Corticosteroids for acute respiratory distress syndrome

Acute respiratory distress syndrome (ARDS) is a major health problem, with the annual incidence in the United States approaching 200 000 cases.1 In the accompanying paper, Peter and colleagues present a systematic review and meta-analysis of nine randomised controlled trials (1073 patients) of corticosteroids for prevention and treatment of ARDS.2 ARDS is a form of severe respiratory failure resulting from direct pulmonary insults (for example, aspiration or pneumonia) or indirect systemic causes (for example, sepsis or trauma).3 This syndrome often has devastating consequences, such as the prolonged need for mechanical ventilation, a high probability of death, and long term physical and psychological sequelae in survivors. Treatment is unlikely to be successful in least developed countries because of limited critical care resources.4

No effective drug treatments are available for ARDS,5 but corticosteroids have attracted attention because they have anti-inflammatory properties that are relevant to ARDS pathology. They reduce both leakage of fluid through the alveolar-capillary membrane and the adhesion of neutrophils to the capillary endothelium, and they modulate the balance between proinflammatory and anti-inflammatory genes.6 However, these physiological benefits are tempered by concerns about side effects, including infections and neuromuscular weakness.

Peter and colleagues analysed outcomes using Bayesian methods, which may be unfamiliar to some readers. Standard frequentist analysis of observed data generates a 95% confidence interval. This implies that if the study were repeated many times, 95% of the confidence intervals generated should contain the true (but unknown) population effect. Unlike these standard confidence intervals, which despite their name do not tell readers how “confident” to be in the observed results, Bayesian methods produce a 95% “credible interval,” which has a 95% probability of containing the population treatment effect.7 These methods also estimate the probability that the population treatment effect exceeds a specified threshold or falls within a specified interval.

The Bayesian approach thus helps readers quantify the statistical uncertainty of the results, but at the cost of requiring a prespecified range of prior beliefs about the treatment effect.8 Clinicians’ prior beliefs—ranging from “enthusiastic” to “non-informative” (no prior belief) to “sceptical”—stem from clinical experience and interpretation of existing studies. In practice, Bayesian analysis of a clinical trial using non-informative priors gives similar results to standard statistical methods.7 However, the advantage of Bayesian results is that their interpretation is more intuitive than the confidence interval.9

The systematic review by Peter and colleagues included trials of moderate quality; steroids had opposite effects depending on the clinical scenario. When prescribed to patients at risk beforethe onset of ARDS in four trials, steroids increased both the odds of developing ARDS (odds ratio 1.55, 95% credible interval 0.58 to 4.05) and subsequently dying from ARDS (1.52, 0.30 to 5.94). Both credible intervals spanned 1 and thus could not exclude a null effect.

Bayesian analysis then answers a question not dealt with by frequentist statistics—what is the probability of harm given the available data? Neutral or sceptical clinicians would conclude that the probability of prophylactic corticosteroids causing ARDS (86.6%) and increasing mortality from ARDS (72.8%) is high. Such clinicians will probably decide that steroids are harmful and that they should be avoided, while even steroid enthusiasts may become less convinced.

Conversely, when used for established ARDS, corticosteroids seemed to decrease mortality (five trials; 0.62, 0.23 to 1.26) and increase days both alive and off mechanical ventilation (three trials). Reassuringly, they did not lead to new infections, except possibly at high doses. Although a null effect on mortality could not be excluded, the probability of a positive effect was high (93.2%).

Many neutral doctors will conclude that corticosteroids are probably helpful for established ARDS, although it would also be useful to know the probability that the reduction in odds ratio exceeds a clinically important threshold. Sceptics will correctly observe that while Bayesian meta-analysis expresses statistical uncertainty more clearly it does not deal with concerns related to bias or other limitations of the primary trials. Indeed, trials were few and clinically heterogeneous, with limited data on other side effects.

Finally, although steroid enthusiasts may be reassured, optimal delivery of this intervention remains unclear. Peter and colleagues found no strong evidence for associations between treatment effect and variables at the study level of total corticosteroid dose or duration of ARDS before treatment, although in the largest trial corticosteroids seemed to increase mortality when prescribed more than 14 days after the development of ARDS.10 In addition, trials differed in precautions to reduce steroid related complications, including actively checking for infection, limiting use of muscle relaxants, slow tapering of the steroid dose, and avoiding hyperglycaemia.

Consequently, for most doctors without strong prior beliefs, this systematic review provides moderately strong evidence for avoiding prophylactic corticosteroids in ARDS, and weak evidence for their therapeutic use. Such doctors will view these results as hypothesis generating and will await results of additional trials. In contrast, previously enthusiastic doctors will probably remain so, at least for the treatment of ARDS. In practice, clinicians have largely abandoned steroids for prevention and early treatment, but many still prescribe them for non-resolving ARDS.11

Given current data, should additional randomised trials of corticosteroids to treat ARDS be conducted? We think so, because relatively few patients have been studied; the corticosteroid regimen (dosing, timing of initiation and discontinuation, monitoring) is unclear; and these drugs are inexpensive and therefore useful in low resource settings if effective. Such trials should be informed by further pilot work to refine the study question. Sample size calculations should consider plausible treatment effects in ARDS (for example, ~9% absolute reduction in mortality with tidal volume limitation)12 and the minimum difference worth detecting that would be important to clinicians and patients13; Bayesian methods may also be used. In addition, clinical predictors of responsiveness to corticosteroids in patients with ARDS would also be desirable, given the side effects. Stated otherwise, the most compelling reason for additional research may simply be that clinicians’ beliefs after the systematic review still depend largely on their prior beliefs.