This is session one, starting with the basics about the artificial intelligence umbrella. And again, Mike and I will be moderating this session. And the first presentation is going to be given by Sravanti Parasa, and her topic is applications of artificial intelligence in gastroenterology. Sravanti is a practicing gastroenterologist who's based in Seattle, and she's had specific training in epidemiology and biostatistics, but has a real passion for the applications of AI as it relates to GI and endoscopy, and is a very valuable member of the ASGE AI Task Force. So welcome, Sravanti, and looking forward to hearing your thoughts. Hello, good morning. I'm Sravanti Parasa, and I'm a gastroenterologist based in Seattle. Thank you all for joining in today morning. We have a lot of ground to cover, so let's get started. In this talk, I'll be discussing applications of artificial intelligence to gastroenterology. And here is my disclosure. I will provide an overview of my talk today. So I'll talk about a little bit of artificial intelligence and its introduction, computer vision, electronic health records, natural language processing, machine learning for gut microbiome drug discovery, robotics, augmented reality, and virtual reality as well. Now the concept of artificial intelligence in medicine is not new. Artificial intelligence is over 70 years old, and early clinical decision support systems have actually been around since the 1970s. The early systems that were clinical decision support aids really never took off. The first computerized clinical decision support system was made around 1972 in the UK. And what the computer did is it calculated the likely cause of acute abdominal pain based on a patient's symptoms. So what they would do is feed the symptoms into the computer, and it would try to classify what the cause of the patient's acute abdominal pain was. The system was doing really well, and was actually more accurate than the junior doctors at the hospital, and almost as accurate as the most senior consultants, who obviously had more experience. But it took overnight to give the diagnosis, which is probably a little long when you are dealing with an acute abdomen. If you look at artificial intelligence, it was really very slow progress for decades. Now some of you might be wondering, why is this leap in progress that was about five years ago? If you look at this graph, there's a data set called ImageNet, and what it is, is it's an image database organized according to the WordNet hierarchy, which is basically nouns. So this could be a database that has 10,000 images of cats, or dogs, or camels. This data is then used to train the computers to identify a cat versus a dog versus a camel, etc. Now in this image, you can see that from 2011 to 2017, the progress that has been made. If you look at the red dashed line, that's the human performance, and the green line is the performance of the computer. Now even humans are not perfect. We have a 0.05 rate of error. Now in 2016 and 2017, actually the computers were outperforming humans in this simple image recognition. So we have had advancements in computer capacity, and also in computer hardware. And then the advances in machine learning, along with the development of these deep neural algorithms have really helped progress this field of artificial intelligence. Now the way I see AI is that it is a tool. This is a very nice quote by Dr. Ezioni, who is the CEO for Allen Institute for Artificial Intelligence in Seattle. Now what you do with the tool is up to you and your creativity. And the problem that you would like to solve using AI is called a use case. AI has been used in several innovative ways, and in my talk, I will discuss some use cases. The list keeps expanding in GI, and we are just starting our journey. The first one that comes into our mind when we think about AI and its applications in gastroenterology, and specifically endoscopy, is computer vision. Computer vision is seeing through the eyes of a machine. In the development cycle for artificial intelligence, the first thing you need to do is come up with what is it that you want to do. In this case, I'm going to play around with chihuahuas. So this is our use case for the next one minute or so. So to achieve this, we have to find thousands and thousands of images and then actually label them as chihuahua or no chihuahua or a blueberry muffin. And that step is called data engineering, and that annotation process results in what is called the AI data. And that data is now smarter than what it was before. And then you apply data science. Now, data science is the stuff that everybody talks about. That is the convolutional neural networks. In the past, to run the algorithm, it used to take months and months. However, now a computer can do it in hours or minutes because of the massively available compute. The next step is the application where you plug in into an edge device like your phone, and then it just gets deployed. So essentially, it's a four-step process. The idea, the data engineering, the data science, and the application itself. Now apply the same analogy to medicine. In GI, prototypes exist to detect and characterize pre-cancerous lesions, et cetera. There are several use cases, and you will learn more about it in the computer vision section of the summit. Moving on, electronic health record. It is the backbone of digital health transformation. We see a lot of buzz in the media about digital health transformation and how COVID has changed a lot of things. Now, if you apply this specifically to gastroenterology, this can be used for high-fidelity risk prediction, risk prognosis models, as well as to kind of predict what the patient outcomes would be. It would be similar to large database analysis that we typically use in medicine, but the methods used are machine learning methods. You will hear more about this during the course of this conference as well. Now moving on, NLP. We keep hearing this once in a while about NLP. Now what is the significance of NLP? Only 20% of the data that we need to do clinical research using electronic health records is in a meaningful way that the computers can actually use. The other 80% of it is in the form of text like clinic notes, PDF documents for pathology reports, et cetera. This 80% data is extremely valuable, but due to its size and structure, it's largely invisible to analytical teams. Thanks to this field of AI called natural language processing, computers can now analyze and understand textual data. So NLP is essentially an interdisciplinary field of computer science and linguistics. It's the ability of computers to understand human language. Now speech recognition. The first thing that comes to my mind is, hey Siri, can you find this? So how does Siri do it? So an acoustic input is turned to a deep recurrent neural network and a text output is out. Some applications of NLP are AI powered voice recognition, which can be used in AI scribes. And that's how I dictate my clinic note. We could also have AI chatbots to aid healthcare consumers. One example is to have a chatbot that can give colonoscopy instructions. Now another important area of NLP that I would like to bring to your attention is AI powered literature search. Now it is noted that the literature published in medicine has been doubling at rates never seen before. As you can see in this graph, the current doubling rate of knowledge that we know in medicine is 73 days. Now another concrete example of data overload in medicine was very well noted during the COVID-19 crisis. Now this is a technology feature published in Nature, which discussed the different artificial intelligence tools, which are aimed to tame the COVID-19 virus literature. One such example is the COD-19 database by Semantic Scholar. And we are very fortunate to have the first author on that paper, Dr. Lucy Lu Wang, at this conference. Several applications have been built using this dataset to help boost discovery and enhance comprehension. One way we used this application was to perform an initial search using AI powered literature search for our meta-analysis alongside manual extraction of information, both from published as well as preprint servers like BioMed Archive or MedArchive. This paper was published in JAMA Open. Now how would you like it if we had a Google-like search engine for medical literature where I could say, hey, GIA, can you please pull up images of endoscopic appearance of neuroendocrine tumors of the diurnum from the journal GIE? This idea was the basis for MediCAD dataset, and we just published this article in findings of EMNLP. MediCAD basically addresses challenges in figure retrieval and figure to text alignment and help with automatic search of figures from literature. This is the algorithm that we described in this article. For example, you see chest CTs in this image, and then you have the inline reference somewhere in the PDF document, and then you have the captions. The goal of this algorithm is to clearly delineate and link all three of these in a searchable format. Moving on, omics research, precision medicine. This is a common buzzword again. Now how is it relevant to GI? Gut microbiome. Gut microbiome is certainly a big data problem. The millions of microbes in each teaspoon of your stool and its microbial DNA would literally take a ton of DVDs to store. So you can think about it the next time you're taking a dump. It's indeed a big data dump, and this data is really important for understanding a range of health conditions. For example, it's now well established that human microbiome in your gut affects how you metabolize different kinds of drugs, including painkillers, anti-cancer agents, and increasingly we are able to understand how changes in your microbiome affect changes throughout the rest of the body, including your brain. This has implications to managing disorders such as IBD and IBS. Now the same thing holds true for the role of nutrition in the different diets, its interaction with human genome, disease state, and all this is huge data, omics, whether it's metabolomics, genomics, microbiomics, or radiomics. So integration of this data to disease state and identifying that unique signal is basically precision medicine. Moving on to robots. Now AI-infused smart pills is on the horizon. Neural robots, which can be ingestible, are being used on experimental basis for detection and treatment, essentially using them as vehicles of drug delivery. Now some of you might be following this area of research, augmented reality and virtual reality. If you are into gaming or if you are looking at the new iPhone. So I'm going to switch gears and talk about augmented reality and virtual reality and its implications in GI. Virtual reality is slowly emerging as a therapeutic tool in gastroenterology, especially for patients with IBS. It has already proven success in chronic pain patients and in treatment of anxiety. Now you all might be wondering why I put this picture of the latest iPhone 12 up here. You know, the latest iPhone 12 comes with the LiDAR scanner. What a LiDAR scanner does is it shoots invisible beams of light across the light spectrum and helps us get a sense of the physical dimensions and motion of objects in its vicinity. This is iPhone giving you the power of better AR experience. Now think of it as a tool that would be available to all the iPhone 12 users and what apps can go with it. Perhaps we can have a LiDAR scanner built into our endoscopes and we can really solve the problem of polyps size. Overall, in the next five years, we'll see deep neural networks making significant strides in speech, vision, language, search, robotics. Perhaps my queries for the future would be, hey, GI AI, is this a tubular adenoma or SSA? Or maybe pull up images of neuroendocrine tumor of pancreas and transcribe my note and send instructions to the patient in Spanish. How will you all use this AI tool? You can reach out to us on the ASGE webpage. Thank you all for your attention and I hope you enjoy the rest of the conference.

artificial intelligence

gastroenterology

computer vision

electronic health records

natural language processing