The NMA was conducted according to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020 statement and the PRISMA extension for NMA (Multimedia Appendices 1-3) [31,32]. The protocol was registered with PROSPERO (International Prospective Register of Systematic Reviews; registration number CRD420251266932). No deviations from the registered protocol occurred.

We systematically searched PubMed (via MEDLINE), Cochrane Library (via Wiley), Embase (via Ovid), and Web of Science (Core Collection) from inception to July 25, 2026, for RCTs comparing AI-assisted colonoscopy with standard white-light colonoscopy for polyp detection. The search strategies combined MeSH terms, Emtree terms, free-text words, and Boolean operators, structured around 3 core concepts: population (colonoscopy), intervention (AI), and study design (RCTs), following Cochrane Handbook recommendations. No language or publication date restrictions were applied during the initial database search. The complete search strategies for each database, including all MeSH terms, Emtree terms, field modifiers, Boolean operators, and filters, are provided in Multimedia Appendix 4. Additionally, we performed backward citation searching (screening reference lists of all included studies and relevant reviews) and forward citation searching (using Web of Science and Google Scholar) to identify additional eligible trials. Clinical trial registries (ClinicalTrials.gov and World Health Organization International Clinical Trials Registry Platform) were also searched. The search was updated on July 25, 2026, to capture any new studies published since the initial search. No additional eligible studies were identified.

Inclusion criteria were RCTs comparing AI-assisted colonoscopy systems with standard white-light colonoscopy, adult patients (≥18 years of age) undergoing colonoscopy for screening, surveillance, or diagnostic indications, reported outcomes stratified by polyp size (diminutive: ≤5 mm; small: 6-9 mm; and large: ≥10 mm) and English-language full-text articles with sufficient methodological detail for data extraction and risk-of-bias assessment. Studies were grouped for synthesis by polyp size (3 groups).

Exclusion criteria included inflammatory bowel disease or hereditary polyposis syndromes, studies without size-stratified data, nonrandomized designs, and conference abstracts with insufficient methodological details.

Two reviewers independently screened titles, abstracts, and full texts. Disagreements were resolved by discussion or consultation with a third reviewer. No automation tools were used for screening.

Using standardized forms, 2 reviewers independently extracted data from each eligible study: first author, publication year, country, study design, sample size, patient demographics (age and sex), AI system name, and mean number of polyps stratified by size (diminutive: ≤5 mm; small: 6-9 mm; and large: ≥10 mm). For studies reporting multiple AI arms, each AI system was treated as a separate intervention node. Disagreements were resolved by consensus. We did not contact study authors for additional data because all necessary information was available in the published reports. We confirmed that no double counting of participants occurred across studies, as each study contributed independent participants to a single comparison.

For each outcome (mean polyp detection count by size), we gathered all relevant measures, time points, and analyses. When multiple results existed, we prioritized the fully adjusted estimate. We also collected data on participant demographics, intervention details, comparators, and funding sources. Missing or unclear information was considered unreported, and no data were imputed.

Two reviewers independently assessed the risk of bias in the included studies using the Cochrane Risk of Bias 2.0 (RoB 2.0) tool, which is the current gold-standard tool for assessing bias in randomized trials [33]. The RoB 2.0 tool evaluates bias across 5 domains: bias arising from the randomization process, bias due to deviations from intended interventions, bias due to missing outcome data, bias in measurement of the outcome, and bias in selection of the reported result. For each domain, risk was judged as “low risk,” “some concerns,” or “high risk.” The overall risk of bias for each study was determined based on the domain-level judgments: “low risk” if all domains were low risk, “some concerns” if at least 1 domain had some concerns but no domain was high risk, and “high risk” if any domain was high risk or if multiple domains had some concerns that substantially lowered confidence in the result. Disagreements between reviewers were resolved by discussion or consultation with a third reviewer. Given the nature of endoscopic interventions, blinding of participants and personnel was anticipated to have some concerns or high risk in most studies, which is a common limitation in this field. We acknowledge that our PROSPERO protocol (CRD420251266932) originally specified RoB 1.0; however, we opted to use RoB 2.0 for the present analysis to align with current methodological standards and to strengthen the methodological rigor of our review, as recommended by the editor and recent Cochrane guidance.

Primary outcome rationale: we selected mean polyp detection count stratified by size as the primary outcome because it was the only size-stratified measure consistently reported as a continuous variable across included trials. Patient-level detection outcomes such as ADR or PDR were rarely reported in a size-stratified manner, precluding NMA. Standardized mean differences (SMDs) with 95% CIs were therefore calculated. Positive SMD values indicate higher mean polyp detection counts with AI-assisted colonoscopy. We acknowledge that SMD is a relative measure and does not convey absolute incidence; consequently, interpretations regarding surveillance intervals, cost-effectiveness, or platform selection are drawn cautiously and considered exploratory. For ratio measures, we analyzed on the logarithmic scale, but the primary analysis used SMD as all included studies reported continuous outcomes. Across all comparisons, positive SMD values indicate higher mean polyp detection counts in the AI-assisted colonoscopy group relative to standard colonoscopy, reflecting improved detection.

Because the standard DerSimonian-Laird method can result in inflated type I error rates when the number of studies is small, or heterogeneity is present, we used the HKSJ method for random-effects meta-analyses [34]. In the context of our NMA, the HKSJ method was applied at the pairwise comparison level. For each pairwise comparison between an AI system and standard colonoscopy, the HKSJ method was used to calculate the random-effects pooled estimate and its 95% CI, as implemented in the rma() function of the metafor package (version 4.4.3) in R (R Foundation for Statistical Computing) with restricted maximum likelihood (REML) estimation and test=“knha.” This approach accounts for both within-study and between-study variance, providing more conservative CIs when the number of studies is small, or heterogeneity is substantial—conditions that characterize our network, which includes only 13 studies and multiple single-study nodes. Additionally, the HKSJ method was used to calculate 95% PIs for each pooled estimate, quantifying the expected range of true effects in future studies and providing a clinically interpretable measure of heterogeneity impact.

All cross-platform comparisons are indirect (via the common comparator of standard colonoscopy) and were estimated using the netmeta package in Stata MP (version 14; Stata Corp); the HKSJ adjustment was applied consistently across both direct and indirect estimates to ensure uniform handling of uncertainty. Critically, this 2-stage HKSJ workflow is statistically valid for our star-shaped evidence network with a single shared standard colonoscopy comparator. All individual AI-vs-control pairwise SMD values were first adjusted via HKSJ within R’s metafor package (rma() function, test=“knha,” REML τ² estimation). These fully corrected pairwise effect sizes were then imported into Stata’s netmeta module to generate indirect cross-AI comparative effects. This sequential pipeline retains the conservative CI advantage of HKSJ even for indirect network outcomes, addressing the inflated type I error limitation of traditional DerSimonian-Laird pooling when trial counts are sparse and inter-study heterogeneity is substantial. All 95% PIs reported in this manuscript were exclusively derived from HKSJ-adjusted pooled summary estimates.

The HKSJ method provides more conservative and accurate CIs compared to the standard DerSimonian-Laird approach, particularly when the number of studies is limited, as is the case in our network. Between-study heterogeneity variance (τ²) was estimated using REML. To quantify the clinical impact of heterogeneity, we calculated 95% PIs for the summary effect estimates of each comparison (where at least 5 studies were available). The PI provides the range within which the true effect of a future study is expected to lie. We prepared data for synthesis by converting all means, SDs, and sample sizes into SMDs using the metafor package in R (version 4.4.3). No missing summary statistics required imputation. We used tabular displays (tables of study characteristics) and graphical methods (forest plots, network plots, SUCRA bar charts, funnel plots, and league tables) to present results. Frequentist NMAs were conducted using the netmeta package in Stata MP 14.

The network geometry was visualized with node size proportional to sample size and edge thickness to the number of studies. Consistency was evaluated using a global consistency test via a design-by-treatment interaction model and local consistency via a loop-specific approach. Inconsistency was defined as P<.10. Because the network was star-shaped with no direct head-to-head trials between AI systems, the consistency assumption could not be formally tested for loops; we therefore relied on the global test across designs (all P>.05). SUCRA was estimated to provide exploratory treatment hierarchies. Because the network is star-shaped with no direct head-to-head trials, SUCRA values are presented as exploratory only and are not interpreted as evidence of superiority. To explore possible causes of heterogeneity, we performed prespecified subgroup analyses by polyp size (the primary grouping). Meta-regression was not performed due to the limited number of studies. We also conducted sensitivity analyses excluding the Tianjin YuJin system (due to potential conflict of interest). Using alternative effect measures (risk difference) to test robustness (results not materially different, data not shown).

We assessed funnel plot asymmetry to examine small-study effects, which may arise from publication bias, selective outcome reporting, true heterogeneity, or chance [35]. Funnel plots and Egger linear regression test were used when ≥10 studies were available. Asymmetry was considered significant at P<.10. Importantly, funnel plot asymmetry should not be equated with publication bias alone, as small-study effects can result from other factors such as differences in study methodology, true variation in effect estimates across study sizes, or random variation. We therefore interpreted the results of the Egger test cautiously, recognizing that significant asymmetry indicates the presence of small-study effects rather than definitively confirming publication bias.

We assessed the overall certainty of evidence for the primary outcomes (mean polyp detection count stratified by size) using the GRADE framework (Multimedia Appendix 5). The certainty was rated as high, moderate, low, or very low based on 5 domains: risk of bias, inconsistency (including width of the 95% PI), indirectness, imprecision, and publication bias. GRADE assessments were performed independently by 2 reviewers, with disagreements resolved by consensus. Data and code availability: all extracted data are provided in Multimedia Appendix 6. The R code for HKSJ meta-analysis, PIs, and figures is available in Multimedia Appendix 7.

This systematic review and NMA used published data from previously approved RCTs; no primary data collection or patient contact was required. The protocol was registered with PROSPERO (CRD420251266932) and conducted according to PRISMA 2020 guidelines.

The literature search identified 3247 records, of which 13 RCTs comprising 4156 participants were included in the NMA (Figure 1). After removing duplicates, 2 reviewers screened titles and abstracts, followed by full-text assessments. A total of 13 RCTs with 4156 participants were included in the NMA. Only English-language full-text articles met the inclusion criteria, as non-English records lacked sufficient methodological detail or were conference abstracts with incomplete data. A total of 7 studies were conducted in China, 2 in the United States, and 1 each in Japan, the United Kingdom, and international multicenter settings. Overall, 8 AI-assisted systems were evaluated: EndoScreener (n=3), EndoAngel (n=2), CAD EYE (n=2), AI-assisted pooled systems (n=2), and single studies for Wision AI, Tianjin YuJin, GI Genius, and Olympus. Standard white-light colonoscopy served as the common comparator in all studies; no head-to-head trials directly compared AI systems (Table 1). An updated search performed on July 25, 2026, identified 1247 new records. After full-text review, no additional studies met the inclusion criteria.

The network formed a star-shaped structure with standard colonoscopy as the central common comparator. A total of 8 AI-assisted systems connected directly to standard colonoscopy, with no direct connections between AI systems (Figure 2). Global consistency tests revealed no significant inconsistency across size-stratified networks (all P>.05), but local consistency could not be evaluated due to the lack of closed loops, relying instead on clinical plausibility.

A random-effects NMA using the HKSJ method with REML estimation was performed. Exploratory SUCRA values were calculated for hypothesis generation, but these rankings should not be seen as evidence of clinical superiority due to the sparse network and indirect comparisons. It is important to emphasize that these SUCRA rankings are derived solely from indirect comparisons and should not be interpreted as evidence of superiority of one AI system over another. The rankings are provided for hypothesis-generating purposes only, to help prioritize which AI systems should be compared in future head-to-head trials. No clinical recommendations can be made based on these rankings.

This NMA enables indirect comparative efficacy evaluation across all 8 AI colonoscopy systems, an analytical output that cannot be obtained from separate pairwise meta-analyses alone. NMA of 11 study arms comparing 7 AI systems vs standard colonoscopy demonstrated significant heterogeneity in AI system efficacy for increasing diminutive polyp counts (I²=86.6%; τ²=0.0332). The pooled SMD indicated a slight advantage for AI-assisted colonoscopy (SMD 0.21, 95% CI 0.07-0.35), but the wide 95% PI (–1.12 to 1.54) crosses the null threshold and includes negative SMD values, demonstrating that AI-assisted colonoscopy may deliver no polyp detection benefit, or even inferior performance compared to standard white-light colonoscopy in specific patient cohorts, endoscopic workflows, or operator groups. EndoScreener was associated with a statistically significant increase in mean polyp detection count compared with standard colonoscopy. Indirect comparisons between AI systems revealed no statistically significant differences (Figure 3 [18-30]). SUCRA values are presented for hypothesis generation only and should not be interpreted as evidence of clinical superiority due to the absence of direct head-to-head comparisons between AI systems. Rankings are based solely on indirect comparisons via standard colonoscopy. Olympus ranked highest (100%), followed by Wision AI (72%), AI-assisted pooled systems (67.9%), EndoScreener (54.1%), GI Genius (49.9%), Tianjin YuJin (33%), EndoAngel (31.8%), and CAD EYE (0%; Figure 4). Values represent SMDs (95% CIs) for row intervention minus column intervention. All comparisons are indirect via standard colonoscopy; no head-to-head trials exist between AI systems. Nonsignificant comparisons are indicated by CIs crossing zero (Figure S4 in Multimedia Appendix 8). According to GRADE, the certainty of this evidence was very low, downgraded for risk of bias (unblinded operators in 12/13, 92.3% of studies), serious inconsistency (I²=86.6%, 95% PI crossing the null), and indirectness (predominantly Asian populations). All pooled estimates were obtained using the HKSJ random-effects method with REML estimation.

This NMA enables indirect comparative efficacy evaluation across all 8 AI colonoscopy systems, an analytical output that cannot be obtained from separate pairwise meta-analyses alone. For small polyps (6-9 mm), effect estimates were near null, and SUCRA rankings varied, with Olympus again leading. The lack of direct trials and minimal pooled effect prevent definitive conclusions on platform performance for small polyps. We advise against viewing high exploratory SUCRA values for any system as evidence of clinical superiority for small polyps, as this may be due to higher baseline detection rates, varying size definitions, operator learning, or statistical uncertainty (Figure 5 [18-30]). Olympus ranked highest (100%), followed by Tianjin YuJin (88.9%), EndoScreener (76%), Wision AI (67.2%), EndoAngel (51.1%), AI-assisted pooled systems (48.7%), GI Genius (48.5%), and CAD EYE (0%). These rankings should not be interpreted as evidence of clinical superiority (Figure 6). All comparisons are indirect via standard colonoscopy. The minimal effect sizes and overlapping CIs across all comparisons indicate no significant differences between AI systems for small polyp detection (Figure S6 in Multimedia Appendix 8). The certainty of this evidence was low according to GRADE, downgraded for risk of bias and indirectness. The 95% PI was –0.03 to 0.07, confirming that the near-null average effect is consistent across settings. The narrow 95% PI (–0.03 to 0.07) also spans the null value of zero, meaning consistent polyp detection superiority of AI cannot be guaranteed across diverse real-world clinical settings; null or inferior detection outcomes are plausible in some trial populations.

This NMA enables indirect comparative efficacy evaluation across all 8 AI colonoscopy systems, an analytical output that cannot be obtained from separate pairwise meta-analyses alone. For large polyps (≥10 mm), no AI system showed significant improvement over standard colonoscopy (pooled SMD 0.01, 95% CI 0.00-0.02). Tianjin YuJin had the highest SUCRA (100%), followed by Olympus (79.6%) and GI Genius (76.9%), but this ranking is based on a single small study (n=128) and should not be seen as proof of clinical superiority. A sensitivity analysis excluding Tianjin YuJin, due to potential bias, did not change the overall finding, confirming no meaningful AI advantage for large polyps. The minimal SUCRA separation and overlapping CIs suggest that AI assistance adds limited value for large, more visible lesions (Figure 7 [18-30]). Tianjin YuJin ranked highest (100%), followed by Olympus (79.6%), GI Genius (76.9%), EndoScreener (35.8%), EndoAngel (32.3%), Wision AI (29%), AI-assisted pooled systems (26.5%), and CAD EYE (0%). Note that the top ranking of Tianjin YuJin is derived from a single small study (n=128) and does not constitute evidence of clinical superiority (Figure 8). The near-zero effect estimates and wide CIs preclude definitive conclusions regarding relative efficacy of different AI systems for large polyp detection (Figure S5 in Multimedia Appendix 8). The certainty of this evidence was low according to GRADE, downgraded for imprecision (few studies contributing data) and indirectness. The 95% PI was –0.01 to 0.03, indicating that the absence of benefit is consistent across settings. The 95% PI (–0.01 to 0.03) crosses zero, consistent with the near-null pooled SMD; AI assistance provides no reliable incremental polyp detection advantage for large lesions across heterogeneous clinical environments.

Excluding the Tianjin YuJin system did not significantly alter findings across polyp sizes. The pooled SMD for diminutive polyps remained stable, and estimates for small and large polyps were unchanged (Figures S1-S3 in Multimedia Appendix 8). The sensitivity analysis excluding Tianjin YuJin confirmed that the exclusion did not materially change any pooled estimates, suggesting that the potential conflict of interest did not introduce systematic bias. However, the GRADE certainty remained unchanged (very low for diminutive, and low for small and large polyps), as the primary limitations—unblinded operators and heterogeneity-were inherent to the included studies rather than driven by a single trial.

No new significant indirect comparisons between AI platforms were found. Risk of bias was assessed using the Cochrane RoB 2.0 tool across 5 domains for each included study (Figure 9 [18-30] and Table 2). Across the 13 included trials, the overall risk of bias was judged as low in 2 (15.4%) studies, as having some concerns in 5 (38.5%) studies, and as high in 6 (46.2%) studies. The most common source of bias was the “deviations from intended interventions” domain, which was rated as high risk or having some concerns in the majority of studies due to the inherent inability to blind endoscopists and participants to the AI intervention. This is a methodological constraint intrinsic to AI-assisted colonoscopy trials rather than a marker of poor study conduct. The “measurement of the outcome” domain was predominantly low risk, as polyp detection counts were objectively measured and histopathologically confirmed. The “selection of the reported result” domain showed some concerns in 5 (38.5%) studies, primarily due to insufficient preregistration of analysis plans. Funnel plots and Egger test were used to assess small-study effects, which could result from factors like publication bias or true heterogeneity (Figure 10 [18-30]). The limitations identified in the RoB 2.0 assessment were fully considered in the GRADE certainty assessment and in the interpretation of all findings.

We assessed the certainty of evidence for the primary outcomes using the GRADE framework (Table 3). For diminutive polyps (≤5 mm), the certainty was rated as very low, downgraded for risk of bias (unblinded operators in most studies), substantial inconsistency (I²=86.6%, wide 95% PIs), and indirectness (predominantly Asian populations). For small polyps (6-9 mm), certainty was low, downgraded for risk of bias and indirectness. For large polyps (≥10 mm), certainty was low, primarily due to imprecision from sparse data and indirectness. No outcome achieved moderate or high certainty, reflecting the methodological limitations of the current evidence base.

Egger test showed significant small-study effects for small polyps (P=.03), borderline for large polyps (P=.09), and none for diminutive polyps (P=.36). With only up to 13 studies and a fragmented network, these tests have low statistical power, and the observed asymmetry should be interpreted cautiously. As discussed in the Methods section, funnel plot asymmetry is indicative of small-study effects, which may reflect publication bias, selective outcome reporting, true heterogeneity, chance, or differences in methodological quality across studies. Therefore, while we cannot exclude the possibility of reporting bias, the observed asymmetry may also be attributable to the limited number of small studies, true variation in effect estimates, or chance. We did not interpret these results as definitive evidence of publication bias (Figure 10).