Ryan Stidham, MD, on the Application of Artificial Intelligence in IBD

Dr Stidham, from the University of Michigan, outlines the present-day application of artificial intelligence in the diagnosis and care of patients with inflammatory bowel disease, and looks ahead to promising advances that could offer improvements in efficiency.

Ryan Stidham, MD, is an associate professor of Gastroenterology & Hepatology at University of Michigan Medicine in Ann Arbor, Michigan.

TRANSCRIPT:

I’m Ryan Stidham from the University of Michigan in Ann Arbor, Michigan, and I'm here at AIBD 2023. And in the last session, I talked about artificial intelligence applications for inflammatory bowel disease, focusing on where we are today and also speculating on where AI may be in our practice in the years to come. In this session, we focused on a couple of the key areas where AI is making an impact, particularly on automated image analysis as well as the beginnings of machines being able to understand text to have reading comprehension.

So some of the major, highlights in the last few years, particularly in image analysis, have been we’re moving from an area of replicating expert knowledge towards actually expanding expert knowledge with artificial intelligence. So we reviewed a lot of the work that's been done to show artificial intelligence and neural network systems are able to take centralized reading from individuals who are adept at grading endoscopy, histology, cross-sectional imaging, and actually being very fast at replicating their knowledge base for doing that scoring. And the value of this is beyond the efficiency and speed of machines being able to do these analyses in IBD; it is the capability of having more reproducibility, less bias, having everyone in the entire IBD space able to use the same measuring instrument and ideally doing this at extremely low cost and high speed. So in this way, machines are already giving us the ability to take that expert's disease assessment and make it available may to anyone, anywhere, at any time. So this is exciting, and it's available right now. And there's many academic groups as well as commercial entities that are bringing those capabilities directly to the IBD clinicians’ clinic.

But more excitement that we have is actually on how can these technologies expand the way that we assess diseases. How can it tell us more? Can it give us some insights that help us in patient care that are maybe not apparent to the naked eye or even the specialist eye? And we're seeing some glimmers of that. So, particularly in histology assessment, where artificial intelligence and neural network-based approaches are able to grade mucosal biopsies and full thickness sections in IBD and actually do a better job at predicting what will be the outcome of that individual patient as compared to the type of scoring you would get from a pathologist or a radiologist in cross-sectional imaging. It's also helping us in endoscopy by giving us totally new metrics of how we will quantify disease activity. This includes new measures in ulcerative colitis and Crohn's disease that help us have more detail and personalization in characterizing and assessing therapeutic response and disease severity. Those technologies are developing now, and they have great promise, and we should expect to see them.

So the second half of what we talked about was actually moving away from imaging, which gets a lot of attention, and into text, which many of us believe is really going to be the big breakthrough in how AI is going to impact direct clinical care and really help make decisions for patients with IBD. Natural language processing is a collection of machine learning techniques that enable a machine to understand text, but not just find keywords, but to understand the context of the phrase. It's not enough to know that the word arthritis was used in the document, but you also want to know, well, what was the context of that word? When you said arthritis in your last note a few weeks ago, were you referencing yourself? Were you referencing the patient, were you referencing a family member, or were you talking about maybe a potential side effect or risk of medication? So these natural language processing systems are able to understand that now in far better ways.

So how is this being put into practice right now? Well, NLP tools are being baked into most of our electronic medical records such that no longer will a practitioner may be forced to spend many, many hours reviewing an IBD patient's prior history records, which is often extensive and takes a lot of time. But the machine can actually read those records for you and then summarize it to the granular elements that you need. So NLP technologies are enabling clinicians to now save time and cognitive energy by being able to automate that information extraction process, to really understand all the detail about a patient with an extensive IBD history without having to do quite as much work. These types of technologies are also lending themselves to automated documentation where ambient conversation documentation systems are actually listening to your encounter between the clinician and the patient, recording that transcript, and then turning that transcript not only into a document that summarizes your encounter and your recommendations, but also even automatically entering the orders in the electronic medical record for you. So this is an efficiency step that may help us may save a little bit more time and apply it to other parts of our practice that are needed.

But perhaps the most exciting NLP technologies are called large language models. These are models that can generate text and understand what text means in a broader context. And many computer scientists, digital neuroscientists, and some philosophers believe that these systems actually show true intelligence, which is something that's sobering to think about. We've seen these large language models such as GPT 4 already do excellent on standardized testing such as the United States medical licensing examination, performing at a level that matches the may 10th percentile or top 10 percentile of physicians taking that test. It hasn't quite proven to be excellent at gastroenterology yet. It's has failed a couple of GI standardized tests, but we expect that this will improve over time. And increasingly, these systems are demonstrating that they are able to understand and integrate all the information in the medical record, take the imaging information we talked about earlier and make better clinical decisions just for that patient.

So all of these tools are certainly, coming. Some of them are ready now. A few will take a couple more years to become available, but these will transform our practice, we think, in in great ways, and we certainly believe that AI is going to continue to require supervision. No one should be worried that this is going to threaten your job or career, And it should just enhance our practice and hopefully, the least, let us take care of more patients, but the best to take better care of our patients with Crohn's disease and ulcerative colitis.