This NMA protocol has been registered on the PROSPERO international prospective register of systematic reviews (ID = CRD42020165093) [24]. This protocol is written following the Preferred Reporting Items for Systematic Review and Meta-Analysis Protocols (PRISMA-P) guidelines and the PRISMA checklist for Reporting Systematic Reviews incorporating Network Meta-analyses [25 - 27]. The PRISMA-P checklist for this study is included in Additional file 1.
Criteria for included studies
Participants
Adult patients (older than 18 years) with CS using VA-ECMO with or without LV unloading will be included. We will not apply restrictions about gender, ethnic origin, or other characteristics.
Interventions and comparators
We will include any LV unloading strategies adjunct to VA-ECMO covering surgical LV unloading strategy, IABP, Impella, and TandemHeart [12, 28 - 30]. Despite the common comparator is VA-ECMO alone, any LV unloading technique adjunct to VA-ECMO could be intervention or comparator in pairwise and network meta-analyses. The network of all possible pairwise comparisons among the eligible interventions is shown in Fig 1. As the surgical techniques of LV venting could be achieved by many approaches, we define the surgical LV unloading strategy as follows: (1) implanting LV venting cannulation of the ventricle apex or through the mitral valve from the left atrium (LA); (2) implanting a catheter across the aortic valve percutaneously; (3) implanting LV venting surgical cannulation through the right superior pulmonary vein, LA roof or interatrial groove into LA; (4) transseptal LA cannula; (5) an interatrial septostomy (septostomy usually with ballooning or stent); (6) the surgical or percutaneous pulmonary artery cannulation; (7) simultaneous left and right atrial drainage with the multistage cannula coming from the femoral vein and positioned transeptally [12, 17, 28, 31].
Outcomes measures
Primary Outcome
The primary outcome was all-cause in-hospital mortality (death of ECMO withdrawal due to futility, of patients unable to be weaned off ECMO, and of patients who died before hospital discharge despite successful ECMO weaning).
Secondary outcomes
The secondary outcomes will be neurological complications, hemolysis, bleeding, limb ischemia, renal failure, gastrointestinal complications, sepsis, duration of mechanical ventilation, length of intensive care unit, and hospital stays.
Study design and publication types
All published clinical studies investigating VA-ECMO and reporting data on LV unloading strategies will be evaluated for inclusion in this meta-analysis. Random controlled trials and prospective or retrospective cohort studies, reporting at least 10 adult CS patients, that compare different LV unloading techniques during VA-ECMO will be included in this study. Case-control and cross-sectional studies, case series studies, reviews and meta-analyses, letters to the editor, case reports, expert opinions, and animal studies will be excluded.
Information source and search strategy
PubMed, Embase, the Cochrane Library, and the International Clinical Trials Registry Platform (ICTRP) will be explored from their inception to 31 December 2020. We will use a combination of MeSH, EMTREE and free-text terms: 'extracorporeal membrane oxygenation', 'extracorporeal life support', 'intra-aortic balloon pumping', 'counterpulsation', 'impella', 'TandemHeart' 'transaortic catheter', 'transseptal left atrial cannula', 'decompression', 'venting', 'unloading'. The search strategy will be implemented by two experienced scholars of information retrieval (D.B.L and Y.N.L). Any potentially-relevant article will be retrieved for review. Besides, references of included studies and narrative reviews and meta-analyses will be considered for additional potential studies. There will be no restrictions on date limit, country, the language of publication, publication status, or year of publication. The search strategies are shown in Additional file 2.
Study screening and selection
All studies to be screened will be managed by Endnote X9 (Thomson-Reuters; 2018, New York, USA). Firstly, it will be used to classify and organize the preliminary literature and exclude repeated literature. Secondly, following the prespecified inclusion criteria, two independent reviewers (P.B.Z and S.L.W) will screen the title and abstract of each study independently and identify relevant studies. Thirdly, they will obtain and review the full text of all potential studies, then, they will make decisions independently and compare their selection of studies. Any discrepancies will be resolved by consensus. If consensus cannot be reached, the third authors (D.B.L and Y.N.L) will serve as an arbitrator. And if a discrepancy is caused by insufficient information of the literature, it is necessary to classify it into the category of waiting for evaluating and then decide whether it should be included after adding sufficient additional information. If studies have duplicate data, only the study with a larger sample size or longer follow-up time will be included. The proposed flow diagram of studies selection is illustrated in Fig 2.
Data extraction and management
Microsoft Excel (V.2019; Microsoft Corporation, USA) will be used to extract data from the included studies by two reviewers (P.B.Z and S.L.W) independently, using a standardized data extraction form. Missing data will be requested from study authors. Any discrepancies will be resolved by consensus. If consensus cannot be reached, the third authors (D.B.L and Y.N.L) will serve as an arbitrator. The characteristics of the extracted data items are shown in Table 1.
Risk of bias assessment
The risk of bias will be assessed using the Cochrane risk of bias tool or Newcastle-Ottawa Quality Assessment Scale (NOS) for random controlled trials or observational studies separately [32, 33]. The Cochrane Collaboration’s risk of bias assessment tool includes the following seven domains: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessment, incomplete outcome data, selective reporting, and other bias. Each item will be classified into one of three categories as follows: unclear, high or low risk. The NOS will be used to assess each included observational study using “star system”. Each study will be judged on three broad perspectives: the selection of the study groups, the comparability of the groups, and the ascertainment of the outcome of interest. A total score of 5 or less is considered low, 6 or 7 is considered moderate, and 8 or 9 is deemed of high quality. The judgments will be made by two review authors (P.B.Z and S.L.W) independently. Any discrepancies will be resolved by consensus. If consensus cannot be reached, the third authors (D.B.L and Y.N.L) will serve as an arbitrator.
Data synthesis and analysis
When quantitative analysis cannot be conducted, we will narratively describe the results. If the quantitative analysis is feasible, statistical analyses will be conducted using Stata (V.16, StataCorp, College Station, TX) and ADDIS (V.1.16.5 Aggregate Data Drug Information System, http://drugis.org/addis). The binary outcomes and continuous outcomes will be presented as risk ratios (RR) or mean difference (MD) with their 95% confidence intervals (CIs) respectively.
Pairwise meta-analyses
All the direct comparisons will be performed using Stata software and random-effects model if no less than two studies. Methodological and clinical diversity always exist in the pairwise meta-analysis, statistical heterogeneity is inevitable [34]. Cochran’s Q tests, I-squared statistic, and visual inspection of forest plots will be used to assess heterogeneity levels. If significant heterogeneity existing (I-squared ≥ 50% or p < 0.1), subgroup or sensitivity analysis or meta-regression will be used to explain the source of heterogeneity. Moreover, if there is considerable heterogeneity, especially when the direction of the effect is inconsistent, we will do a general statistical description.
Indirect and mixed comparisons of interventions
A random-effects network meta-analysis within the Bayesian framework will be applied. Interactions among all included studies will be shown in the network geometry, and the contribution plot for the network will show the contributions of direct comparisons [35]. In the network diagram, the dots will represent every intervention, and the size of the dot will mean the number of participants. The lines will indicate direct comparisons between different interventions and the thickness of the line will mean the number of studies [36]. For each outcome, we will present the contribution plots, which exposit the contribution of each direct comparison to the entire network as well as for each network estimate [37]. The main characteristics of NMA are ranking analysis having the ability to rank the various treatments for each outcome. The cumulative probability will be used to rank the included LV unloading strategies.
Assessment of inconsistency
Inconsistency is the statistical manifestation of the violation of the transitivity assumption. It is the differences between indirect and direct effect estimates for the same comparison, includes loop inconsistency and design inconsistency [38]. We will use the node-splitting method to evaluate the inconsistency between direct and indirect evidence locally. If p > 0.05, it suggests consistency between direct and indirect evidence. We will also investigate possible sources of inconsistency using the inconsistency factor (IF) among studies in each closed loop. If the 95% CIs of IF values include zero, it indicates that there is no significant inconsistency.
Subgroup analysis and sensitivity analysis
If there are sufficient data, we will conduct prespecified subgroup analyses for outcomes based on following: (1) etiology of cardiogenic shock (postcardiotomy shock (PCS); acute myocardial infarction (AMI); myocarditis; mixed etiologies); (2) quality of study; (3) the time of LV unloading; (4) the mechanism of LV decompression. Besides, the sensitivity analysis will also be conducted to validate the robustness of the results by excluding each study.
Assessment of publication bias
To assess small study effects and publication bias, a funnel plot will be used in pairwise meta-analyses when at least 10 studies would be analyzed. The comparison-adjusted funnel plot will be employed to identify possible small-study effects including publication bias at the network level [37]. And Egger’s test will be used to assess the symmetry of the funnel plot [39].
Quality of evidence
The Grading of Recommendations Assessment, Development, and Evaluation (GRADE) guidelines will be used to assess the quality of direct and indirect evidence for the main outcomes [40]. Five factors can reduce the quality of evidence: study limitations (risk of bias), inconsistency, indirectness, publication bias, and imprecision. Correspondingly, three factors can improve the quality of evidence: residual confounding, dose-response gradient, and large magnitude of the effect. The quality of evidence will be graded in four levels: very low, low, moderate, high.