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

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 O₂ flow rather than those of arterial hyperoxemia. In addition, in these patients the PaO₂ 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 O₂ pressure (PaO₂). For defining a condition of exposure to hyperoxia, any cutoff value of PaO₂ 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 O₂ fraction (FiO₂) were excluded if not guided by any assessment of PaO₂. Patients not exposed to hyperoxia constituted the comparator group. The primary outcome of interest was hospital mortality from any cause.

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 hand-searched 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.

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.

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 single-center), 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 PaO₂, 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.

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 follow-up’ 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 FiO₂. 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.

Data were synthesized using meta-analytic methods [26],[27], and statistically pooled by the standard meta-analysis 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 FiO₂. If the authors analyzed the unadjusted or adjusted association between mortality and increasing quartiles/quintiles/deciles of PaO₂, we considered patients as ‘hyperoxic’ if they fell in the upper stratum. When adjusted data were not reported or PaO₂ 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]. I² 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).

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 PaO₂ as the one associated with the lowest PaO₂/FiO₂. Conversely, Eastwood et al. [20] defined the worst PaO₂ in patients with an FiO₂ ≥0.5 as that associated with the ABG providing the highest arterial-alveolar (A-a) gradient; for patients with an FiO₂ <0.5, the lowest PaO₂ was recorded. If arterial blood gases (ABGs) were taken in patients in whom FiO₂ <0.5 and ≥0.5 were both recorded during the first 24 hours, the value of PaO₂ taken when the FiO₂ ≥0.5 was used. In one study [2] hyperoxia was defined on the basis of a time-weighted SpO₂ >98%; this parameter was calculated as the mean SpO₂ 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 SpO₂ 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.

Six retrospective cohort studies (three multicenter [17],[21],[31], three single-center [35]-[37]) evaluated patients resuscitated from cardiac arrest (n = 19,144). Different PaO₂ measures were used to define hyperoxia exposure (Table 1). Most studies used a PaO₂ cutoff of 300 mmHg [17],[21],[31],[37]. In the study by Janz et al. [35], the relationship between mortality and PaO₂ 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 PaO₂ 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).

The pooled ES shows an association between hyperoxia exposure and increased in-hospital mortality (OR = 1.42 (95% CI 1.04 to 1.92), P = 0.028, number of studies (k) = 5), in the presence of a moderately high heterogeneity (Q (4) = 12.4, P = 0.015; I ² = 67.73) (Figure 3). 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 PaO₂. 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.

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 PaO₂ ≥300 mmHg as being exposed to hyperoxia. Young et al. [23] evaluated the independent association between mortality and deciles of PaO₂, with the upper decile (PaO₂ >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 (PaO₂ ≤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).

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, PaO₂ 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 heterogeneity (Q (4) = 11.28, P = 0.024; I ² = 64.54) (Figure 5). The funnel plot indicated no obvious publication bias (Additional file 4). The exclusion of the study by Asher et al. [33] (single-center, time of PaO₂ assessment beyond the first 24 hours) in the sensitivity analysis did not substantially change the combined ES (OR = 1.46 (95% CI 1.08 to 1.98)) nor did it decrease heterogeneity (I ² = 67.29). Results of the moderator analyses are shown in Table 4, with studies stratified based on design, PaO₂ value and cutoff used for defining hyperoxia, time of PaO₂ assessment and comparator group. Study quality as defined by the NOS score, mean age and gender did not influence the ES.

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 PaO₂ levels may simply reflect the clinicians’ attempts to optimize O₂ delivery by administering a higher FiO₂; thus PaO₂ 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 (PaO₂ 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.

Given the widespread use of O₂ therapy in critical care, clinicians should be aware of the potentially deleterious effects of excessive O₂ administration. Several studies reported that FiO₂ is rarely adjusted for arterial hyperoxia, especially when this occurs at lower FiO₂ settings [1],[2]. The rationale for giving supplemental O₂ to nonhypoxemic patients should be reconsidered as there is insufficient evidence of benefit [43]. When hemoglobin is fully saturated, additional O₂ only marginally increases O₂ transport capacity; conversely, a paradoxical decrease in regional O₂ 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 PaO₂ targets for titrating FiO₂, rather than comparing two arbitrarily selected FiO₂ 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 PaO₂ 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 PaO₂ as a continuous variable, instead of stratifying patients on the basis of an arbitrary PaO₂ cutoff value. Finally, further studies should address other specific categories of critically ill patients, such as sepsis, polytrauma, post-operative cases and hemorrhagic shock.