This protocol has been written in accordance with the Preferred Reporting Items for Systematic Review and Meta-Analysis Protocols 2015 statement (PRISMA-P) and its extension for NMA.(33, 34) This protocol is also registered in the International prospective register of systematic reviews (PROSPERO, CRD42019121755).
Criteria for included studies
Participants and settings
We will include any studies that enrolled adult patients (18 years of age or older) with AHRF, defined by the new onset of clinical signs (tachypnea, increased work of breathing), radiologic signs (unilateral or bilateral chest x ray opacities) and gas exchange alterations (different degrees of hypoxemia). We will consider studies including patients treated in the intensive care unit, intermediate care unit, medical wards and emergency department. Studies primarily focused in the treatment of acute exacerbations of chronic obstructive pulmonary disease (AE-COPD) or congestive heart failure (CHF) will be excluded. Specifically, we will exclude such studies if fifty percent or more of the included population are patients with either AE-COPD or CHF, acknowledging that a lower threshold will exclude too many studies in the field. Our pre-planned sensitivity analysis will specifically address this issue. The rationale for exclusion is based in the already shown superiority of non-invasive ventilation for these conditions. Furthermore, we will also exclude studies including patients after extubation for major thoracic surgery.
Interventions and comparators
All interventions considered in this review are non-invasive oxygenation strategies. HFNC oxygenation refers to oxygen delivered through a nasal cannula, at a FIO2 up to 1.0 and a flow rate that can be as high as 60 liters / minute. NIV provides positive pressure ventilation that can be delivered with different modalities: continuous positive airway pressure, bi-level positive airway pressure or Pressure Support Ventilation. The interfaces used can be either a face mask (FM-NIV) or a helmet (H-NIV). Given the potential differences in reported outcomes, these interfaces will be considered as two different interventions in the network. SOT (low flow systems) comprise traditional nasal cannula, Venturi system masks or non-rebreather masks. All of these modalities can provide oxygen at low flow (<15 liters / minute) with varying levels of FIO2.
The primary outcome of this study will be all-cause mortality, defined as the longest available in the first 90 days after randomization. The secondary outcome will be receipt of invasive mechanical ventilation (longest available at 30 days). Given that the criteria for endotracheal intubation might vary between studies, we will collect information on whether decisions to intubate were based on pre-specified criteria or not.
Study design and publication types
We will include only randomized controlled trials focused on the treatment of acute hypoxemic respiratory failure if a combination of any of these interventions was assessed: HFNC, FM-NIV, H-NIV and SOT.
Information sources and search strategy
The following seven electronic bibliographic databases will be searched using a comprehensive search strategy developed by an information specialist (MR): (1) Ovid MEDLINE, (2) Ovid EMBASE, (3) PubMed (Non-Medline records only), (4) Ovid EBM Reviews - Cochrane Central Register of Controlled Trials, (5) EBSCO CINAHL Complete, (6) Web of Science, and (7) LILACS. We will also search ClinicalTrials.gov, WHO International Clinical Trials Registry Platform, and International Standard Randomized Controlled Trial Number Registry for all registered clinical trials and randomized controlled trials. The search strategy will be structured according to the Peer Reviewed Electronic Search Strategies (PRESS) 2015 Guidelines42 (Refer to Supplementary File for full search strategy). We will use a validated search filter for randomized controlled trials from the Cochrane Handbook for Systematic Reviews of Interventions Version 5.1.0, Section 6.4.1141. to screen Ovid MEDLINE, Ovid EMBASE and PubMed. We will adapt a pre-tested search filter for randomized controlled trials from the Scottish Intercollegiate Guidelines Network43 for EBSCO CINAHL Complete and Web of Science. No limits will be applied to language, publication year, gender or race. We will manage all references and duplicates using EndNote X8 citation management software.
Two reviewers (BLF and FA) will screen independently the titles and abstracts retrieved from the search strategy and the additional sources in order to identify those meeting the mentioned eligibility criteria. Subsequently, we will obtain full texts of the articles meeting these pre-specified criteria and review again in a second stage. Any disagreement between the reviewers will be discussed and referred to final decision by a third investigator (LM, HW, DS).
The two reviewers (BLF and FA) will perform data extraction independently in a pre-piloted data extraction form created in Excel (Microsoftâ). Any differences will be resolved by consensus or discussion with a third author (LM). Abstracted data will include study characteristics (trial design, size and funding source), patients’ characteristics (age, sex, etiology of AHRF, immunocompromised status and the presence of severe AHRF (PaO2:FIO2 ratio <200)), details of the interventions (location of application, duration of the exposure to each of the oxygenation strategy, type of non-invasive ventilation modality), outcome data for each endpoint of interest.
Risk of bias assessment
The risk of bias will be assessed independently using the tools specified in the Cochrane Handbook for Systematic Reviews of Interventions for RCTs. [27-30] The following aspects will be assessed: (1) Random sequence generation (selection bias); (2) Allocation concealment (selection bias); (3) Blinding of participants and personnel (performance bias); (4) Blinding of outcome assessment (detection bias); (5) Incomplete outcome data; (6) Selective reporting (reporting bias); (7) Other sources of bias .
Patient and public involvement
This research will be done without patient involvement. Patients will not be invited to comment on the protocol design nor consulted to develop patient relevant outcomes.
Data synthesis and analysis
We will summarize the included studies based on study and patient characteristics, outcomes measures and risk of bias. We will perform a series of pairwise Bayesian meta-analyses with a random-effects model, followed by a network meta-analysis using a Bayesian framework to derive head-to-head treatment effect estimates comparing all interventions. Analyses will be based on Markov chain Monte Carlo methods using prior distributions for event rates derived from previous literature, minimally informative treatment effect estimates, and informative prior distributions for heterogeneity estimates derived from external evidence.(35) We will report pairwise and NMA treatment effect estimates as risk ratio (RR), estimating summary treatment effect estimates from the median and corresponding 95% credible intervals (CrIs) from the 2.5th and 97.5th percentile of the posterior distribution. We will also rank interventions according to their apparent effectiveness, and will calculate the probability that interventions have risk reductions or increases greater compared to standard oxygen therapy. We will quantify heterogeneity in treatment effects between studies using the posterior distribution τ². Inconsistency (incoherence) between direct and indirect comparisons will be estimated using the node-splitting approach contrasting estimates from both direct and indirect evidence.(36, 37) We will visually assess model convergence using the Brooks-Gelman-Rubin diagnostic, trace plots, and auto-correlation plots. We will assess the goodness-of-fit of our final models by comparing the mean residual deviance with the number of contributing data points, calculating the percentage of standardized node-based residuals within 1·96 of the standard normal distribution, and visually inspecting the distribution of residuals on Q–Q plots. We will perform all analyses in R v3.6 and OpenBUGS.
Assessment of publication bias
We will assess for the presence of publication bias by examining the comparison-adjusted funnel plot.(38) We will examine the shape of the funnel plot, and assume that there is a risk of publication bias if its shape is asymmetrical. We will formally test for asymmetry in the funnel plot by performing the Harbord’s test.(39)
Where information is available, we will assess for credible subgroup effects using a random-effects NMA meta-regression Bayesian model to determine whether the estimated treatment effects are affected by the following factors: (1) Severe AHRF (PaO2:FIO2 ratio <200) vs less severe AHRF (PF ratio >200), hypothesizing that HFNC is better than SOT for patients with severe AHRF; and (2) Immunocompromised patients (hematologic malignancies, solid tumors with active chemotherapy, treatment with immunosuppressant drugs and solid organ transplant recipients) vs non-immunocompromised patients, hypothesizing that NIV may be more effective in immunocompromised patients.
We will perform sensitivity analysis for the treatment effect estimates by excluding studies with high risk of bias in each of the assessed items. Furthermore, we will conduct a sensitivity analysis by excluding those studies that included any patients with AE-COPD or CHF, and with non-informative prior distributions for the heterogeneity parameters.
Grading of recommendations
We will rate the quality of each direct, indirect and NMA estimates based on the four-step approach suggested by the Grading of Recommendations Assessment, Development and Evaluation (GRADE) Working Group.(40) The network effect estimates for each of the potential comparisons and every outcome, along with the GRADE rating for certainty in the evidence will be depicted in a summary of findings table.