AI-assisted colonoscopy and detection/miss of adenomas and sessile serrated lesions: An updated systematic review and meta-analysis.

4

Background: AI-assisted colonoscopy may improve real-time lesion recognition across multiple quality indicators, including adenoma detection rate (ADR), miss rates, and sessile serrated lesion (SSL) outcomes. Additional trials and endpoints have emerged since prior syntheses, warranting an updated estimate. Methods: PubMed, Scopus, Embase, and the Cochrane Library were searched from inception through 8 January 2026 for studies comparing AI-assisted versus standard colonoscopy. Prespecified outcomes were adenomas per colonoscopy (APC), ADR, adenoma miss rate, SSL detection rate, and SSL miss rate. Two reviewers independently screened and extracted data. Effects were pooled as RRs (dichotomous) or MDs (continuous) with 95% CIs using random-effects models, with fixed-effect models as sensitivity analyses. Heterogeneity was assessed by I², τ², and Q; robustness by leave-one-out, influence diagnostics (Baujat/radial/GOSH), and funnel plots. Results: Twenty-eight studies (N=27,188; AI n=13,393, control n=13,795) were included. ADR improved with AI assistance (28 studies; RE RR 1.19, 95% CI 1.13–1.25; I²=66.6%), including among expert endoscopists (9 studies; n=8,866; RE RR 1.17, 95% CI 1.09–1.25; I²=37.4%). Adenoma miss rate was lower with AI (6 studies; n=2,096 adenomas; RE RR 0.45, 95% CI 0.37–0.54; I²=14.8%). SSL detection improved (19 studies; n=20,603; RE RR 1.15, 95% CI 1.01–1.31; I²=41.8%). APC showed no significant difference (21 studies; n=18,839; RE MD 0.14, 95% CI −0.05 to 0.32; I²=98.6%), with discordant fixed-effect results (MD −0.16, 95% CI −0.19 to −0.12). SSL miss rate evidence was imprecise (4 studies; 107 SSLs; RE RR 0.47, 95% CI 0.15–1.48; I²=65.4%). Influence diagnostics and GOSH analyses confirmed robustness of primary estimates; funnel plots indicated outliers in sparse outcomes. Conclusions: AI-assisted colonoscopy improved ADR, adenoma miss rate, and SSL detection, including among expert endoscopists. SSL miss rate evidence remains limited to 4 studies with sparse events. Further trials are needed to clarify variability across platforms and settings.