I want to congratulate critique and Mike for organizing this third summit they've once again put together a fabulous lineup of speakers and I want to thank all of the speakers, as well as all of you for participating in engaging today in this summit. I'd like to introduce our first speaker Tyler Burzin who is well known to all of you in the AI field Tyler is co director of GI endoscopy and director of the advanced endoscopy fellowship that Beth Beth Israel and assistant professor of medicine at Stanford Medical School he's authored many papers in the areas of us and er CP quality and safety and GI endoscopic anesthesia, and an AI was just reviewing the randomized tandem study that was co authored by Tyler just the other night for ASG journal scan so we're delighted to have Tyler get us kicked off his talk will be called understanding the basics of AI as it pertains to GI welcome Tyler. Good morning, I'd like to thank Dr Sharma Dr Wallace and the leadership for investing so significantly in the future of our field. My goal in the next 15 minutes is to provide an introduction to some of the concepts that will spend the rest of the conference exploring. I'm not a computer scientist or an AI expert really so what I can do is to provide a physician level understanding of this very exciting field that I think is really at the cusp of transforming our practice in gastroenterology. These are my disclosures. I'd like to explore in the next 15 minutes. The first is just to get us comfortable with some of the key definitions regarding AI will explore to very specific AI applications computer vision and natural language processing. And then I also want to give you a couple of warnings about some of the potential flaws of AI that we may encounter as we begin to incorporate this into our practice. First, the definitions. The term AI in general refers to programming computer systems to perform some type of task which we would normally think of as requiring human intelligence and that might mean speech recognition on our iPhone or computer vision a Tesla recognizing the road and driving across the right path. The complexity here involves overlapping terminology and the challenge is that you'll hear some of these terminologies use interchangeably all the time. Machine learning happens to be a subset of artificial intelligence. It's just a particular type of way to train an algorithm to learn pattern recognition. And then deep learning is a subset of machine learning. It's a different type of algorithm that looks a little bit more like a neural network that can do some more complex tasks, even beyond machine learning. And then the other terminology that will be using multiple times is this concept of computer vision technology that can see and interpret visual content. And this diagram on the right situates that I think in a in the correct and accurate way, which is the computer vision actually pre existed the use of AI for computer vision. There are ways to perform computer vision before AI existed. Modern computer vision, modern polyp detection in particular relies on AI algorithms. Let me show you what computer vision looked like before AI. So this is an example of a 2003 paper where a gastroenterologist and programmer computer programmers worked together to try to help a computer to identify polyps. And the concept here was using a hand programmed feature recognition. It was basically essentially the computer scientists sitting down with a gastroenterologist and saying, all right, so what does a polyp look like? And the GI doc would say, well, it's round, maybe it's a little bit pink, the color is a little bit different. And the programmer would say, okay, find things that are pink, that have a slightly different color, etc, etc. And this type of hand program feature recognition worked to some degree, but had some major limitations in 2003. And the limitations were that the accuracy in general was not nearly as high as our current AI based systems. Most importantly, the algorithm and the computer processor were not fast enough for live endoscopy or clinical care in general. And this was also just an image classifier. It said, is there a polyp in this still image or not? Yes or no. It didn't actually localize where the polyp was. And this is a good way to understand the difference between traditional programming, which I just showed you there, and machine learning. Traditional computer programming relied on a programmer writing a computer program so that the software could, for instance, recognize polyps compared to non polyps. And machine learning really reverse engineers this, it sort of turns things on its head. And so here, we actually have an algorithm that can learn some aspect of computer vision, the algorithm itself is good at recognizing things or prepared to recognize things. And then we just give it 100,000 or 10,000 polyp images. And then the result is a new program effectively, which can recognize polyps. It's very, very data hungry, but it performs generally a heck of a lot better than having a programmer say, identify things that are round or identify things that are red. So this is sort of the new model of computer programming. The reason we're talking so much about this and hearing so much about it is that we've reached a very exciting moment for AI, which is that for very specific tasks, AI can meet or exceed human performance. And that includes playing chess, it includes recognizing certain visual images, it includes speech recognition. But I want to again emphasize that this is task specific AI. We often sort of joke about and worry about this dystopian future of a generalized AI, a computer that is becoming self aware, making its own decisions, doing surgery on its own, as gastroenterologists or, you know, lose our jobs. And this is far, far away from where we are. So all the examples we're going to be talking about today, and I think in the entire conference are task specific AI, things that are trained by humans to do certain specific tasks. What is this going to mean for gastroenterology for all of us? Well, there are a handful of areas to pay attention to. In the top left and right corners are computer aided detection and diagnosis, finding the polyp, recognizing then what it is. These are two of the hottest areas we'll pay attention to. The central slide is predictive modeling. This is allowing us to potentially to predict the course of, say, our patient admitted with an upper GI bleed. Can we use AI to do something better than modern scoring systems to predict as the patient can end up in the ICU or rebleed? Similarly, our Crohn's patient, should we put them on interferon? Should we put them on infliximab or some other medication? Might we be able to predict using AI at which medication might be most effective for that patient? The other areas on the bottom are areas that will also impact our practice from quality assessment to analytics to a term called natural language processing, which I'll introduce. So let's talk first about computer vision and natural language processing as the two key applications for us to explore today. Computer vision is faced with some pretty challenging tasks as it seeks to interpret the world that we live in, which is actually a very complex visual world. Human beings are adapted to navigate this world very efficiently. And despite the challenge of the images in front of you, dogs that look like certain food items, you don't struggle or sweat as a human being to distinguish between the two. But for computers, this is actually a pretty challenging task. And again, it's only recently using computer vision powered by AI that computers could succeed at these tasks in a meaningful way. The tasks that all of us deal with on a daily basis is a similar challenge, which is dealing with these subtle polyps. And if you're not a gastroenterologist, you may not be able to even see half of the polyps that are on the screen here. But we, through training, are able to detect them, hopefully reasonably well. But it's still a major challenge. And this is one of the areas, hopefully, that AI can assist us with. One of the big challenges that hopefully we'll explore more today is the fact that many of these algorithms are incredibly data hungry. They need 5,000, 10,000, 20,000 images in order to begin performing at a level that is clinically acceptable. And so if you think about the data management challenge behind that, a 15 minute endoscopic procedure, which is about 30 high definition frames per second, that equals 27,000 images generated just from 15 minutes of endoscoping. And thinking about how we're going to handle and, in fact, label these images is a really core data challenge in our field. This is an example from the radiology literature of a tool that allows radiologists to label, for instance, the liver or various other organs, so that an AI system can learn how to do this. But the back end of this is still a human being, and generally an expert level human being, piece by piece, dot by dot, outlining the liver. And that's incredibly, incredibly time intensive. And we face similar challenges in our field of GI. This is an annotation tool that allows the gastroenterologist to label where a polyp is on the screen. And you can imagine if you had to label 27,000 images, this would be untenable. There are technologies and innovations that aim to try to make this a little bit more efficient by this particular process is labeling one out of every 10 or 20 frames and then interpolating the labels between to generate a lot more data just from a few labels. But even despite that, you can imagine how painful this process is to step through on a Saturday night if you have 30 videos to label. And I think the ASG has very much recognized that our focus cannot just be on thinking about cool new use cases, polyp detection or polyp diagnosis, but we really need to support and develop the data science aspect of gastroenterology. And the bottom right here, how are we going to manage images and videos? How are we going to label these in order to grow our AI space and take a leadership position as data scientists in this field? I want to move for a moment to the concept of natural language processing. These are algorithms which can take unstructured data, a dictated report, a colonoscopy report, a path report, and then use that for computation and analysis to produce helpful analytics. This, I think, is one of the really exciting areas where we'll be able to start engaging meaningfully with our own practice data in a way that we really can't do easily now. You can imagine if you could walk into your office and say, hey, Alexa or Google, show me the ADR trend for first time screening colonoscopies during the last 12 months and upload that to the quality registry. The math of ADR is not complicated. This is a numerator and a denominator, basically, but extracting that rich data from the chart is really challenging. And that's the case with many of our analytics. The math is not complicated, but extracting the data is the tricky part. I think there's real promise for natural language processing to assist with this. This is one of the best example papers in GIE in the last year where the research team looked at a complex EHR system across multiple hospitals using Epic Probation PowerPath and I think actually one more EHR. And they used a technology to combine all of that data seamlessly, including some of it which was scanned in PDF quality. And they were able to output adenoma detection rate, polyp detection rates that were equivalent to a human doing this by hand. I think one of the profound statements in this particular article was that the natural language processing algorithm in under 30 minutes could extract all of the data for all colonoscopy procedures ever done at their institutions since the introduction of EHRs. And if you did this by manual data collection, which they did for the data as well, it took six to eight minutes per patient, which was 160 man hours for annotating data from just 600 patients. So this is really one of the areas, I think, of big promise for us in leveraging AI technologies for our GIE practices. I want to take a moment to also recognize some of the flaws in AI. For all the excitement that we're going to be talking about today, we have to face some of the risks and barriers to these systems performing as we want them to. One really important concept is the concept of fragility. Human intelligence is incredibly robust, and AI systems are in fact not robust at all in many circumstances and can get thrown off in ways that humans can't. The picture on the bottom is a really good example of this. This is an example of an AI system that is trained to recognize bananas and slugs and snails and a variety of other images, and it recognizes the banana on the top with very high certainty level. But Google researchers put a type of sticker called an adversarial patch next to the banana, and it fooled the system into thinking that this picture was now a picture of a toaster, and in fact, with almost 100% certainty, it thought it was a toaster. A human being doesn't get thrown off by this. You think it's a banana with a weird sticker next to it, but it's still a banana. But this really demonstrates nicely the fragility of AI systems, and there can be similar fragility in our clinical world. Examples that have existed in the past are chest X-ray algorithms that performed great at the university hospital, and then they were brought down the road. An example was in Stanford to a local affiliate hospital, which used a slightly different radiology system, and all of a sudden, the AI classifier for chest X-ray nodules didn't work at all. A radiologist doesn't have the same weakness. You can walk from one hospital to the next, and even if it's a different X-ray system, the radiologist will still be able to perform. Same with the gastroenterologist. You give me a Fujiscope, Pentaxscope, and Olympiscope, I'll be able to identify polyps, but it's not a guarantee that AI systems will detect polyps or work diagnostically in the same way across different systems. So we have to be aware of this type of fragility, and we have to really interrogate the fragility in our trials. Related to the fragility is this concept of overfitting. When we train algorithms, sometimes they can perform incredibly well on a training set. This is a training set of red and blue dots, and this model can detect every single red dot in the training set with incredible accuracy, but the reality is, in the real world, you don't want this type of performance. You want systems that can generalize a little bit and recognize patterns. This is what humans do. So we don't want a model that necessarily performs 100% perfectly on the data set and then falls apart in the real world. We want a model that generalizes a little bit, and sometimes that's at the cost of some errors. But we have to recognize that overfitting is an issue, and there are 1,001 examples of systems, particularly in various programming competitions, polyp detection competitions, chest x-ray AI competitions, where a system is trained to work incredibly well on the competition data set, but in the real world, it may not work as well. And then one other concept which I'd like to identify and introduce for the purposes of our conference is the risk that AI can amplify existing health inequities. This is a key flaw of AI. We often think of AI as unbiased and sort of perfect in certain ways, but in fact, there are 100 different ways where bias and health inequities in particular can creep into our algorithm development. That can range from identifying the problems that we want to focus on as researchers to the data collection, which populations we may include in our research trials, to how we identify outcomes and well beyond. So this is something we have to pay close attention to, and we cannot assume that just because the computer is a computer that it is unbiased because it's affected very much by how we design trials and how we identify clinical problems to attack. I will end there and again thank the conference organizers and I look forward to the additional talks and the rest of the conference today. Thank you.

AI basics

gastrointestinal practices

machine learning

computer-aided detection

health inequities