Awake prone positioning in non-intubated patients with acute hypoxemic respiratory failure due to COVID-19: a systematic review and meta-analysis

Background: Awake prone positioning (APP) has been advocated to improve oxygenation and prevent intubations of patients with acute hypoxemic respiratory failure due to coronavirus disease 2019 (COVID-19). This paper aims to synthesize the available evidence on the ecacy of APP. Methods: We performed a systematic review and meta-analysis of observational studies to compare oxygenation parameters in-hospital intubation rate in patients treated with APP or with standard care. Results: A total of 46 published and 4 unpublished observational studies that included 2994 patients were included. APP was associated with signicant improvement of various oxygenation parameters in 19 studies (n=381) that reported this outcome. The intubation rate was 27% (95%CI, 19 to 37%) in the 870 patients treated with APP, as compared to 30% (95%CI, 20 to 42%) in the 852 patients treated with usual care (p=0.71). Conclusions: On the basis of the available evidence, it is not possible to demonstrate ecacy of APP for patients with COVID-19 acute respiratory failure, as assessed by the need for invasive ventilation. Routine implementation of APP outside of a clinical trial is not supported by current evidence. Randomized controlled clinical studies are urgently needed to denitively assess the utility of APP in these patients. Fifteen studies reported patients’ tolerability to APP, varying from 47% to 100%[9,10,12,14,28–30,32,33,35,36,40,43,45,46]. Eight papers reported on patient's discomfort while in prone position[9,12,14,32,36,40,43,45], including back pain, dyspnea, and general discomfort. The daily duration of APP was reported in 17 papers (n=366)[8– 10,12,14,28,29,31,34,36,38,40,41,45–47]. In 9 papers (n=201), patients tolerated APP for less than 4 hours daily[8– 10,14,28,31,38,40,41]. A single paper reported on a cohort of 55 patients who were able to achieve APP for more than 16 hours daily[47].


Introduction
The coronavirus disease 2019 (COVID- 19) pandemic has led to a sudden surge of hospital admissions for acute hypoxemic respiratory failure (AHRF). A signi cant proportion of patients who are hospitalized for COVID-19 ful ll the criteria for the acute respiratory distress syndrome (ARDS) [1], and require prolonged mechanical ventilation.
In the United Kingdom, among patients who were admitted to the intensive care unit (ICU), 72.1% received advanced respiratory support, de ned as mechanical ventilation (MV), or extracorporeal membrane oxygenation (ECMO), or endotracheal intubation [2], during their stay, with a median duration of MV of 13 days, and mortality rate of 48%. Similar outcomes were reported elsewhere [3][4][5].
This surge of severely ill patients, who required high levels of care for protracted periods of time, threatened to overwhelm healthcare systems worldwide, especially in those regions with lower ICU capacity [6]. It is in this unusual setting in which both ICU beds and ventilators were in short supply, that clinicians turned to awake prone positioning (APP) as a means of avoiding --or, at least, delaying --endotracheal intubation, and thus lessening the pressure on the overburdened ICUs [7-support, the details of APP intervention, tolerability and outcomes. Any disagreements were resolved by consensus in the presence of all four investigators.
If the outcomes of intubation rate and mortality were not reported, or if it was not clear whether the patients received APP and for what duration, the corresponding authors were contacted for clari cations.
To enlarge the sample size, unpublished data provided by the investigators' institutions (BM, JJ, WZ, DR) was also included in the meta-analysis. Ethical approval was obtained at each institution prior to data collection.
Pre-planned statistical analyses The main outcomes are and the proportion of physiological "responders" to APP and the in-hospital intubation rate. In conformity with established custom in the ARDS literature, responders were de ned by an increase of PaO 2 /FiO 2 ratio ≥ 20% [21]. When the PaO 2 /FiO 2 was not reported, an increase of SpO 2 /FiO 2 ratio ≥ 20% was considered as a response, given the linear relationship between the two ratios [22,23]. In-hospital mortality rate is a complex outcome that is modulated by multiple individual and population-level confounders, and is the reported as an exploratory secondary outcome.
For dichotomous outcomes, we pooled proportions using a logit transformation with 95% con dence intervals (CI). We assessed statistical heterogeneity by visual inspection of the forest plots and by calculating the Q and I² statistics, which were interpreted according to conventional thresholds. For all analyses, we implemented random-effects models with inverse variance weighting, providing that at least three studies were available. Potential sources of heterogeneity or inconsistency include baseline disease severity in terms of PaO 2 /FiO 2 at the initiation of therapy, duration of APP, the timing of APP initiation, and the type of respiratory support (conventional oxygen therapy, HFNC, NIV). We investigated the distributions of these characteristics across groups and studies.
We pre-speci ed 3 characteristics in the protocol to be subject to subgroup analyses on the probability of intubation and mortality. When the information was available we restrained the analysis to the studies with PaO 2 /FiO 2 <150 mmHg vs. ≥ 150 mmHg and according to respiratory support devices (HFNC vs. CPAP/NIV). The cut-off value of PaO 2 /FiO 2 <150 mmHg was based on the previously described signal of survival bene t when these patients are managed with early intubation, as compared to a non-invasive strategy [24]. The third subgroup analysis was restrained to studies in the group of APP, in which we analyzed the relationship between APP duration and the probability of intubation and mortality. For multiple comparisons, the Bonferroni correction was used to establish the threshold of statistical signi cance, thus all pvalues were compared to the threshold of 0.0125. We also performed a sensitivity analysis by excluding all data obtained from unpublished sources.
We did not formally assess bias of included studies, as all of them were observational, and inherently highly biased. We did not produce a funnel plot, as this method is inaccurate for meta-analyses of proportion studies [25].

Post hoc comparator groups
While collecting data, and before carrying out any analyses, we realized that only a minority of identi ed papers reported on "pure" populations in which either all patients were subjected to APP, or none were. We therefore decided to group patients into three groups a priori: (1) "APP" when all patients were proned, (2) "some APP" when some (at least 10%) but not all patients were proned, and (3) "no APP" when no patients were proned (less than 10%). Papers that focused on APP were classi ed as APP, regardless of the number or proportion of patients that were able to remain in PP. We compared patients treated with APP (group 1) with those not treated with APP (group 3), and we reported the p-value associated with the test for subgroup differences between group 1 and group 3. We also performed sensitivity analyses including group 2 as a separate subgroup in the analyses. While some patients were treated with conventional oxygen therapy in the "APP" group, only HFNC, CPAP and NIV were used in the "no APP". We therefore performed a sensitivity analysis by excluding patients treated with conventional oxygen therapy. We elected not to exclude these patients from our primary analysis, as the use of HFNC and CPAP or NIV was often prohibited early in the pandemic due to disease transmission considerations, and patients treated with conventional oxygen only were not necessarily less sick than those treated with other modalities.
Finally, in order to test for geographical variation of care, we performed sensitivity analyses with restriction to high-or low-GDP countries.
All analyses were performed in R version 3.6.3 [26], with the help of meta package [27].

Results
Our search strategy identi ed 173 publications on the subject of APP ( Figure E1 in the online supplement), and 271 papers on the subject of non-invasive oxygenation modalities ( Figure E2 in the online supplement) in severe COVID-19. Thus, a total of 444 potentially relevant publications were identi ed, and 440 were screened for inclusion after removal of duplicates ( Figure 1). After full-text review, 46 published studies [8][9][10]12,14, and data from 4 unpublished datasets were included in the nal review, with a combined 2994 subjects: 921 patients treated with APP, 870 patients treated without APP, and a group of 1203 patients in whom a signi cant proportion were treated with APP ( Figure 1, Table 1, and Tables E1 and E2 in the online supplement). Clari cations and supplemental data were obtained from 18 corresponding authors.
Physiological response to awake prone positioning.
Subgroup analyses, with strati cation according to the duration of APP (<4h daily vs ≥4h daily), the device (HFNC vs CPAP vs NIV), and the severity of the ARDS (PaO 2 /FiO 2 <150 mmHg vs PaO 2 /FiO 2 ≥150 mmHg) did not demonstrate any signi cant difference in intubation rate between patients who were treated with APP and those who were not ( Figure 3). There was a very high level of heterogeneity across all studies. Sensitivity analyses with the inclusion of the subgroup of patients with some exposure to APP, and with exclusion of unpublished data demonstrated a similar lack of bene t ( Figures E3 through E11), as did the analyses according to country GDP ( Figures E23 and E24).

Discussion
Our systematic review and meta-analysis demonstrated that APP improved oxygenation but did not change the frequency of intubation or mortality in patients with AHRF secondary to COVID-19. The main strength of our study was the large sample size, with a total of 921 patients treated with APP. However, given the observational nature of included studies, high heterogeneity, and broad con dence intervals, our results should be interpreted as demonstrating absence of evidence, rather than high quality evidence of an absence of bene t.
We found that APP improved oxygenation parameters, and this improvement was sustained even after the patients returned to the supine position in three studies [42,45,46]. APP was also associated with reduced respiratory rate, and good tolerability was reported with the use of various modalities of respiratory support, including conventional oxygen therapy, HFNC, and CPAP or and NIV that was delivered through either a helmet or full face mask. Improvement in oxygenation with APP can be explained by the correction of ventilation/perfusion mismatch [11], better lung recruitment, and reduction of alveolar strain [17]. However, improvements in oxygenation do not guarantee better clinical outcomes. For instance, improvements of PaO 2 /FiO 2 ratio do not correlate with mortality in intubated patients subjected to prone positioning [69].
More physiological and clinical studies are needed to delineate the relationship between improvement of oxygenation parameters and clinical outcomes in patients with COVID-19.
Contrary to previous reports [19,34], we did not nd that APP reduced intubation rates. Several reasons can be advanced to explain this lack of e cacy. First, intubation criteria were not uniformly de ned across studies, and involved the treating physician's subjective judgment. During the pandemic, the recommended respiratory support strategies evolved from early aggressive intubation to strategies of respiratory support designed to prevent intubation [8,55,[70][71][72]. Second, the timing of APP initiation, either as an "adjunctive" (early) or "salvage" (late) therapy may in uence intubation rate. The use of APP at an early stage (PaO 2 /FiO 2 ratio >150mmHg) may be better tolerated, result in better oxygenation, and protect patients from self-induced lung injury (SILI), and thus prevent further disease progression [73,74]. However, in our meta-analysis, we did not detect a signal of bene t of APP in the subgroup of patients with PaO 2 /FiO 2 ratio >150mmHg. Third, the duration of APP might have a dose-response relationship, and it is possible that a reduction in the rate of intubations could be seen only in patients who were subjected to longer periods of APP. Our subgroup analyses did not demonstrate signi cantly lower intubation rates for patients who remained in PP for longer periods of time, but it could be argued that our analysis was underpowered, as only two studies (n=65) reported daily APP periods >16h [12,47]. Fourth, intubation might be inevitable as the disease progresses, despite initial and sustained improvement in oxygenation. It has been argued that intubation rates are lower in patients who experience sustained improvement in oxygenation after APP, the so-called "responders" [46]. However, this nding has not been replicated in other retrospective studies [14], and could be the result of simple reverse causality, with patients "responding" to APP because of their already favorable clinical course. Finally, an unknown proportion of patients with do-not-intubate orders were included in both groups, which could have diluted any possible bene t of APP.
We did not demonstrate a signal of reduced mortality with APP. Given the complex relationship between disease severity, individual co-morbidities, socio-economic status, and variable access to quality care during a pandemic, this nding should be interpreted as exploratory. Due to the retrospective nature of included studies, selection biases are very likely. The type of respiratory support (conventional oxygen therapy, HFNC, CPAP/NIV delivered through a conventional mask vs a helmet) was not balanced between patients treated with APP and those who were not. Analyses with strati cation by the type of respiratory support device did not demonstrate signi cant subgroup differences in mortality. These subgroup analyses were severely limited by the fact that we had access only to overall group statistics, not individual patient data, and a proportion of patients were treated with various devices through the course of their disease.
The use of APP has been widely promoted as the new standard of care, and even included as such clinical care guidelines [76], on the basis of physiological plausibility and low quality evidence from case reports. It has been suggested that studies of this intervention are not warranted and that investigators lack equipoise. Our meta-analysis of available low quality data does not support this unbridled enthusiasm, and we strongly caution against uncritical inclusion of unproven therapies in routine patient care. The critical care literature is replete with examples of interventions that improved surrogate outcomes, yet ultimately failed to improve clinical outcomes in randomized controlled trials [77][78][79]. Implementing an ineffective therapy is not benign, as, at best, this diverts nite clinical resources and attention. At worst, it could result in unnecessarily delayed intubations in di cult conditions, self-induced lung injury, and potentially higher mortality [80]. Unfortunately, there are numerous precedents of physiologically sound therapies that unexpectedly resulted in higher mortality, such as aggressive uid resuscitation [81] or liberal red-cell transfusions [82].

Limitations
Our study has several signi cant limitations. First, data were available only from a group of relatively heterogeneous observational studies. Signi cant levels of inclusion bias are also likely to be present. Without individual patient data, we could not account for the many uncontrolled differences between patients treated with APP, and those who received usual care. Some patients were subjected to APP in extremis after failing usual care, and could have been sicker than patients included in cohorts without APP. Conversely, in other reports, only patients who could self-prone were treated with APP, and these were likely less sick than those in the control group. Second, a variety of respiratory support devices, including helmet CPAPs, were used in both groups. It is not known whether the choice of the device has an impact on outcomes in patients with severe COVID-19. Third, outcomes were highly heterogeneous, which likely re ects populations with various disease severities, various co-morbid conditions, as well as geographical variations of care for patients with ARDS [75]. Fourth, we included unpublished, non-peer-reviewed data. However, our ndings remained robust with the exclusion of unpublished data. Fifth, the mortality rate in our studies is lower than reported in other large cohorts [2,4,5], which suggests selection and publications bias, which would be expected to be in favor of APP. Sixth, we were not able to control for the use of evidence-based treatments such as corticosteroids. However, all included reports nished enrolment before the bene t of corticosteroid was demonstrated [83] and when their use was indeed actively discouraged. Seventh, only a minority of patients were able to tolerate longer periods of APP, and it can be argued that the duration of APP was not su cient to generate a clinically meaningful change in outcomes. However, a physiologically effective, but clinically intolerable intervention would remain ineffective overall. Lastly, data for other important outcomes, such the number of ventilator-free days or the length of ICU stay, were not available for analysis.

Conclusions
In summary, available evidence from observational studies suggests that awake prone positioning improves oxygenation, but these improvements do not appear to translate into reduced rates of intubation. We did not nd any obvious signals of harm, and we did not see any worrisome signal in mortality.
The high selectivity of patients, the inconsistency in the application of prone positioning in published reports and the heterogeneity of outcomes emphasizes the need for randomized controlled trials, as a clinically signi cant bene t cannot be excluded based on available low-quality data.

Declarations
Ethics approval and consent to participate All published and unpublished reports included in the meta-analysis have undergone appropriate ethical approval.

Consent for publication
Not applicable Availability of data and materials Not applicable Table   Due to technical limitations, table 1 xlsx is only available as a download in the Supplemental Files section. Table 1. Summary of included studies of patients severe COVID treating with, and without the use of awake prone positioning.
Abbreviations, in order of appearance: P/F, PaO 2 /FiO 2 , i.e., the ratio of arterial oxygen partial pressure (PaO 2 in mmHg) to fractional inspired oxygen (FiO 2 ); PaCO 2 , arterial carbon dioxide partial pressure; SpO 2 , peripheral oxygen saturation; ED,   Association between awake prone positioning and intubation, in each report, and overall Association between awake prone positioning and intubation, within subgroups de ned by the du-ration of proning, the type of respiratory support device, and the PaO2/FiO2 ratio at enrolment.

Figure 4
Association between awake prone positioning and mortality, in each report, and overall. Association between awake prone positioning and mortality, within subgroups de ned by the duration of proning, the type of respiratory support device, and the PaO2/FiO2 at enrolment.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.