Introduction
The use of prone positioning during acute respiratory distress syndrome (ARDS) ventilation has a robust scientific ground and was evaluated in numerous randomised controlled trials (RCTs). Despite significant and sustained increase of oxygenation, prone positioning had no impact on mortality [
1‐
4]. Most of these studies were underpowered, however, and meta-analyses intended to overcome the effects of inadequate sample sizes in individual RCTs failed to uncover any robust trend toward improved overall survival using prone positioning [
5‐
9]. Yet from the first RCT evaluating prone ventilation (Prone-Supine Study), Gattinoni and colleagues highlighted in a
post-hoc analysis that prone positioning reduced the 10-day mortality of patients with the highest disease severity (Simplified Acute Physiology Score II ≥50) [
1]. A similar message is conveyed by selected analysis of the most severe patients in study-level meta-analyses [
7,
8]. These findings were recently reinforced by the conclusions of the Prone-Supine II Study suggesting that the most severe ARDS patients (defined by PaO
2/FiO
2 ratio <100 mmHg) could derive beneficial effects from prone ventilation with reduced mortality [
10]. Consequently, recent meta-analyses of individual patient data obtained either from all published RCTs or from the four largest published RCTs showed unquestionably that the subgroup of the most severe patients (those with PaO
2/FiO
2 ratio <100 mmHg) had a significant reduction in mortality with prone ventilation [
11,
12].
Meta-analysis of individual patient data helps to avoid ecological bias, allows sufficiently powered subgroup analysis, and even allows powerful and reliable evaluation of treatment effects across individuals [
13]. This type of meta-analysis, however, does not solve all problems encountered in study-level meta-analyses. Indeed, the accuracy of individual patient data depends on the quality (conduct) and similarity (design) of primary studies, and heterogeneity might still be present if trials are not sufficiently similar or carry potential sources of bias [
13]. Moreover, individual patient data meta-analysis has frequently been shown to disclose divergent results from those of study-level aggregate meta-analysis [
13,
14].
In a previous aggregate meta-analysis we emphasised the substantial clinical (rather than statistical) heterogeneity of primary studies, making it difficult to conduct a study-level meta-analysis evaluating prone ventilation [
5]. This heterogeneity resulted merely from ecological bias, which is caused by confounding across trials [
15]. Ecological bias usually arises from within-group variability in covariates that may influence the outcome. In the particular setting of early studies on prone ventilation, ecological bias consisted of variable prone duration, mixed severity of acute lung injury, variable time-lapse between lung injury onset and inclusion, and lack of standardisation of co-interventions such as the lack of protective lung ventilation. Early studies were also vulnerable to treatment contamination, by allowing for crossover from one trial arm to another.
Given the large sample sizes of the initial studies, the heterogeneity in terms of severity as well as patient management heavily impacted the study-level meta-analyses [
5‐
9]. Of note, the most recent RCTs - which learned from the shortcomings of early studies, and were able to incorporate recent knowledge advances regarding lung-protective ventilation - reached a consistent design that was sharply different from that of the large RCTs published earlier in the decade. Indeed, careful examination of these trials shows that they share the following common features: inclusion of the most severe patients (ARDS only, excluding acute lung injury (ALI) non-ARDS patients), and control for the most relevant confounders - that is, proning duration (usually >17 hours/day) and use of lung-protective ventilation [
10,
16,
17]. Interestingly, each of these studies reported a substantial reduction in absolute risk of mortality varying between 9 and 15% but lacked power to reject a type II statistical error [
10,
16,
17]. The possible minimisation of ecological bias therefore makes these new studies an interesting opportunity for a new updated study-level meta-analysis.
In the present article, we update our recent meta-analysis of the effects of prone positioning on intensive care unit (ICU) mortality. Along with a global meta-analysis, a subgroup meta-analysis is performed on the group of studies that restricted inclusion to only adult ARDS patients. We also explore the effects of proning duration.
Discussion
The current meta-analysis shows that global analysis of RCTs assessing ventilation in the prone position in ALI/ARDS patients does not show a significant benefit on ICU mortality. The subgroup analysis stratified on the type of included patients in primary studies, however, disclosed a statistically significant reduction in mortality in the studies that restricted inclusion to only patients with ARDS, and not in those also enrolling patients with less disease severity. The comparison of the mean effect size between subgroups was close to significance (
P = 0.06), however, which does not allow one to ensure that the effects of proning were significantly different between subgroups. Another confounder may also be the daily duration of ventilation in the prone position (
P = 0.06). Prone positioning was not associated with an increase in major airway complications. The current study-level meta-analysis confirms and reinforces recent findings of individual patient data meta-analyses made by Sud and colleagues and Gattinoni and colleagues [
11,
12].
In many meta-analyses, the inclusion criteria are so broad that a certain amount of diversity among studies is inevitable. A study-level meta-analysis should anticipate this diversity and interpret the findings according to the results dispersion across the primary studies. We therefore applied the random-effects model, and computed a summary effect in subgroups of studies enrolling patients of variable lung injury severity, yielding important information on the peculiar effects of prone ventilation in the most severe patients.
A way to fully account for the ecological bias inherent to diversity of designs in primary studies is the performance of a meta-analysis using individual patient data [
13]. Indeed, previous inferences on prone ventilation benefits for the most severe hypoxemic patients were recently confirmed by the meta-analyses from Sud and colleagues and from Gattinoni and colleagues showing reduced mortality rate in patients with PaO
2/FiO
2 ratio <100 mmHg [
11,
12]. This threshold was considered prospectively only in the study by Taccone and colleagues [
10], however, while separation on this threshold basis was mostly retrospective for the other trials. Owing to increased risks of untoward effects, the authors recommended considering prone ventilation only in the most severe hypoxemia (despite a significant benefit up to PaO
2/FiO
2 ratio = 140 mmHg).
Our study used a different meta-analysis approach and stratified subgroups of studies according to the disease severity of included patients, rather than performing a subgroup analysis of included patients. This study reached the same conclusions as individual patient data meta-analyses, although our findings suggest that the benefits can go beyond the recommended threshold and concern all patients meeting ARDS criteria. A study-level meta-analysis like ours could therefore be an alternative for clinicians to detect true intervention effects (signals) despite differences among studies regarding participants, interventions, and co-interventions (noise) [
22]. We should, however, recognise that such meta-analysis necessarily suffers some shortcomings - such as mixing in the same subgroup the early study by Gattinoni and colleagues [
22], which included almost 93% ARDS patients, and that by Guerin and colleagues [
22], which included only 30% of ARDS patients.
It is also difficult to control for important confounders such as the differences in prone duration, ventilation strategy, or associated treatments. Indeed, studies that included only ARDS patients also implemented lung-protective ventilation and longer prone duration, making it difficult to ascribe the observed reduction in ICU mortality to only one of these variables. Lung-protective ventilation has proved to lessen ventilation-induced lung injury and to reduce mortality, while longer prone duration helps to increase lung recruitment and enhances gas exchange [
23,
24]. Following Gattinoni and colleagues, however, we should admit that a strong physiological rationale underlies the fact that only the most severe forms of ALI (namely patients with ARDS) have physiological conditions for proning efficacy and might derive clinical benefit from prone ventilation [
12]. Patients with ARDS indeed have a higher percentage of potentially recruitable lung, greater amounts of lung oedema, and a small portion of aerated lung [
25]. Our working hypothesis prompting stratification of included studies according to the severity of acute lung injury (ARDS studies versus ALI/ARDS studies) therefore seems the most likely to account for the observed reduction in mortality in the ARDS subgroup.
The fact that the test of interaction yielded only a trend to different mean effect size of prone ventilation in the subgroup of ARDS patients when compared with studies that included all ALI is not surprising given that studies including ALI patients also enrolled a substantial proportion of patients with ARDS. Without specific studies enrolling only ALI non-ARDS patients, this type of effect comparison may be difficult. Apart from a type II statistical error, the nonsignificant test of interaction might also reflect a true lack of heterogeneity of prone ventilation effects. The use of confidence intervals is helpful to solve this uncertainty [
26]. The 95% CI actually represents the range within which the true treatment effect falls 95% of the time. In the subgroup of studies enrolling only ARDS patients, the CI around the point estimate suggests that the reduction of mortality by prone ventilation could not be less than 1%. Similarly, the CI boundaries for the effect of prone ventilation in ALI/ARDS studies do not exclude a reduction by 18% in the mortality in such patients.
Our cumulative meta-analysis shows that beneficial effects of prone ventilation have progressively become apparent as new studies have been published. This finding suggests that the gradual incorporation of research advances (protective lung ventilation, inclusion of homogeneous groups of severity, standardisation of length of proning, and so forth) influenced the trend toward an apparent benefit from prone positioning. This cumulative meta-analysis also shows that the size effect of prone ventilation on mortality has become almost constant since 2006 following the study by Mancebo and colleagues [
16]. Subsequent studies have merely contributed to improve precision of this effect as reflected by a progressive narrowing of the confidence interval. Increased precision rather than substantial alteration in size effect is probably what would be added by any new study on prone ventilation. Furthermore, such a study would be difficult to complete given inclusion barriers encountered by most of the recent RCTs. Meanwhile, the present aggregate meta-analysis and the recent individual patient data meta-analyses provide compelling evidence to recommend prone ventilation in ARDS patients.
Our meta-analysis did not disclose a statistically significant increase in major airway complications of prone positioning. The most recent RCT (Prone-Supine II Study), however - which should be regarded as the most reliable reflection of real-life practice - recorded a higher incidence of adverse events associated with prone positioning [
10]. This concerned not only airway complications but also the need for increased sedation, transient desaturation or hypotension, and displacement of vascular lines. Accordingly, caution should be kept during the manoeuvre - which should be applied only in the most severe patients.
The survival difference between ALI/ARDS studies and only ARDS studies might have additional possible contributors, other than the disease severity. The ALI/ARDS studies are the older studies, characterised by several methodological differences such as the absence of relevant co-treatments (lung-protective mechanical ventilation strategy), other criteria of enrolment (time window between ARDS criteria and enrolment), and so forth. The main difference is that the length of the proning treatment - which may constitute an important determinant of the survival benefit - is profoundly different between older studies (shorter duration) and newer studies (longer duration). Indeed, alveolar recruitment in the prone position is a time-dependent phenomenon [
23]. Our study therefore cannot ascertain whether the enrolment criteria by themselves explain the results, and suggests that proning duration also played a role. We addressed the practical issue of the optimal proning duration by a meta-regression analysis. We found only a trend towards an interaction between longer proning duration and reduction in mortality. The initial studies by Guerin and colleagues [
2] and Gattinoni and colleagues [
1] had the greatest impact on the slope of the regression line. The subgroup of studies including only ARDS patients also applied the longest proning durations (17 to 24 hours/day). Hence, although proning duration seems to play a role in the outcome effect, the present analysis cannot definitely confirm this effect.
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
FA conducted the literature searches, selected studies, extracted data, assessed study quality, prepared initial and subsequent drafts of the manuscript, and integrated comments from other authors into revised versions of the manuscript. LO-B, FD, and IO screened abstracts, selected studies meeting inclusion criteria, extracted data, and assessed study quality. FA and LO-B carried out the statistical analyses with input from IO, FD and LB. LB provided methodological guidance on drafting the manuscript. All authors read and approved the final manuscript.