Artificial intelligence–assisted colonoscopy and adenoma detection: An updated systematic review and meta-analysis of 42 studies.

Abstract

2 Background: Colonoscopy misses a meaningful proportion of adenomas, limiting preventive effectiveness. We updated the evidence to evaluate AI-assisted versus standard colonoscopy on adenoma detection rate (ADR), adenomas per colonoscopy (APC), adenoma miss rate (AMR), and withdrawal time. Methods: PubMed, Scopus, Embase, and the Cochrane Library were searched from inception to January 9, 2026 (PRISMA-compliant). Risk ratios (RRs) and mean differences (MDs) with 95% CIs were pooled using random-effects models. Heterogeneity was assessed by I² and τ²; prediction intervals (PIs) calculated. Robustness was evaluated by leave-one-out analyses, Baujat plots, Cook's d, and GOSH diagnostics; small-study effects by funnel plots. Results: Across 42 studies (n=34,699), AI increased ADR (RR 1.19, 95% CI 1.13–1.27; I²=67.2%; PI 0.96–1.56); leave-one-out estimates were stable (RR 1.21–1.23 across all omissions). Baujat and Cook's d identified Aniwan 2023 (RR 0.76; sole negative-direction study) and Gong 2020 (RR 2.11) as most influential; GOSH confirmed distributed heterogeneity. Funnel asymmetry was right-sided, consistent with small-study effects. APC increased across 29 studies (n=26,163; MD 0.21, 95% CI 0.15–0.27; I²=77.5%; PI −0.05 to 0.46); leave-one-out estimates were stable (MD 0.20–0.22). Lau 2024 and Tiankanon 2023 were the most influential studies; Wallace 2022 (MD −0.59) was the sole negative outlier but its omission did not shift the pooled estimate (MD 0.22; I²=77.3%). AMR was reduced across 6 studies (n=2,273; RR 0.53, 95% CI 0.36–0.78; I²=77.4%; PI 0.16–1.76); Lui 2023 (RR 1.47) was identified as the dominant heterogeneity source across all influence metrics — its omission shifted the pooled RR to 0.46 (0.36–0.59) with I² reducing to 58.4%, and GOSH showed a bifurcated distribution attributable to this study. The remaining five AMR studies were directionally consistent. Withdrawal time was longer with AI across 34 studies (n=29,156; MD 0.48 min, 95% CI 0.29–0.67; I²=85.8%; PI −0.54 to 1.49); Thiruvengadam 2024 omission produced the largest shift (MD 0.43; I²=84.5%); Baujat identified Gong 2020 and Thiruvengadam 2024 as most influential; GOSH confirmed distributed heterogeneity; funnel asymmetry was right-sided. Conclusions: AI-assisted colonoscopy improved ADR, APC, and AMR across 42 studies, with a modest increase in withdrawal time. Prediction intervals for all outcomes crossed the null, indicating benefit is not uniform across settings. Standardized trials stratified by endoscopist experience and practice volume are needed to identify which populations derive the most consistent benefit.