The study was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) reporting guidelines. The protocol was registered in the Prospective Register of Systematic Reviews (ID: CRD 420251180897, https://www.crd.york.ac.uk/PROSPERO/view/CRD420251180897). We systematically searched the PubMed, Web of Science, Embase and Cochrane Library databases from inception to November 10, 2025. To minimize the risk of bias, we supplemented our search by examining the ClinicalTrials.gov registry and manually searched conference abstracts from the American Society of Clinical Oncology (ASCO), the European Society for Medical Oncology (ESMO), and the Society of Gynecologic Oncology (SGO) to identify unpublished gray literature. The comprehensive search keywords included (Cervical Cancer) AND (Drug Therapy) AND (Clinical Trial). The specific search strategy is detailed in the Supplementary Material 1.

This study included patients with histologically confirmed R/MCC, focusing on randomized controlled trials (RCTs) that compared drug treatment regimens. Trials had to compare any drug regimen (including chemotherapy, targeted therapy, or immunotherapy) against a concurrent control arm (such as standard treatment or placebo) and must report the hazard ratios (HR) for overall survival (OS) and/or progression-free survival (PFS), or provide sufficient data to calculate these HR. Studies with non-randomized designs, those involving non-metastatic or non-recurrent cervical cancer patients, and those where the intervention was non-drug therapy or lacked a concurrent control group were excluded. Furthermore, Phase I clinical trials, non-clinical studies, duplicate publications, and studies for which the full text could not be obtained, which did not report relevant survival outcome HR and from which relevant data could not be extracted, were also excluded.

Literature screening was performed independently by two researchers, followed by cross-validation. Any disagreement was resolved by consulting a third party. Initial screening was performed by reading titles and abstracts, and further screening involved reading the full text in conjunction with the inclusion and exclusion criteria. Data were extracted by two researchers. Extracted items mainly included: first author and publication year, characteristics of the included population, sample size, interventions in the treatment and control groups, median OS and PFS, and the HR and confidence intervals (CIs) for each intervention. Additionally, if a study did not report the HR but provided the corresponding Kaplan-Meier survival curves, we used the WebPlotDigitizer software to extract individual patient data, and utilized the IPDfromKM R package developed by Liu et al. to calculate the HR and its standard deviation (6).

Two reviewers independently assessed the risk of bias using the Cochrane Risk of Bias Tool 2 (RoB 2.0). Each study was rated as low risk, high risk, or some concerns, with disagreements resolved through discussion. RoB 2.0 covers five domains: randomization process, deviations from intended interventions, missing outcome data, measurement of the outcome, and selection of the reported result (7). The robvis R package was used to visualize the assessment results (8).

Given that the NMA model requires good connectivity and needs to avoid over-complication, we merged interventions with similar therapeutic effects: (1) Regimens combining placebo with other drugs were considered the same node as the single drug regimen, as placebo does not affect efficacy; (2) Cisplatin or carboplatin plus paclitaxel regimens were simplified to cisplatin plus paclitaxel. For instance, cadonilimab plus cisplatin or carboplatin + paclitaxel was simplified to cadonilimab plus cisplatin + paclitaxel. This decision was based on two considerations: First, in trials comparing platinum-doublet chemotherapy with other therapies, the specific platinum agent was chosen based on clinical practice, and the randomization process itself ensured baseline comparability between groups, thus balancing the potential impact of the choice of platinum agent. Second, a head-to-head comparison study (JGOG-0505) in R/MCC had already demonstrated that carboplatin plus paclitaxel was non-inferior to cisplatin plus paclitaxel regarding OS and PFS (9).

In addition, the study selected the HR as the effect size and conducted NMA using both Bayesian and Frequentist approaches. Specifically, Bayesian network meta-analyses were performed using the gemtc package in R, applying a random-effects model under the consistency assumption, with four Markov chain Monte Carlo (MCMC) chains run. For the Bayesian inference, non-informative (vague) priors were specified to ensure that results were data-driven: relative treatment effects ( dik) were assigned a normal distribution N(0,1002), while the study-to-study standard deviation τ) was assigned a uniform distribution U(0,om.scale). Four Markov chain Monte Carlo (MCMC) chains were run, each undergoing 5,000 burn-in iterations followed by 50000 formal iterations with a thinning interval of 5, yielding 40000 effective posterior samples.

Network inconsistency was assessed via both local and global approaches. Local inconsistency was evaluated using the node-splitting method, while global inconsistency was assessed by fitting an unrelated mean effects (UME) model (the design-by-treatment interaction test). Model fit and consistency were appraised by comparing the Deviance Information Criterion (DIC) values between the consistency and UME models.

Heterogeneity was quantified not only by the I2 and Cochran's Q statistics but also by reporting the posterior estimates of the between-study variance ( τ2) derived from the random-effects model. To ensure the robustness of the results, a prior sensitivity analysis was conducted by comparing the primary model with an alternative specification using a half-normal prior ( τ∼HN(0,1)). Finally, the SUCRA method was used for treatment ranking, and a Frequentist analysis using the netmeta package was performed for cross-validation.

A total of 5376 records were retrieved through the established search strategy. After removing duplicates, 3406 records remained for screening by title and abstract. Following a careful full-text review of the remaining 102 articles, and based on the inclusion and exclusion criteria, 15 RCTs were finally included. The detailed screening process is shown in Figure 1.

The 15 included studies comprised 4 phase II clinical trials and 11 phase III clinical trials, involving a total of 4,588 patients with R/MCC. Detailed characteristics of the included studies are presented in Table 1. 11 studies were rated as having a low risk of bias, and 4 studies were rated as having some concerns. Across the included clinical trials, the patient populations were predominantly composed of squamous cell carcinoma cases (accounting for 70% to 85%), and all patients had stage IVB R/MCC. Moreover, the reported median age ranges were highly consistent, falling between 46 and 55 years. This high degree of homogeneity in baseline characteristics provides a solid foundation for the transitivity assumption required in the subsequent network meta-analysis. The detailed risk of bias and clinical characteristics for the included studies were provided in the Supplementary Material 1.

All studies (9–23) were utilized for the NMA of OS, involving 21 interventions. The comparison relationships among these interventions are shown in Figure 2. Cisplatin plus paclitaxel served as the central node, directly comparing with various other treatment regimens. Simultaneously, the network map exhibited closed loops, allowing for the assessment of inconsistency.

As shown in Figure 3A, multiple treatment regimens demonstrated a significant OS benefit compared to cisplatin plus paclitaxel under the Frequentist method. These primarily included immunotherapy regimens such as pembrolizumab plus chemotherapy (with or without bevacizumab), cadonilimab plus chemotherapy (with or without bevacizumab), and atezolizumab plus chemotherapy and bevacizumab. Among these, the regimen of pembrolizumab plus chemotherapy and bevacizumab showed the superior effect (HR: 0.45, 95%CI: 0.30–0.67), suggesting it may be the optimal treatment strategy for improving survival in this patient population. However, it is noteworthy that these differences were not significant under the Bayesian method (Figure 3C). This may be attributed to the Bayesian analysis providing a more complete quantification of uncertainty compared to the Frequentist method, particularly in handling random effects and model parameter uncertainty (24).

Concurrently, in the pairwise comparisons under the Frequentist framework (Table 2), single-agent cisplatin showed a higher risk compared to cisplatin plus paclitaxel (HR: 1.35, 95%CI: 1.08–1.69). Other doublet chemotherapy regimens (e.g., topotecan plus paclitaxel) did not show a benefit or a significant benefit compared to single-agent cisplatin. Notably, the carboplatin plus paclitaxel regimen showed a small and non-significant difference in OS benefit compared to the cisplatin plus paclitaxel regimen (HR: 0.99, 95%CI: 0.75–1.31). Detailed pairwise comparison results are available in the Supplementary Material. This suggests that the cisplatin plus paclitaxel regimen is likely the best chemotherapy backbone for R/MCC, with the carboplatin plus paclitaxel regimen showing minimal difference.

Furthermore, the ranking of treatment regimens obtained via SUCRA (Figure 3D) also indicated that adding pembrolizumab and bevacizumab to the cisplatin plus paclitaxel chemotherapy regimen (87%) was most likely to increase the OS benefit. This was followed by the cadonilimab plus cisplatin and paclitaxel regimen and the atezolizumab plus chemotherapy and bevacizumab regimen (84% vs 83%). Other competitive regimens included cadonilimab plus cisplatin and paclitaxel and bevacizumab (73%), and pembrolizumab plus cisplatin and paclitaxel (71%). It is worth mentioning that the treatment regimen rankings were nearly consistent between the Frequentist (Figure 3B) and Bayesian methods(Figure 3D), suggesting strong robustness of the results.

A separate NMA was conducted for PFS, including 14 studies (9–15, 17–23). Isolated treatment regimen nodes (including regimens such as atezolizumab) were removed due to lack of key data, ultimately covering 16 treatment regimens.

The results, as shown in Figure 4A, indicated that the cadonilimab plus chemotherapy regimen demonstrated a significant benefit in PFS compared to cisplatin plus paclitaxel (HR: 0.46, 95%CI: 0.32–0.66), ranking first by SUCRA (Figure 4D)with a score of 90%. This was followed by the regimen of carboplatin plus paclitaxel combined with cediranib, which showed a benefit compared to cisplatin plus paclitaxel but was not statistically significant (HR: 0.60, 95%CI: 0.32–1.14), with a SUCRA score of 78%. Ranking third was the pembrolizumab plus cisplatin and paclitaxel regimen, which demonstrated a significant benefit compared to cisplatin plus paclitaxel (HR: 0.69, 95%CI: 0.50–0.95), with a SUCRA score of 74%. The pairwise comparisons between all interventions can be found in the Supplementary Material. Similarly, under the Bayesian framework, none of the treatment regimens showed a statistically significant advantage compared with the cisplatin plus paclitaxel regimen (Figure 4C), whereas under the Frequentist method, some regimens demonstrated statistically significant benefits; nevertheless, the treatment rankings were nearly identical between the two approaches (Figures 4B, D).

Safety Profiles Overall, ICI-based combinations exhibited higher toxicity compared to chemotherapy alone. As summarized in Table 3, the incidence of Grade≥3 treatment-related AEs was highest in the quadruple/triple regimens (ICI + chemotherapy + bevacizumab), ranging from 74% to 85%. Notably, the discontinuation rate for any treatment component in these groups reached up to 40.8% (KEYNOTE-826). In contrast, ICI plus chemotherapy without bevacizumab showed a more manageable safety profile (Grade≥3 AEs: 60.4%). Immune-related AEs (irAEs) were predominantly low-grade, with hypothyroidism being the most frequent (11.7%–33.0%). Bevacizumab-containing arms were characterized by distinct toxicities, including hypertension and proteinuria. Traditional chemotherapy backbones primarily resulted in hematological toxicities, consistent with historical data.

The heterogeneity of the NMA was low across both frameworks. Under the Bayesian random-effects model, the posterior estimate of the between-study standard deviation was  τ=0.256 (95% CrI: 0.012–0.661), corresponding to a variance ( τ2) of approximately 0.065. The I2 statistic ranged between 0% and 5% for both Frequentist and Bayesian methods, suggesting high similarity and minimal heterogeneity among the included RCTs. Convergence of the Bayesian model was confirmed using Gelman–Rubin diagnostics and trace plots, with all potential scale reduction factors (PSRF) close to 1.0, indicating adequate mixing (see Supplementary Material 1). Consistency was rigorously assessed through both global and local approaches. The global design-by-treatment interaction test showed that the DIC for the consistency model (41.50) was nearly identical to that of the unrelated mean effects (UME) model (41.95). A ΔDIC < 5 indicates that the consistency model effectively captured the data variability without significant global inconsistency. Furthermore, local node-splitting results (refer to Supplementary Material 1) revealed no significant differences between direct and indirect comparisons ( P > 0.05), demonstrating strong coherence across the treatment network. The risk of publication bias was assessed using comparison-adjusted funnel plots (Figure 5A). The results indicated no significant publication bias, as evidenced by the symmetrical distribution in the funnel plots and an Egger’s test p-value > 0.05. Sensitivity analyses were conducted to evaluate the robustness of the findings. A prior sensitivity analysis was performed by comparing the primary model with an alternative specification using a half-normal prior ( τ∼HN(0,1)). The results were highly consistent, with a negligible change in DIC (41.68) and stable SUCRA rankings. Cross-validation between the Frequentist and Bayesian approaches showed that while some treatment regimens reached statistical significance under the Frequentist framework, their 95% credible intervals crossed the null under the Bayesian framework. This divergence may be attributed to the Bayesian model’s more conservative integration of the full posterior uncertainty of τ. To investigate potential outliers and their impact on the model, we performed an influence analysis using a deviance-leverage plot (see Figure 5B). The contributions to deviance and leverage for all 15 trials clustered near 1, suggesting that the results were robust and not disproportionately driven by any individual study. Meanwhile, the ranking scores and the identification of the most efficacious regimens remained nearly identical between the two methods, confirming the overall robustness of the study's conclusions. Finally, we conducted a subgroup network meta-analysis based on PD-L1 CPS ≥ 1, while removing one study with moderate risk (Mountzios, 2009) and an older study (GOG-110) for sensitivity analysis. The results confirmed that the relative efficacy ranking remained stable (see Supplementary Material 1).