Higher Versus Lower Oxygenation Strategies in The General Intensive Care Unit Population: A Systematic Review, Meta-Analysis and Meta-Regression of Randomized Controlled Trials.

Oxygen therapy is vital in intensive care unit (ICU) patients, but it is indistinct whether higher or lower oxygen targets are favorable with regards to patient centered outcomes. Our aim was to systematically evaluate the ndings of randomized controlled trials (RTCs) comparing higher and lower oxygen therapy strategies in mechanically ventilated patients. Methods A meta-analysis and meta-regression analysis of RCTs published between 2015 and 2021 was conducted. Electronic databases of MEDLINE, EMBASE, and Web of Science were systematically searched. Publications comparing the effects of higher or lower oxygen therapy in adult ICU patients were eligible. Risk of bias was assessed using the Cochrane Risk of Bias tool. The primary outcome was mortality at day 90; secondary outcomes include serious adverse events (SAE), ICU and hospital length of stay and support-free days. Results We included nine RCTs enrolling 5807 patients. No signicant difference was observed for mortality at day 90. A signicantly lower odds ratio (OR) was found for SAEs, favoring lower oxygenation (OR 0.86 [0.77, 0.96], P=0.009). Our meta-regression analysis shows that the effect sizes for 90-day mortality may be inuenced by both the between-group difference in achieved oxygenation and the severity of hyperoxia in the higher group (ß=-0.0022, P=0.1). In the general ICU population, there was no signicant difference in 90-day mortality between higher and lower oxygenation strategies. However, a lower incidence of serious adverse events was found for the lower oxygenation groups. These ndings may have clinical implications for practice guidelines, yet it remains of paramount importance to continue conducting clinical trials, ideally comparing groups with a clinically relevant contrast for specic patient groups. PROSPERO


Introduction
Oxygen therapy has been successfully used in the acute care setting for over a century (1). Most critically ill patients are at risk for hypoxemia which may cause tissue damage, organ failure or even death. Owing to these risks, the professional norm among health care specialists is to attentively avoid and sometimes even overcompensate hypoxemic events by liberally administering oxygen or deliberately inducing supranormal arterial oxygen levels (2,3).
Oxygen has proven to be very effective in the treatment of hypoxemic patients, but may not be bene cial in all patients. The deleterious properties of oxygen are increasingly acknowledged. Harmful effects can include cerebral and coronary vasoconstriction, reduction of cardiac output, absorption atelectasis, acute lung injury and central nervous system toxicity (4). In addition, studies repeatedly showed a negative correlation between hyperoxia and patient centered outcomes (5)(6)(7). Accordingly, newer guidelines on oxygen therapy generally recommend a more conservative approach (8). However, not all cautions with regards to hyperoxia have been conclusively justi ed as observational studies and randomized controlled trials (RCTs) comparing higher versus lower oxygen targets show heterogeneous results.
Few systematic reviews have been conducted in order to provide guidance for administering oxygen in a safe and e cient manner to intensive care unit (ICU) patients. The results were unequivocal (5,9), but new studies have recently been published (10). Furthermore, important data concerning vital secondary outcomes, such as ischemic events, have not been aggregated in detail before. By systematically combining all available evidence from RCTs comparing higher and lower targeted oxygen strategies in mechanically ventilated patients, we aim to provide conclusive insights into favorable oxygen therapy.

Methods
This study was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines for systematic reviews and meta-analyses (see Additional le 1) (11). The protocol of this study was registered on PROSPERO (CRD42021286372). Studies were included when meeting the following inclusion criteria: RCTs comparing higher (liberal) and lower (conservative) oxygen therapy strategies in the general ICU population of which the majority is mechanically ventilated.

Information sources and search
After consulting a librarian, the electronic databases of Medline , EMBASE (1970EMBASE ( -2021 and Web of Science (1970-2021) were searched. This search was supplemented by manually screened reference lists of relevant studies and reviews. Full search terms and search strategies are presented in Additional le 3 of the Additional les. Main MeSH headings and key words that were used are "oxygen inhalation therapy", "hyperoxia", "hypoxia", "oxygen" and "respiration, arti cial". The literature search was last updated October 20 th 2021.
Study selection and Data collection process Two authors (L.I.W, H.J.F.H) independently screened articles on title and abstract. Differences in this process were resolved by consensus. No language restrictions were applied. Studies were excluded using the following criteria: patients younger than 18 years, animal studies, extracorporeal life support, perioperative settings and hyperbaric oxygen therapy. Studies that solely focused on one speci c patient group (e.g. myocardial or cerebral infarction) were excluded to improve the comparability of the included study population. Duplicates were removed by using the method of Bramer et al (9).

Data analysis and outcomes
Data abstraction was done by two content area experts using an electronical standardized data abstraction sheet. Extraction was reviewed by two review authors independently. Disagreements were resolved by consensus. If no consensus could be reached, a third co-author would resolve the issue. Data was collected as dichotomous and continuous outcome data. The primary outcome of interest was mortality at day 90; and mortality at day 28, day 180 and ICU and hospital mortality were also analyzed. Secondary outcomes were adverse events, support-free days at day 28 and length of stay (LOS).
If data was available on intended primary or secondary outcomes, studies were included in the meta-analysis. Corresponding authors were contacted to clarify important missing data or for further trial details. The Grading Of Recommendation, Assessment, Development and Evaluation (GRADE) approach was used to grade the certainty of evidence of the included papers (12). Heterogeneity between studies and between subgroups was assessed by using either Tau 2 , Chi 2 or I 2 statistics, or a combination of the latter and presented as either a p-value, percentages or both. A low p-value was considered as evidence of heterogeneity of intervention effects. An I 2 of 0 to 40% was considered not important, 30 to 60% was considered moderate, 50 to 90% was considered substantial and 75 to 100% considerable (13).
For individual effect estimates, pooling and formal meta-analysis of the data odds ratios (ORs) were used. Pooled estimates are displayed in forest plots and summarized as OR with a 95% CI. For dichotomous data a random-effects model according to Mantel and Haenszel (M-H) was used; continuous data was analyzed by the use of a random-effects inverse variance model. A random-effects model was chosen due to existing heterogeneity. The effects of hyperoxia by achieved oxygenation in the randomized groups were analyzed using a meta-regression framework (16). For this meta-regression analyses we calculated a combined score in order to assess the effects of the achieved difference in PaO 2 (contrast) between the oxygenation groups in combination with the degree of hyperoxia that was achieved in the higher oxygenation group. Hence, the combined score was calculated as: between group difference in achieved PaO 2 plus the achieved PaO 2 in the highest group. The ORs were based on the 90-day mortality. Our hypothesis was that a higher between-group difference and a more severe hyperoxic target in the higher group, increase the effect size for 90-day mortality. Since Yang et al did not report on 90 day mortality, this study was not included in the meta-regression analysis.
In order to assess bias, the Cochrane risk of bias tool for randomized trials was used (14). A funnel plot was used to graphically display reporting bias, using standard errors of the intervention effect estimate. Funnel plot asymmetry was tested by using Egger's exact test.

Results
Study selection and study characteristics Figure 1 depicts the study owchart. Our search strategy resulted in 1311 studies considered for inclusion after deleting duplicates. In total, 19 full-text articles were assessed for eligibility after title and abstract screening ( Figure 1). The most important exclusion reasons were study types other than RCTs, post hoc analyses or lack of comparison between higher and lower oxygenation . For the nal analysis, nine studies, in a total of 5807 patients, were included (Table   1) (10,(15)(16)(17)(18)(19)(20)(21)(22). In the studies included, data collection took place between 2010 and 2020; the study reports were published between 2015 and 2021. All included studies were RCTs comparing a higher versus a lower oxygenation in mechanically ventilated patients focusing on the general ICU population. Either PaO 2 , SpO 2 , FiO 2 or a combination of these parameters were used to pursue oxygenation targets. The duration of the interventions ranged from 24 hours to 90 days ( Table 1).

Risk of bias in studies
Overall, the risk of bias was moderate to low, except for blinding and early stopping bias (see Figure 2 and 3 and Additional le 2 of Additional les). Due to the design of the trials, it was essentially unfeasible to blind clinicians for the assigned treatment group. However, patients were blinded in the majority of the trials. Also, three trials scored high on early stopping bias. The funnel plot regarding publication bias is displayed in Figure 4. Egger's exact test for funnel plot asymmetry was based on Additional 4 of the Additional les. A z-score of 0.83 and a corresponding p-value of 0.41 was found, suggesting no funnel plot asymmetry.

Primary endpoint
Mortality at day 28, day 90, day 180, in the ICU and in the hospital were assessed separately and are listed in a forest plot ( Figure 5). No effect of different oxygenation strategies was found for either mortality at day 28 (P=0.75), day 90 (P=0.91), day 180 (P=0.70), ICU mortality (P=0.55) or in-hospital mortality (P=0.53).

Secondary endpoints
Adverse events were categorized in the following subgroups: myocardial ischemia, intestinal ischemia, ischemic stroke, respiratory infection, systemic infection, shock, organ failure, renal replacement therapy and arrythmias ( Figure 6 Support-free days were analyzed as ventilator-and vasopressor-free days at day 28 (Figure 7). In general, the support-free days did not show a signi cant difference (OR 0.20 [-0.45, 0.85] P=0.55). Heterogeneity was moderate between all studies (I 2 =33%, P=0.16) and not important between subgroups (I 2 0%, P=0.65).
The analysis of the length of stay (LOS) was categorized according to two subgroups: hospital LOS and ICU LOS (Figure 8). No signi cant differences were observed in the different subgroups.

Sensitivity analysis
All included studies in this meta-analysis were assessed using the GRADE approach (12). In most cases, the certainty of evidence was high to moderate ( Table  1). Despite methodological differences, we mainly observed low heterogeneity and low risk of bias. Due to a variety in the chosen targets of the studies, we performed a meta-regression analysis that compared the odds for lower oxygenation on 90-day mortality for different achieved oxygenation targets. The regression was performed for both the achieved high and low oxygenation groups (see Additional le 5 of Additional les) and for the combined score. The combined score is calculated by combining the achieved difference between the high and low oxygenation targets and the achieved higher oxygenation target ( Figure 9). Figure 9 shows that this combined score is in the same order of magnitude for the majority of the trials (10, 15, 17-19, 21, 22) and the OR of mortality in the lower group approximates 1. One trial (16) shows a combined score of 260 mm Hg in combination with a lower OR for mortality in the lower group. Taken together, the risk of mortality after 90 days in the lower group may be dependent on the combined score, i.e. the combination of the difference of achieved oxygenation and the severity of achieved hyperoxia in the higher group. However, the beta associated with this meta-regression analysis did not reach statistical signi cance.

Discussion
This systematic review and meta-analysis shows no difference in mortality at day 90 when aggregating data from RCTs comparing lower and higher oxygenation targets. For secondary outcomes, a signi cant effect favoring lower oxygenation targets was identi ed with regards to serious adverse events..
The following study strengths and limitations should be considered. First, this meta-analysis includes important secondary patient centered outcomes such as support-free days and serious adverse events. Second, in order to ensure comprehensiveness of the data, corresponding authors were contacted for additional data. Also, established guidelines, such as the PRISMA and GRADE approach, were used to ensure the quality of our methodology and certainty of evidence.
A key limitation is that all trials on oxygenation in ICU patients used different targets. Several varying thresholds were used for PaO 2 targets. Moreover, some trials used SpO 2 targets rather than PaO 2 targets and one study managed liberal oxygenation by applying an FiO 2 of 1.0 (irrespective of SpO 2 ) during the rst 24 hours while not using different SpO 2 or PaO 2 targets afterwards. A PaO 2 of 90 mmHg can correspond with a SpO 2 of 100% but also 93%, partially depending on the underlying disease. Therefore, a higher PaO 2 cannot be consistently translated into a xed SpO 2 (23). As thresholds differ, a patient could be categorized in the higher oxygenation group in one trial, whereas this could be the lower oxygenation group in the other trial.Moreover, in some studies chosen oxygenation targets can overlap (15,18), suggesting there may not be a true comparison between a 'high' and a 'low' group.
Another limitation is the heterogeneity of the ICU population in combination with the heterogenous treatment effect that can be expected from oxygen therapy in certain subgroups. For example, in vasodilatory septic shock, arterial hyperoxia may be bene cial due to antibacterial properties and the counteraction of vasodilation (24,25), while in ischemia and reperfusion injuries, such as myocardial infarction, hyperoxia may have detrimental effects (26). Recent reviews explored the optimal oxygen targets per subgroup by underlying disease (27,28) and no optimal oxygen target per subgroup could be identi ed, though it seems justi able to avoid both hypoxemia and excess hyperoxemia. Hence, the key question remains whether we should settle for a one-strategy-that-ts all (optimal) oxygen therapy approach, or whether optimal oxygen strategies can be applied per subgroup.
None of the included trials were blinded, which can be considered a limitation since the latest literature could have imposed bias towards the bene cial effect of lower oxygenation targets (29,30). Therefore, clinicians may be more prone to adhere to the lower targets, making it more di cult to create contrast in oxygenation between two different groups. However, owing to the design of the trial, it was essentially unfeasible to blind clinicians for the assigned treatment group. Despite lack of blinding and early stopping bias, an overall low risk of bias (Figure 2 and 3) and low heterogeneity was observed with a moderate to high certainty of evidence (Table 1). Taken together, this provides a reasonable degree of certainty that there is no or little evidence for important differences in mortality between compared oxygenation targets.
Our ndings are in line with recent systematic reviews showing that different oxygenation strategies did not have a signi cant impact on mortality (9,31).
However, the ndings are in contrast to previously published reviews (5,32,33), that support a conservative oxygen strategy. A simple explanation for these contradictions might be that patients either simply do not bene t from a lower oxygenation strategy or that the achieved lower and higher PaO 2 in both groups lack su cient contrast to be able to detect a difference. In the included trials it has proven to be di cult to accomplish a clinical contrast between the intervention and the control group. The majority of the trials that reported on the achieved oxygenation show a difference of 10 to 20 mmHg (15,(17)(18)(19)21).
Our sensitivity analysis using a meta-regression framework (Figure 9 and additional les) shows that trials with a smaller achieved difference (10-20 mm Hg) (15,(17)(18)(19)21) and studies with a larger achieved difference (25-70 mm Hg) (10,16,22) both show heterogenous results. It should be noted that achieved differences are in the same order of magnitude for most studies (10-30 mm Hg) despite one outlier (70 mm Hg). When a large difference is achieved there is a sign that patients may bene t from a lower oxygenation target. Though, due to lack of signi cant results, this may also be originated by chance. Furthermore, when speci cally targeting a very high or low target a signi cant clinical difference may be achieved but neither the intervention nor the control group may then represent usual care. Accordingly, the present study may demonstrate that a broad range of less extreme achieved oxygenation falls within a fairly safe category.
The different results amongst included trials can be explained by secondary factors such as early stopping bias, subgroup analysis and not choosing a truly hyperoxic target. Taking all included trials that reported on achieved targets together, an average higher oxygenation around 110 mm Hg was achieved, with an individual maximum of 185 mm Hg (16). The hypothesis that 110 mm Hg is not a truly hyperoxic target is supported by earlier literature that showed a signi cant increase in mortality in the hyperoxic group, where hyperoxia was de ned as PaO 2 > 300 mmHg (26). Our meta-regression analysis ( gure 9) shows that when a hyperoxic target of 185 mm Hg is achieved (16), patients may have a lower risk of mortality in the lower oxygenation group. In line, these results are not signi cant and the more severe the higher target, the less it represents usual care and the higher the chances of mortality.
A new nding in our meta-analysis resulted from studying serious adverse events. We found that adverse events are more likely to occur in the higher oxygenation groups. As in previous studies, serious adverse events should be critically reviewed to evaluate whether the event is consistent with the natural history of the critical illness (34). If a large difference is observed, similar to the difference found in our meta-analysis, it might be attributable to the different interventions. As serious adverse events can highly impair patient health and quality of life the potential negative impact of higher targets may also be a compelling argument to adhere to a lower oxygenation strategy. However, the results on adverse events are dominated by one study (18) and a low number of studies reported on the individual adverse events groups. Even though this nding, concerning adverse events , is an important signal for clinical practice guidelines, more robust data is needed for a compelling conclusion.

Conclusion
In the present meta-analysis comparing higher and lower oxygenation targets we found no difference in 90-day mortality. Importantly, we did nd a signi cant difference in serious adverse events favoring lower oxygenation targets. Differences in methodology, oxygenation targets and primary and secondary endpoints may hamper a comparison of studies, even though the certainty of evidence in the majority of studies was graded relatively high and heterogeneity was generally low. Robust future clinical trials remain of paramount importance, ideally adequately separating the intervention groups based on achieved oxygenation. The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Competing interest
All authors declare to have no competing interests.

Not applicable
Authors' contributions LIW and HH conducted the search, reviewed the articles for eligibility and performed the statistical analysis. LIW wrote the manuscript. All authors contributed to the writing, critical revision of the manuscript and approved the nal version.     Funnel plot for publication bias based on mortality at day 28, day 90, day 180, ICU mortality and hospital mortality.  Forest plot adverse events. Patients from each study are counted once in the test for overall effect. Figure 7