We performed an endoscopist-level meta-analysis and meta-regression of randomized controlled trials to clarify the role of baseline endoscopist performance in CADe effectiveness in colonoscopy and guide evidence-based implementation

We systematically searched MEDLINE, Embase, CENTRAL, Web of Science, and clinical trial registries from inception to December 2024 for RCTs comparing real-time computer-aided detection (CADe) with high-definition white-light colonoscopy in adults. Trials were eligible if endoscopist-level adenoma detection rate (ADR) data were available; when pre-trial baseline ADRs were not reported, control-arm ADR was used as a proxy. Secondary outcomes included SSLDR, PDR, APC, NNLPC (non-neoplastic lesions per colonoscopy), and IT (inspection time). Endoscopists performing ≥10 colonoscopies per arm were stratified into baseline ADR quintiles (Q1 <25%, Q2 25–34.9%, Q3 35–44.9%, Q4 45–54.9%, Q5 ≥55%). Meta-analyses were conducted separately within each quintile, including only endoscopist-specific study data belonging to that quintile. Risk ratios (RR) with 95% CI were calculated for categorical outcomes and mean differences with 95% CI for continuous outcomes. Meta-regression assessed whether CADe benefit varied with baseline ADR using (1) log-RR as a function of ADR quintile and (2) ΔADR as a function of continuous baseline ADR. Analyses were performed using OpenMeta, RevMan, and Python enabling random-effects modeling, robust variance estimation, visualization, and influence diagnostics.

A total of 25 studies with 256 endoscopists were included (11532 patients in CADe arm, 11238 in HD-WL arm). When stratified by baseline ADR, CADe demonstrated a clear gradient of effect (n=256 endoscopists). The greatest benefit was observed in Q1 (pooled RR 1.54 [95% CI 1.37–1.73]), declining across Q2–Q4 (RRs 1.23 [1.10–1.38], 1.16 [1.06–1.26], 1.11 [1.05–1.18]). In Q5, CADe conferred no significant benefit (RR 0.95 [0.89–1.02]). PDR and APC benefits mirrored ADR across Q1–Q4, with no increase in Q5 (n=190 endoscopists). Higher-quintile endoscopists (Q3–Q5) spent more time inspecting the colon with CADe (n=174). NNLPC was similar across arms (n=110). Meta-regression treating baseline ADR as ordinal predictor showed that each one-quintile increase reduced CADe benefit by ~9% (β = −0.0957, p < 0.0001). Using ADR as a continuous predictor, each 1% higher baseline ADR corresponded to ~0.29% smaller improvement with CADe (β = −0.00288, p < 0.001).

These meta-analyses and meta-regression of quintile-level data derived from 256 endoscopists’ baseline ADR across 25 studies demonstrate that CADe is most effective for endoscopists with low-to-moderate baseline ADR, substantially improving detection and possibly equalizing quality. For highest detectors, incremental benefit is minimal and may be offset by efficiency trade-offs. Future research should assess long-term interval cancer outcomes, cost-effectiveness, and optimal integration into training and quality assurance frameworks.