Association of noninvasive respiratory support with mortality and intubation rates in acute respiratory failure: a systematic review and network meta-analysis

Background Noninvasive respiratory support devices may reduce the tracheal intubation rate compared with conventional oxygen therapy (COT). To date, few studies have compared high-flow nasal cannula (HFNC) use with noninvasive positive-pressure ventilation (NPPV). We conducted a network meta-analysis to compare the effectiveness of three respiratory support devices in patients with acute respiratory failure. Methods The Cochrane Central Register of Controlled Trials, MEDLINE, EMBASE, and Ichushi databases were searched. Studies including adults aged ≥ 16 years with acute hypoxic respiratory failure and randomized-controlled trials that compared two different oxygenation devices (COT, NPPV, or HFNC) before tracheal intubation were included. A frequentist-based approach with a multivariate random-effects meta-analysis was used. The network meta-analysis was performed using the GRADE Working Group approach. The outcomes were short-term mortality and intubation rate. Results Among 5507 records, 27 studies (4618 patients) were included. The main cause of acute hypoxic respiratory failure was pneumonia. Compared with COT, NPPV and HFNC use tended to reduce mortality (relative risk, 0.88 and 0.93, respectively; 95% confidence intervals, 0.76–1.01 and 0.80–1.08, respectively; both low certainty) and lower the risk of endotracheal intubation (0.81 and 0.78; 0.72–0.91 and 0.68–0.89, respectively; both low certainty); however, short-term mortality or intubation rates did not differ (0.94 and 1.04, respectively; 0.78–1.15 and 0.88–1.22, respectively; both low certainty) between NPPV and HFNC use. Conclusion NPPV and HFNC use are associated with a decreased risk of endotracheal intubation; however, there are no significant differences in short-term mortality. Trial registration PROSPERO (registration number: CRD42020139105, 01/21/2020) Supplementary Information The online version contains supplementary material available at 10.1186/s40560-021-00539-7.

There is a widespread application of NPPV in patients with ARF before tracheal intubation and IMV [6][7][8] which decreases the need for IMV rather than the use of COT [11,12]. NPPV potentially increases the risk of complications, including aspiration pneumonia, facial skin breakdown, eye irritation, interface intolerance, and patient discomfort from the inability to communicate or eat during therapy [13,14], which limits NPPV application in the clinical setting. HFNC can deliver highconcentration humidified oxygen via nasal cannulas without NPPV-related complications and is increasingly used in critically ill adult patients despite contradictory results from several clinical trials [9,10]. However, there is a paucity of evidence on pre-IMV HFNC use in patients with ARF.
Systematic reviews and meta-analyses that compared two of the three respiratory support devices (COT, NPPV, and HFNC) [15][16][17][18][19][20][21][22] showed that HFNC use reduced the tracheal intubation rate compared with COT, albeit without significance between-group differences when compared with NPPV. There was no intergroup difference in mortality between the use of any two of the three respiratory support devices. Several studies in those systematic reviews compared HFNC use with COT and NPPV with COT, although a few studies have compared HFNC use with NPPV. Small sample sizes possibly affected the results of the abovementioned systematic reviews. To overcome these limitations, we performed a systematic review and network meta-analysis (NMA) to compare the effectiveness of three supplemental respiratory support devices in studies that compared at least two of the three techniques (COT, NPPV, and HFNC use) in patients with ARF.

Protocol and registration
This systematic review was designed according to the Preferred Reporting Items for Systematic review and Meta-Analyses extension statement for reviews incorporating network meta-analyses (details in e- Table 1 in Additional file 1) [23], and the protocol is registered with PROSPERO (CRD42020139105).

Eligibility criteria Type of studies
We included all randomized-controlled trials (RCTs) reported in English and Japanese regardless of publication status (published, unpublished, and academic abstracts). Randomized crossover, cluster-randomized, and quasiexperimental trials were excluded.

Type of participants
This review included adults (age ≥ 16 years) with acute hypoxic respiratory failure, defined by any of the following criteria: ratio of arterial oxygen partial pressure to fractional inspired oxygen (P/F ratio) < 40.00 kPa; SaO 2 or SpO 2 < 94% on room air or > 95% with > 6 L/min; and PaO 2 < 8.00 kPa with room air or < 10.67 kPa with O 2 . This meta-analysis excluded studies in which more than half of the patients had post-extubation respiratory failure, acute exacerbation of chronic obstructive pulmonary disease (COPD), acute exacerbation of asthma, hypercapnia (> 6.00 kPa), tracheostomy, post-surgical status, trauma, and do-not-resuscitate orders. The exclusion criteria were limited to factors that were judged clinically appropriate for exclusion by the participating clinicians.

Types of interventions and comparators
We included RCTs comparing two of the following three methods before tracheal intubation: 1. COT: Low-flow nasal cannula, face mask, and venturi mask (with no limit on the flow rate). 2. NPPV: The type of mask and mode, duration of ventilation, and methods of weaning were not limited. 3. HFNC: The flow rate and F I O 2 were not limited.

Type of outcomes
The outcome measures included a primary outcome of short-term mortality at the end of the follow-up period (< 90 days), ICU discharge, and hospital discharge. The secondary outcome was the rate of intubation during ICU stay.

Information sources
We searched for eligible trials in the following databases: The Cochrane Central Register of Controlled Trials (CENTRAL); MEDLINE via PubMed; EMBASE; and Ichushi, a database of Japanese research papers. Additionally, we searched for ongoing trials in The World Health Organization International Clinical Trials Platform Search Portal. For cases with unknown data, the authors were contacted.

Study selection
Two of the three physicians (TM, HO, and HY) screened the title and abstract or the full text at the first and second screenings, respectively, for relevant studies and independently extracted data from the included studies into standardized data forms. Disagreements, if any, were resolved by discussion with one of three physicians who did not screen that particular study; original authors were contacted for clarification as required. For abstract-only studies that could not be evaluated for eligibility based on our review criteria, we attempted to contact the authors. Discrepancies between two reviewers were resolved by mutual discussion or discussion with a third reviewer as needed.

Data collection process and data items
After identifying studies in the second screening, data were extracted from each study by the reviewers (TM, HO, and HY) using two tools: the Cochrane Data Collection Form (RCTs only) [24] and Review Manager (RevMan) software V.5.3.5 (Cochrane Collaboration) [25]. We extracted the following study characteristics: (1) Methods: study design, total study duration, number and locations of study centers, study setting, withdrawals, and date of study initiation (2) Participants: number, mean age, age range, sex, severity of condition, diagnostic criteria, and inclusion/exclusion criteria (3) Interventions: treatment approaches and comparison methods (4) Outcomes: primary and secondary outcomes that were specified and collected, and the timepoints reported

Risk of bias within individual studies
The risk of bias of primary outcomes in the included studies was independently assessed by two of the three authors (TM and HO) using the Cochrane Risk of Bias tool 1.0 (Cochrane Collaboration) [26,27] in seven domains: (a) random sequence generation, (b) allocation concealment, (c) blinding of participants and personnel, (d) blinding of outcome assessors, (e) incomplete outcome data, (f) selective outcome reporting, and (g) other sources of bias. The risk of each bias was graded as low, unclear, or high. Discrepancies between the two reviewers were resolved through discussion among themselves or with a third reviewer as necessary.

Statistical analyses Direct comparison meta-analysis
A pairwise meta-analysis was performed using RevMan 5.3 (RevMan 2014). Forest plots were used for metaanalysis, and the effect size was expressed as relative risk (RR) with the 95% confidence interval (CI) for categorical data and as weighted mean differences with the 95% CI for continuous data. Outcome measures were pooled using a random effect model for study-specific effects in measures. For all analyses, a two-sided p value < 0.05 was considered statistically significant. Study heterogeneity between trials for each outcome was assessed by visually inspecting forest plots and with an I 2 statistic to quantify inconsistency [28] (RevMan; I 2 = 0-40, 30-60, 50-90, and 75-100% indicated minimal, moderate, substantial, and considerable heterogeneity, respectively). When heterogeneity was identified (I 2 > 50%), we investigated the reason and quantified it using the Chi-square test.
We planned to use a funnel plot, Begg's adjusted rank correlation test, and Egger's regression asymmetry test to investigate publication bias if ≥ 10 studies were available (RevMan) [29]. As < 10 studies were included, we did not test for funnel plot asymmetry.

Network comparison meta-analysis
Data synthesis A network plot was constructed to determine the number of studies and patients included in this meta-analysis. An NMA using netmeta 0.9-5 Rpackage (version 3.5.1) was performed via a frequentistbased approach with multivariate random-effects metaanalysis, and effect size was expressed as the RR (95% CI). Covariance between two estimates from the same study shows the variance of data in the shared arm, as calculated in a multivariable meta-analysis performed using the GRADE Working Group Approach for an NMA [30,31].
Transitivity The transitivity assumption underlying the NMA was evaluated by comparing the distribution of clinical and methodological variables that could act as effect modifiers across treatment comparisons.
Ranking Ranking plots (rankograms) were constructed based on the probability that a given treatment had the highest event rate for each outcome. The surface under the cumulative ranking curve (SUCRA), which is a simple transformation of the mean rank, was used to determine treatment hierarchy [32] and was constructed using standard software (Stata 15.0, Stata, TX, USA).
Risk of bias across studies The assessment of the risk of bias across studies followed considerations of pairwise meta-analysis, and conditions associated with "suspected" and "undetected" across-study bias were determined by the presence of publication bias on a direct comparison.
Indirectness We evaluated the indirectness, classified as "no", "some", or "major" concern, of each study included in the NMA based on its relevance to the research question, including the study population, interventions, outcomes, and study setting. The study-level judgments could be combined with the percentage contribution matrix.
Imprecision The approach to imprecision involved comparing the range of treatment effects included in the 95% CI with the range of equivalence. We assessed the heterogeneity of treatment effects for a clinically important risk ratio of < 0.8 or > 1.25 in the CIs.
Heterogeneity To assess the level of heterogeneity, we compared the posterior and predictive distributions of the estimated heterogeneity variance [33]. Concordance between assessments based on CIs and prediction of intervals, both of which do not capture heterogeneity, were used to assess the importance of heterogeneity of treatment effects for a clinically important risk ratio of < 0.8 or > 1.25 in prediction intervals.

Assessment of inconsistency
The inconsistency of the network model was estimated using inconsistency factors and their uncertainties. We statistically evaluated consistency using the design-by-treatment interaction test [34]. For inconsistency, p values less than 0.05, between 0.05 and 0.10, and otherwise were classified as "Major concerns", "Some concerns", and "No concerns", respectively.

Additional analyses
If there were sufficient data, we conducted a subgroup analysis of the severity of respiratory failure (P/F < 200) and the cause of respiratory failure (immunocompromised patients, excluding congestive heart failure (CHF)/acute exacerbation of COPD patients) to investigate the impact of risk of bias and assess the heterogeneity in participants in each study.

Study selection
The comprehensive search yielded 5507 records (e- Fig 1 in Additional file 1), of which 27 studies were included in this NMA [9,. These 27 studies included two three-group studies that directly compared NPPV with HFNC use and COT. The final analysis included 19, 7, and 5 studies that compared NPPV with COT, HFNC use with COT, and HFNC use with NPPV, respectively. The network structures of each outcome are shown in Fig. 1a and b.

Study characteristics
The protocols and characteristics of each study included in the final dataset of the meta-analysis are summarized in Table 1 Fig. 1 Network plots for the association of noninvasive oxygenation strategies with short-term mortality and intubation: a short-term mortality b Intubation hypoxic respiratory failure was pneumonia, followed by cardiopulmonary edema from CHF. Seven of the 27 included studies comprised many immunocompromised patients. The reported average P/F at randomization differed among the studies (range 95-249).   (37) AHRF Acute hypoxic respiratory failure; ALI Acute lung injury; ARDS, Acute respiratory distress syndrome; CAP Community-acquired pneumonia; CHF Congestive heart failure; COPD Chronic obstructive pulmonary disease; COT Conventional oxygen therapy; CPE Cardiogenic pulmonary edema; HFNC High-flow nasal cannula; ICU Intensive care unit; NPPV Noninvasive positive pressure ventilation Continuous data were shown as mean and standard deviation, except for data labeled with " a ". a Data were reported as median and IQR (interquartile range).

Network meta-analysis
The

Short-term mortality
In the analysis of short-term mortality (including 20 studies), compared with COT, NPPV (RR, 0.88 [95% CI, 0.76-1.01]; low certainty) and HFNC use (RR, 0.32 [95% CI, 0.80-1.08]; low certainty) showed trends for lower mortality risk (Fig. 2a) Table 2). Confidence in the RR of each comparison and shortterm mortality assessed by the GRADE system is shown in Table 3. Incoherence between direct and indirect RRs was observed for all three comparisons determined by p values of inconsistency. All comparisons (NPPV vs. COT, HFNC use vs. COT, and HFNC use vs. NPPV) showed "Major" concerns. The heterogeneity of all three comparisons resulted in "Major" concern outcomes due to the 95% CI of the predicted risk ratio.
A ranking analysis revealed that the hierarchy for efficacy in reducing short-term mortality was HFNC use (SUCRA 72.1), followed by NPPV (SUCRA 68.8) and ultimately, COT (SUCRA 9.0) (Fig. 3a). The summary of findings of the NMA for short-term mortality is shown in Table 2. The estimate and certainty of the evidence of direct, indirect, and network comparisons are summarized in e- Table 3 in additional file 1.
Confidence in the RR of each comparison and intubation assessed according to the GRADE system (Table 3) showed incoherence between direct and indirect RRs for all three comparisons determined by the p value of inconsistency. All comparisons (NPPV vs. COT, HFNC use vs. COT, and HFNC use vs. NPPV) showed "Major" concerns. The heterogeneity of all three comparisons resulted in "Major" concern due to the 95% CI of the predicted risk ratio. The ranking analysis revealed that the hierarchy for efficacy in reducing intubation was HFNC use (SUCRA 79.4), followed by NPPV (SUCRA 70.2), and ultimately, COT (SUCRA 0.4) (Fig. 3b). Table 4 summarizes the NMA findings for intubation; e- Table 3 in additional file 1 summarizes the estimate and certainty of the evidence of direct, indirect, and network comparisons.

Results of additional analyses
Of the 27 RCTs included in this study, only 14 focused on a single cause: 5, immunocompromised status; 4, pneumonia; 3, CHF; 1, post-transplant solid tumors; and 1, interstitial pneumonia. Therefore, the sensitivity analyses performed included the subgroup analysis of the severity of respiratory failure (P/F < 200) and the cause of respiratory failure (immunocompromised patients, excluding CHF/acute exacerbation of COPD patients). In all subgroup analyses, both short-term mortality and intubation rates were similar to those in the main analysis (e- Table 4 in Additional file 1).

Discussion
This systematic review and NMA showed that NPPV and HFNC use were associated with lower risks of intubation compared with COT rather than an improved mortality risk. The SUCRA values of intubation for HFNC use and NPPV in the NMA showed similar effects to those for NPPV and HFNC use. These results are consistent with those of previous systematic reviews (non-NMA) [18,21,61]. A systematic review by Zhao et al. [21] that included 11 studies (n = 3459) compared HFNC use with COT or NPPV and found that unlike NPPV, HFNC use reduced the intubation rate compared with COT. Those studies differ from this study as they included many post-extubation studies. Despite the results of the previous review being consistent with those of our study-that HFNC use reduces the intubation rate compared with COT-the inclusion of studies examining the prevention of re-intubation after extubation would result in a deviation from the clinical question in this study. The sample size was inadequate due to the lack of RCTs that directly compared NPPV and HFNC use, and previous reviews have not shown a significant difference between NPPV and HFNC use. To complement the limitations of the existing studies, a systematic review using an NMA was necessary. This NMA is the second study to demonstrate the effectiveness of NPPV and HFNC use in ARF. Ferreyro et al. [62] first reported an NMA describing the effects of noninvasive oxygenation strategies (e.g., NPPV and Second, they included studies wherein patients with COPD constituted < 50% of the population while excluding studies that had a majority (> 50%) of patients with COPD. As noninvasive oxygenation strategies are useful in COPD [63], these studies with populations mostly consisting of patients with COPD may have influenced the NMA results. Additionally, the inclusion of patients with COPD increased the heterogeneity of the Fig. 3 Surface under the cumulative ranking of each noninvasive oxygen strategies for short-term mortality and intubation: a short-term mortality, b intubation study population. Third, the NMA by Ferreyro et al. included the studies of ARF that occurred after abdominal surgery and chest trauma-associated respiratory failure, while excluding studies of patients after major cardiovascular surgery. Following abdominal surgery, diaphragmatic dysfunction and decreased vital lung capacity can cause atelectasis, resulting in hypoxemic respiratory failure (HRF). However, in pneumonia, which is the main cause of acute HRF, respiratory failure is caused by the decreased functional residual capacity due to inflammatory leachate in the alveoli and ventilator-perfusion mismatch. Analyzing these distinct pathogeneses of respiratory failure in a similar way is problematic and will affect the interpretation of results.

Implications
For patients, respiratory management without intubation is obviously more comfortable. Despite the superiority of HFNC use and NPPV being inconclusive in our study, noninvasive respiratory management is useful because per 1000 patients, utilizing NPPV and HFNC will help avoid intubation in 57-70 patients compared with COT. IMV is associated with various adverse events (e.g., VILI and VAP) and needs specific skills for comfortable management. Respiratory management that avoids tracheal intubation can reduce such complications, and the patient is relieved of the intubation discomfort. For the hospital manager, the costs of respiratory management may decrease due to the lower rate of intubation as daily ventilation costs increase healthcare costs by 59% compared with non-ventilation costs [64].

Limitations
This study had several limitations. First, there may be heterogeneity among the studies included, which may have affected the results. The NMA assumption is that the individual trials enrolled similar populations, and the intervention protocol was similar across different studies. Statistical heterogeneity is affected by a consequence of clinical or methodological diversity. Although statistical heterogeneity was a major concern in the mortality and intubation results, the clinical heterogeneity of diseases, including heart failure, pneumonia, and COPD, which were evaluated in this NMA, may not be considered high because of the difficulty in distinguishing between those diseases in the early stages of real-world clinical practice. The studies included in this NMA could not be clearly categorized by the cause of acute hypoxic respiratory failure because they did not include patients with a single cause. Therefore, sensitivity and subgroup analyses could not be performed. Second, all RCTs included had high risks of performance bias due to the dramatic differences between HFNC use, COT, and NPPV, which made blinding impossible.
Third, although the statistical analysis was based on the assumption that there was no effect modifier, the inclusion of patients with various degrees of respiratory failure and different outcomes may have influenced the results due to undetectable effect modifiers. However, even for outcomes with different baseline risks, the relative effects of the interventions can remain consistent [65]. A sensitivity analysis based on respiratory failure severity should have been conducted; however, it was not performed because of the possibly similar severity (e.g., P/F~200) of the patients in most studies included.
Lastly, a few studies compared NPPV and HFNC use, and thus, the sample size was insufficient to compare NPPV and HFNC use. However, a trend in the direction of a difference is present, which may be further evaluated in the future if more studies compare these two noninvasive oxygenation strategies.

Conclusions
The results of this NMA show that both NPPV and HFNC use are associated with lower risks of endotracheal intubation; however, no significant differences in short-term mortality exist between these respiratory support devices.
Additional file 1: e- Table 1. PRISMA NMA Checklist of Items to Include When Reporting A Systematic Review Involving a Network Meta-analysis. e- Table 2. Search strategy. e- Table 3. Estimate and certainty of the evidence of direct, indirect, and network comparison: (a) Short-term mortality. (b) Intubation. e- Table 4. Summary of sensitivity analysis for the association of all interventions with outcomes: (a) Short-term mortality.