandari. I'm very happy to be with you. I really enjoyed all the three talks. And I'm supposed to start off the discussion. So far, I haven't seen any questions from the audience. And I would like to discuss a bit further, first of all, with the second speaker with Dr. Kan, because I think maybe you can lead us a little bit here. As you discussed, I think, well, if we have the payment issues, I think they're very interesting. You said, you said a little bit about we might improve productivity. But what I didn't hear, which might be different from your job as compared to gastroenterologists, that, for example, in the polyp field, it might mean we reduce, for example, a lot of other costs, like the typing time for the endoscopy report, or a polyp diagnosis by a pathologist. So, isn't it, shouldn't it be regarded a bit more in balance, also, the costs you spare versus the costs you have to invest? Yeah, that's a great question. It's really challenging, in fact, to figure out where some of these savings might lie. And I'll just maybe give you a couple of examples. I do ultrasound, so thyroid ultrasounds, I would love to give over to an AI. Personally, I would just love to give that up. But the question is, you know, what can be done? So there have been a number of things. And I think some, actually, kind of my prediction for at least in radiology, and I don't know the space in gynecoscopy as well, but in radiology, probably most of the savings that we're going to achieve are going to be not the image interpretation, it's not the computer vision applications, it's the management applications, it's managing no-shows, it's doing some of the other things that go around it. I'll give you just one example of what I consider a real success. We had one of our, one of my colleagues at Penn named Michonne Bilello, who's a neuroradiologist, and created a tool that does MS, multiple sclerosis plaque measurement, and it does it as an automated tool. It's important because in our practice, we do something like 20 to 30 MRs a day of multiple sclerosis patients, very active, it's a tremendously active practice. And these people get brain, cervical spine, thoracic spine, lumbar spine, without contrast, and then they go back and have it with. Using this tool, we're able to limit 80, we're able to eliminate giving contrast to 87% of those patients, which is great for the patient, because it means they don't have to get injected, they don't have to, you know, go through the second half of the scanning, it means we get the more scanner time, which is one of our limiting resources. It's easier for the radiologist, it saves money for the insurance company, kind of a win all the way around. But, you know, it really points up to that you have to focus the application of these tools on things that really are going to have high impact in your practice. And that's proving to be one of the great challenges, you know, and to some extent, if it were easy to squeeze more juice out of all the lemons, we would, you know, we would be doing it already to some extent. And so it really is, it's challenging to find those applications where we really are gaining. But I'll just give you one other instance. And that is that a number of the companies that sell things like systems to detect pulmonary embolism, radiologists are interested in detecting pulmonary embolism. But in fact, they is, you know, some of them had kind of given up on radiology to make their, they're actually going to the pulmonologists, with the notion that, like our PE response team would be the ones who would be interested in purchasing the system and getting it integrated into radiology, because it would tend to bring things to them. And I think it's kind of the same with some of the discussion that was had earlier about, you know, allowing you to get to more polyps and get, you know, get more resection time and that, but it'll make, you can make, be more productive and do things that are more reimbursable. Hi, very interesting discussion, guys. I was listening and thinking a lot about how AI will get implemented, especially CAD-E and CAD-X, which are already available by various manufacturers. So my question was to Dr. Thakkar about detection, we know how it can be implemented in terms of it will help us detect. And there are issues about finances and logistics, but other than that, in clinical practice, detection is very easy to implement. How is characterization going to be implemented? What is your thought process? Is it just going to tell us this is hyperplastic, this is adenoma, and we carry on like what we do now, or we're going to go to resect and discard? And if you're going to go to resect and discard, are we ready for that? Is patient's going to accept this? I don't think I've ever seen any data on patient acceptability of that. So what's your thought process? Could you tell us how this is going to happen? Yeah, that's a very good question. It was discussed a little bit in the talk earlier about what the patient perspective would be on a resect and discard or a diagnose and leave strategy. And we don't have that. My feeling is, what the ASG has also said is that when you can have a predictability of greater than 90% or greater, that the resect and discard strategy should be implemented. It would save a significant cost amount to the utilization of pathology and healthcare in general. I think we're seeing, the fortunate thing that we have is that we're seeing it happen around the world. Right now, there are other countries where the diagnose and leave strategy are being used or the resect and discard are being used and it's being accepted. And I think that this is a great opportunity for us to utilize things like endocytoscopy to better advance the healthcare and reduce the waste that's happening. So it will be an interesting strategy nonetheless, but I think it is a part of the evolution that is happening with artificial intelligence. And so long as we're doing things like eliminating bias, then there's a great opportunity ahead to reduce that waste. Thank you so much. So I think this is a great discussion. And then I think we could move also to Dr. Parasa because she indeed very nicely explained us about how to consider the different AI strategies and the biases there. So going on in this discussion, who decides in the end in which situations we can use which systems, for example, for pull-up characterization, let's say. Because I think you really nicely point out the difficulty in interpreting the data we have so far, Dr. Parasa, can you comment on that? So I'll put the question back. So let's say you have maybe a new CT device or an MRI device and somebody says, you're able to detect something with much more accuracy or detect new things that you have never seen before. Right now, forget that it's AI, right? What do we do? We try to see why is it detecting things more? Who is making that decision, whether you want to send this patient for the regular MRI or the enhanced MRI or some other new technology, right? We as clinicians make that decisions. We as societies think about it and see that there is value to this. So essentially, I think that's why it's very important as clinicians who are delivering the care and deciding what treatment choices or what diagnostic tests go in, we need to understand and make that decision, whether it's pull-up characterization algorithm or detection algorithm or gastric cancer or whatever it is. That's how I would leave it. Can I ask a very simple question to Dr. Parasa again about bias that we talked about? And I was very interested to find out that by x-rays, you can find out the race, maybe by looking at the polyps, we can find out the race of the patients. I'm not aware right now, any of the companies I'm not aware right now, any of the companies balancing their polyps by ethnicity. Is that a relevant thing? Is that something the manufacturers should start looking into that we need to collect equal number of polyps from blacks and whites? Will that make a difference? Well, we don't know yet. So that's the question. So a lot of times one of the, you know, what we call close nearby associations and so forth, based on the images that you can see and the milieu around the polyp or around the colon or the stomach, you are able to, AI basically uses that technology to predict what it is. It's just not looking at what you are looking at. It's not the nice classification of the polyp that it's going in a rule-based fashion, right? It's making its own patterns and trying to correlate it and come up with an idea. So we don't, I don't know the answer to colon polyp and race if that has been backtracked, but there has been studies on diabetic retinopathy where you have retinal images, which are not too different in terms of, you know, not having the recognition of your face or other things. You would imagine that you can't identify race or ethnicity just based on a retinal image, but they could kind of come back and classify that and that's papers already published. So we just have to maybe do that study with colonoscopy. But I think that's, that's where we are headed. But isn't that interesting? One more question. Evelyn, go ahead. Yeah. So I think you, you're raising a very interesting point because you clearly said that the AI system is using his or her algorithms to understand without knowing anything about NICE or about, about any classification, for example, it just uses the data in, but we also need to put data in for the computer to have our part, right? As long as the computer doesn't know the race of the patients, he doesn't, we, we will never know whether there's a correlation there, same for many other data. So how can we take care of that issue? I think what you're trying to allude here is we are right now in a supervised learning fashion. So we are inputting the data, we are labeling the data, but what we are trying to do is there are some people out there in the garage somewhere who can use this data and backtrack and get to that race or not. Now, how accurate that could be, we don't know unless there is a lawsuit. And, you know, just to add a comment to that, this is a limitation right now of ADR, right? I mean, ADRs are going to be different by provider based on the populations they serve. And, you know, this may actually be a benefit to AI when you have that ability to better understand and better interpret what it means to have an ADR in the population. Okay. Thank you, Evelyn and Pradeep. Good to see both of you and thank you for joining us on a Saturday evening from your home. So wonderful. And again, I wanted to thank Irving, Shyam, Chuck, and Sravanti for this wonderful session. So we'll bring this to an end right now. Thank you.

AI in gastroenterology

productivity

cost reduction

clinical practice

biases in AI algorithms