Arterial hyperoxia and mortality in critically ill patients: a systematic review and meta-analysis

Introduction The safety of arterial hyperoxia is under increasing scrutiny. We performed a systematic review of the literature to determine whether any association exists between arterial hyperoxia and mortality in critically ill patient subsets. Methods Medline, Thomson Reuters Web of Science and Scopus databases were searched from inception to June 2014. Observational or interventional studies evaluating the relationship between hyperoxia (defined as a supranormal arterial O2 tension) and mortality in adult intensive care unit (ICU) patients were included. Studies primarily involving patients with exacerbations of chronic pulmonary disease, acute lung injury and perioperative administration were excluded. Adjusted odds ratio (OR) of patients exposed versus those not exposed to hyperoxia were extracted, if available. Alternatively, unadjusted outcome data were recorded. Data on patients, study characteristics and the criteria used for defining hyperoxia exposure were also extracted. Random-effects models were used for quantitative synthesis of the data, with a primary outcome of hospital mortality. Results In total 17 studies (16 observational, 1 prospective before-after) were identified in different patient categories: mechanically ventilated ICU (number of studies (k) = 4, number of participants (n) = 189,143), post-cardiac arrest (k = 6, n = 19,144), stroke (k = 2, n = 5,537), and traumatic brain injury (k = 5, n = 7,488). Different criteria were used to define hyperoxia in terms of PaO2 value (first, highest, worst, mean), time of assessment and predetermined cutoffs. Data from studies on ICU patients were not pooled because of extreme heterogeneity (inconsistency (I2) 96.73%). Hyperoxia was associated with increased mortality in post-cardiac arrest patients (OR = 1.42 (1.04 to 1.92) I2 67.73%) stroke (OR = 1.23 (1.06 to 1.43) I2 0%) and traumatic brain injury (OR = 1.41 (1.03 to 1.94) I2 64.54%). However, these results are limited by significant heterogeneity between studies. Conclusions Hyperoxia may be associated with increased mortality in patients with stroke, traumatic brain injury and those resuscitated from cardiac arrest. However, these results are limited by the high heterogeneity of the included studies. Electronic supplementary material The online version of this article (doi:10.1186/s13054-014-0711-x) contains supplementary material, which is available to authorized users.


Introduction
Oxygen (O 2 ) administration is the most widely prescribed therapy in critically ill patients and frequently represents a life-saving intervention. Since hypoxemia is generally viewed as deleterious and moderate levels of arterial hyperoxia as benign, health care practitioners are more likely to accept supranormal arterial O 2 levels as this provides a wider safety buffer [1,2].
The use of supplemental O 2 in various medical emergencies is supported by many guidelines [3][4][5]. One hundred percent O 2 is commonly administered during cardiopulmonary resuscitation from cardiac arrest [6]. Normobaric hyperoxia is touted as a potential therapeutic strategy for patients with traumatic brain injury or stroke [7], with an underlying rationale of increased brain O 2 delivery [8] and protection of the ischemic penumbra through inducing redistribution of blood from normal to ischemic areas [9]. However, these potential benefits must be weighed against the injurious effects of high-dose supplemental O 2 . In both animal and human studies there are reports of pulmonary toxicity [10][11][12], increased vasoconstriction with a fall in cardiac output [13], free radical-mediated damage to various organs [14], and a marked reduction in coronary blood flow and myocardial O 2 consumption [15].
Clinical data regarding the relationship between arterial hyperoxia and outcome are contradictory. An association between arterial hyperoxia and mortality has been reported in disparate patient populations (mechanically ventilated [16], post-cardiac arrest [17], traumatic brain injury [18], stroke [19]) but not confirmed by other studies [20][21][22][23]. Therefore, the question whether exposure to supranormal arterial O 2 tensions (PaO 2 ) is safe in critically ill patients remains unanswered. We thus performed a systematic review and meta-analysis of studies describing the relationship between arterial hyperoxia and mortality in critically ill patients.

Materials and methods
This report adheres to the Preferred Reporting Items for Systematic reviews and Meta-Analysis (PRISMA) standards for reporting systematic review and meta-analysis studies [24].

Eligibility criteria
Observational (prospective or retrospective cohort or case-control studies) or randomized controlled trials (RCTs) investigating the relationship between arterial hyperoxia and mortality in critically ill patients were eligible for inclusion. Participants were required to be adult patients admitted to a critical care unit for any reason. We excluded studies involving patients with an acute exacerbation of chronic obstructive pulmonary disease (COPD) or acute lung injury (ALI). As hypoxemia is the main problem in these patients, studies on the impact of hyperoxemia were likely to be uncommon. We expected that studies on patients with COPD/ALI may have explored the effects of excessive O 2 flow rather than those of arterial hyperoxemia. In addition, in these patients the PaO 2 range defined as normal/acceptable could have been lower than that applied in patients with preserved respiratory function. Studies on surgical patients were excluded unless they were exposed to hyperoxia during a post-operative admission to a critical care unit. Hyperoxia was defined by the measurement of supranormal values of arterial partial O 2 pressure (PaO 2 ). For defining a condition of exposure to hyperoxia, any cutoff value of PaO 2 and time of assessment were deemed acceptable. Studies where patients were defined as 'hyperoxic' solely on the basis of exposure to a predetermined increase in inspired O 2 fraction (FiO 2 ) were excluded if not guided by any assessment of PaO 2 . Patients not exposed to hyperoxia constituted the comparator group. The primary outcome of interest was hospital mortality from any cause.

Search strategy
Studies were identified by searching Medline (PubMed), Scopus and Thomson Reuters Web of Science databases from their inception. The main search was run on 28 March 2014 and updated weekly until June 2014. The keywords 'hyperoxia', 'hyperoxemia', 'arterial oxygen', 'oxygen saturation', 'critically ill', 'acutely ill', 'intensive care', 'critical care', 'mechanically ventilated', 'cardiac arrest', 'cardiopulmonary resuscitation', 'traumatic brain injury', 'head trauma', 'stroke', 'sepsis', 'septic shock', 'trauma', 'post-operative', 'post-surgery', 'cardiac failure', 'heart failure', 'myocardial infarction', 'shock', 'mortality', 'survival', 'death', 'outcome' were typed in various combinations using Boolean operators. Queries were limited to those involving human subjects. The detailed search strategy applied to Medline (PubMed), and adapted for the other databases, is described in Additional file 1. Hand searches of reference lists of articles and relevant literature reviews were used to complement the computer search. Content pages of the main critical care medicine journals were handsearched to find any relevant in-press articles. The search was not limited by language, but focused solely on articles published in peer-reviewed journals to enhance the methodological rigor of the studies examined and the conclusions drawn.

Study selection
Two independent reviewers (ED and EA) screened all identified records (title and abstract) and performed the eligibility assessment of the selected full-text articles in an unblinded standardized manner. Disagreements were resolved through discussion or arbitration by a third reviewer (AD). Interobserver agreement was assessed by kappa statistics.

Data extraction
Two independent investigators (ED and EA) extracted descriptive, methodological and outcome data from all eligible studies using a predefined data extraction form. Disagreements were resolved through consensus. The datasheet included study design (RCT, retrospective or prospective observational study, multicenter or singlecenter), country, years of enrollment, publication year, primary endpoint, sample size, mean age (as a continuous variable), gender distribution (as a percentage of males), category of critical illness (mechanically ventilated, post-cardiac arrest, traumatic brain injury, stroke, other), criteria for the definition of hyperoxia exposure (time of assessment, selection of the first/highest/worst/ mean PaO 2 , cutoff value to define hyperoxia), prevalence of hyperoxia, overall in-hospital mortality, and prevalence of chronic cardiovascular and/or respiratory disease. Additional data were extracted for studies on post-cardiac arrest patients: average delay to return of spontaneous circulation (minutes); prevalence of initial shockable rhythm (ventricular tachycardia/fibrillation); prevalence of out-of-hospital cardiac arrest; prevalence of therapeutic hypothermia. Unadjusted outcome data (number of survivors and nonsurvivors to hospital discharge in hyperoxic and nonhyperoxic patients) and adjusted odds ratio (OR) (95% confidence interval) describing the association between hyperoxia exposure and mortality were extracted for calculation of effect size (ES). In-hospital mortality was chosen as the primary endpoint of our analysis since it was the outcome measure most frequently reported. When in-hospital mortality was not stated, we extracted data reporting the longest-term mortality available. The study authors were contacted to request additional information whenever a study did not report data necessary for calculation of the ES.

Study quality assessment
The Newcastle-Ottawa Scale (NOS) for cohort studies was used to assess the quality of the included studies [25]. The item 'representativeness of the exposed cohort' was fulfilled if ≤10% of patients had been excluded because of missing data. The item 'completeness of followup' was fulfilled if ≤10% of patients had been excluded because of missing mortality data. For assessment of comparability of cohorts, two confounding factors were defined a priori: illness severity (as defined by any of the following severity scores: Acute Physiology and Chronic Health Evaluation (APACHE), Simplified Acute Physiology Score (SAPS), Sequential Organ Failure Assessment (SOFA), Injury Severity Score (ISS)) and FiO 2 . A study was considered to adequately control for each of these factors if it either demonstrated balance between groups for the confounder, or adjusted for it in the statistical analysis.

Statistical analysis
Data were synthesized using meta-analytic methods [26,27], and statistically pooled by the standard metaanalysis approach, that is studies were weighted by the inverse of the sampling variance. Whenever possible, calculation of the ES of individual studies was based on the adjusted OR of the association between hyperoxia exposure and mortality as compared to nonexposure. When the authors reported results from more than one multivariate model, we extracted data deriving from either the model that included the maximum number of covariates, or the model that included severity of illness and FiO 2 . If the authors analyzed the unadjusted or adjusted association between mortality and increasing quartiles/quintiles/deciles of PaO 2 , we considered patients as 'hyperoxic' if they fell in the upper stratum. When adjusted data were not reported or PaO 2 was analyzed as a continuous variable, unadjusted ORs were reconstructed from binary raw data (number of survivors/nonsurvivors in hyperoxia exposed/not exposed). When normoxia and hypoxemia were considered as two separate categories, only normoxic patients were included in the comparator group. Overall ES was expressed as OR and its corresponding 95% confidence interval (CI). The DerSimonian and Laird random effects model was used as a conservative approach to account for different sources of variation among studies. Forest plots were constructed to graphically represent the results. Q statistics were used to assess heterogeneity among studies. A significant Q value indicates a lack of homogeneity of findings among studies [26,27]. Inconsistency analysis (I 2 ) statistics were then used to quantify the proportion of observed inconsistency across study results not explained by chance [28]. I 2 values of <25%, 50% and >75% represent low, moderate and high inconsistency, respectively [28]. Sensitivity analyses were planned a priori to assess the impact of potential outliers (based on statistical significance of the standardized residuals) and sources of heterogeneity. Several variables were identified and their effects on outcome examined. Categorical variables were treated as moderators and the strength of the association between hyperoxia and mortality assessed and compared across subgroups formed by these moderators. Continuous variables were examined as covariates using random effects meta-regression. Meta-regression was performed to assess the effect of study quality (NOS score) on the calculated estimate. The presence of publication bias was investigated through funnel plots both visually and formally by trim and fill analysis and Eggers's linear regression method [29]. A P value less than 0.05 was used to indicate statistical significance. All analyses were conducted using a computer software package (ProMeta Version 2, Internovi, Cesena FC, Italy).

Results
From the 2,389 articles that were initially identified ( Figure 1), 70 potentially relevant original articles were examined in full text (κ = 0.87 (95% CI, 0.85 to 0.90)). Seventeen studies eventually met our inclusion criteria (κ = 0.91 (0.88 to 0.94)). The study by Kilgannon et al. [30] was excluded as it was a subgroup analysis of the same patient population previously described by the same group [17]. For the study by Ihle et al. [31] we extrapolated outcome data related to the years 2010 to 2011, as the authors used the same database as Bellomo et al. [21] with study populations overlapping for the years 2007 to 2009. Two studies were identified in which hyperoxia exposure was defined on the basis of a peripheral O 2 saturation (SpO 2 ) >98% [2,32]; although it is questionable whether these patients were really hyperoxic to a significant degree, we decided to include these studies in the analysis as the reported time-weighted PaO 2 values were above the upper normal limit of 100 mmHg in both cases [2,32].

Mechanically ventilated ICU patients
The four studies including general populations of mechanically ventilated ICU patients (number of participants (n) = 189,143) were highly heterogeneous in terms of design, outcome measure, criteria for defining hyperoxia exposure, and statistical method applied for analysis (Table 1). de Jonge et al. [16] defined the worst PaO 2 as the one associated with the lowest PaO 2 /FiO 2 . Conversely, Eastwood et al. [20] defined the worst PaO 2 in patients with an FiO 2 ≥0.5 as that associated with the ABG providing the highest arterial-alveolar (A-a) gradient; for patients with an FiO 2 <0.5, the lowest PaO 2 was recorded. If arterial blood gases (ABGs) were taken in patients in whom FiO 2 <0.5 and ≥0.5 were both recorded during the first 24 hours, the value of PaO 2 taken when the FiO 2 ≥0.5 was used. In one study [2] hyperoxia was defined on the basis of a time-weighted SpO 2 >98%;     this parameter was calculated as the mean SpO 2 between consecutive time points multiplied by the period of time between these time points, with the sum of such time-weighted values being divided by total time to obtain a time-weighted average. For the before-after study by Suzuki et al. [32], we considered the 'conservative' group as not exposed to arterial hyperoxia and the 'conventional' group as exposed to arterial hyperoxia (showing a time-weighted average SpO 2 of 98.4%). Extreme heterogeneity was found among the study findings (Q (3) = 91.85, P <0.001; I 2 = 96.73), with an ES ranging from 0.73 to 2.86. Individual ES values (95% CI) are shown in Figure 2. A pooled estimate was not calculated in view of the insufficient homogeneity.

Patients resuscitated from cardiac arrest
Six retrospective cohort studies (three multicenter [17,21,31], three single-center [35][36][37]) evaluated patients resuscitated from cardiac arrest (n = 19,144). Different PaO 2 measures were used to define hyperoxia exposure (Table 1). Most studies used a PaO 2 cutoff of 300 mmHg [17,21,31,37]. In the study by Janz et al. [35], the relationship between mortality and PaO 2 as a continuous variable was evaluated by multivariable regression analysis; to limit heterogeneity in both hyperoxia definition and analysis between this and the other studies, we reconstructed the unadjusted binary OR by stratifying patients between hyperoxic and nonhyperoxic groups based on a 300 mmHg PaO 2 cutoff and extracting the number of survivors/nonsurvivors in the two groups. The study by Lee et al. [36] was not included in the quantitative synthesis because of the substantial differences found in comparison to the other reported studies (different criteria for defining hyperoxia).

). A funnel plot indicates no obvious publication bias (Additional file 3).
Results of the moderator analyses are shown in Table 3. A significant overall ES was found among multicenter (k = 3) but not single-center (k = 2) studies. A significant association with mortality was indicated only in the study that used the first available PaO 2 . An association of borderline statistical significance was shown only for studies in which adjusted data were used. Meta-regression analyses showed a progressively weaker association with increasing prevalence of chronic cardiovascular disease (k = 4, range 12 to 36%). No significant moderator effect was shown by the following variables: mean age (k = 5; range 60.5 to 64 years); gender (k = 4; range 54 to 80% of males); hospital mortality (k = 5; range 32.4 to 66%); average delay to return of spontaneous circulation (k = 3; range 15 to 25.7 minutes); prevalence of initial shockable rhythm (ventricular tachycardia/fibrillation, k = 3; range 40 to 100%); prevalence of out-of-hospital cardiac arrest (k = 5; range 43 to 100%). Study quality as defined by the NOS score had no effect on the ES.

Patients with stroke
Two multicenter retrospective cohort studies [19,23] evaluated the relationship between in-hospital mortality and exposure to arterial hyperoxia in the first 24 hours of ICU admission in patients with stroke (n = 5,537). Rincon et al. [19] defined patients with the first PaO 2 ≥300 mmHg as being exposed to hyperoxia. Young et al. [23] evaluated the independent association between mortality and deciles of PaO 2 , with the upper decile (PaO 2 >341 mmHg) used as the reference category. To make the two studies comparable, for [23] we considered patients in the upper decile as those being exposed to hyperoxia and reconstructed the unadjusted OR for mortality; patients in the first decile (PaO 2 ≤69 mmHg) were excluded. The pooled ES indicates an association between hyperoxia exposure and increased hospital mortality (OR = 1.23 (95% CI 1.06 to 1.43), P = 0.005; Q (1) = 0.04, P = 0.844, I 2 = 0) ( Figure 4).

Patients with traumatic brain injury
Four multicenter [18,22,34,38] and one single-center [33] retrospective cohort studies evaluated patients with traumatic brain injury (n = 7,488). Different criteria were used to define hyperoxia in terms of time of assessment, PaO 2 selection and cutoff value used (Table 1). All studies reported the adjusted ORs for hospital mortality ( Table 2).
The pooled ES indicates an association between hyperoxia exposure and increased mortality (OR = 1.41 (95% CI 1.03 to 1.94), P = 0.032) in the presence of significant  Table 4, with studies stratified based on design, PaO 2 value and cutoff used for defining hyperoxia, time of PaO 2 assessment and comparator group. Study quality as defined by the NOS score, mean age and gender did not influence the ES.

Discussion
The main findings of our systematic review and metaanalysis may be summarized in three points. First, the majority of studies that explored the relationship between arterial hyperoxia and mortality were retrospective observational investigations, with only one interventional prospective before-after study comparing a conventional ventilation strategy against a more conservative strategy. RCTs comparing two different PaO 2 /SpO 2 -targeted ventilation strategies are lacking. Second, high heterogeneity was found between studies in the criteria used for defining hyperoxia exposure (PaO 2 value and cutoff selected, time of assessment) and the statistical method applied for analysis; this  limits comparison between the study results. Third, while studies in general populations of ICU patients gave highly inconsistent results, an association between arterial hyperoxia and increased hospital mortality was found in critically ill patient subsets (post-cardiac arrest, stroke, traumatic brain injury). Four studies evaluated general populations of mechanically ventilated patients [2,16,20,32] and gave highly inconsistent results. Two large multicenter retrospective studies [16,20] reported a U-shaped relationship between PaO 2 levels and mortality by unadjusted analysis. After adjusting for potential confounders including severity of illness, the association between higher PaO 2 levels and mortality was confirmed only by de Jonge et al. [16], while Eastwood et al. [20] showed a protective effect of hyperoxia. Differences in the methods applied for the analysis (PaO 2 stratified in quintiles/deciles, different reference categories) make these studies difficult to compare. A pilot before-after trial was the only interventional study that compared conventional management to a conservative strategy using an SpO 2 target between 90 and 92% [32]. Although this study was underpowered to demonstrate a difference in mortality, it supported the feasibility and safety of a restrictive O 2 therapy, which led to a marked reduction in O 2 exposure without being associated with major clinical and physiological adverse effects.
Our analysis showed a significant association between hyperoxia exposure and mortality in patients resuscitated from cardiac arrest. This is consistent with the result of a recent meta-analysis by Wang et al. [39]. In a meta-analysis of animal trials, the administration of 100% O 2 therapy following resuscitation from cardiac arrest was associated with worse neurological outcomes as compared with lower O 2 concentrations [40]. An association between hyperoxia exposure and poor neurological outcomes has been reported by several authors [17,35,36,41], but not confirmed by a recent multicenter observational study [42] that instead highlighted PaCO 2 as a possible confounding factor. The adverse effects of hyperoxia may be due to enhanced oxidative stress, which may be particularly deleterious during the early reperfusion phase after cardiac arrest, and a vasoconstrictor effect that may paradoxically lead to a net reduction in local O 2 delivery to tissues including the myocardium and brain [43]. Mechanisms by which hyperoxia causes vasoconstriction include an inhibition of vasodilator (prostaglandins, nitric oxide) by reactive O 2 species [43]. Of note, the strength of association between hyperoxia and mortality was inversely related to the prevalence of chronic cardiovascular disease. We speculate that the response to a high O 2 tension may be blunted in the presence of an underlying endothelial dysfunction, where nitric oxide levels may be Figure 4 Forest plot showing individual and pooled odds ratios for mortality of studies on patients with stroke. Odds ratios >1 (right side of the plot) indicate an association between hyperoxia and higher mortality. Heterogeneity was Q (1) = 0.04, P = 0.844, I 2 = 0. The size of the boxes is inversely proportional to the size of the result study variance, so that more precise studies have larger boxes. ES, effect size; CI, confidence interval; W, weight; Sig., P value. Odds ratios >1 (right side of the plot) indicate an association between hyperoxia and higher mortality. Heterogeneity was Q (4) = 11.28, P = 0.024; I 2 = 64.54. The size of the boxes is inversely proportional to the size of the result study variance; more precise studies have larger boxes. ES, effect size; CI, confidence interval; W, weight; Sig., P value. chronically low [44]. However, this hypothesis can be challenged by data showing a deleterious effect of high O 2 in patients with severe coronary artery disease and myocardial infarction [45,46]. An alternative hypothesis may be that hyperoxia exposure had a minor role on mortality when the prevalence of cardiovascular comorbidity was higher.
The exposure to arterial hyperoxia in the first 24 hours of ICU admission was associated with higher mortality in patients with stroke, although this result is limited by the low number of studies included. Previous RCTs showed only transient radiological (magnetic resonance imaging) and clinical improvement [47], no benefit [48] or worse outcomes [49] in stroke patients receiving supplemental O 2 during their initial management. The Stroke Oxygen Study RCT, due to report in early 2016, is comparing three-day continuous or nocturnal O 2 administration with no supplemental O 2 in 6,000 patients (ISRCTN52416964, www.controlled-trials.com). Likewise, arterial hyperoxia was associated with increased hospital mortality in patients with traumatic brain injury. This should, however, be interpreted with caution given the heterogeneous characteristics of the studies included. The rationale for giving high O 2 concentrations to patients with traumatic brain injury is to improve brain O 2 delivery and metabolism [50]. While studies using indirect measures of brain metabolism provided promising results [7], these were not subsequently confirmed by a study that found no improvement in brain O 2 utilization measured by positron emission tomography one hour after ventilation with 100% O 2 [51].
The different criteria used for defining hyperoxia exposure were the main source of heterogeneity among the analyzed studies. The selection of 'worst' PaO 2 based on the A-a gradient [20][21][22][23]31] has been questioned by several authors, as this gradient does not correlate in a linear fashion with PaO 2 as the FiO 2 increases [35] and this method may reduce the probability of finding any association between hyperoxia and mortality [52]. However, a subanalysis by Bellomo et al. [21] on a sample of 100 patients showed that the worst PaO 2 was more representative of mean PaO 2 than was the first PaO 2 measured after ICU admission. Different PaO 2 measures may have different pathophysiological consequences. If even short periods of hyperoxia were dangerous, then the highest PaO 2 would represent the most sensitive approach to identify patients at risk; conversely, if the overall effect depended on the total amount of excess O 2 received, then the mean PaO 2 or a time-weighted measure would be a better choice. Alternatively, the first Table 4 Moderator analyses for studies in patients with traumatic brain injury (k = 5)  [17,19,38] found an independent association between hyperoxia exposure and hospital mortality, while studies using other measures showed more inconsistent results. This may suggest that exposure to high O 2 tensions in the early phase of the critical illness may be particularly associated with worse outcomes. Interestingly, in a subgroup analysis by de Jonge et al. [16], exposure to hyperoxia as defined by higher mean PaO 2 values during the entire ICU stay was not independently associated with mortality. All studies that considered a timespan longer than the first 24 hours for assessing oxygenation status did not find any significant association between arterial hyperoxia and worse outcomes [2,33,36].
In most of the studies, patients were categorized as hyperoxic or nonhyperoxic based on an arbitrarily predetermined PaO 2 cutoff value. Similarly, our analysis was based on binary ORs of hyperoxia exposure versus nonexposure. This approach may be limited by poor resolution in describing the relationship between increasing arterial O 2 tensions and outcome. Several studies in which PaO 2 was analyzed as a continuous variable showed a linear relationship between increasing arterial O 2 tensions and mortality, without a clear threshold for harm [30,35]. Furthermore, there is no consensus on the PaO 2 cutoff value to use for defining hyperoxia exposure, which varied markedly across the analyzed studies. This is likely to influence the associations observed. In a meta-regression analysis on the overall set of studies, the association between hyperoxia and worse outcome appeared to become stronger when the PaO 2 cutoff value used for defining exposure increased (data shown in Additional file 5).

Strengths and weaknesses
This is the first systematic review on the relationship between arterial hyperoxia and mortality in critically ill patients that gathers together data from a large number of subjects within several distinct disease categories. In our quantitative data syntheses, every effort was made to control for possible sources of heterogeneity and confounding factors. The authors were contacted if additional unpublished data were needed; any overlap between study populations was avoided. Whenever possible, adjusted outcome data were used and/or hypoxic patients excluded. Moderator analyses were performed to analyze the impact of several sources of heterogeneity (definition of hyperoxia exposure, study design). Study quality was assessed by means of a standardized scale and its impact on the studied association was explored. A random-effects model was used to pool data to account for unmeasured confounding factors and sources of heterogeneity.
Our analysis has several limitations. First, the studies were mainly observational investigations that cannot directly support any causal relationship between hyperoxia exposure and worse outcome. Higher PaO 2 levels may simply reflect the clinicians' attempts to optimize O 2 delivery by administering a higher FiO 2 ; thus PaO 2 becomes a marker of illness severity rather than being directly responsible for the outcome. Second, the included studies used different criteria for defining hyperoxia exposure and applied different statistical methods for analysis (PaO 2 as an ordinal/continuous variable; different multivariable regression models). Third, hypoxemic patients could not always be excluded from the analysis: these patients are likely to be responsible for an increased mortality in the subgroup of those not exposed to hyperoxia and might thus have blunted the studied association. Finally, we did not include unpublished studies, dissertations, or conference abstracts. We decided to consider only published material to ensure that only higher quality, peer-reviewed studies were included in the analysis.

Clinical implications and directions for future research
Given the widespread use of O 2 therapy in critical care, clinicians should be aware of the potentially deleterious effects of excessive O 2 administration. Several studies reported that FiO 2 is rarely adjusted for arterial hyperoxia, especially when this occurs at lower FiO 2 settings [1,2]. The rationale for giving supplemental O 2 to nonhypoxemic patients should be reconsidered as there is insufficient evidence of benefit [43]. When hemoglobin is fully saturated, additional O 2 only marginally increases O 2 transport capacity; conversely, a paradoxical decrease in regional O 2 delivery could be caused by vasoconstriction [53].
An urgent need exists for adequately designed studies to provide conclusive answers regarding the safety of hyperoxia in critically ill patients. Only RCTs can confirm a causal relationship between hyperoxia exposure and higher mortality. These trials should evaluate ventilation strategies using different PaO 2 targets for titrating FiO 2 , rather than comparing two arbitrarily selected FiO 2 targets. A pilot before-after study has supported the safety and feasibility of a conservative oxygen therapy [32]. RCTs comparing current liberal ventilation practices to more restrictive approaches in ICU patients are currently ongoing (ClinicalTrials.gov, NCT01319643 and NCT01722422).
There is an imperative to identify the best criteria that define hyperoxia exposure in observational studies. The relationship between hyperoxia and mortality should be evaluated using different criteria for defining hyperoxia exposure, comparing different PaO 2 measures and the time of assessment. This analysis may have important pathophysiological implications and could clarify whether early exposure to hyperoxia during critical illness is more deleterious. In future research, it would also be more useful to assess and report the relationship between mortality and PaO 2 as a continuous variable, instead of stratifying patients on the basis of an arbitrary PaO 2 cutoff value. Finally, further studies should address other specific categories of critically ill patients, such as sepsis, polytrauma, post-operative cases and hemorrhagic shock.

Conclusions
The majority of studies that have explored the relationship between arterial hyperoxia and mortality in critically ill patients are retrospective observational investigations, with only one prospective before-after study supporting the safety of a more conservative strategy. A quantitative data synthesis was not possible for studies on general populations of mechanically ventilated ICU patients because of differences in design, definition of hyperoxia, and the outcome measure reported. Hyperoxia exposure may be associated with mortality in patient subsets (post-cardiac arrest, stroke and traumatic brain injury). However, these results must be interpreted cautiously given the heterogeneity in criteria used for defining hyperoxia exposure and a significant inconsistency between study findings. Nevertheless, these data provide the rationale for future RCTs comparing conventional practice against more restrictive oxygenation targets.