Ryan W. Stidham
Like fire, the wheel, the printing press, and electricity, artificial intelligence (AI) will transform human existence. AI is ushering in new possibilities, excitement, and disruption across every industry, including gastroenterology. For the last few years gastroenterology has seen the potential of AI in proof-of-concept demonstrations aiming to improve diagnostic accuracy on endoscopy, better predict outcomes, and ease administrative burdens in practice. The speed of AI deployment in gastroenterology was not expected; AI is everywhere. In 2015 doctorial computer scientists and specialized bioinformaticians were needed to design emerging neural networks and AI tools. Today, a motivated GI fellow can develop custom AI applications with the aid of low-code tools and accessible cheap compute. We are accustomed to hearing about the promise of AI. Gastroenterologists are now eager to see the reality of what AI can do to meaningfully improve care. In this special issue of The American Journal of Gastroenterology, we present a collection of studies examining today's AI capabilities (or lack thereof) across gastroenterology, with a few examples of early groundbreaking work that will transform what it means to be a gastroenterologist. The original research curated in this issue highlights several practical AI applications for the gastroenterologist. Studies examine how computer vision and large language models (LLMs) can be used to extract information and provide reasoning to build tools that standardize expert assessments, model outcomes, educate patients, and automate common practice decisions. Computer aided diagnosis (CAD) in endoscopy, particularly polyp detection, is the most recognizable AI tool in gastroenterology, but questions remain on accuracy, value, and ultimately impact on outcomes. Does CAD help detect polyps? It depends on who you ask. In this issue you'll read a real-world experience from Mayo Clinic in 4,000 patients showing evidence that both polyp and adenoma detection were improved when CAD AI-assistance was used (1). However, a companion 1,600 subject multi-center study from Germany found no difference in adenoma detection when using AI (2). Yes, AI improves finding polyps, but how many more meaningful lesions are being detected as a result of AI assistance? One could argue that using AI could only help; are we sure this is correct? A Canadian group led by Daniel Von Renteln share a report on how CAD polyp detection tools can contribute to “alert fatigue” that reduced adenoma detection throughout a full day of performing screening colonoscopy (3). The AI value proposition is for more than colorectal cancer screening alone. Does reviewing video capsule endoscopy still need to take hours? An international prospective real-world study showed AI-assisted video capsule endoscopy surpassed the quality of gastroenterologist interpretation, reviewing complete studies in just 4 minutes (4). Can AI go beyond replicating expert opinion to see more than an unaided specialist? It depends on the problem. Brenner et al. share a brief report on the first Food and Drug Administration approved AI colonoscopy quality assessment and how computer-generated metrics could change our perspective on what constitutes an adequate colon examination (5). Help is on the way for patients with defecatory disorders, as Michael Camilleri's group present an AI model for anorectal manometry interpretation to enhance clinical decision making (6). But do not discount the power of the human visual cortex, which still has 100,000 times more synaptic connections than the largest computer models. If gastroenterologists can't see a finding often computer models cannot either. When trained on endoscopic images, AI tools no better predicted polyp histology than gastroenterologists; both had poor performance (7). This special issue of the Journal also highlights the variety of scenarios where LLMs can use in-silico reasoning to standardize professional decisions. Two groups present different LLM-based methods for automating colonoscopy surveillance recommendations, with both showing excellent performance recommending appropriate follow up (8,9). LLMs are proving capable completing tasks traditionally needing a medical assistant or nurse's attention. In hepatology, we see how LLMs and modern machine learning are performing practical tasks such as summarizing outside records, more accurately identifying patients with encephaoloaphy than standard coding, and identifying decompensated cirrhosis (10–12). Other studies demonstrate how LLM-powered AI tools can help coach patients through colonoscopy based on stool images and even diet guidance for celiac disease (13,14). Modern machine learning modeling of clinical data and imaging parameters are showing promise for predicting clinical outcomes in defacatory disorders and pancreatic cancer (15,16). Perhaps you are not impressed with these AI capabilities, seeing only marginal incremental value of AI-agents completing very narrow specific tasks on our behalf. As AI technology improves, little imagination is needed to anticipate a dramatic broadening in the complexity and repertoire of roles AI can take on in gastroenterology. A few years ago, we could unquestionably say AI was no match for a human gastroenterologist knowledge base, with AI failing ACG self-assessments (17). Apologies, but modern LLMs have closed the knowledge gap as Ibrahim et al. (18) report that LLMs can now easily pass GI Board Examinations. Yes, AI can acquire knowledge and answer questions, but can AI intubate the terminal ileum or complete a polypectomy? Automated endoscopy is not here today, but Keith Obstein and colleagues (19) share their phase 1 study of magnetically driven robotic colonoscopy, realizing the beginnings of AI in gastroenterology by adding an AI-brain to modern endoscopic robotics. Enthusiasm to quickly realize the benefits of AI also stands to introduce new potential harms, responsibilities, and conflicts in gastroenterology. Megan Adams and Andrew Feld offer a thoughtful overview on ethical and legal AI considerations, commenting on the adapting regulatory environment, new concepts of medico-legal responsibilities, unanticipated economics, and the harms of provider deskilling (20). We also share an ACG professional society consensus on the state of AI, our needs, and concerns as practicing gastroenterologists (21). Finally, patient perspectives and comfort on using AI in gastroenterology care are explored in a multicenter study survey from Jahagirdar et al. (22), highlighting their concerns about AI reliability, privacy, and impacts on healthcare cost. AI is about to get very (very) real in our daily practice. AI will offer improvements in care efficiency, quality, and standardization. AI will begin to help with new scientific discoveries and contribute to better fundamental understanding of these conditions we treat. We will all need to become conscientious consumers of AI, be comfortable evaluating AI value and performance, and remain active in steering how AI use evolves in gastroenterology. CONFLICTS OF INTEREST Guarantor of the article: Ryan W. Stidham, MD, MS. Specific author contributions: Sole author wrote the editorial. Financial support: None to report. Potential competing interests: R.W.S.: Associate Editor at AJG; served as a consultant or on advisory boards for AbbVie, Bristol Myers Squibb, CorEvitas, Eli Lilly, Exact Sciences, Gilead, Janssen, Merck, Pfizer, and Takeda; and holds intellectual property and equity on medical imaging and endoscopic analysis technologies licensed by the University of Michigan to PreNovo, LLC; AMI, LLC; and PathwaysGI, Inc.