This NMA has provided no clear evidence that any one intervention or treatment regimen is better than any other across the spectrum of outcomes. The use of glucocorticoids in the critically ill is commonplace, but its targets and true indications are hazy. There have been three indications for the use of glucocorticoids in sepsis over the last century. First, they were given as immunosuppressants; large doses of glucocorticoids that have significant immune effects such as dexamethasone and methylprednisolone were given until studies in the late 1980s [
23‐
27] showed a trend towards increased rates of superinfection and these high dose studies were halted. Second, there have been attempts to identify and treat a group of patients who are thought to be relatively deficient in glucocorticoids during their critical illness (critical illness-related corticosteroid insufficiency [
1]), although sensitive and specific diagnostic tests for this remain to be found [
28‐
31]. Last, and more recently, there has been a movement towards giving all patients who are unresponsive to, or on high-dose vasopressors, so called low-dose hydrocortisone (<400 mg/day) [
32‐
34], in effect using it as a non-catecholamine vasoconstrictor. It is this approach that is currently recommended in the Surviving Sepsis guidelines [
35]. This variation of study design intention has led to significant heterogeneity between the studies in this area. To try and reduce this, we excluded studies specifically designed with respiratory function as the outcome [
12‐
16]. The hypothalamic-pituitary-adrenal axis of children is different from that of adults and develops across childhood and puberty [
36]. Including the outcomes of children would further increase heterogeneity if assimilated with adult studies, and therefore data from children were also excluded for this analysis. In health [
6], major surgery [
7] and critical illness [
8], adrenocorticotropic hormone (ACTH) and cortisol are secreted in a pulsatile manner generated by the positive feedforward and negative feedback of these two hormones. This pulsatility is critical for patients with absolute cortisol deficiency [
37,
38] to maintain normal glucocorticoid receptor (GR) signalling and optimal physiological function. The pulsatile secretion pattern also translates into the pattern of receptor binding [
39], with continuous and intermittent binding yielding both quantitative and qualitative differences in gene transcription [
40]. In health, much of the effect of glucocorticoids is mediated by the mineralocorticoid receptor, whereas at higher (stress) levels, the effects are mediated mainly through the GR. This, coupled with the knowledge that cortisol binding globulin is saturable [
41] at plasma cortisol levels of around 400 nmol/L, means that GR may be continually activated during critical illness rather than the intermittent activation of health. Intermittent delivery of hydrocortisone under these conditions may not yield the benefits seen under non-stressed conditions. There is some early animal evidence that there are no pattern dependent effects (continuous versus pulsatile) seen at glucocorticoid-responsive genes when high-dose plasma corticosteroids are used [
42]. More studies are required to clarify the differential effects of these different glucocorticoid patterns.
The drive behind this NMA was to gain information that may aid in the design of improved treatment regimens. This analysis shows that hydrocortisone is likely to lead to shock reversal, but that this does not yet translate into improved mortality outcomes. Design of studies in this area should now be focused on elucidating the optimal dose and regimen for glucocorticoid treatment using hydrocortisone, yielding the benefits of improved cardiovascular reactivity and capillary permeability, with the lowest risk of hyperglycaemia, superinfection and GI bleeding.
Comparison with other studies
There have been previous meta-analyses of steroids within the critical care environment [
2‐
4], but, to our knowledge, this is the first using an NMA approach and the first analysis focussing on the effect of the therapeutic agent and regimen. Although all corticosteroids possess the immune, metabolic and fluid homeostatic features of their group, marked differences in the activity of each drug in these features exists. Authors of the previous meta-analyses have always used pairwise techniques to compare steroids versus placebo. In two cases, this was using a frequentist approach [
2,
4]. In the third case, the different effects of high-dose (>1000 mg/day) versus low-dose (<1000 mg/day) hydrocortisone equivalents in synthetic ACTH responders and non-responders were analysed using a Bayesian approach [
3]. The results of all three were also inconclusive in terms of mortality, as our analysis is. The Bayesian meta-analysis, like our analysis, did show a high probability of treatment efficacy of low-dose hydrocortisone treatment in terms of shock reversal. This again was not at the expense of higher complications, and this pattern persists regardless of the method of analysis [
2‐
4].
Strengths and limitations
The strengths of this study are that it takes the robust, structured approach of a Cochrane review and applies a different statistical approach to the data. NMA allows multiple pairwise comparisons and therefore may allow a more granular approach to answering the clinical question than a generic ‘steroids versus placebo’ question. In principle, for a network analysis to hold true, the interventions should be ‘jointly randomisable’. That is, a participant in any trial meeting the inclusion criteria could have been randomised to any of the interventions [
43]. This is true for this analysis and the underlying principle regarding why patients included in the trials designed to look at respiratory function and children were excluded. The study included data from over 3000 patients, and therefore it may appear that its ability to discriminate small differences in outcome should be reasonable. However, this ability is significantly reduced by dividing the patients into multiple treatment groups, whereby the number of patients in each group is fewer.
The limitations of the study include that there were few direct comparisons of treatment regimens (all but two studies were intervention versus placebo). This means that the strength of inference made in an NMA between different interventions is not as robust as it could be and that consistency between direct and indirect evidence could not be assessed. This study used data from the last 50 years. Improvements in intensive care over the last half-century are myriad, and the breadth of patients, in terms of both age and co-morbidities now treated within a critical care environment, has increased. This means that bias will have inevitably crept in. Whilst the inclusion criteria for studies may not have changed, the population from which they are drawn will. Many patients not considered appropriate for critical care in the 1970s and 1980s are the mainstay of work for many critical care units in developed countries [
44]. As mentioned previously, the glucocorticoids used in trials have changed with time from high-dose immunosuppressive agents towards lower-dose hydrocortisone. Thus, although at an individual level the included patients are ‘jointly randomisable’, between the groups there are likely to be significant differences. We also did not compare the effect of the dose of corticosteroids. The network is already rather sparse, and therefore our ability to differentiate between different doses and different agents is very low. Other limitations included that many of the studies reported different outcome measures for the same headline outcome. This made performing NMA on these outcomes impossible in many circumstances. This was most striking for ‘hyperglycaemia’, where the definition varied between ‘total insulin dose’ [
32] blood glucose >150 mg/dl on any day [
45], an increase of >200 mg/dl [
46] and others, hence its exclusion as an outcome. Standardisation of outcome definitions for critical care trials would improve this problem going forward.