Introduction

Acute respiratory distress syndrome (ARDS) is defined as acute respiratory failure due to non-hemodynamic pulmonary edema caused by inflammatory cytokines, and it is associated with a high mortality rate [1, 2]. Inflammatory cytokines were modulated by a pro-inflammatory transcription factor, nuclear factor-kappa B (NF-κB) [3, 4]. Steroids are expected to treat ARDS by attenuating NF-κB by the action of the glucocorticoid receptor α.

The most recent systematic review and meta-analyses (SR/MA) on ARDS have shown that the use of steroids decreases mortality and increases ventilator-free days (VFD) [5]. Another recent SR/MA suggested that early and longer administration of steroids in patients with ARDS might reduce mortality [6]. In these SR/MA, different types and dosages of steroids, such as high-dose methylprednisolone [7], low-dose methylprednisolone [8,9,10], hydrocortisone [11], and dexamethasone [12, 13] were used. In addition, the biological half-life and anti-inflammatory potency of steroids vary depending on its type. Though the differences in the type and dosage of steroids might affect outcomes, the optimal steroid regimen in terms of the benefit-risk ratio in ARDS remains unclear because these pairwise SR/MA assessed over all efficacy and outcomes of steroid by comparing groups of patients that did and did not use steroids.

We hypothesized that the steroid type influences the outcome in ARDS patients and sought to verify this by performing a network meta-analysis (NMA). Therefore, we conducted this systematic review and NMA to identify the optimal steroid therapy among patients with ARDS.

Methods

Protocol

We conducted this SR according to the Preferred Reporting Items for Systematic review and Meta-Analyses extension statement for reviews incorporating network meta-analyses (Online Resource 1) [14].

Eligibility criteria

We included randomized control trials (RCTs) that compared steroids for established ARDS. We did not search for quasi-randomized studies, cohort studies, case–control studies, and case series. ARDS was diagnosed according to the American European Consensus Conference (AECC) [15] definition (1994) or Berlin (2012) diagnostic criteria [16]. During our search for studies published before the establishment of the AECC, we included patients with acute respiratory failure, which was defined as follows: an acute onset of hypoxemia with a PaO2 to FiO2 ratio (P/F ratio) of ≤ 200 mmHg, bilateral noncardiogenic pulmonary edema, no clinical evidence of increased left atrial hypertension, and a pulmonary artery wedge pressure ≤ 18 mmHg in the presence of a pulmonary artery catheter. Patients who had been previously treated with corticosteroids and immunosuppressive drugs or had cardiogenic pulmonary edema and hypercapnia without hypoxia were excluded. As we investigated the effects of steroids in established ARDS patients, we excluded patients with oxygenation of 200 < P/F ratio ≤ 300 and no ventilation. We also excluded patients who were administered inhaled steroids because of the limited effects on the lungs without collapse.

We categorized interventions according to the received four types of steroids based on previous studies [7,8,9,10, 12, 13, 17, 18] and compared the results obtained from the following five groups: high-dose methylprednisolone, low-dose methylprednisolone, hydrocortisone, dexamethasone, and no steroid. We defined high-dose methylprednisolone as ≥ 1000 mg/day on the first day, 1000 mg/day on the second and third day, and tapering from the fourth day. We defined low-dose methylprednisolone as a methylprednisolone dose of 1.0–2.0 mg/kg/day after loading dose of 1.0–2.0 mg/kg/day from the first day to 14th day, which was administered for at least 3 weeks with tapering. We defined hydrocortisone as 200–300 mg/day for 7 days. We defined dexamethasone as any dosage for 10 days. We defined no steroid as a placebo or conventional therapy.

The primary outcome of this SR was hospital mortality. The secondary outcomes were incidence of infection and number of VFD. When data on hospital mortality were not available, we collected data on the longest mortality within 28–60 days after randomization.

Information sources, search strategy, study selection, and data collection process

The searched databases were MEDLINE via PubMed, the Cochrane Central Register of Controlled Trials (CENTRAL), and Igaku Chuo Zasshi (ICHUSHI). The ICHUSHI Web is the largest database of Japanese medical journals, containing approximately 10 million manuscripts from 6000 journals. We comprehensively searched these databases for relevant studies published from the earliest records to March 2021. We also searched for ongoing and unpublished trials in the following trial registers: ClinicalTrials.gov and the World Health Organization International Clinical Trials Platform Search Portal (WHO ICTRP) in March 2021. The list of search terms is shown in Online Resource 2. Additionally, we manually searched reference lists of included studies, previous SR/MA [6, 19,20,21,22], and articles citing these studies (based on citation information from the Web of Science).

When the screened records did not contain the necessary information, we inquired the authors for details. Two or more reviewers independently screened the titles and abstracts of the studies in the first screening. Articles included in the first screening were assessed for eligibility according to the inclusion criteria by reading the full texts. When the reviewers disagreed with the inclusion, conflicts were resolved by consensus. Two or more reviewers independently extracted data from full manuscripts. We extracted data pertaining to study characteristics such as study design, locations of the study center, sample size, information of participants (age, diagnostic criteria, and main etiology of ARDS), interventions, comparisons, low tidal ventilation as conventional therapy, and outcomes. We contacted the authors of the studies to collect sufficient information if necessary. The relevant data of the included studies were summarized using Microsoft Excel and Cochrane Statistical Package Review Manager (version 5.3; Cochrane Collaboration, Copenhagen, Denmark).

Geometry of the network

We constructed network plots for each outcome in this NMA. Each node corresponded to a treatment strategy. The edge presented direct comparisons between interventions and comparators and added to the number of included studies.

Abstraction of data and assessment of the risk of bias

Two or more reviewers independently classified the risk of bias for each included study as high, low, or unclear. We assessed the following seven domains for risk of bias using the Cochrane risk of bias assessment tool [23]: random sequence generation, allocation concealment, blinding of participants and personnel, blinding of outcome assessor, incomplete data outcome, selective reporting, and other bias. When reviewers disagreed with the decision on the risk of bias, conflicts were resolved by discussing with a third reviewer.

Data synthesis

We conducted a pairwise meta-analysis using the random-effects method to assess for direct associations between interventions and outcomes. We integrated the dichotomous variables as risk ratios (RR) with 95% confidence intervals (CI) and continuous variables as mean differences (MD) with 95% CI. We assessed heterogeneity using visual inspection of forest plot and I2 statistics. We assessed indirectness by checking patient characteristic of each included studies. We assessed publication bias by checking the number of studies that had not been published on ClinicalTrials.gov and WHO ICTRP.

We conducted an NMA using the Bayesian random-effects method to derive direct and indirect estimates comparing all interventions and comparisons. We integrated the dichotomous variables as odds ratios (OR) with 95% credible intervals (CrIs) and continuous variables as MD with 95% CrIs. We assessed intransitivity by checking patient characteristic of each direct comparison as the source for indirect estimates. We assessed imprecision by each 95% CrIs crossing clinically important intervals of OR defined as 0.8–1.25. We assessed incoherence using the node-splitting method [24]. We derived rank probability for each intervention and comparison from the results of the NMA within a Bayesian framework and decided treatment hierarchy using surface under the cumulative ranking curve (SUCRA) [25].

We assessed the certainty of evidence (CoE) for each estimate using the Grades of Recommendation, Assessment, Development, and Evaluation Working Group (GRADE) approach for NMA [26, 27]. First, we assessed the CoE for direct estimate using five considerations (risk of bias, heterogeneity, indirectness, publication bias, and imprecision). Second, we assessed the CoE for the indirect estimate by considering the CoE for the direct estimate of the common comparator related to the indirect estimate and using two considerations (intransitivity and imprecision). Finally, we assessed the CoE for network estimate by considering the CoE for direct and indirect estimates and using two considerations (incoherence and imprecision).

We defined statistical significance as 95% CIs and CrIs that did not cross the value 1.0. We used the Cochrane Statistical Package Review Manager (version 5.3; Cochrane Collaboration, Copenhagen, Denmark) for the pairwise meta-analysis. We performed NMA using JAGS and R (version 4.0.4; The R Foundation, Wien, Austria) software, whereas the rjags and gemtc packages were used for the Bayesian approach. We then created the table showing the summary of findings [28].

Results

Study selection

A flow diagram of this study is shown in Fig. 1. After eliminating the duplicates, we assessed 2810 records through databases and other source searches. We excluded 2714 records by screening the titles and abstracts and 96 records using full texts. On assessing the remaining 54 records for eligibility, we found two articles [8, 29] published from the same study. Subsequently, we identified 15 eligible studies. Among the 15 eligible studies, we excluded six studies with unpublished outcomes (Online Resource 3). Two of the six excluded studies had completed recruitment, but we did not get a reply from the authors regarding the outcome data, proportion of ARDS, and diagnostic criteria for ARDS. Finally, we included nine studies (n = 1212) for quantitative synthesis, the details of which are provided in Table 1. The network of eligible comparisons for the meta-analysis is shown in Fig. 2.

Fig. 1
figure 1

Flow diagram of this study. CENTRAL Cochrane Central Register of Controlled Trials, ICHUSHI Igaku Chuo Zasshi, ICTRP World Health Organization International Clinical Trials Platform Search Portal

Full size image
Table 1 Details for included studies
Full size table
Fig. 2
figure 2

Network plot. A Hospital mortality B incidence of infection C Ventilator-free days. Each node corresponded to a treatment strategy. The edge presented direct comparisons between interventions and comparators and added to the number of included studies

Full size image

Study characteristics

One study compared high-dose methylprednisolone with placebo [7]. Three studies compared low-dose methylprednisolone with placebo [8,9,10], two studies compared hydrocortisone with placebo [11, 30], and three studies compared dexamethasone with no steroid and were not used as placebo [12, 13, 31]. Direct pairwise comparisons are provided in Online Resource 4.

The most common etiology was pneumonia 57.1% (692/1,212). Ventilation strategy of six studies was low tidal ventilation [9, 11,12,13, 31, 32]. We grouped 28-day mortality for two studies [12, 30], 60-day mortality for three studies [9, 11, 31], and 45-day mortality for one study [7] to assess hospital mortality. VFD were measured in eight studies, while VFD of one study [8] were obtained from the author. We converted 95% CI of VFD to a standard deviation in one study [12]. Data on infections were collected from nine studies. One study [13] assessed new infections in the ICU, while another study [31] assessed pneumonia in the ICU. The observation period for infection was 28 days in two studies [9, 12], 7 days in one study [7], 1 year in one study [10], and up to discharge in one study [11]. Two studies [8, 30] did not report the observation period for the infection. Two studies [12, 31] did not use a placebo and had unplanned termination of the recruitment of participants. We received no reply about the data on 60-day mortality and infection from the authors of one of the studies [31].

Risk of bias and CoE

We assessed the overall risk of bias of four studies [7,8,9, 11] as low, that of four studies [10, 12, 30, 31] as unclear, and that of one study [13] as high using the Cochrane risk of bias assessment tool. Domains of risk of bias for each outcome are shown in Table 2. We showed processes of assessment for CoE in Online Resource 5, and the results concerning CoE in Online Resource 6. As we could not analyze direct comparisons for each steroid, we could not assess incoherence between direct and indirect estimates.

Table 2 Risk of bias summary
Full size table

Hospital mortality

Network estimate for hospital mortality was created from nine RCTs including 1212 patients (Table 3). High-dose methylprednisolone, low-dose methylprednisolone, hydrocortisone, and dexamethasone were not associated with a reduction in hospital mortality compared with no steroid. Using dexamethasone as a reference comparator, there was no statistical difference in the association with hospital mortality when comparing high-dose methylprednisolone (OR 1.16 95% CrI [0.08, 12.9], 36 more patients per 1000 patients, CoE: very low). Low-dose methylprednisolone (OR 0.66 95% CrI [0.07, 3.04], 95 less patients per 1000 patients, CoE: very low), and hydrocortisone (OR 0.72 95% CrI [0.06, 4.35], 76 less patients per 1000 patients, CoE: very low) were not associated with the reduction of hospital mortality compared with dexamethasone. Using hydrocortisone as a reference comparator, high-dose methylprednisolone (OR 1.62 95% CrI [0.13, 31.0], 113 more patients per 1000 patients, CoE: very low) was not associated with increasing hospital mortality. There was no statistical difference in the association with hospital mortality when comparing low-dose methylprednisolone with hydrocortisone (OR 0.93 95% CrI [0.11, 7.42], 16 less patients per 1000 patients, CoE: low). Using low-dose methylprednisolone as a reference comparator, high-dose methylprednisolone (OR 1.74 95% CrI [0.18, 30.0], 118 more patients per 1000 patients, CoE: very low) was not associated with increasing hospital mortality. The rank probability of hospital mortality was in the order of low-dose methylprednisolone, hydrocortisone, dexamethasone, high-dose methylprednisolone, and no steroid (Fig. 3).

Table 3 NMA-SoF table
Full size table
Fig. 3
figure 3

Rank probabilities in the network meta-analysis

Full size image

Incidence of infection

Network estimate for incidence of infection was created from nine RCTs including 1212 patients (Table 3). High-dose methylprednisolone, low-dose methylprednisolone, hydrocortisone, and dexamethasone were not associated with a reduction in the incidence of infection compared with no steroid Using dexamethasone as a reference comparator, high-dose methylprednisolone (OR 2.26 95% CrI [0.55, 10.2], 174 more patients per 1000 patients, CoE: very low), and hydrocortisone (OR 1.32 95% CrI [0.49, 3.84], 53 more patients per 1000 patients, CoE: very low) were not associated with increasing incidence of infection compared with dexamethasone. There was no statistical association when the reduction of incidence of infection compared low-dose methylprednisolone with dexamethasone (OR 0.74 95% CrI [0.32, 1.77], 49 less patients per 1000 patients, CoE: very low). Using hydrocortisone as a reference comparator, high-dose methylprednisolone (OR 1.73 95% CrI [0.33, 8.45], 95 more patients per 1000 patients, CoE: very low) was not associated with increasing incidence of infection. There was no statistical association when the reduction of incidence of infection compared low-dose methylprednisolone with hydrocortisone (OR 0.56 95% CrI [0.19, 1.60], 71 less patients per 1000 patients, CoE: low). Using low-dose methylprednisolone as a reference comparator, high-dose methylprednisolone (OR 3.06 95% CrI [0.70, 14.1], 272 more patients per 1000 patients, CoE: very low) was not associated with increasing incidence of infection. The rank probability of the incidence of infection was in that order of low-dose methylprednisolone, dexamethasone, hydrocortisone, no steroid, and high-dose methylprednisolone (Fig. 3).

VFD

Network estimate for VFD was created from seven RCTs including 1094 patients (Table 3). Low-dose methylprednisolone was associated with increasing VFD compared with no steroid (MD 6.00, 95% CI [3.39, 8.61]; CoE: high). Other comparisons with no steroid showed that the other treatments were not associated with increasing VFD. Using dexamethasone as a reference comparator, low-dose methylprednisolone (MD 2.40 95% CrI [− 2.87, 8.70], CoE: very low) was not associated with increasing VFD. There was no statistical difference in association with VFD when hydrocortisone was compared with dexamethasone (MD − 1.54 95% CrI [− 8.34, 4.87], CoE: very low). Using hydrocortisone as a reference comparator, low-dose methylprednisolone (MD 3.99 95% CrI [− 1.90, 11.1], CoE: moderate) was not associated with increasing VFD. The rank probability of VFD was in the order of low-dose methylprednisolone, dexamethasone, hydrocortisone, and no steroid (Fig. 3).

Discussion

The present study revealed that although there were no significant differences in hospital mortality and incidence of infection among all comparisons, including steroids and no steroid, low-dose methylprednisolone resulted in a greater improvement of VFD compared with no steroid. Though the CoE of most comparisons was very low, the anticipated absolute effect and rank probability indicated that the effect of steroids on the outcome in ARDS patients might vary depending on the type and/or dose of steroids.

The analysis of individual patients’ data from RCTs showed that low-dose methylprednisolone reduced mortality and VFD compared to placebo, although this SR/MA did not compare each steroid regimen [32]. Prolonging any steroid was suggested to improve mortality of patients with ARDS regardless of the type of steroids [33]. In addition, recent SR/MAs was demonstrated for early and longer steroids in patients with ARDS and its effectiveness on mortality was verified by synthesizing studies on low-dose methylprednisolone and dexamethasone [6]. One recent retrospective propensity-matched cohort study suggested that the pulsed methylprednisolone therapy worsened the 60-day mortality and VFD [18]. In contrast to previous SR/MAs, our direct pairwise comparisons show that any of the steroids were not associated with reducing hospital mortality and incidence of infection compared with no steroid. Based on these results, any type of steroids might not build up a steadfast position.

Though previous studies provided results of a pairwise comparison with placebo, we created indirect estimates of comparisons among steroids by NMA and verified estimates of hospital mortality, the incidence of infection, and VFD among each steroid. This NMA for patients with ARDS revealed that low-dose methylprednisolone or hydrocortisone relative to dexamethasone may reduce hospital mortality (95 per 1000 patients fewer, 76 per 1000 patients fewer, respectively), high-dose methylprednisolone relative to hydrocortisone may increase hospital mortality (113 per 1000 patients more), and high-dose methylprednisolone relative to low-dose methylprednisolone may increase hospital mortality (118 per 1000 patients more). Low-dose methylprednisolone relative to hydrocortisone may reduce incidence of infection (71 per 1000 patients fewer), while high-dose methylprednisolone relative to other steroids may increase incidence of infection (95–272 per 1000 patients more). Low-dose methylprednisolone relative to hydrocortisone and dexamethasone may increase VFD (6.06, 2.40 higher, respectively). In addition, the rank probability showed that low-dose methylprednisolone may be the optimal treatment, whereas high-dose methylprednisolone and no steroid may be the inferior to other treatment in terms of hospital mortality, incidence of infection, and VFD. Though the evidences were uncertain, considering these results and point estimates, our NMA suggested that the effect of steroids on outcome vary depending on the types of steroids.

Our NMA has several limitations. We could not use EMBASE for our database search. However, we consider studies that were included as appropriate, as we manually searched the reference list of the included articles and their reference lists and checked previous SRs/MAs. Although we performed an NMA, each comparison included only a few studies. In addition, NMA for VFD did not include studies using high-dose methylprednisolone. Conventional therapy strategies besides steroids may vary among eligible studies. In particular, compliance with the low tidal ventilation strategy, which was the strategy to improve mortality, was low in 1999 or earlier. Though the definition of infection was not reported in some eligible studies and was discrepant between studies, the definition of each study conformed to the general criteria. There may be an unexplained heterogeneity among eligible studies other than different definitions of infection because a placebo was not used in studies on dexamethasone. However, we made efforts to minimize heterogeneity by including patients with mechanical ventilation and checking age, diagnosis criteria, and etiology. Finally, we could not deny the possibility that not only the types of steroids but also the duration of administration might affect our results. We have defined the types and protocols of steroids based on previous studies, as previous RCTs did not verify different durations among steroids; however, we could not stratify results by the duration of steroids. Considering past studies, as low-dose methylprednisolone was administered the longest duration among the types of steroids, favorable outcomes of low-dose methylprednisolone may be obtained. Therefore, we could not draw conclusions on the optimal steroid agent. Further studies including these above studies must confirm our results.

In conclusion, our results indicate that the effect of steroids on outcome might vary depending on the type of steroids in ARDS patients. Further studies are needed to identify the optimal type and dosage.