This study was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement and the PRISMA Extension Statement for NMA (eMethods in the Supplementary material) (10, 11). The study protocol, including methods and analyses, was prespecified and registered on PROSPERO (CRD 42024585931). To identify all RCTs on CH treatment, we systematically searched PubMed, Embase, the Cochrane Central Register of Controlled Trials, Web of Science, ClinicalTrials.gov, CNKI, Wanfang, Vip, China Biomedical Literature, and the ChiCTR from inception to September 24, 2024. We also searched reference lists of key reviews and meta-analyses to supplement identified citations. Detailed search strategies for each database are provided in the eMethods in the Supplementary material.

For the study design, we included RCTs comparing at least two different surgical methods (Figure 1). The inclusion criteria required patients to have a diagnosis of CH, including hydrocephalus secondary to trauma, hemorrhage, or inflammation, and idiopathic normal pressure hydrocephalus (iNPH). We excluded studies that explicitly included cases of non-CH (obstructive hydrocephalus) or mixed-type hydrocephalus and those that did not clearly specify the type of hydrocephalus.

Two independent researchers selected studies, extracted relevant data, and assessed the risk of bias. The extracted information included the study characteristics (journal name, year of publication, authors, and country of origin), study design, participant characteristics (age, sex, and other demographic details), etiology of hydrocephalus, surgical interventions, study outcomes, follow-up duration and results. For missing data in the original publication, we contacted authors and trial investigators to offer incomplete information. Discrepancies were double-checked and resolved through discussions with other review team members.

The primary outcomes were efficacy (favorable outcome) and safety (complications) (more details in eMethods in the Supplementary material). Overall, the favorable outcome was defined as improvement in clinical symptoms, requiring no further surgical intervention, with or without reduction of the ventricular system on imaging. Secondary outcomes included revision surgery, infection, seizures, operation time (minutes), and duration of hospitalization (days). Whenever possible, we prioritized the 6 month postoperative time for outcome assessments. Given the extremely poor outcomes of nonsurgical treatments (conservative or delayed surgical interventions) (8), we designated the most widely used VPS as the standard treatment (control group).

Bayesian NMAs were conducted using the GeMTC (version 1.0.2) package in R (version 4.3.2) with Markov Chain Monte Carlo (MCMC) method (12). Uninformative prior distributions were applied for the treatment effects, with a minimally informative prior used for the common standard deviation parameter. Four MCMC chains were set for the initial value, with each chain undergoing 50,000 iterations. To exclude initial value bias, we discarded the first 10,000 annealings and commenced sampling from iteration 10,001. Trace, density and Brooks-Gelman-Rubin diagnosis plots were utilized to visually examine convergence (13). Model fit was evaluated through the posterior total residual deviance and unconstrained data points. Both random-effects (RE) and fixed-effects (FE) models were used to pool network results, with model selection regarding deviance information criterion (DIC) (14). Details of the Bayesian model specifications were provided in eMethods.

Network graphs were scaled based on the number of studies and patients for each treatment node. Summary RRs were estimated for dichotomous outcomes and mean differences (MDs) for continuous outcomes, with 95% CrIs in the NMAs. League tables, two-dimensional graphs, and forest plots were used to visualize relative treatment effects in network estimations. Treatment rank probabilities were calculated, and SUCRA values were generated to display cumulative ranking probability plots for the interventions. Higher SUCRA values indicate better intervention effects.

Following the Cochrane Handbook for Systematic Reviews, statistical heterogeneity in NMA and each pairwise comparison was assessed using I² statistics and visual inspection of forest plots (15). We appraised the clinical and methodological characteristics of the included studies to determine the appropriateness of the transitivity assumption. Inconsistencies between direct and indirect evidence were assessed using global and local approaches. Global inconsistency was assessed by comparing residual deviance and DIC between the unrelated study effects and consistent models (14). Local inconsistency was assessed using the node-splitting approach where relevant head-to-head trials were available (16). Small-study effects and publication bias were assessed using funnel plots if at least 10 studies were available (17). Conventional pairwise meta-analyses were conducted for comparisons involving head-to-head studies. To compare the overall efficacy and safety of lumbar versus cranial surgical interventions, we consolidated LPS and LPS + LS into the unified lumbar approach, while merging VPS and ETV with or without CPC into the cranial approach. Statistical significance was set at p < 0.05, and all p-values two-tailed.

We assessed the risk of bias in individual studies using the Cochrane Risk of Bias tool for Randomized trials version 2 (RoB2) (18). The certainty of evidence was evaluated using the Confidence in Network Meta-Analysis (CINeMA) framework (19) and Grading of Recommendations, Assessment, Development, and Evaluation (GRADE) (20, 21).

To assess potential heterogeneity in treatment effects, we performed several sensitivity analyses stratified by: (1) age: pediatric patients (age <18 years) versus adults (age ≥18 years); (2) statistical model selection: fitting both RE and FE models; (3) etiology: reporting outcomes separately for post-hemorrhagic hydrocephalus, post-infectious hydrocephalus, and iNPH; (4) evidence quality: excluding studies with a high risk of bias; (5) follow-up duration: excluding studies with less than 3 months of follow-up.

We identified 4,358 citations through database searches and relevant reviews, with 1,445 duplicates removed. After screening 2,913 titles and abstracts and 504 full texts, a total of 34 RCTs involving five surgical interventions between 2013 and 2024 were included in this NMA (Figure 2) (22–55).

The 34 eligible studies included 2,528 participants (33–300 per trial). Study characteristics are in Table 1. Among the 34 trials, the mean age ranged from 3.1 months to 71 years (for trials that reported mean or median age), with six studies (17.6%) exclusively focusing on the pediatric population. The proportion of male patients ranged from 50.0–76.7%. Most of the trials (n = 32, 94.1%) recruited patients from Asia. A single trial (2.9%) examined three treatment groups. The interventions included VPS, ETV, LPS, LPS with laparoscope (LPS + LS), and ETV + CPC. The risk of bias assessment for eligible trials for each outcome is provided in Supplementary Table 1. Twenty-five (73.5%) trials were judged as having a low risk of bias, eight (23.5%) as having some concern, and one (2.9%) as having a high risk of bias due to missing outcome data.

A network of eligible comparisons for the primary and secondary outcomes is presented in Figure 3 and Supplementary Figure 1. The geometry of the evidence network indicated that all surgical interventions were evaluated in at least one RCT, with most comparisons involving VPS.

Figure 4 presents the NMA league table and SUCRA scores for the primary outcome. To ensure the certainty of the evidence, we incorporated CINeMA assessments into the league table. The quality of evidence graded by CINeMA ranged from low to high, with low-quality evidence primarily attributed to major concerns in imprecision and incoherence (Supplementary Table 2). We found no evidence of funnel plot asymmetry (Supplementary Figure 2). The results of the model fit statistics for the FE consistency, RE consistency, and inconsistency models are presented in Supplementary Table 3. FE models were selected for most outcomes except operation time and duration of hospitalization. The leverage plots are provided in Supplementary Figure 3. No evidence of global inconsistency was found (Supplementary Table 3). However, node-splitting analysis revealed some local inconsistencies in complications and seizures (Supplementary Figure 4). We verified the data for errors, violations of homogeneity or transitivity, and other sources of inconsistency. No errors were found in data extraction or statistics, and no significant variables differed across comparisons.

In terms of favorable outcome (31 RCTs, 2,390 patients), LPS (RR 1.18, 95% CrI 1.13–1.23, high certainty) and LPS + LS (RR 1.27, 95% CrI 1.18–1.39, high certainty) were more effective compared to VPS (Figure 4A). Additionally, both LPS and LPS + LS outperformed ETV with or without CPC (RR range 1.16–1.48; high to moderate certainty). Figure 4B and Supplementary Figure 5 present the SUCRA curve score, ranking each surgical intervention. LPS + LS ranked highest (SUCRA 99.2%), followed by LPS (75.5%), ETV (38.7%), and VPS (30.5%), with ETV + CPC having the lowest score (6.2%).

In the NMA of complications (23 RCTs, 1,783 patients), LPS, LPS + LS, and ETV significantly reduced adverse events compared to VPS, with RR ranging from 0.20–0.40, supported by high-certainty evidence (Figure 4A). LPS + LS had fewer complications than LPS (RR 0.49, 95% CrI 0.29–0.79; moderate certainty). The SUCRA curve (Figure 4B; Supplementary Figure 5) suggests that LPS + LS has the lowest risk of adverse events (SUCRA 93.3%), followed by ETV (77.1%) and LPS (53.7%), while ETV + CPC has the highest risk (7.6%).

Figure 5 shows two-dimensional graphs of efficacy versus safety in all studies and head-to-head studies. Overall, both analyses yielded similar results. In all studies, LPS + LS was the optimal surgical intervention for efficacy and safety. In head-to-head comparisons, LPS and LPS + LS were superior to VPS in both efficacy (LPS: RR 1.20, 95% CI 1.15–1.26; LPS + LS: RR 1.29, 95% CI 1.17–1.42) and safety (LPS: RR 0.34, 95% CI 0.25–0.45; LPS + LS: RR 0.32, 95% CI 0.20–0.52). ETV was superior to VPS only in safety (RR 0.28, 95% CI 0.13–0.61) (Supplementary Figure 6).

The pooled analysis comparing lumbar and cranial approaches was shown in Figure 6, with certainty of evidence assessed using the GRADE approach (Supplementary Table 4). Due to the absence of head-to-head comparisons between LPS or LPS + LS and ETV or ETV + CPC, the cranial approach in our analysis effectively comprised only VPS. Across various follow-up time points from 1 week to 12 months, lumbar surgeries were superior to cranial surgeries in both efficacy (RR 1.23, 95% CI 1.19–1.28; moderate certainty) and safety (RR 0.33, 95% CI 0.26–0.43; moderate certainty) (Figure 6; Supplementary Table 4). Moreover, compared with cranial access, lumbar approach significantly reduced the incidence of infection (RR 0.28, 95% CI 0.17–0.46), revision (RR 0.28, 95% CI 0.16–0.49), and seizures (RR 0.41, 95% CI 0.18–0.94), and significantly shortened both operative time (MD −23.07, 95% CI −25.42, −20.73) and duration of hospitalization (MD −8.73, 95% CI −9.40, −8.06) (Figure 6).

Results of secondary outcomes are presented in Supplementary Figures 5, 7. For infection (27 RCTs, 1,991 patients), ETV, LPS, and LPS + LS significantly reduced infection rates compared to VPS, with RR ranging from 0.06–0.29, supported by moderate-to-high certainty of evidence. In seven RCTs (808 patients), LPS and LPS + LS significantly reduced the risk of revision compared to VPS, ETV, and ETV + CPC (LPS: RR 0.27–0.39; LPS + LS: RR 0.10–0.37, moderate certainty). In nine trials (717 patients), only LPS significantly reduced seizure incidence compared to VPS (RR 0.29, 95% CrI 0.09–0.81, moderate certainty). In nine trials (750 patients), the operation time for LPS, LPS + LS, and ETV were significantly shorter than for VPS, with MD ranging from −19.78 to −34.77 (moderate evidence). In eight RCTs (690 patients), compared to VPS, both ETV (MD −4.41, 95% CrI −8.55 to −0.29, moderate evidence) and LPS (MD −7.23, 95% CrI −9.74 to −4.79, moderate evidence) significantly shortened the duration of hospitalization.

Subgroup and sensitivity analyses based on age, statistical model selection, etiology, evidence quality, and follow-up duration did not reveal substantial variation in the results (Supplementary Figure 8). For infant CH, no RCTs evaluated LPS or LPS + LS. Among the available interventions, VPS ranked highest in both efficacy and safety, followed by ETV and ETV + CPC. For adult CH, LPS + LS demonstrated the highest efficacy, followed by LPS, ETV, and VPS. In safety, LPS + LS also ranked first, with ETV, LPS, and VPS following in descending order. No RCTs evaluating ETV + CPC were identified for adult CH. For iNPH, LPS + LS was again the most efficacious intervention, followed by LPS, VPS, and ETV, though no RCTs reported comparative safety outcomes for this specific population (Supplementary Figure 8).

Our study had some limitations. First, the lack of a standardized grading system for hydrocephalus led to imprecise outcome definitions. However, we detailed the definitions of favorable outcomes in each original study (eMethods in the Supplementary material) and synthesized them based on expert consensus. The I² statistics indicated no statistical heterogeneity. Second, this same lack of a standardized grading system also limited the precision of severity assessment across studies. Consequently, subgroup analyses based on severity indices (such as Evans’ index or clinical severity scores) could not be performed. These limitations constrain the direct clinical applicability of our findings. Therefore, while our analysis provides a robust evidence base, the conclusion suggesting the superiority of lumbar approaches should be interpreted with caution in clinical practice, particularly in severe cases with significantly elevated intracranial pressure or marked ventricular dilation. Third, the included studies originated from Asia, Africa, and South America, with the majority from Asia. This geographic distribution did not result from the deliberate exclusion of trials conducted in North America or Europe. Rather, it reflects the current global research landscape in this specific field. To ensure a rigorous comparison of surgical strategies for CH, we restricted inclusion to RCTs that (1) employed random allocation, (2) directly compared different surgical procedures, and (3) enrolled patients with a clearly defined diagnosis of CH. Based on comprehensive and systematic searches across multiple major databases, no trials from North America or Europe were identified. Given known regional differences in surgical practice patterns, healthcare systems, and patient characteristics, caution is warranted when extrapolating these findings to Western populations, and further high-quality RCTs from diverse geographic regions are needed to enhance external validity. Fourth, including both adult and pediatric CH may have influenced the transitivity assumption. To address this concern, we performed prespecified subgroup analyses for adults and infants, and the results remained unchanged. Fourth, data on several less common surgeries including VAS, VPlS, VSSS, and laparoscopy-assisted VPS, were unavailable due to the lack of eligible RCTs. However, our study covers current primary clinical treatment scenarios for CH. Fifth, CINeMA assessed some comparisons as low or very low quality. Censored adverse event data below a pre-specified study-dependent threshold may bias the estimated incidence in these comparisons. We integrated the certainty of evidence into our results and interpreted them with caution. Future RCTs should place greater emphasis on censored adverse event including secondary tonsillar descent or Chiari formation to provide more comprehensive evidence.