15-05-2025
AI in Ulcerative Colitis: Enhancing Clinical Workflow
Ryan W. Stidham, MD, MS
Artificial intelligence (AI) is used in ulcerative colitis to assist in the assessment, monitoring, and management of disease. To explain how this technology is being applied in the clinical setting, Janelle McSwiggin, MSN, RN, spoke with Ryan W. Stidham, MD, MS, associate professor in the Division of Gastroenterology at University of Michigan Health System. Read on to learn more.
In terms of real-time clinical applications, how can AI-based systems assist healthcare providers during an endoscopy?
There are several ways in which AI is improving endoscopy for IBD using computer vision technologies. For instance, can our existing disease scoring be standardized and perfectly replicated as if it were performed by an expert? Can machines be used to detect, measure, and count all the disease features, like ulcers, erythema, or polyps, for new scores and evaluation tools that would be informative but too impractical and tedious for clinicians to perform? AI can do all of these things during endoscopy and may reshape how we diagnose and monitor IBD by powering new ways to describe mucosal injury.
AI during endoscopy is also proving to help in deciding when and where to biopsy. Commercial systems are already helping detect polyps in real time during the procedure. Experimental systems are also showing promise in predicting whether a polyp is a precancerous adenoma or a benign lesion. In the near future, AI will determine whether the lesion is an adenoma, and rather than sending it to a pathologist for confirmation, it can simply be discarded. Alternatively, suspect polyps that are confidently determined to be benign may simply be left in place potentially, without resection. Similarly, in IBD, there is hope that AI will help detect high risk precancerous tissue that historically has been difficult to see.
How is AI improving endoscopic evaluation in ulcerative colitis, and what are the benefits of using AI over traditional scoring methods?
During a colonoscopy, the clinician is looking at the mucosa to assess the degree of ulceration, erythema, scarring, and even polyps to determine disease severity and quality. Established scores like the Mayo endoscopic score summarize severity with a 0-3 score, although there have been challenges in standardizing scoring, as even experts may disagree on exact grade. AI is already helping automate and standardized the familiar endoscopic scoring systems. Multiple groups have shown the ability to replicate the Mayo endoscopic score and other scores, including the Ulcerative Colitis Endoscopic Index of Severity (UCEIS). These scoring systems form the core of not only standardized endoscopic assessment but also a key endpoint in therapeutic clinical trials. New commercially available technology is digitally recording endoscopic video and providing automated IBD scoring in the background, which helps in objective disease grading, understanding UC population health, and helping to identify patients for clinical trials.
However, there is so much more detail and nuance in disease descriptions than a 0, 1, 2, 3 can capture. The appearance of ulceration, the distribution of features, the changes over time are all factored in a seasoned clinician's perspective for describing disease. Quantifying the detail and interacting features considered by experts is difficult to convert into a simple score. Our group has used AI to develop the Cumulative Disease Score (CDS), where IBD features are detected and quantified every 1-2 centimeters. CDS and similar approaches will help better quantify disease to separate the patient with a small patch of severe disease from someone with extensive or severe disease. Other groups look at the same issues differently and are using AI to develop new severity rating definitions. One study in Japan had gastroenterologists look at thousands of colonoscopy videos and asked them to rate the videos on a 0-10 scale to determine severity. The AI determined the components that led to most experts determining that a patient has severe disease, is completely normal, or is somewhere in between.
Natural language processing (NLP) is a new segment of AI designed to analyze human text and automate clinical health records. How is NLP being used to support patients with ulcerative colitis?
Comprehending the meaning of text requires more than knowing the definitions of words; it's also about understanding grammar, temporal reference, co-reference, negation, assertion — it's an amazing human skill. Today, AI has the same ability and it's starting to understand medical text. The natural language component also includes the ability for machines to generate natural conversational language. The large language models (LLMs) have exploded on the scene as chatbots that provide lifelike responses that are meaningful.
Direct NLP applications that are being used in IBD and the UC space currently focus on helping with administrative tasks. Ambient documentation systems are now able to listen to a clinician-patient conversation, understand the meaning of the conversation, and then generate good- to very good-quality documentation. Office notes, telephone encounters, letters to insurers, even letters to patients and their families can be automated using LLM technology.
We are in the early days of exploring what LLMs can and cannot do, but the possibilities are exciting. Some electronic medical record vendors and other companies are now providing tools that read patient portal messages and then generate a draft reply. This can address major issues for providers, such as of lack of time and burnout resulting from increased communication responsibilities. However, the reliability of these automated 'patient reply' systems has not been rigorously studied, and at the moment they are far from ready to operate without close supervision from healthcare experts. Soon, AI will interpret emails, charts, and phone call records to order medication refills and interpret disease status.
What do you foresee as the next steps in the near future of AI in IBD?
We should expect that all image analysis, whether endoscopy, MRI and CT, or pathology, will soon be primarily assessed by AI. Image analysis systems are maturing quickly, and these systems approach or exceed human reliability, reproducibility, and objectivity. The gastroenterologist role will no longer be assessing images but instead interpreting the clinical meaning of images. I don't really want to measure the bowel wall thickness of the entire colon; let the machine do it and I will tell you what it means for the patient. Increasingly, such AI analytics will be built into imaging equipment (eg, the colonoscopy processor). This will enable a new degree of standardization in endoscopy and UC treatment decisions.
In addition, we should expect that administrative tasks increasingly will be replaced by AI. Documentation will soon be almost fully automated. LLMs will scan notes and patient records to determine appropriate billing and diagnosis codes. Scheduling will be managed by an interactive chatbot that can not only triage patients but also reach out to patients who are waiting for appointments when they become available.
Over the next decade, we will experience major transitions in IBD care as AI ability increasingly comes to understand disease management. We are already seeing examples of LLMs and chatbots acing standard tests, such as the United States Medical Licensing Examination for general medicine. While a few years ago ChatGPT-3 and ChatGPT-4 failed the American College of Gastroenterology self-assessment test, it's only a matter of time before LLMs prove able to understand specialty IBD care questions. This will probably mean that diagnosis and even management plans will be provided by AI tools that have access to patient records, medical literature, guidelines, and some training from experts.
Is there anything else you'd like to discuss related to AI and IBD?
AI capabilities are truly astounding, but we need to be thinking about what we want them to do and the consequences of deploying these tools in care. How does the structure of healthcare delivery change in the post-AI world? Will IBD patients still need return visits with a clinician or can the AI chatbot and LLMs provide all monitoring? What is the role of the clinician in that scenario? If LLMs are managing routine, low-complexity, stable patients, human gastroenterologists could become overwhelmed with a schedule full of maximum-severity patients. I would speculate that over the next decade, medicine will move toward population-level care, with expert clinicians managing many more patients with the help of armies of AI agents to assist.
Regulatory aspects of AI also remain unclear. The FDA and other regulators are thinking hard on balancing safety and innovation in regard to AI in medicine, but we are all learning as we go. What happens if two different FDA-approved AI decision-support systems disagree? What are the consequences of not using AI for decision support, particularly when there is a poor outcome? Who is paying for these AI tools to be developed and maintained? Which will be more valuable to patients: unlimited access to knowledgeable AI IBD care agents, or the seasoned human gastroenterologist?
AI is an exciting revolution in specialty medical care like IBD. While we are still separating the hope, help, and hype of AI, rest assured that changes are coming. We should all be directly involved in this evolution of care to best ensure that the future is one designed to benefit both patients and clinicians in IBD.
