Although the role of artificial intelligence continues to expand in gastroenterology practice, recent research highlights the need to balance AI and human expertise, according to experts speaking at the 6th Annual Global AI in GI Summit, in Washington, D.C.

AI systems can only process what the endoscopist sees on the screen, noted speaker Dennis L. Shung, MD, an assistant professor of medicine and the director of digital health and applied AI at Yale School of Medicine, in New Haven, Conn. “AI cannot correct for poor procedural technique to ensure mucosal exposure.” At this point, Dr. Shung told Gastroenterology & Endoscopy News, “we are not anywhere near AI-driven endoscopy. At most, it is AI-assisted, though sometimes it’s AI-annoying.”

Standardization and Automation Benefits

Nevertheless, standardization of expert-level quality examinations is likely among the greatest benefits of AI in gastroenterology, according to Michael B. Wallace, MD, a professor of medicine in the Division of Gastroenterology and Hepatology at Mayo Clinic, in Jacksonville, Fla., “AI technology offers the potential to encode expert-level detection and classification into the software to improve the quality of endoscopy across the spectrum of providers,” Dr. Wallace told Gastroenterology & Endoscopy News.

For colon polyp detection, the use of AI has been rigorously studied in many randomized controlled trials in the United States and Europe, said Dr. Wallace, who served as the co-chair of the summit with Prateek Sharma, MD. “These trials demonstrated that AI technology significantly improved the detection of precancerous polyps and reduces the miss rate when compared to standard colonoscopy alone.”

image

Beyond the use of AI-enhanced lesion detection and classification in colonoscopy, Dr. Wallace said that “other near-term applications include detection of dysplasia and Barrett’s esophagus or ulcerative colitis, early gastric cancer, and classification of inflammatory bowel disease.” Potential applications of AI to the routine workflow include assessing the quality of the examination, including bowel preparation scores and withdrawal time, he added.

Dr. Shung noted that as more sophisticated systems are integrated across care encounters, other potential benefits of AI include simplifying documentation, providing recommendations in real time, and allowing for more face-to-face discussions with patients regarding endoscopy findings and next steps.

For example, “large multimodal models have the potential to automate documentation directly from endoscopic images, and also to start describing endoscopic findings in greater detail and granularity, from early gastric neoplasia to inflammation severity,” Dr. Shung said. These models, ultimately, may have the potential to act as second or third opinions in providing clinical decision support to patients and providers, he said.

But, he cautioned, “This is an area where there are currently no guardrails, so regulations and safety efforts are necessary to limit the harm of misinformation for our community.”

Barriers to Uptake

Trust and accountability remain barriers to the use of AI in endoscopy, he explained, noting that “providers have to be convinced to turn the AI devices on, which is no simple feat.” When the systems make an error, human endoscopists require a clear explanation of what occurred so they know when to be extra vigilant. Cost is a factor, as well, especially for under-resourced endoscopy centers and underprivileged populations, he added. “Efforts to ensure equity in distribution and access are key from the society level but must be supported by partners in industry and government funding. As has been seen in other fields, expertise is critical in using AI effectively.”

Experts can work even if the AI system malfunctions and can navigate the failure modes of AI to ensure that both patient and provider get through safely. However, because AI cannot ensure adequate technique, Dr. Shung noted that “additional research should focus on how to best focus endoscopist attention when it matters.”

Regulatory issues, user acceptance and cost are additional barriers added Dr. Wallace. In the United States, regulatory approval for AI usually falls to the FDA under the term “software as medical device.” However, he noted that once a predicate device is established, subsequent devices generally need only to prove equivalence to the predicate device.

In the clinical setting, the box prompt that appears to flag lesions is annoying to some users, and false alerts can be bothersome, but most are short-lived and only minimally distracting, Dr. Wallace said. As for cost, AI systems are not currently reimbursed as a separate device. However, he added, the cost is somewhat offset by the direct gains in polyp detection and other lesion detection and by the potential reduction in workload that may improve the efficiency of an endoscopy unit.

Looking ahead, “a key area for research is developing a user interface to minimize distraction and improve uptake and the translation of clinical research evidence into clinical practice, and [determining] how this will impact training,” Dr. Wallace said. “We do not yet know how fellows trained with AI-enhanced endoscopy will develop and maintain skills with and without AI enhancement.”

—Heidi Splete

Dr. Shung reported no relevant financial disclosures. Dr. Wallace reported relationships with Boston Scientific, Cook Medical, Cosmo/Aries, Endiatx, Fujifilm, Medtronic, Microtek, Ninepoint Medical, Olympus, Surgical Automations, Synergy, Verily and Virgo.

This article is from the February 2025 print issue.