Incorporating artificial intelligence (AI) into colonoscopy led to a twofold reduction in the miss rate of pre-cancerous polyps, a randomized trial found.
Adenoma miss rate (AMR) was significantly lower in those who first received a computer-aided colonoscopy compared to those who first received a standard colonoscopy (16% vs 32%; adjusted odds ratio [aOR] 0.38, 95% CI 0.23-0.62), according to researchers led by Michael B. Wallace, MD, PhD, of the Mayo Clinic in Jacksonville, Florida, and Sheikh Shakhbout Medical City in Abu Dhabi, United Arab Emirates.
The AI-first group had a lower AMR for diminutive (≤5 mm) polyps (16% vs 36% with standard colonoscopy) and non-polypoid lesions (17% vs 46%, respectively), and for lesions in the proximal (18% vs 33%) and distal (11% vs 32%) colon, the group reported in Gastroenterology.
“Utilizing some sort of AI does seem to be where the industry will move,” said Allen Kamrava, MD, of Cedars-Sinai Medical Center in Los Angeles, who was not involved in this study. “The concept of an AI platform overlay that can plug into existing hardware as a software upgrade is compelling.”
Two prior parallel studies from Wallace’s group showed that the AI technology used in the current study, GI Genius, led to increased adenoma detection rates, and the technology was authorized by the FDA last year to help detect polyps or suspected tumors during colonoscopy.
Neil Hyman, MD, of the University of Chicago, sounded a note of caution about the time and expense in using AI-assisted technology. “You can’t use all the [appointment] slots on one person,” he said.
“The next step would be to show this has biologic importance,” said Hyman, who was not involved with the research. “The gold standard is showing patients that have colonoscopy with AI have a lower risk of cancer.”
“Adenoma detection rate and adenoma miss rate are intermediate (secondary) endpoints,” he noted. “Sometimes we have to remember that an adenoma never hurt anybody unless they become a cancer.”
For their study, Wallace and colleagues enrolled 230 patients undergoing screening or surveillance colonoscopy across eight centers in the U.S., U.K., and Italy from February 2020 to May 2021. Patients were randomized to receive two consecutive white-light colonoscopies on the same day. The only difference was in the order: either standard colonoscopy followed by another colonoscopy with AI (n=114) or AI colonoscopy followed by standard colonoscopy (n=116). Participants were adults ages 45 and up who were at average risk for colorectal cancer.
AMR was defined as the number of “histologically confirmed” adenomas detected during the second colonoscopy divided by the total number of adenomas from the first and second colonoscopies.
Patients were a mean age of 64 years, about two-thirds were men, and nearly all were white (93%). About half were receiving a surveillance colonoscopy.
Overall, 493 adenomas were detected and removed in both groups, 246 in the AI-first group and 247 in the standard colonoscopy first group. In the AI-first group, 38 additional adenomas were detected on standard colonoscopy, and in the standard colonoscopy first group, another 80 adenomas were detected during the AI-based colonoscopy.
The AI-first colonoscopy group had a lower false negative rate (7% vs 30%) as well as a lower mean number of adenomas and carcinomas (0.33 vs 0.70) detected on second colonoscopy.
Adverse events (AEs) included gastrointestinal disorders, general disorders, and administrative site conditions. There were 17 AEs overall — six in the AI-first group and 11 in the standard colonoscopy first group. Two patients in the standard group discontinued due to AEs, but none were judged to be serious.
Limitations to the data included that only endoscopists with 1,000 completed colonoscopies and an adenoma detection rate ranging from 30% to 70% were included. Also, too few sessile serrated lesions were detected to calculate their miss rate.
Study funding was provided by Cosmo Artificial Intelligence.
Wallace reported relationships with Cosmo Pharmaceuticals, Fujifilm, Olympus, Verily, and Virgo. A coauthor reported relationships with Fujifilm and Medtronic.