All arterial blood gases during the first 24 hours of ICU admission were collected and entered into a standardized data collection system which automatically selects the appropriate high and low simultaneous fraction of inspired oxygen (FiO₂) and partial pressure of arterial oxygen (PaO₂) measurements and deletes other oxygenation data. Using the APACHE II and III methodology for intubated patients with FiO₂ ≥0.5, the PaO₂ associated with the arterial blood gas with the highest alveolar-arterial (A-a) gradient is selected as the index of worst oxygenation. For nonintubated patients or intubated patients with FiO₂ < 0.5, the lowest arterial blood gas PaO₂ level is recorded. The ratio of PaO₂ to FiO₂ (P/F ratio) is also used as an index of illness severity.

We recorded the sizes, types and locations of the hospitals. At the patient level, we extracted the following variables: demographics, comorbidities according to APACHE II and III classifications, hospital and ICU admission source, intubation, treatment limitation, year of admission, physiological and arterial blood gas parameters over the first 24 hours in the ICU, vital status at hospital discharge, hospital discharge destination and an APACHE III risk of death score 8 . As a marker of severity of illness independent of arterial oxygenation, we calculated an adjusted APACHE III index of illness severity (AP3no-ox), in which the oxygen component of the APACHE III scoring system was removed.

All analyses were performed using SAS version 9.2 (SAS Institute Inc., Cary, NC, USA). Continuous data are presented as means ± standard deviations or as medians with interquartile ranges (IQRs), depending on the underlying data distribution. Categorical data are reported as proportions. We categorized oxygenation levels into the same three groups as in the EMShockNet study 6 , defined by the worst PaO₂ and P/F ratio obtained in the first 24 hours of ICU admission. Thus, we divided patients into three groups according to worst PaO₂ or A-a O₂ gradient in the first 24 hours after admission. We defined 'hyperoxia' as a PaO₂ 300 mmHg or greater, 'hypoxia/poor O₂ transfer' as either PaO₂ < 60 mmHg or a P/F ratio <300, 'normoxia' as any value between hypoxia and hyperoxia and 'isolated hypoxemia' as PaO₂ < 60 mmHg regardless of FiO₂ level.

The primary outcome measures were in-hospital mortality and survival time, which are reported as ORs (95% confidence interval (95% CI)) or hazard ratios (HRs) (95% CI), respectively. To determine functional recovery, we also considered discharge to home as a secondary outcome. We compared outcomes between groups using the χ² test with the Bonferroni correction. We conducted multivariate analysis using logistic regression for mortality and Cox proportional hazards regression for survival time, with models constructed using both stepwise selection and backwards elimination procedures. To increase robustness and model validity, we used a P value of 0.01 for variable inclusion. We applied several models to the statistical analysis of the independent relationship between oxygenation and patient outcome. We constructed an initial model for mortality in accordance with the EMShockNet model (see Additional file 1, Statistical appendix, Model 1) 8 . We then applied a second model to improve discriminatory power using AP3no-ox as a marker of severity (see Additional file 1, Statistical appendix, Model cluster 2). Finally, we conducted further sensitivity analysis inclusive of propensity analysis 9 , Cox proportional hazards modelling, testing of different cutoff points for hyperoxia, analysis of subgroups contemporaneous with the EMShockNet cohort and assessment of PaO₂ according to deciles (Additional file 1, Statistical appendix, Model cluster 3).

As our database contained only the worst recorded oxygenation in the first 24 hours after ICU admission, we explored its relationship with that of the first PaO₂ measurement after ICU admission (as in the EMShockNet study) and the mean oxygenation on ICU admission days 1, 2 and 3 by selecting 100 of the database patients and obtaining additional data from all of their hospital arterial blood gas records during their ICU stay (see Additional file 1, Statistical appendix, Model cluster 3).

The proportion of living patients with hyperoxia (PaO₂ > 400) was 5% (n = 280). Comprising 5,140 patients who lived and 6,968 patients who died, this study had 93% power to detect a change of 1.5% (5% versus 6.5%) in the proportion of patients with hyperoxia (PaO₂ > 400) with a two-sided P value of 0.05.

There were 625 patients in the data set with hyperoxia (PaO₂ > 400). Comprising 11,483 patients without hyperoxia, this study had 90% power to detect a difference in mortality of 7% (55% versus 62%) between groups with a two-sided P value of 0.05. Given an observed difference of 14% (55% versus 69%) in the EMShockNet study between hyperoxic patients (PaO₂ > 400) and nonhyperoxic patients, we felt that this study was adequately powered to detect a relationship between mortality and PaO₂ > 400. In our study, the mortality rate in the hyperoxia group (PaO₂ > 400) was only 0.5% higher than that in the nonhyperoxia group (54.7% versus 55.2%, P = 0.22).

There were 531 hyperoxia survivors (PaO₂ > 300). With 4,609 nonhyperoxia survivors, this study had 80% power to detect a difference between groups of 6% (64% versus 58%) regarding the proportion discharged to home with a two-sided P value of 0.05. Given an observed difference of 6% (38% versus 44%) in the EMShockNet study between hyperoxic survivors (PaO₂ > 300) discharged to home and discharged nonhyperoxic survivors, we again felt that this study was adequately powered to detect a relationship between hyperoxia and discharge to home. In our study, there was no observed difference in the proportion of patients who were discharged to home between the hyperoxia and nonhyperoxia groups.

We conducted a large, multicentre, cohort study of patients admitted to ICUs in ANZ after resuscitation from cardiac arrest to examine the relationship between hyperoxia and patient outcome. We initially found that hyperoxia was relatively uncommon and had only a weak relationship with risk of death. This relationship was significantly reduced by the addition of illness severity scores. In addition, once Cox proportional hazards modelling of survival, sensitivity analyses using deciles of hypoxemia, time period matching and defining hyperoxia in keeping with experimental studies 10 11 12 13 as PaO₂ > 400 mmHg, hyperoxia had no independent association with mortality. Finally, after adjustment for FiO₂ and relevant covariates, PaO₂ was no longer predictive of hospital mortality. Thus, hyperoxia was relatively uncommon, and it had no robust and consistently reproducible independent relationship with mortality.

Until very recently, concerns about the possible risks associated with hyperoxia during and after recovery from cardiac arrest were based on animal experiments 10 11 12 13 . In this regard, several experimental studies have suggested that hyperoxia can increase oxidative stress 14 , induce more severe histopathological changes 10 and worsen neurological injury 15 . On the other hand, two studies have failed to confirm such findings 16 17 , and two other often-quoted major studies did not actually assess animals after cardiac arrest 10 12 . Nonetheless, despite the lack of human data, the International Liaison Committee on Resuscitation moved to advocate the avoidance of arterial hyperoxia. The committee instead advocated the targeting of arterial oxygen saturation not exceeding 94% to 96% 18 . In response to these issues, in June 2010, the EMShockNet investigators reported that, in a cohort of 6,326 USA patients who had survived nontraumatic cardiac arrest and were admitted to the ICU, hyperoxia was independently associated with increased risk (OR 1.6) of in-hospital mortality. This was the largest clinical study to date of the association between hyperoxia after cardiac arrest and mortality. Our findings should therefore be evaluated in direct comparison to the EMShockNet study and have been specifically configured to facilitate such comparison.

Several important observations emerge from such comparison. First, the baseline characteristics of the U.S. and ANZ patients appear almost identical, although no information was available on the APACHE scores for the USA cohort. Despite such similarities, there were striking differences in the lowest body temperatures recorded (ANZ 34.9°C, USA 36°C). These differences may reflect greater uptake of therapeutic hypothermia in ANZ and make the observations from our cohort more relevant to current recommended practice 4 5 . However, we cannot determine whether greater use of therapeutic hypothermia accounts for the difference in the proportion of survivors discharged to home (65% in ANZ as compared to 44% in the USA, approximately a 50% relative increase in favourable outcome).

In the ANZ cohort, hyperoxia occurred in only 10.6% of patients as compared with 18% in the USA, and mortality in the hyperoxic group was identical to that in the hypoxia/poor O₂ transfer group, instead of being much greater. Importantly, the relationship between hyperoxia and in-hospital mortality appeared much weaker using the same modelling used by the EMShockNet investigators. Even more importantly, this relationship could not be confirmed when a different threshold for hyperoxia was applied which mimicked that reported in experimental studies 11 12 13 (rather than using a seemingly arbitrary cutoff point of 300 mmHg), when a Cox proportional hazards model was used, when PaO₂ was split into deciles, when FiO₂ was taken into account or when the same time period (2000 to 2005) was used for analysis. In the aggregate, these observations suggest that the relationship between hyperoxia and mortality is dependent on the individual healthcare system, the statistical model used, the time period examined and the definitions used. Such features are not consistent with a robust and reproducible biological phenomenon.

Our findings imply that it is incorrect and premature to conclude that hyperoxia is an independent risk factor for mortality in patients resuscitated from nontraumatic cardiac arrest. In particular, we contend that hyperoxia implies the administration of high FiO₂ fractions, making it more likely for hyperoxia to actually be a marker of illness severity than a biological toxin. This notion is supported by the significant decrease in ORs for mortality once APACHE scores were added to the model and the disappearance of significant ORs once FiO₂ was added to the model. Moreover, the definition of 'hypoxia' used by the EMShockNet investigators (reproduced here to facilitate comparison under the term 'hypoxia/poor O₂ transfer') included patients with a P/F ratio <300, together with patients with PaO₂ < 60 mmHg. This approach conflates physiologically relevant lack of oxygen at the tissue level (true hypoxia) with a gas transfer problem. When we examined isolated hypoxemia (PaO₂ < 60 mmHg), we found that it was nearly as frequent as hyperoxia. The risks of cerebral injury associated with hypoxemia are well known 19 20 21 , and hypoxemic patients in our cohort had particularly poor outcomes. Importantly, after adjustment for FiO₂ and relevant covariates, PaO₂ was no longer predictive of hospital mortality. These observations suggest that the association seen in the models that do not include FiO₂ may simply reflect the fact that hyperoxia is an indirect marker of higher FiO₂ (that is, the higher the FiO₂, the greater the PaO₂) and that a higher FiO₂ is a marker of illness severity (that is, the sicker the patient is perceived to be, the greater the FiO₂ administered in an emergency situation). However, the link between FiO₂ and outcome is independent of APACHE score. Thus, FiO₂ cannot be considered simply a marker of disease severity. The physician can lower PaO₂ and FiO₂ levels at the same time and avoid inducing hyperoxia. Only interventional studies can clarify whether the association between oxygenation and outcome is truly a causal relationship.

All the above observations have potential clinical relevance. For example, emergency responders may not have access to pulse oximetry or blood gas analysis, or such techniques may be unreliable immediately following cardiac arrest because of decreased peripheral perfusion. If fear of hyperoxia led emergency responders, in the absence of adequate monitoring, to limit FiO₂ levels, logically more people would likely be exposed to the risk of hypoxemia. In our study, >40% of patients with hypoxemia might have had correction of their hypoxemia had a higher FiO₂ level been induced. Thus, if concerns about the alleged ill effects of hyperoxia or high FiO₂ administration were taken into the clinical arena to avoid a condition whose association with mortality is uncertain, more patients might be exposed to a condition whose adverse cerebral effects are well established. Given our findings, we counsel against implementing policies of deliberately induced decreases in FiO₂ unless accurate continuous pulse oximetry monitoring is in place. Importantly, in no way do we advocate, promote or justify hyperoxia in this setting. However, lowering FiO₂ is justified only if good transcutaneous or arterial oxygenation monitoring is available.

This study has several strengths. It involved more than 12,000 patients from 125 ICUs in two countries, making it the largest study of its type conducted so far and making its findings reflective of all ICUs in ANZ 22 23 . It included a multifaceted assessment of the independent relationship between hyperoxia and outcome using multiple models and adjusting for illness severity. However, like other studies of association using a large database, it is limited by the nature of the data available and by the fact that no causal inferences can be drawn. The assessment of oxygenation status in the first 24 hours was based on the 'worst' possible arterial blood gas result, while the EMShockNet study used the 'first' ICU arterial blood gas measurement for evaluation. Thus, patients may have been exposed to hyperoxia and may not have been identified in our study. However, using a random sample of 100 patients, we found that the measurement used in our study was more closely representative of overall mean oxygenation status in ICU patients during the first 24 to 48 hours after admission (when reperfusion injury occurs) than the first set of blood gas measurements obtained in the ICU. In our study, data were missing for only 5.4% of patients compared with 27.6% in the EM ShockNet study, making selection bias in our study less likely. One-third of patients had a lowest body temperature <34°C. Clinical knowledge (confirmed by the EMShockNet data) that such severe spontaneous hypothermia is uncommon suggests that many patients were therefore treated with induced hypothermia as is common in ANZ 24 25 26 27 28 . This finding distinguishes our study from the US investigation because it is in keeping with current recommendations. Unfortunately, however, our database does not enable us to identify which patients had induced versus spontaneous hypothermia. Finally, we are unable to comment on the causes of death or consider other potential confounding variables that were not collected as part of the ANZICS-APD.

More investigations appear necessary, perhaps using other national databases 29 . Prospective investigations with focused data collection are also needed. If such studies confirmed a postive association, interventional strategies should be tested; if not, interventional studies would not seem justified.