Hybrid endoscopic submucosal dissection enhances efficiency with comparable curative outcomes: a Bayesian meta-analysis, systematic review, and Monte Carlo simulation of randomized controlled trials

Abstract

Background/Aims

Gastrointestinal (GI) cancers accounted for 3.24 million deaths worldwide in 2022. Endoscopic submucosal dissection (ESD) enables en bloc resection of early neoplasms but is time-intensive and technically challenging. Hybrid ESD (H-ESD), which combines dissection with snare polypectomy, may improve efficiency while maintaining efficacy. This Bayesian meta-analysis of randomized controlled trials (RCTs) compares the safety and efficacy of H-ESD with conventional ESD (C-ESD).

Methods

We searched the PubMed, Embase, Cochrane, Scopus databases for RCTs comparing hybrid and C-ESD for superficial GI neoplasms. Bayesian modeling in Just Another Gibbs Sampler used normal-normal hierarchies for procedure time and logRR for binary outcomes with weakly informative priors. Frequentist analyses were performed using Der-Simonian-Laird τ². Heterogeneity was assessed using I²/τ².

Results

H-ESD showed comparable en bloc (RR, 0.97; 95% [CrI], 0.92–1.03; I²=22.8%) and R0 resection rates (RR, 0.97; 95% CrI, 0.90–1.04; I²=0%). Procedure time decreased significantly with H-ESD (MD, –18.05; 95% CrI, –27.77 to –8.33 minutes; I²=66.5%). Adverse events (RR, 1.08; 95% CrI, 0.44–2.63) and perforations (RR, 1.05; 95% CrI, 0.34–3.19) were similar between the two procedures.

Conclusions

H-ESD matches C-ESD in curative resection while reducing the procedure time, offering a safer and more efficient alternative. Larger RCTs are needed to confirm the long-term benefits.

Graphical abstract

INTRODUCTION

Gastrointestinal (GI) cancers exact a profound global toll, encompassing 4.78 million new cases and 3.24 million deaths in 2022—23.9% and 33.2% of all cancers, respectively. Their incidence is projected to surge beyond 5.2 million cases and 3.5 million deaths by 2025 amid aging populations and escalating risk factors like obesity and processed diets.1,2 This growing burden, particularly from colorectal, gastric, and esophageal cancers, heightens the urgency for early detection and minimally invasive treatments that can spare patients from the rigors of open surgery while preserving their quality of life.

Over the past two decades, endoscopic submucosal dissection (ESD) has become a cornerstone for managing early GI neoplasms, notably early gastric cancer and large colorectal lesions, offering precise en bloc resection, which reduces recurrence rates to as low as 1%–2%.3 However, conventional ESD (C-ESD) demands considerable skill, often stretching procedure times beyond 60–120 minutes, and carries risks like perforation and bleeding, challenges that are especially pronounced in the colon’s delicate anatomy.4,5 These hurdles not only strain endoscopic resources but also increase patient discomfort and healthcare costs.6

Hybrid ESD (H-ESD) offers a promising evolution, blending submucosal dissection with snare polypectomy to shorten procedures by up to 50% while maintaining en bloc margin-negative outcomes.7 Although observational studies and meta-analyses have suggested colorectal benefits and safety equivalence of H-ESD, a definitive synthesis of randomized evidence remains elusive.7,8 Through this Bayesian meta-analysis of randomized controlled trials (RCTs), we sought to harness probabilistic modeling to clarify the efficacy and safety of H-ESD, potentially shaping more efficient patient-centered endoscopic approaches to tackle the global GI cancer challenge.

To further bridge statistical findings to clinical practice, we employed Monte Carlo simulations using Bayesian posteriors to estimate the probabilistic effects on time, cost, and safety in a high-volume endoscopy setting.

METHODS

This Bayesian meta-analysis of RCTs was conducted in accordance with Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines.9 We aimed to compare H-ESD with C-ESD for the following key outcomes: en bloc resection rate, R0 resection rate, procedure time, overall adverse events, and perforation.

Search strategy

A comprehensive literature search was performed using PubMed, Embase, the Cochrane Central Register of Controlled Trials, Scopus, and Web of Science from inception to 30th June 2025. The keywords included “hybrid endoscopic submucosal dissection,” “conventional ESD,” “randomized controlled trial,” “en bloc resection,” “R0 resection,” “procedure time,” “adverse events,” and “perforation,” etc. with Medical Subject Headings terms and Boolean operators (Supplementary Material 1). Gray literature was searched via ClinicalTrials.gov and conference abstracts (e.g., Digestive Disease Week).

Study selection and data extraction

We included RCTs comparing H-ESD and C-ESD in adults with superficial GI neoplasms (gastric, esophageal, and colorectal cancer), and binary outcomes (en bloc/R0 rates, adverse events, perforation) reported as events/population, and continuous outcomes (procedure time) reported as mean±standard deviation (SD). The exclusion criteria were non-RCTs, animal studies, abstracts without data, or duplicate publications. Two reviewers independently screened titles, abstracts, and full texts, and resolved disagreements by consensus or a third reviewer. To assess potential sponsorship bias, intervention details, outcome measures, and funding sources for each included trial were independently extracted by two reviewers using an Excel sheet (Supplementary Table 1, Supplementary Material 2). For binary outcomes, event counts and totals were extracted; for procedure time, mean, SDs, and sample sizes were extracted. For this analysis, H-ESD was strictly defined as planned H-ESD (circumferential mucosal incision and partial submucosal dissection, followed by snare resection). Studies evaluating salvage H-ESD (snare resection attempted only after C-ESD became technically unfeasible) were excluded to maintain procedural consistency and to avoid selection bias toward difficult lesions.10

Risk of bias assessment

The risk of bias in RCTs was assessed independently by two reviewers by using the Cochrane Risk of Bias 2.0 (RoB 2.0) tool and supplemented by funnel plots and contour-enhanced assessments for publication bias (Supplementary Fig. 1), which assesses five different domains of risk of bias: (1) randomization process; (2) deviations from intended interventions; (3) missing outcome data; (4) measurement of the outcome; and (5) selection of the reported result. Studies were classified as having low risk, some concerns, or a high risk of bias.11 Any disagreements were discussed and resolved. This tool evaluates five domains: randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result. A detailed breakdown of the risk of bias for each specific domain per study is provided in Supplementary Figures 2–6. Funnel plots were obtained.

Bayesian random-effects meta-analysis and statistical analysis

We conducted a Bayesian random-effects meta-analysis using R ver. 4.3.1 (R Foundation for Statistical Computing) with the meta package (ver. 8.2-1) for frequentist methods and rjags (ver. 4.3.1) with coda for Bayesian modeling. For binary outcomes (en bloc/R0 rates, adverse events, perforation), risk ratios (RRs) and odds ratios (ORs) were pooled using the inverse variance method with both common and random-effects models (Der-Simonian-Laird estimator for τ²), while for the continuous outcome (procedure time), mean differences (MDs) were calculated. Heterogeneity was evaluated using I², τ², and the Q-test, with I² >50% indicating substantial variability; a continuity correction of 0.5 was applied for zero-event studies. The Bayesian random-effects model employed a normal likelihood for logRR or MD, with priors set as μ ~ Normal(0, 2) and τ ~ Half-Normal(0, 1), and sensitivity analysis used a more informative prior (μ ~ Normal(0, 0.5)); four Markov chain Monte Carlo (MCMC) chains ran 1,000 iterations post-1,000 burn-in, with convergence assessed using trace plots, autocorrelation plots, and the Gelman-Rubin statistic (<1.1). Diagnostic tools included posterior densities, predictive intervals, Q-Q plots, and Baujat plots, with sensitivity analyses testing before robustness, all implemented using the R packages ggplot2, bayesplot, car, and MASS.

Given the limited number of available RCTs, a Bayesian framework was specifically chosen over traditional frequentist methods alone because Bayesian updating is more robust in low-sample scenarios and provides exact probabilistic statements rather than relying on large-sample approximations.

Due to the inclusion of only three RCTs, a formal subgroup meta-analysis stratified by anatomical location (gastric vs. colorectal) could not be adequately powered. Instead, the influence of anatomical location was evaluated using a leave-one-out sensitivity analysis.

Monte Carlo simulation

We conducted Monte Carlo simulations to propagate uncertainty from Bayesian posteriors into clinical scenarios for a hypothetical endoscopy center that performs 100 H-ESD procedures annually. Using R, we drew 10,000 samples from approximated distributions: normal for procedure time MD (mean: –18.05 minutes, SD: 4.96 derived from a 95% credible interval [CrI] of –27.77 to –8.33) and log-normal for RRs (adverse events: mean log(1.08), SD: 0.456 from a CrI of 0.44–2.63; perforations: mean log(1.05), SD: 0.526 from a CrI of 0.34–3.19). The baseline C-ESD adverse event rate was 5% and the perforation rate was 1%, based on pooled RCT data and literature. The cost of the endoscopy suite was assumed to be 15 United States dollars (USD)/min. The outputs included the total time savings (positive for H-ESD efficiency), delta adverse events/perforations, and cost savings. Distributions were visualized as histograms.

The protocol for this study is registered on PROSPERO with ID-CRD420251082552.

RESULTS

Study selection

A detailed PRISMA flowchart that comprehensively explains the study selection process is shown in Supplementary Figure 7. All forest plots are shown in Supplementary Figure 2-6.

Outcome: en bloc resection

Analysis of the plots for en bloc resection rates in H-ESD vs. C-ESD revealed no significant differences between the groups. The frequentist forest plots for RR showed a pooled RR of 0.98 (95% confidence interval [CI], 0.94–1.02) in the common-effect model and 0.97 (0.92–1.03) in the random-effects model, with low heterogeneity (I²=22.8%, τ²=0.0006, p=0.2737) (Fig. 1). The pooled OR was 0.43 (95% CI, 0.15–1.26), also with negligible heterogeneity (I²=0%, p=0.6131). The Baujat plot indicated minimal influence of individual studies on heterogeneity, with Esaki et al.12 contributing the most, but still low. The funnel plot was symmetrical, indicating a lack of publication bias. Bayesian posterior density of μ sharply peaked near 0, with predictive density confirming uncertainty around zero (Fig. 2). The Q-Q plot supported the normality of the residuals, and the autocorrelation plots showed rapid decay, indicating good MCMC convergence. Sensitivity analysis demonstrated robustness to prior changes. Overall, these findings suggested comparable en bloc rates, with no evidence of bias or heterogeneity.

Fig. 1.

Frequentist forest plots showing en bloc resection rate. OR, odds ratio; CI, confidence interval; RR, risk ratio.

Fig. 2.

Starting at the top, the figures from left to right show the Bayesian plots for en bloc resection rates: (A) Bayesian forest plot; (B) autocorrelation plots; (C) joint posterior Mu-Tau; (D) trace Plots; (E) posterior density Tau; (F) posterior density Mu.

Outcome: R0 resection rate

The plots for R0 resection rates comparing H-ESD and C-ESD indicated no significant differences between the techniques. Frequentist forest plots showed a pooled RR of 0.97 (95% CI, 0.90–1.04) under common-effects and random-effects models, with zero heterogeneity (I²=0%, τ²=0, p=0.821), and a pooled OR of 0.66 (95% CI: 0.29-1.49) also with negligible heterogeneity (I²=0%, p=0.8613) (Fig. 3). The Baujat plot revealed minimal study influence on heterogeneity, with Esaki et al.12 contributing the most, but still low overall. The normal Q-Q plot confirmed the normality of the residuals as the points aligned closely with the reference line. The Bayesian posterior density of μ sharply peaked near zero, suggesting no effect, while the density of τ was skewed to low values, consistent with low variability. The posterior predictive density was centered near zero with a narrow spread, and the sensitivity analysis showed prior robustness with overlapping distributions. The autocorrelation plots demonstrated rapid decay, confirming the convergence of the MCMC (Fig. 4). Overall, these visuals supported equivalent R0 rates, absent bias or substantial heterogeneity.

Fig. 3.

Frequentist forest plots for the R0 resection rates. RR, risk ratio; CI, confidence interval; OR, odds ratio.

Fig. 4.

Starting from top left, the figures from left to right show the Bayesian plots for R0 resection rate. (A) Bayesian forest plot; (B) autocorrelation plots; (C) joint posterior Mu-Tau; (D) trace plots; (E) posterior density Mu; (F) posterior density Tau.

Outcomes: procedure time

The plots for procedure time comparing H-ESD and C-ESD demonstrated a significant reduction in time with the hybrid methods. The frequentist forest plot for MD showed a pooled MD of –18.05 (95% CI, –27.77 to –8.33) minutes, favoring H-ESD, with moderate heterogeneity (I²=66.5%, τ²=48.9007, p=0.0507) (Fig. 5). Study-specific MDs ranged from –13.20 (Bae et al.13) to –29.30 (Esaki et al.12), with Yang et al.14 at –13.20. The Baujat plot indicated that Yang et al.14 contributed the most to heterogeneity but had a minimal overall influence, whereas Esaki et al.12 had a higher influence on the results. The normal Q-Q plot confirmed the normality of the residuals, with the points aligning well within the envelope. Bayesian posterior density of μ peaked sharply at –18, indicating shorter times for H-ESD, and the density of τ was right-skewed near zero, indicating low variability (Fig. 6). Sensitivity analysis showed overlapping distributions, confirming the robustness of the priors. The posterior predictive density centered at –18 with a moderate spread, and autocorrelation plots showed rapid decay across chains, verifying the convergence. Overall, these visuals supported the efficiency of H-ESD in reducing procedure time with moderate heterogeneity but without substantial bias.

Fig. 5.

Frequentist forest plots of the procedure time in minutes. SD, standard deviation; MD, mean difference; CI, confidence interval.

Fig. 6.

Starting from top left, the figures from left to right show the Bayesian plots for procedure time in minutes. (A) Bayesian forest plot; (B) autocorrelation plots; (C) joint posterior Mu-Tau; (D) trace plots; (E) posterior density Mu; (F) posterior density Tau.

Outcomes: overall adverse events & perforation risk

The plots for overall adverse events in H-ESD versus C-ESD indicated no significant differences between the techniques. Frequentist forest plots revealed a pooled RR of 1.08 (95% CI, 0.44–2.63) and OR of 1.08 (95% CI, 0.40–2.87) under both common- and random-effects models, with zero heterogeneity (I²=0%, τ²=0, p≈0.80) (Fig. 7). Study-specific estimates showed wide CIs, particularly for Esaki et al.,12 with few events. The Baujat plot (not shown but inferred low) indicated minimal influence of the study on the results or heterogeneity. The normal Q-Q plot confirmed the residual normality, with points aligned within the envelope. Bayesian posterior density of μ peaked around –1 to 1, favoring no effect, while the density of τ skewed right near zero, indicating low variability. The posterior predictive density centered near –3 to 3 with moderate spread, and the sensitivity analysis demonstrated prior robustness through overlapping distributions. The autocorrelation plots showed a quick decay to zero by lag five across chains, verifying MCMC convergence (Fig. 8). Overall, the visuals supported comparable adverse event rates without evident bias or substantial variability.

Fig. 7.

Frequentist forest plots for overall adverse events. RR, risk ratio; CI, confidence interval; OR, odds ratio.

Fig. 8.

Starting from top left, the figures from left to right show the Bayesian plots for overall adverse events: (A) Bayesian forest plot; (B) autocorrelation plots; (C) joint posterior Mu-Tau; (D) trace plots; (E) posterior density Mu; (F) posterior density Tau.

The forest plots for the perforation rates comparing H-ESD and C-ESD also indicated no significant differences between the techniques. Frequentist forest plots showed a pooled RR of 1.05 (95% CI, 0.34–3.19) and OR of 1.05 (95% CI, 0.32–3.45) under common- and random-effects models, with zero heterogeneity (I²=0%, τ²=0, p≈0.62) (Supplementary Fig. 6).

Frequentist and Bayesian leave-one-out sensitivity analyses were performed to evaluate the robustness of pooled estimates by sequentially excluding each study. For procedure time, excluding the Esaki et al.’s study12 reduced the MD to –13.2 (95% CI, –20.5 to –5.9) minutes and heterogeneity (I²=0%), suggesting its influence on variability. This indicates that anatomical location is a primary driver of procedural time variance, with hybrid techniques potentially offering different time-saving magnitudes in gastric versus colorectal anatomies. Other outcomes (en bloc, R0, adverse events, and perforation) showed no significant changes in pooled estimates or heterogeneity, reinforcing the robustness across both frequentist and Bayesian models (Supplementary Table 2).

Monte Carlo simulation results

For 100 H-ESD procedures, simulations estimated median total time savings of 30.1 (95% prediction interval [PI], 13.8–46.4; mean, 30.1; SD, 8.3) hours, with the probabilities of >20 hours of savings at 89% and >30 hours at 51% (Fig. 9, approximately normal distribution centered at 30 hours). Median delta for adverse events was 0.4 (95% PI, –4.6 to 6.5; mean, 1.0; SD, 2.9), with a 43% probability of fewer events in H-ESD (Fig. 10, positively skewed distribution with peak near zero and tail to +30). The median cost savings were 27,094 (95% PI, 12,390–41,718; mean, 27,091; SD, 7,464) USD, mirroring the time distribution. The median delta for perforations was 0.05, consistent with safety equivalence (Figs. 9–11).

Fig. 9.

Distribution of total time saved (hours) from the Monte Carlo simulations.

Fig. 10.

Distribution of total cost saved (United States dollar).

Fig. 11.

Distribution of delta adverse events (H-ESD–C-ESD). ESD, endoscopic submucosal dissection; H-ESD, hybrid ESD; C-ESD, conventional ESD.

DISCUSSION

In the escalating fight against GI cancers, which claim over 3 million lives annually and are projected to exceed 3.5 million deaths by 2025,1 H-ESD offers a pivotal advancement by reducing procedure times without compromising the curative potential of modern endoscopy.15 Our Bayesian meta-analysis of three RCTs showed that H-ESD yields en bloc and R0 resection rates equivalent to C-ESD, enabling reliable margin-negative resections for early gastric and colorectal neoplasms.12,13 This parity, supported by low heterogeneity, ensures oncologic outcomes comparable to those of C-ESD, with symmetric funnel plots ruling out publication bias.14 Critically, H-ESD delivers a statistically credible reduction in procedure, a boon that could transform clinical workflows by minimizing anesthesia exposure, alleviating patient discomfort, and optimizing resource allocation in overburdened endoscopy suites.7 This time-saving advantage is particularly resonant in high-volume centers, where procedural delays exacerbate wait times for early cancer interventions, potentially worsening outcomes in a disease where timeliness is paramount.16 Moreover, with no credible increase in adverse events or perforation risks, H-ESD maintains the safety profile of C-ESD, addressing the longstanding concerns regarding perforation and bleeding in thin-walled colonic lesions.17,18 Practitioners can potentially embrace this technique for larger lesions or resource-limited contexts, aligning with guidelines like European Society of Gastrointestinal Endoscopy (ESGE) to enhance patient care. Nevertheless, further studies with subgroup analyses are needed to address the issue of differences related to the anatomy of gastric and colorectal lesions, which can lead to different perforation risks.6

Clinically, these results advocate H-ESD for larger lesions or resource-scarce settings, potentially reducing costs by 20%–30% per case through shorter room occupancy and fewer consumables.19 For patients, reduced times indicate less discomfort and a faster return to daily life, improving adherence to screening programs amid rising colorectal cancer rates. Endoscopists also benefit from H-ESD’s hybrid design, which blends dissection with snare assistance to ease the steep learning curve of C-ESD and foster adoption in Western centers, where expertise lags that in Asia.20,21 Bayesian predictive intervals can further guide real-world implementation and forecast time savings in future trials.

The sensitivity analyses underscored the overall robustness of our findings, particularly for efficacy outcomes such as en bloc and R0 resection rates, where exclusions did not meaningfully alter the conclusion of equivalence between H-ESD and C-ESD. However, for procedure time, the influence of Esaki et al.12—a gastric-focused study—was evident, since its exclusion eliminated heterogeneity and moderated the time savings to –13.20 minutes, suggesting that anatomical differences (gastric vs. colorectal) may contribute to variability in efficiency gains. Therefore, while pooled data suggest overall equivalence, endoscopists must cautiously interpret these findings, recognizing that the safety and efficiency profiles of H-ESD in thick-walled gastric tissue (Esaki et al.12) may differ mechanically and clinically from those in thin-walled colonic tissue (Yang et al.14, Bae et al.13). This aligns with prior observations of higher technical challenges in colorectal ESD, potentially amplifying the hybrid benefits in such settings. Safety outcomes remained consistent despite wide intervals, reflecting low event rates and limited power with only three RCTs. These results reinforce H-ESD as a reliable alternative but highlight the need for larger, anatomically stratified trials to refine estimates and address residual heterogeneity, ultimately guiding its broader clinical adoption.

This Bayesian meta-analysis stands out for its rigorous synthesis of RCTs, providing robust evidence for ESD across key outcomes, such as en bloc resection, R0 rates, procedure time, adverse events, and perforations. The integration of both frequentist and Bayesian methods enhanced the statistical power and uncertainty quantification, particularly with small sample sizes. The use of advanced diagnostics ensured reliability, whereas the professional color palettes and high-resolution figures met the highest professional standards. Sensitivity analyses further validated the robustness, and the focus on the efficiency of H-ESD offered actionable clinical insights.

The Monte Carlo simulations provided actionable insights by quantifying uncertainty in scaled scenarios, revealing high probabilities of time (89% >20 hours) and cost savings (27,094 USD median) for 100 procedures, while confirming near-equivalence of safety (43% chance of fewer adverse events). The normal distributions for time and cost reflect robust MD estimates, whereas the skewed adverse delta values underscore low event uncertainty. These support the adoption of H-ESD in resource-limited settings, aligning with cost analyses19 and calls for efficiency,6 though assumptions (e.g., 15 USD/min cost) warrant sensitivity testing in future work.

Despite its strengths, this study was constrained by the inclusion of only three RCTs, which limited its statistical power and broad generalizability. Although observational studies were considered, we restricted our analysis to RCTs to prevent the introduction of unmeasured confounding variables inherent to procedural outcomes. Our use of Bayesian modeling specifically mitigated the effects of a small sample size by leveraging probabilistic predictive intervals, although future large-scale trials are essential. The moderate heterogeneity in procedure time indicated variability across studies, potentially due to unmeasured factors such as operator experience. The differences between gastric versus colorectal neoplasms across studies warrant cautious interpretation.

In conclusion, this Bayesian meta-analysis of three RCTs demonstrated that H-ESD achieved en bloc and R0 resection rates comparable to those of C-ESD, with negligible heterogeneity. Notably, H-ESD significantly shortened the procedure time by an MD of –18.05 min, although with moderate heterogeneity. The safety profiles of the two approaches were equivalent, with no differences in overall adverse events or perforation rates, and both exhibiting zero heterogeneity. Bayesian modeling, which incorporated weakly informative priors and robust sensitivity analyses, confirmed these findings with posterior densities centered near zero for efficacy and safety outcomes, underscoring low between-study variability, although the small number of studies limited model reliability and generalizability. Overall, H-ESD has emerged as a viable alternative to C-ESD, preserving oncologic efficacy while enhancing procedural efficiency.

Larger multicenter RCTs are urgently needed to confirm these benefits across diverse populations, particularly in non-Asian patients where ESD adoption lags, and to track long-term outcomes such as recurrence-free survival and cost savings.5,22 Innovations such as artificial intelligence-driven dissection or traction tools could tackle procedure time variability (I²=66.5%), possibly halving durations further.23 The Bayesian approach presented here paves the way for adaptive trials that synthesize evidence in real time, even with limited data.24 These steps could make ESD a global standard, bridging equity gaps in GI cancer care and prompting updated ESGE guidelines to prioritize hybrid techniques, ushering in a new era of faster and safer endoscopic treatment worldwide.

Supplementary Material

Supplementary Material 1. Search strategy.

Supplementary Material 2. PICOS framework.

Supplementary Table 1. Baseline characteristics of included studies.

Supplementary Table 2. Leave-one-out sensitivity analysis results.

Supplementary Fig. 1. Risk of bias assessment.

Supplementary Fig. 2. For outcome-en bloc resection.

Supplementary Fig. 3. For outcome R0.

Supplementary Fig. 4. For outcome–procedure time.

Supplementary Fig. 5. For overall adverse events.

Supplementary Fig. 6. For perforation.

Supplementary Fig. 7. Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flowchart of study selection.

Supplementary materials related to this article can be found online at https://doi.org/10.5946/ce.2025.439.