The main finding of our study is that exposure to hyperoxemia after initiation of VA ECMO is significantly associated with almost doubled incidence of mortality or poor neurological outcome, (ORs of 1.81 and 1.97, respectively). Notably, the magnitude and the direction of our findings (suggesting a negative role of exposure to hyperoxemia in VA ECMO patients) were confirmed in the subgroup analyses according both to the indication for starting VA ECMO (CS or eCPR) or to the cut-off used to define hyperoxemia (200 vs. 300 mmHg).
Several considerations strengthen our results: 1) findings were confirmed in most of the sensitivity analyses; 2) the analysis pooled a large amount of data (over 10000 patients for mortality, and over 1300 for the neurological outcome); 3) the post-hoc analysis, including the seven studies using the same timepoint for PaO2 collection (24 hours after initiation of ECMO), showed an even greater OR for mortality (2.35) with lower heterogeneity; 4) most of the included studies reported adjusted ORs, and the analysis conducted using these ORs confirmed the findings of the primary analyses. Further, five VA ECMO studies that we could not include due impossibility to extrapolate data on exposure to hyperoxemia, suggest negative impact of high PaO2. In particular, three of these studies were conducted in the eCPR population. Bonnemain et al[35] found significantly higher mean PaO2 values over the first 24 hours in non survivors as compared to survivors (306 mmHg vs. 164 mmHg); Hong et al[36] found that patients in the highest tertile of mean PaO2 (> 160 mmHg) had significantly higher mortality. Tonna et al[37] found that higher PaO2 values within 24 hours significantly higher mortality (OR 1.45, p < 0.001). Similarly, the other two studies (conducted in unselected VA ECMO populations) showed significant association between higher PaO2 values and mortality[38, 39]. Therefore, despite the GRADE suggests low certainty of evidence due to the observational nature of the studies included, the sample size and the quality of the data suggests robustness of our findings. Hence, even if our results are not derived from randomized controlled trials (RCTs), the findings are supported by the correction for possible confounders and the large sample size. To randomize patients for different levels of oxygenation, including manifest hyperoxemia, would be certainly unethical considering that a not negligible body of literature suggests harmful effects of hyperoxemia in critically ill patients[40–46].
We conducted subgroup analyses according to the cut-off used by the authors for the definition of hyperoxemia. Whilst most studies have used 300 mmHg as cut-off to define severe hyperoxemia [18, 19, 25–29, 33], others have used more conservative threshold (i.e. 200 mmHg[30, 31] ), and sporadically also a higher one (400 mmHg) has been reported [27]. In our study, the association between mortality and exposure to hyperoxemia were confirmed in both subgroup analyses (200 and 300 mmHg), although the vast majority of data (8 studies out of 10) are focused on the higher cut-off. In our opinion, the use of a uniform cut-off for defining severe hyperoxemia (> 300 mmHg) by most of the included studies had the value of reducing the clinical heterogeneity of our findings; however, this leave more uncertain whether a lower degree of hyperoxemia (> 200 mmHg) is truly harmful in the VA ECMO population. Interestingly, a recent analysis on data from a RCT in patients resuscitated after CA (TTM-2 trial), confirmed a U-shape effect of PaO2 levels on mortality; regarding the harmful effects of hyperoxemia, the authors identified 197 mmHg as the best cut-off for predicting an increase in mortality[46], a value very close to the 200 mmHg cut-off used by two studies included in our metanalysis. Moreover, in a population undergoing eCPR, Chang et al. [32] identified the PaO2 range between 77 and 220 mmHg (OR. 2.29) as the best interval for predicting favorable neurological outcome, again suggesting that a threshold around 200 mmHg is likely to be correct as compared to a more conservative definition of hyperoxemia at 300 mmHg. More data would be certainly desirable to better understand the more reliable cut-off that clinicians dealing with VA ECMO should keep in mind to adjust the inspired fraction of oxygen (FiO2) at the gas-blender (also known as FbO2).
We also analyzed subgroups according to the indication for starting rescue VA ECMO support, namely CA or CS. In this case the subgroup analysis was conducted for mortality only, because studies on CS did not report data on neurological outcome. The results of the subgroup analyses point in the same direction, although the association between mortality and exposure to hyperoxemia remained uncertain in the eCPR population (p = 0.07). Our results provide further support to the recent guidelines from the European Resuscitation Council that suggest the avoidance of hyperoxemia after return of spontaneous circulation and a target of oxygen saturation during the post-resuscitation period around 94–98% [45]. Further, another large metanalysis of observational data suggested harmful effects of hyperoxemia in CA patients, both in terms of survival and of neurological outcome [42], rising the importance of adapting ventilatory targets after CA [47].
We believe our meta-analysis is very relevant as it represents the first attempt to meta-analyze the impact of hyperoxemia episodes after VA ECMO cannulation on both mortality and neurological outcome in patients experiencing CA or CS. Our results suggest that clinical efforts should be made to avoid hyperoxemia whilst offering VA ECMO support. Whilst it is difficult to clearly define a threshold of PaO2, it seems reasonable to avoid prolonged exposure to values above 200 mmHg and certainly to keep the PaO2 below 300 mmHg. One of the simple actions that could be practically suggested is to avoid the setting of very high FiO2 at the time of VA ECMO cannulation, unless concomitant severe respiratory failure is present. Two ongoing large RCTs on patients supported by VA ECMO (BLENDER, NCT03841084 [44]; ECMOxy NCT04990349 [48]) are designed to study the role of different targets of post-oxygenator saturations, achieved by modifying the FiO2 on the ECMO blender; however, these studies are not directly targeting high levels of PaO2 and therefore will not study the effect of exposure to hyperoxemia.
Although it is not intention of this manuscript to review the effects of hyperoxemia in critically ill patients, in virtually all trials conducted so far, the comparison between strategies of liberal and conservative oxygenation targets brought conflicting results [49]. The little or no effect seen on clinical outcomes in most trials is likely due to small differences in the oxygenation and/or saturation targets. Our metanalysis was conducted in a selected population of critically ill patients supported by VA ECMO, and it cannot evaluate the best oxygenation targets but it confirms the detrimental effects of exposure to high level of PaO2. Hence, we add knowledge to the growing body of literature suggesting that high levels of PaO2 are harmful on both neurological outcome and survival of critically ill patients. Notably, the harmful effects of hyperoxemia have been also suggested for the pediatric populations of critically ill patients [50] and in those supported by VA ECMO[51] .
Limitations
Our metanalysis has some limitations. First, as we pooled results from observational studies this resulted in a low certainty of evidence as assessed by the GRADE. However, as mentioned, several studies provided OR adjusted for confounding factors and we analyzed a large pool of data; moreover, the sensitivity analyses mostly supported the results obtained by the primary analysis. It should be considered that observational studies on this specific topic probably remain the only available and ethical approach to gather information on whether severe hyperoxemia may be harmful, as it would not be ethically reasonable to randomize patients to severe hyperoxemia. As mentioned, at least five studies that we could not include in our metanalysis have provided additional evidence on the harmful effects of hyperoxemia [35–39].
Second, we investigated the impact of episodes of severe hyperoxemia experienced on patient’s neurological outcome and mortality; unfortunately, we could not evaluate the impact of the duration of exposure to severe hyperoxemia nor if a very early exposure (i.e. within 6 hours) is more harmful than exposure happening at later period. Notably, seven of the included studies referred to the exposure to hyperoxemia according to PaO2 obtained at a single timepoint (24 hours after ECMO cannulation) and the post-hoc analysis of these studies only showed an even greater OR for mortality (2.35).
Third, several factors may act as potential confounders when investigating the influence of hyperoxemia on the outcomes of interest in patients supported by VA ECMO. In this regard, the peripheral cannulation and the underlying cardiac contractility may determine differences in PaO2 according to the site of blood sampling. Not all studies confirmed whether they used the right radial artery for arterial blood gas sampling.
Fourth, as most results were reported as OR rather than as number of events on the entire population, a trial sequential analysis to assess the robustness of our results was not feasible[24].