Study Selection
Our search strategy yielded 5361 citations after de-duplication. We reviewed full texts for 1523 articles for a detailed evaluation and included 85 articles in our qualitative assessment (Figure 1). The final selection of 85 articles included 30 RCTs (35.3%) and 55 observational studies (64.7%). 19,963 patients were enrolled in the included studies. Among the 85 included studies, 9 studies investigated an adjunctive therapy versus control, and thus only the control arm was included in the primary analysis. But we included both arms of the 9 studies in a sensitivity analysis which consisted of 21,328 patients. We also performed a sensitivity analysis of studies reporting hospital mortality. This analysis consisted of 15 RCTs (32.6%) and 31 observational studies (67.4%), totaling 12,168 patients.
Study Characteristics
46 (54.2%) studies used the AECC criteria to diagnose ARDS, 36 studies (42.3%) used Berlin Criteria, 2 (2.3%) studies used either criteria, and 1 study (1.2%) used the Chinese critical care medicine definition. The mean age was reported in 70 studies (82.4%). The mean baseline PaO2/FiO2 ratio was reported in 72 studies (84.7%), mean APACHE II score was reported in 38 studies (44.7%), and mean SOFA score was reported in 32 studies (37.6%). Tidal volume in ml/kg was reported in 56 studies (65.9%), PEEP was reported in 60 studies (70.6%), Pplat was reported in 49 studies (57.7%), and we were able to calculate the driving pressure in 44 studies (51.8%). The use of inhaled vasodilators, NMB, HFOV, PP, and ECMO was reported in 23 (27%), 22 (25.8%), 9 (10.7%), 25 (29.4%), 20 (23.5%) studies, respectively (Table 1). There were no differences in the baseline characteristics, severity of initial illness and use of ventilatory strategies between the RCT’s and observational studies (Table 1). The lack of asymmetry on the Funnel plot and the result of the Egger’s test imply that publication bias did not alter the results (e-Figures 5-6).
Quality assessment
The risk of bias for RCTs was low in twelve studies, moderate in fourteen RCTs, and high in four RCTs (e-Table 2). The high risk of bias was driven mainly by deviation from intended intervention. For observational studies the risk of bias low in two studies, moderate in forty-one studies, and high in twelve studies (e-Table 3). The high and moderate risk of bias was driven mainly by confounding given the nature of study design.
Mortality
Hospital mortality (46 studies) was most commonly reported in the included studies. In the remaining 39 studies it was substituted with ICU mortality (8 studies), 90-day mortality (3 studies), 60-day mortality (9 studies), and 28/30-day mortality (19 studies). The weighted pooled mortality of all 85 studies published from 2009 to 2019 was 38% (95% CI 35,40). Mortality was higher in observational studies [40% (95% CI 37, 42)] compared to RCTs [35% (95% CI 30,39)], (p=0.04) (Figure 2 and Table 1). There was significant heterogeneity among the included studies (I2 = 92.23%, p<0.01). This heterogeneity persisted across both observational studies (I2 = 90.21%, p<0.01) and RCTs (I2 = 93.87%, p<0.01) (Figure 2). Figure 2 depicts the reported mortality from the included studies listed in chronological order according to publication year. The continent where the study was performed failed to show a statistically significant different in mortality as determined by an one-way ANOVA [F(5,79) = 1.5, p=0.20].
The reported mortality did not change based on the sensitivity analysis. The first sensitivity analysis which included both the interventional and control arm of the 9 studies evaluating a particular adjunctive therapy resulted in a pooled mortality of 37% (95% CI 33,40) (e-Figure 7). The weighted pooled mortality was similarly higher in observational studies [40% (95% CI, 37,43)] compared to RCTs [33% (95% CI 29,37)] (p<0.01). The second sensitivity analysis which only included studies reporting hospital mortality (46 studies) resulted in a pooled mortality of 39% (95% CI 36,42) (e-Figure 8). Mortality again remained higher in observational studies [41% (95% CI 37,44)] compared to RCTs [34% (95% CI 28,39)] (p=0.05). The cumulative mortality analysis, which we conducted using the median year of enrollment, displays the evolution of mortality in studies published from 2009 to 2019, excluding one study which did not mention enrollment dates(28) (Figure 3). The median year of enrollment ranges from 2000 to 2018 and on visual inspection the mortality suggests an initial decrease after which it stabilizes around the mean with little change.
Meta-regression
In our meta-regression analysis, the initial PaO2/FiO2 ratio in the included studies was strongly associated with the reported mortality [b coef. -0.0041 (95% CI -.0023, -.0005); p<0.01] (Figure 4). We found significant variability in the reported ventilatory strategies and utilization of adjunctive therapies. The tidal volumes in our systematic review ranged from 5.8 to 8.9 ml/kg with a mean of 7.2 ml/kg. Though not statistically significant, a very clear trend towards a mortality benefit was observed in studies with lower reported tidal volumes [b coef. 0.0337 (95% CI -.0009, .0680); p=0.06], see Figure 5. PEEP ranged from 4.6 to 16.1 cm H2O at the time of enrollment with a mean of 10.2 cm H2O. Reported plateau pressures ranged from 21.0 to 35.1 cm H2O, with a mean of 25.6 cm H2O. Driving pressure ranged from 12.4 to 22.8 with a mean of 15.4. Neither PEEP [b coef. -0.0048 (95% CI -.0200, .0103); p=0.53], Pplat [b coef. 0.0046 (95% CI -.0077, .0169); p=0.45], nor driving pressure [b coef. 0.01382 (95% CI -.0030, .0307); p=0.11] appeared to impact mortality (e-Figures 9-11). Mean age, APACHE II, and SOFA also did not have any impact on mortality in our meta-regression (e-Figures 12-14). Reported adjunctive therapies were significantly variable in the included studies and did not have any impact on the mortality reported in our meta-regression (e-Figures 15-18). In the sensitivity analysis, which only included studies that reported hospital mortality, repeating meta-regression analysis for the same variables mentioned above produced identical results (e-Table 4).