This systematic review was designed and implemented strictly in accordance with the PRISMA 2020 guidelines. This review was not prospectively registered in PROSPERO; however, the protocol was finalized prior to data extraction and has been consistently followed throughout the review process. The research team searched multiple Chinese and English databases through a computer system to collect clinical research literature on the application of simulation training in obstetrics and gynecology surgical education from the establishment of the databases to December 2024. The retrieved databases include MEDLINE (PubMed), Embase, Cochrane Central Register of Controlled Trials (CENTRAL), and Web of Science, Scopus, CINAHL, and ERIC. In addition, the research team also searched the ClinicalTrials.gov clinical trial registration platform to obtain information on unpublished or ongoing studies. These databases were selected to ensure comprehensive coverage of medical education literature: MEDLINE and Embase for biomedical research, Cochrane CENTRAL for clinical trials, Web of Science and Scopus for multidisciplinary coverage, CINAHL for nursing and allied health perspectives, and ERIC for educational research. The search was conducted from database inception to December 2024 to capture all available evidence without temporal restrictions. The English search terms include: “Simulation-based training,” “simulation training,” “virtual reality,” “simulator,” “box trainer,” “surgical” education, surgical training, gynecology, obstetrics, laparoscopy, hysterectomy, vaginal surgery, robotic surgery, surgical skills, clinical competence, OSATS, GOALS, randomized controlled trials, etc. The retrieval strategy adopts a method that combines subject terms (MeSH terms or Emtree terms) with free terms, and makes corresponding adjustments according to the characteristics of each database.

The search strategy combined Medical Subject Headings (MeSH) terms and free-text keywords across four conceptual domains: (1) surgical education and clinical competence; (2) simulation training, virtual reality, and simulator technologies; (3) obstetrics, gynecology, and related surgical procedures; and (4) randomized controlled trial methodology. Boolean operators (AND/OR) were used to combine search terms within and across domains. The complete search strategies for all databases are provided in Supplementary Table 1.

To ensure the comprehensiveness of the search, the research team also adopted a variety of supplementary search strategies. Firstly, the reference list of the included literature was manually retrieved to identify any relevant studies that might have been overlooked. Secondly, references to previously published systematic reviews and meta-analyses were retrieved, including relevant reviews in the Cochrane systematic Review database. Secondly, gray literature sources were retrieved, including dissertation databases, conference proceedings, and professional society websites [such as the American College of Obstetricians and Gynecologists (ACOG), the American College of Gynecological Endoscopists (AAGL), and the Society of Gynecological Surgeons (SGS)]. Finally, the research team also reached out to relevant experts in the field of simulation medical education to inquire whether there were any unpublished studies or ongoing trials.

The literature screening process was completed by two trained independent researchers and managed in a standardized manner using the Covidence literature management system. The screening is carried out in two stages: In the first stage, researchers conduct a preliminary screening by reading the titles and abstracts of the literature and based on the pre-established inclusion and exclusion criteria, excluding literature that obviously do not meet the inclusion criteria; In the second stage, full-text reading and evaluation of the literature that has passed the initial screening will be conducted to determine the final studies to be included. If the two researchers encounter differences during the screening process, they first reach an agreement through discussion. If the dispute still exists, it will be arbitrated by a third senior researcher. The entire screening process is detailedly recorded with the reasons for excluding the literature, and a PRISMA flowchart is drawn to demonstrate the complete process of literature screening. To ensure the quality and consistency of the screening, the research team conducted a pre-screening training before the formal screening. Two researchers independently screened 50 randomly selected documents, calculated the Kappa coefficient to assess the consistency of the screening, and only began the formal screening work after ensuring that the Kappa value reached above 0.8.

This systematic review adopted the PICOS (Population, Intervention, Comparator, Outcomes, Study design) framework to establish inclusion and exclusion criteria to ensure the systematicness and scientificity of literature screening.

Data extraction was carried out using a pre-designed standardized table and was independently completed by two researchers. The extracted content includes: research characteristics (first author, publication year, country, study design type), participant characteristics (sample size, years of service, specialized background, baseline skill level), intervention measures (simulation type, training duration, training frequency, training mode), control measures, surgical type, outcome indicators and measurement tools, follow-up time and adverse events. For continuous outcomes, extract the mean, standard deviation and sample size; for binary classification outcomes, extract the number of event occurrences and the total number of cases. In addition to the outcome indicator data, this study also systematically extracted implementation-related information, including training costs, required equipment, training duration, instructor requirements, trainee compliance, as well as the implementation obstacles and promoting factors reported by the researchers. For the research on the hybrid method providing qualitative data, two researchers independently extracted the topics regarding the feasibility, acceptability and sustainability of implementation. For studies with incomplete data reports, the research team will contact the original authors via email to obtain necessary information.

All statistical analyses were conducted using the metafor package of R software (version 4.3.0) in this study. For continuous outcome measures, the Standardized Mean Difference (SMD) and its 95% Confidence Interval (CI) were adopted as effect size measures to eliminate the influence of dimensional differences among different assessment tools. The interpretation standard of SMD is as follows: SMD < 0.2 is considered to have no effect, 0.2 ≤ SMD < 0.5 is a small effect, 0.5 ≤ SMD < 0.8 is a moderate effect, and SMD ≥ 0.8 is a large effect. For binary outcome measures, Odds Ratio (OR) or Risk Ratio (RR) and their 95% confidence intervals were used for combined analysis.

Considering the possible heterogeneity of the included studies in terms of participant characteristics, intervention measures and study design, this study conducted a meta-analysis using a random-effects model estimated by the Restricted Maximum Likelihood (REML) method. Heterogeneity assessment was conducted using Cochran’s Q test and I² statistics. Among them, I² < 25% indicates low heterogeneity, 25% ≤ I² < 75% indicates moderate heterogeneity, and I² ≥ 75% indicates high heterogeneity. When P < 0.10 or I² > 50%, significant heterogeneity among studies is considered to exist.

The planned subgroup analyses include: stratified analysis by simulation type (high-fidelity vs. low-fidelity), learner’s years of experience (junior vs. senior resident), surgical type (laparoscopic vs. vaginal vs. robotic surgery), training mode (standard training vs. fixed number of training), and surgical complexity (simple vs. complex surgery). Sensitivity analysis assesses the robustness of the results by eliminating included studies one by one and repeating the meta-analysis. When the number of included studies is ≥ 10, funnel plots, Egger’s regression test and Begg’s rank correlation test are used to evaluate publication bias, and a P-value < 0.05 is considered statistically significant publication bias. All statistical tests were conducted using two-sided tests, and the significance level was set at α = 0.05.

Through systematic retrieval of multiple databases, 4,287 relevant literatures were initially obtained. After screening the titles and abstracts, 1,563 duplicate literatures and 2,358 obviously irrelevant literatures were excluded. A total of 366 literatures were initially included for full-text evaluation. After further full-text reading, 336 literatures that did not meet the inclusion criteria were excluded. The main reasons for exclusion included: 142 studies that only reported the outcome measures within the simulator (T1 outcome) without evidence of clinical environment transformation, 89 non-controlled study designs, 61 studies with incomplete or unextractable data reports, and 44 studies whose research subjects did not meet the inclusion criteria. Ultimately, 30 studies were included for systematic review and meta-analysis. The literature screening process is shown in Figure 1.

All 30 included studies were randomized controlled trials, with publication periods ranging from 2012 to 2024. A total of 1,247 participants were included in these studies, among whom 68.3% (852) were resident physicians in obstetrics and gynecology, 23.7% (296) were medical students, and 8.0% (99) were Fellow physicians. The geographical distribution of the studies is extensive, including 12 studies in North America (40.0%), 10 studies in Europe (33.3%), 6 studies in Asia (20.0%), and 2 studies in Oceania (6.7%). The surgical types involved in the included studies included 24 laparoscopic surgeries (80.0%), 3 vaginal surgeries (10.0%), 5 robot-assisted surgeries (16.7%), and 1 hysteroscopic surgery (3.3%). Some of these studies involved multiple surgical types. In terms of simulation training types, high-fidelity virtual reality simulators were applied in 17 studies (56.7%), low-fidelity box trainers were used in 10 studies (33.3%), and the combination of the two was used in 3 studies (10.0%). The training models included 18 studies (60.0%) of proficiency training and 12 studies (40.0%) of fixed-frequency training. In terms of outcome indicator reporting, among 1,247 participants, 27 studies (90.0%) reported the standardized technical skills scores of 1,110 participants, using assessment tools including OSATS, GOALS, and other standardized scoring systems; Twenty-two studies (73.3%) reported the operation time data of 892 participants; Fifteen studies (50.0%) evaluated the learner confidence of 645 participants; Only three studies (10.0%) reported patient-related outcomes such as intraoperative blood loss or complications for 156 participants. The basic characteristics of the included studies are detailed in Table 1. The consistency Kappa value of literature screening by two independent researchers was 0.86, indicating that the screening process had good reliability.

The studies included in the current review, as evident from Table 2, were extremely heterogeneous with regard to types of surgery, types of simulation, and groups of participants. Laparoscopic surgery training was most prevalent in included studies, noted in 24 (80.0%) studies, followed by vaginal surgery in 3 (10.0%), robot-assisted surgery in 5 (16.7%), and hysteroscopy in 1 (3.3%) study. In applying simulation technology, 17 (56.7%) studies utilized high-fidelity virtual reality simulators, 10 (33.3%) studies used low-fidelity box trainers, and mixed methods in 3 (10.0%) studies. Simulation modalities were categorized as follows: (1) High-fidelity simulators included virtual reality (VR) laparoscopic trainers (e.g., LapSim, LapMentor) and robotic surgery simulators (e.g., dV-Trainer) that provide realistic visual and haptic feedback with computer-generated metrics; (2) Low-fidelity simulators included box trainers, pelvic models, and tablet-based devices that offer basic psychomotor skill practice without advanced computer-generated feedback; (3) Combined modalities employed both high- and low-fidelity components within the training curriculum. Geographical distribution depicted a dominance of North American (n = 12, 40.0%) and European (n = 10, 33.3%) hubs, complemented by Asian (n = 6, 20.0%) and Oceanian (n = 2, 6.7%) hubs. A total of 18 studies (60.0%) used proficiency-based training and fixed-repetition training was used in the other 12 studies (40.0%). The vast majority of studies (n = 26, 86.7%) provided technical skill scores from standardized instruments like OSATS or GOALS, while patient-related outcomes were seldom reported, and only 3 studies (10.0%) provided complications or blood loss.

Methodological quality was highly variable among the randomized controlled trials included. As seen in Figure 2A, most studies were at low risk of bias for random sequence generation (n = 22, 73.3%) and allocation concealment (n = 19, 63.3%), suggesting proper randomization procedures. Participant and personnel blinding was a significant issue inherent to surgical simulation research, with 8 studies (26.7%) reporting low risk for this item. Outcome measure blinding was more common, with 18 studies (60.0%) having independent, blinded raters for the scoring of technical skill. Incomplete outcome data and selective reporting issues were low, impacting less than 20% of included studies. The Medical Education Research Study Quality Instrument (MERSQI) analysis provided a mean score of 13.2 (SD = 2.1, range: 9–17) out of a possible 18 points, as indicated in Figure 2B. Two-thirds of studies (n = 20, 66.7%) scored 12–15 on the MERSQI, reflecting moderate to high methodological rigor. More highly rated studies on the MERSQI more often used validated measures, reported interrater reliability, and used appropriate statistical tests (Figure 3). Overall, although the evidence base presented has good methodological soundness, ongoing issues with blinding procedures reflect the practical limitations of simulation-based education interventions.

It searched for and identified 27 studies that contrasted simulation-based training with conventional instruction on the basis of surgical skill scores. Meta-analysis also identified that the students who underwent simulation training performed significantly better on standardized skills testing such as OSATS and GOALS compared to the control group with traditional instruction (overall SMD = 0.82, 95% CI: 0.64–1.00, P < 0.001), and a large effect size as can be seen from Figure 4. Subgroup analysis also identified that the high-fidelity virtual reality simulation training group (17 studies, n = 718) had an effect size of SMD = 0.89 (95% CI: 0.67–1.11, P < 0.001), and there was moderate between-study heterogeneity (I² = 56%, P = 0.003). The low-fidelity box trainer subgroup (10 studies, n = 392) also showed significant effects (SMD = 0.71, 95% CI: 0.48–0.94, P < 0.001), with heterogeneity of I² = 48% (P = 0.04). Statistical contrast indicated that there was no significant difference in the effect sizes between high-fidelity and low-fidelity training (subgroup difference P = 0.28), which indicates both simulation modes to be effective in improving technical skills and with similar outcomes. The heterogeneity observed is likely due to differences in training duration, assessment instruments, student levels of experience, and complexity of procedure between studies. Sensitivity analysis without high risk of bias studies revealed similar results (SMD = 0.78, 95% CI: 0.59–0.97, P < 0.001), with reassurance of the stability of these findings. The combined findings are presented in Table 3 with comparable advantages of simulation-based training irrespective of fidelity level or study group.

A total of 22 studies reported operative time data, with meta-analysis demonstrating that the simulation training group achieved significantly shorter operative times compared to the traditional teaching group (SMD = −0.62, 95% CI: −0.81 to −0.43, P < 0.001), as illustrated in Figure 5. Subgroup analysis revealed that high-fidelity simulation training produced more pronounced improvements in operative time (15 studies, n = 634, SMD = −0.68, 95% CI: −0.89 to −0.47, P < 0.001, I² = 62%), whereas low-fidelity training demonstrated a weaker effect with borderline statistical significance (7 studies, n = 258, SMD = −0.42, 95% CI: −0.71 to −0.13, P = 0.005, I² = 44%). Statistical comparison indicated a significant difference between high-fidelity and low-fidelity training in operative time reduction (subgroup difference P = 0.04), suggesting that higher technological sophistication may confer additional benefits for procedural efficiency.

Learner confidence was evaluated in 15 studies involving 645 participants. Meta-analysis indicated that simulation training was effective in increasing learner self-confidence scores (SMD = 0.71, 95% CI: 0.49–0.93, P < 0.001), which was a moderate effect size as indicated in Figure 6. There was moderate between-study heterogeneity (I² = 62%, P < 0.001), which could possibly be explained by differences in confidence measurement tools and measurement timing between studies. Subgroup analysis indicated that both low-fidelity and high-fidelity simulation modalities were similarly effective at increasing learner confidence with similar effect sizes (SMD = 0.73 vs. 0.68, P = 0.67).

Patient outcomes were only reported in 3 studies (10.0%), including intraoperative blood loss, surgical complications, and hospital length of stay. Meta-analysis was not conducted due to the small number of studies, methodological variations, and enormous heterogeneity. Descriptive analysis showed no differences in patient outcomes between intervention and control groups in the majority of studies, although this result could be accounted for by a lack of statistical power and short follow-up duration. The only study (10) exclusively assessing patient safety for laparoscopic hysterectomy found no statistically significant differences in intraoperative blood loss (144 ± 89 mL vs. 165 ± 112 mL, P = 0.205) or immediate complications (0 vs. 0%) between groups, as presented in Table 4.

Dual Surgical Types Differential simulation-based training effects between and within different learner groups and training modes were investigated in this study using subgroup analyses. Stratification based on learner experience level showed that PGY 1–2 junior residents and medical students gained most from simulation training with a weighted effect size of 1.35 (95% CI: 1.08–1.62, P < 0.001) which represents the massive contribution of simulation toward achieving basic skill. Senior residents (PGY 3–4) and fellows had smaller but statistically significant effect sizes (SMD = 0.87, 95% CI: 0.62–1.12, P < 0.01), which indicates simulation training remains effective for learning advanced skills, though between-subgroup differences were not statistically significant (P = 0.18), as shown in Figure 7. Comparison between training modalities showed no significant difference in overall effect sizes between fixed-repetition training and proficiency-based training (SMD = 1.18 vs. 1.05, P = 0.24). Nonetheless, the proficiency-based training group demonstrated substantially reduced between-study heterogeneity (I² = 32.4 vs. 58.7%), implying that this method might produce more reliable and reproducible training results. The proficiency-based protocols carried an average of 12.3 ± 4.2 h of training time, which was about 3 h longer than the 9.8 ± 3.5 h needed for fixed-repetition training (P = 0.03), though greater time expenditure of this nature was associated with more uniform skill acquisition. During all the surgical procedures, heterogeneity of both the quality of evidence and efficacy in training was high, as reflected in Table 5. Laparoscopic surgery was the most rigorously researched type of surgery, with laparoscopic salpingectomy and tubal ligation having the largest number of studies (12 studies), with uniform technical skills improvement scores (SMD = 1.28, 95% CI: 1.05–1.51, P < 0.001) and decreases in operating time (SMD = −0.72, 95% CI: −0.94 to −0.50, P < 0.001). Laparoscopic hysterectomy training had moderate-quality evidence (6 studies) and simulation enhanced technical expertise and learner confidence. Evidence for vaginal surgery continued to be limited (3 studies), although available data indicated that simulation of individual procedures enhanced procedural knowledge and intraoperative performance. Robot-assisted surgery training showed positive results (5 studies) with outstanding training efficiency as well as remarkable skill transferability. Evidence for hysteroscopic surgery and other obstetric-gynecologic procedures continued to be low, and more high-quality research is needed.

The accuracy of primary results was confirmed by conducting extensive sensitivity analyses. Exclusion of eight high-risk studies produced a pooled effect size of SMD = 0.78 (95% CI: 0.59–0.97, P < 0.001) for surgical skill scores against SMD = 0.82 in the full analysis, with direction of effects and statistical significance unaffected. This minimal reduction of point estimate indicates that some of the high-risk studies overestimated slightly the efficacy of training. When restricted to 20 good-quality studies with an MERSQI score of ≥ 13, the combined effect size was SMD = 0.79 (95% CI: 0.61–0.97), very close to that of the complete analysis and consistent with the results. Exclusion of two extreme outlier trials (SMD > 2.0 or < −2.0) eliminated heterogeneity from I² = 53–42% without changing stable effect estimates (SMD = 0.81, 95% CI: 0.65–0.97), i.e., individual extreme trials did not exert noticeable influence on pooled results (as presented in Table 6).

Regarding publication bias evaluation, the funnel plot of surgical skill scores (27 trials) showed nearly symmetric distribution without clear small-study effects (as presented in Figure 8A). Neither Egger’s regression test (P = 0.24) nor Begg’s rank correlation test (P = 0.31) revealed statistically significant publication bias. Trim-and-fill adjustment added three potential missing studies with negative results, yielding an adjusted pooled effect size of SMD = 0.77 (95% CI: 0.58–0.96), indicating little difference from the initial analysis. The operating time funnel plot (22 studies) was moderately asymmetrical (as depicted in Figure 8B) and hinted at selective reporting, but tests were again not statistically significant (Egger’s P = 0.18, Begg’s P = 0.22). The learner confidence outcome was extremely symmetrical in its funnel plot (as depicted in Figure 8C) with no hint of publication bias (as depicted in Table 7).

The notably symmetric funnel plots and non-significant publication bias tests (all P > 0.18) may appear unusually ideal. This likely reflects our stringent inclusion criteria requiring validated clinical outcomes (excluding 142 simulator-only studies that typically drive asymmetry), the mature methodological standards in simulation research with standardized assessment tools (OSATS/GOALS), and comprehensive search strategies including gray literature. While absence of statistical bias does not guarantee complete absence of selective reporting, sensitivity analyses restricted to high-quality studies (MERSQI ≥ 13) yielded consistent results (SMD = 0.78–0.79), supporting robustness of findings.

GRADE evidence quality assessment revealed substantial variation in certainty of evidence across outcome measures (as shown in Table 8). Evidence supporting the effectiveness of simulation-based training in improving surgical skill scores was rated as moderate quality (SMD = 0.85, 95% CI: 0.68–1.02), downgraded one level due to moderate between-study heterogeneity (I² = 58%) potentially attributable to variations in simulation modalities, training dosage, and assessment instruments. Similarly, moderate-quality evidence demonstrated significant reductions in operative time (SMD = −0.62, 95% CI: −0.81 to −0.43), though imprecision in effect estimates and subgroup differences warranted one-level downgrading. Evidence supporting improvements in learner confidence was rated as low quality (SMD = 0.71, 95% CI: 0.49–0.93), downgraded two levels due to high risk of bias in eight studies lacking assessor blinding and substantial heterogeneity (I² = 62%). Most critically, evidence regarding patient-related outcomes remained extremely limited, with only three studies (n = 156) reporting complications, yielding very low-quality evidence (OR = 0.78, 95% CI: 0.42–1.45). This was downgraded three levels owing to serious imprecision from small sample sizes, indirectness as patient outcomes were not primary endpoints, and high risk of bias. The paucity of high-quality evidence for patient safety outcomes represents a crucial knowledge gap requiring adequately powered prospective investigations with standardized long-term follow-up protocols.

It should be noted that implementation-related data were available from only 8 studies (26.7%), limiting the generalizability of these findings. The following synthesis reflects the limited empirical evidence from included studies rather than comprehensive implementation research. Systematic extraction of implementation-related data from studies included identified that only 8 studies (26.7%) clearly stated the facilitators or barriers to simulation training implementation, whereas the other 22 studies (73.3%) presented merely efficacy results. Of studies that reported implementation data, the highest mentioned barriers were the cost of equipment purchase being high (5 studies, 16.7%), with virtual reality simulators priced between $40,000 and $200,000 and with maintenance costs as a recurring expense. Institutional and logistical difficulties were the most prevalent, with 4 studies (13.3%) citing lack of covered training time in clinical schedules as a primary difficulty, and 3 studies (10.0%) citing insufficient faculty experience in simulation-based education. Two studies (6.7%) identified lack of administrative support and inadequate physical space as secondary difficulties. Facilitators to effective implementation were less systematically reported but included several recurring themes. Three studies stressed the value of well-planned curriculum integration and competency milestones as explicitly defined, as opposed to intermittent training periods. Two studies pointed to the utilization of expert simulation coordinators to oversee logistics and trainee compliance. Low-fidelity box trainers and take-home devices were suggested as cost-effective strategies to enhance accessibility in 3 studies for low-resource settings. Nevertheless, the limited implementation research represents a significant knowledge gap, since knowledge about facilitators and barriers is critical to the translation of efficacy findings into standard educational practice. Subsequent investigation must systematically examine implementation outcomes using valid frameworks like RE-AIM (Reach, Effectiveness, Adoption, Implementation, Maintenance) to inform sustainable incorporation of simulation training into residency curricula.