Impact of artificial intelligence-assisted colonoscopy on gastroenterology fellow performance: a pragmatic randomized controlled trial.

[BACKGROUND AND AIMS] The substantial miss rate during screening and surveillance colonoscopy, particularly for the right side, underscores the need to improve training. The role of artificial intelligence (AI)-assisted colonoscopy in the training environment has not been thoroughly defined. This study explores the impact of AI on colonoscopy performed by trainees in a gastroenterology (GE) fellowship program.

[METHODS] Between March and October 2023, we randomly assigned GE fellows to AI-enhanced versus conventional colonoscopy (CC) rooms daily. Consecutive colonoscopies performed by fellows were included unless there were attending interventions, inadequate bowel preparation, or incomplete colonoscopy. The primary end point was adenoma detection rate (ADR), defined as the proportion of colonoscopies with 1 or more adenomas detected. Additional outcomes included right-sided colon ADR (RADR) and left-sided colon ADR (LADR), the polyp detection rate, and procedure (colonoscope insertion to withdrawal) and withdrawal (cecum to withdrawal) times. Mean ADR differences for the AI versus CC procedures were estimated using generalized linear models.

[RESULTS] A total of 1045 colonoscopies were performed by 16 fellows. The overall ADR was similar for AI (40.5% ± 3.9%) versus CC (35.0% ± 3.6%), with a mean difference of 5.5% (95% CI, -4.3% to 15.3%). The RADR was higher in AI (24.1%) versus CC (16.5%), with a mean difference of 7.6% (95% CI, 1.7%-13.5%). Among 130 screening colonoscopies, ADR for AI was 49.1% versus 26.7% for CC, with a mean difference of 22.3% (95% CI, -2.7% to 47.4%), whereas RADR was higher for AI (AI: 35.1% vs CC: 13.7%), with a mean difference of 21.0% (95% CI, 7.6%-35.2%). This was most pronounced for first- and second-year fellows. There was no difference in procedural or withdrawal time with the addition of AI.

[CONCLUSIONS] This pragmatic randomized controlled trial demonstrates that AI-assisted colonoscopy improves RADR for GE trainees. The overall ADR was not significantly different between groups. We propose a use case via AI-assisted colonoscopy for trainees guiding improvement of adenoma detection in the right side of the colon and standardizing a critically needed colorectal cancer screening quality metric. (Clinical Trials.gov Identification NCT05423964.).