ARTIFICIAL INTELLIGENCE-ASSISTED COLONOSCOPY FAILED TO INCREASE ADENOMA DETECTION RATE: AN IMPLEMENTATION STUDY

Introduction
Artificial intelligence-assisted colonoscopy (AIAC) is a promising technology to improve colonoscopy quality by detecting precancerous polyps that may be missed during unassisted colonoscopy. Multiple studies have found improved adenoma detection rate (ADR) with AIAC. However, a few practical implementation studies have shown equal or decreased ADR with AIAC. Thus, more implementation studies are needed to determine the real-world impact of AIAC.

Methods
In this single center, quasi-experimental, implementation study, we aimed to compare our primary endpoint (ADR) between 2 cohorts: a 6-month historical cohort of patients undergoing standard non-AI colonoscopy versus a 6-month prospective cohort undergoing AIAC. Both cohorts included colonoscopy for screening, surveillance, or positive fecal immunochemical test (FIT) indications. Secondary endpoints included sessile serrated lesion (SSL) detection rate, advanced ADR, hyperplastic polyp resection rate, and other benign, non-adenomatous, non-hyperplastic pathology resection rate (e.g. lymphoid aggregate, normal mucosa). Statistical analysis was performed using Pearson’s chi-square or two sample T-test.

Results
We evaluated 441 non-AI (November 2022-April 2023) and 604 AIAC cases (May-October 2023). Groups were balanced without statistically significant differences in patient demographics, endoscopists involved, AI technology used, procedure time, or mean polyps detected per colonoscopy. ADR and SSL detection rates were higher in unassisted colonoscopy compared to AIAC: 85.3% versus 78.5% (p=0.0054) and 21.3% versus 14.9% (p=0.0072), respectively. Alternatively, rates of benign, non-adenomatous, non-hyperplastic resection were higher in AIAC (31.5%) versus unassisted colonoscopy (22.9%) (p=0.0023) (Table 1). Subgroup analysis by colonoscopy indication revealed surveillance colonoscopies largely accounted for the lower ADR (87.2 vs 78.7%; p=0.0032) and the higher benign, non-adenomatous, non-hyperplastic pathology resection rate with AI (22.9% versus 31.8%, p=0.0093). The reduction in SSL detection rate with AIAC was most notable in FIT positive cases (30.9% versus 10.9%, p=0.0006) (Figure 1).

Discussion
Our real-world implementation study shows lower ADR in AIAC compared to unassisted colonoscopy. While this unexpected result challenges prior randomized controlled trial data, it is in line with recent pragmatic trials that did not show an ADR benefit with AI. Possible contributors may include a high baseline ADR for our endoscopists (historical range 45-78%, average 63%), a lower SSL detection rate in AIAC specifically in FIT-positive cases which may reflect endoscopist overreliance on AI technology, and a shift in focus to benign, non-adenomatous, non-hyperplastic pathology detected with AI that would otherwise go undetected or unresected by endoscopists during unassisted colonoscopy.