Dangers of hyperoxia

Since its discovery [1‐3], oxygen (O₂) has been recognised as “friend and foe” [4]. It is vital for aerobic respiration within the mitochondria, yet mitochondrial respiration also forms reactive oxygen species (ROS) [5], production of which relates to O₂ concentration [6‐8]. Supplemental O₂, i.e. inspiratory O₂ concentrations (F_IO₂) > 0.21, may cause hyperoxaemia (arterial PO₂ > 100 mmHg) and subsequently increased ROS formation [9‐11]. This is particularly pronounced during ischaemia/reperfusion (I/R) and/or hypoxia/re-oxygenation [6‐8]. ROS are as “Janus-headed” as O₂: ROS are vital for host defence, and also toxic [12]. Consequently, O₂ toxicity, especially pulmonary, is a matter of concern [13‐15], and optimal dosing remains unclear in critical care. This review discusses potential harms of O₂ in various underlying critical illnesses. Figure 1 summarises the possible dangers of hyperoxia, highlighting pathophysiological mechanisms and their impact on specific disease conditions. The most important clinical studies are listed in Table 1; “Additional file 1” shows the complete study list.

Oxygen generally exists as di-atomic molecule (O₂); its two atoms bond to each other through single bonds leaving two unpaired electrons. O₂ performs its actions through these unpaired electrons which act as radicals. ROS are even more reactive molecules formed through oxygen’s electron receptivity (e.g. superoxide, peroxide, and hydroxyl anion).

Over 90% of O₂ consumption is utilised by mitochondria, predominantly for ATP production (oxidative phosphorylation), but also for heat generation through uncoupling, and superoxide production. O₂ is the terminal electron acceptor at Complex IV of the electron transport chain (ETC), being reduced to water in this process. For each mole of glucose metabolised, anaerobic respiration (glycolysis) generates only 2 ATP moles compared to approximately 28–30 from oxidative phosphorylation. In health, 1–3% of mitochondrial O₂ consumption is used at the ETC Complexes I and III to generate superoxide, an important signalling molecule [16]. Superoxide is necessary for enzyme processes, e.g. oxidases (catalysing oxidation–reduction reactions) and oxygenases (incorporating oxygen into a substrate). Activated immune cells utilise O₂ for extra-mitochondrial ROS production: NADPH oxidase generates superoxide (“respiratory burst”) for phagocytosis. Unless overwhelmed by ROS over-production, antioxidant capacity (e.g. superoxide dismutase, glutathione, thioredoxin) prevents oxidative damage to DNA, proteins and lipids, and subsequent cell death.

O₂ also affects the inflammatory response. Experimental models and volunteer and patient studies demonstrate that hyperoxia (and hypoxia) can induce pro- and anti-inflammatory responses, with both protective and harmful sequelae [17]. Hyperbaric oxygen is used to aid wound healing and treat gas gangrene, but may cause neurotoxicity. Whether the response to hyperoxia relates to its degree and/or duration, specific cell types, background inflammation, or other factors remains uncertain; clearly, O₂ toxicity can be induced de novo without underlying pathology, predominant organs being lung, brain, and eye.

Pulmonary toxicity was first described by Lorrain Smith: pure O₂ at hyperbaric pressures caused inflammatory pneumonitis [18]. At atmospheric pressure pneumonitis was seen after days in non-human primates breathing 60–100% O₂ [19‐21]. After initially affecting the airways (tracheobronchitis) with reduced mucociliary clearance [22], the lung parenchyma becomes involved. In humans, this occurs especially when the inspiratory PO₂ is significantly enhanced in a hyperbaric environment. Initial complaints are retrosternal chest pain, then coughing and dyspnoea as a pneumonitis develops with pulmonary oedema and diffuse radiological lung shadowing. In healthy volunteers breathing 98–100% O₂, chest pain commenced after 14 h, coughing and dyspnoea between 30 and 74 h [22]. Due to nitrogen washout [23], there may also be atelectasis in lung regions with low ventilation/perfusion ratios [24].

Whether hyperbaric vs. normobaric O₂ toxicity mechanisms and onset are similar is unclear. Pulmonary injury was accelerated by hyperbaric hyperoxia, but was less inflammatory in character and driven by a neurogenic component that could be blocked by inhibiting neuronal nitric oxide synthase or vagal nerve transection [25]. Possible synergistic effects on O₂ toxicity of underlying lung pathology are poorly characterised, especially at the more moderate degrees of hyperoxia inflicted on patients. This is, however, well-recognised with bleomycin toxicity where mild hyperoxia may be damaging [26].

Neurotoxicity was described over a century ago [27]: 3Atm of O₂ produced convulsions and death. Seizures or syncope occurred after 40 min at 4Atm O₂, and within 5 min at 7Atm [28]. This was usually preceded by milder symptoms such as tunnel vision, tinnitus, twitching, confusion, and vertigo. The impact of high concentration normobaric O₂ on neurotoxicity, however, is unclear.

Mitochondrial ROS production increases either with O₂ deficit or excess, but particularly during excess O₂ (hyperoxia). This can occur in sepsis and/or I/R injury, i.e. whole-body (e.g. resuscitation from cardiac arrest or major haemorrhage), or organ-specific (e.g. revascularisation after myocardial infarction or stroke). A similar injury may be induced by acute hypoxaemia followed by rapid correction (hypoxia/re-oxygenation-injury). The impact of reperfusion injury may be as severe as the ischaemic insult. Although preclinical and clinical studies are not consistent [29‐31], reperfusion injury is generally exacerbated by hyperoxia. The hyperoxia effect may be exacerbated by acidification of the hypoxic tissues; the right-shifted oxyhaemoglobin dissociation curve of blood (re-)entering the hypoxic tissue augments O₂ release, with a subsequent increase in superoxide production [31].

Teleologically, the body has not evolved to deal with high tissue O₂ tensions. Tissues not metabolising adequately, e.g. due to toxins or switching off (“hibernating”) in response to hypoperfusion, reduce O₂ utilisation. As a normal protective response, negative feedback signals reduce local blood flow by vasoconstriction to mitigate local build-up of O₂ and subsequent toxicity. Acute hyperoxia thus induces vasoconstriction, reducing local blood flow [32], particularly in the cerebral and coronary vasculature [5‐7]. This vasoconstriction is in part related to reduced release of nitric oxide (NO) from S-nitrosohaemoglobin binding [33]. Vasoconstriction has been shown in patients with and without coronary artery disease, where supplemental O₂ reduced cardiac output and coronary sinus blood flow [34, 35]. Seizures associated with neurological O₂ toxicity occur with paradoxical vasodilation during hyperbaric hyperoxia [6].

The Oxygen-ICU trial was the first major study to suggest clinically important harm from liberal O₂ administration in a general ICU population [36]. This single-centre, RCT included 480 patients expected to stay in the ICU for at least 72 h. ICU mortality was 20.2% with conventional and 11.6% with conservative O₂ therapy. Around two-thirds of patients included were mechanically ventilated at baseline, around a third had shock, and the illness acuity was relatively low. The difference was statistically significant although the study was stopped early after a non-preplanned interim analysis, and the magnitude of the reported treatment effect was larger than hypothesised [36]. Given the variety of mechanisms of death in ICU patients [37], such a high proportion of deaths in a heterogeneous population of ICU patients is unlikely to be attributable to the dose of O₂ therapy used. However, the Oxygen-ICU trial [37] did highlight the need for further investigation.

Subsequently, the IOTA systematic review and meta-analysis [38] reported that conservative O₂ use in acutely ill adults significantly reduced in-hospital mortality. Although these findings were concordant with the Oxygen-ICU trial [36], they provided only low certainty evidence: First, the Oxygen-ICU trial [36] contributed 32% of the weight to the mortality analysis. Second, predominant conditions were acute myocardial infarction and stroke, and a range of O₂ regimens were tested so that the analysis provided only indirect evidence about the optimal O₂ regimen for patients in the ICU. Third, the overall mortality treatment effect estimates were imprecise. Finally, an updated systematic review and meta-analysis found no evidence of benefit or harm comparing higher vs. lower oxygenation strategies in acutely ill adults [39].

The multicentre randomised ICU-ROX trial found that conservative O₂ therapy did not significantly affect the primary end point of number of days alive and free from mechanical ventilation (ventilator-free days) compared with usual (liberal) O₂ therapy [40]. Overall, 32.2% of conservative and 29.7% of usual O₂ patients died in hospital. While these findings provide some reassurance to clinicians about the safety of the liberal O₂ use that occurs in standard practice, they do not exclude clinically important effects of the O₂ regimens tested on mortality risk. Indeed, based on the distribution of data, there is a 46% chance that conservative O₂ therapy increases absolute mortality by more than 1.5% points, and a 19% chance that conservative O₂ therapy decreases absolute mortality by more than 1.5% points [41, 42]. Finally, a recent RCT conducted in ICU patients fulfilling the systemic inflammatory response syndrome criteria, found no significant difference between high-normal and low-normal oxygenation targets for non-respiratory organ dysfunction over the first 14 days, or in Day-90 mortality [43]. Accordingly, the most appropriate dose of O₂ to give to adult ICU patients remains uncertain.

Clinicians should titrate O₂ therapy to avoid both hypoxaemia and hyperoxaemia. While the harmful effects of tissue hypoxia are clearly understood [44], over-correction leads to tissue hyperoxia which may also be deleterious. Hyperoxia injures the lung via ROS production, causing oxidant stress with pro-inflammatory and cytotoxic effects [35, 45, 46]. Pathophysiologic consequences include arterial vasoconstriction [35, 47‐49], alveolar-capillary “leak” and even fibrogenesis [50, 51]. Clinicians use higher F_IO₂ than necessary to correct hypoxia in the critically ill [52], possibly to avoid (occult) tissue hypoxia [53, 54], to provide a “buffer” should rapid clinical deterioration occur, or because the consequences of hyperoxia are considered less severe. The lack of clearly defined targets for PaO₂ and/or SaO₂ is also an issue. The ARDS Network trials targeted a PaO₂  of  55-80 mmHg [55], while the British Thoracic Society suggests a target SpO₂ of  94–98% in acutely ill patients [56].

In ARDS, the potential for hyperoxia to impact outcomes is further complicated by the severity of gas exchange impairment. Specifically, extreme hyperoxaemia (i.e. PaO₂ > 300 mmHg), associated with harm in other critically ill populations, is impossible to achieve in ARDS (see Table 1). However, moderate hyperoxaemia is possible and could be harmful as well [57]. Furthermore, high F_IO₂ can directly injure the lung [58], sensitise it to subsequent injury [59], adversely affect its innate immune response [60], and worsen ventilation-induced injury [61, 62]. It is therefore necessary to distinguish between hyperoxaemia and high F_IO₂ use when assessing the effects of hyperoxia on the lung.

The recent LOCO2 trial in ARDS was stopped early for futility and safety concerns regarding mesenteric ischaemia in the conservative O₂ group. Moreover, 90-day mortality was significantly higher in patients receiving conservative O₂ therapy [63]. The HOT-ICU trial studied ICU patients with acute hypoxaemic respiratory failure and found no difference in 90-day mortality between conservative and liberal PaO₂ targets [64]. In the LUNG SAFE observational cohort study, both systemic hyperoxaemia and excess F_IO₂ use were prevalent, with frank hyperoxaemia (30% of patients) more prevalent than hypoxaemia in early ARDS [65]. Two-thirds of these patients received excess O₂ therapy. Hyperoxaemia did not appear to be used as a “buffer” in unstable patients: frequency was similar in shocked patients. While a similar proportion of patients were hyperoxaemic on day-2, higher F_IO₂ use did decrease. Both hyperoxaemia and excess O₂ use were mostly transient, although more sustained hyperoxaemia was seen. Reassuringly, no relationship was found between the degree and duration of hyperoxaemia, or excessive O₂ use, and mortality in early ARDS.

While these findings contrast with findings in other ICU cohorts, a key differentiating factor is the reduced potential for extreme hyperoxia in ARDS patients. The potential for harm from hyperoxia appears to be related to the severity of hyperoxaemia [54, 61, 66, 67]; those with relatively preserved lung function are at greatest risk [68]. However, no dose–response relation was found between PaO₂ and mortality [67]. Hence, paradoxically, patients with ARDS may be at less risk as they are unable to achieve extreme degrees of hyperoxia. A recent observational study suggested a U-shaped relationship between PaO₂ and mortality in ARDS patients; patients with a time-weighted PaO₂  of  93.8-105 mmHg had the lowest mortality risk [69]. Intriguingly, this range is near identical to the “liberal” target PaO₂ range targeted in the LOCO2 [63]. Hence, much remains to be learned about optimal targeting of PaO₂ in patients with ARDS.

Theoretically, hyperox(aem)ia might help septic patients due to its vasoconstrictor effect, counteracting hypotension [6‐8], and to the antibacterial effects of O₂ [70, 71]. However, hyperoxaemia did not affect cardiac output in septic patients [72]. The number of days with PaO₂ > 120 mmHg was an independent risk factor for ventilator-associated pneumonia (VAP) [73]; however, these patients had other risk factors, e.g. more frequent use of proton pump inhibitors and sedatives, higher incidence of circulatory shock with prolonged and higher catecholamine infusion rates, and more red blood cell transfusion. In an observational study on VAP patients, the same group reported that hyperoxaemia did not affect mortality [74]. The HYPER2S RCT [75] compared standard therapy vs. 100% O₂ over the first 24 h after diagnosing septic shock. Despite a significantly lower SOFA score at day 7, the trial was prematurely stopped due to higher, albeit not statistically significant mortality in the hyperoxia group at Day-28 and Day-90. The hyperoxia group had significantly more serious adverse events, including ICU-acquired weakness (p = 0.06). A post hoc analysis based on Sepsis-3 criteria found increased Day-28 mortality in patients with hyperlactataemia > 2 mmol/L (p = 0.054), but not with normal lactate levels [76]. The authors speculated that a hyperoxaemia-related increase in tissue O₂ availability may have led to excess ROS production and, consequently, oxidative stress-related tissue damage.

The opposite hypothesis, i.e. attenuation of oxidative stress-induced tissue damage by reducing O₂ exposure did not beneficially influence outcome in septic patients either. A post hoc analysis of the ICU-ROX trial [40] of the septic cohort showed no statistically significant difference with respect to ventilator-free days or Day-90 mortality for the “conservative” when compared to the “usual” oxygenation [77]. Point estimates of treatment effects even favoured the latter. Hence, it seems reasonable to avoid PaO₂ > 100-120 mmHg due to the possible deleterious consequences of excess tissue O₂ concentrations in the presence of sepsis-related impairments of cellular O₂ extraction [78].

Increasing F_IO₂ in acutely brain-injured patients, alongside other clinical interventions [79], can improve brain tissue PO₂ (PbtO₂) [80, 81]. The effects of normobaric hyperoxia are less significant in large hypoperfused brain regions [82], but highly relevant in small pericontusional areas [83]. Moreover, incremental F_IO₂ increased cerebral excitotoxicity in severe traumatic brain injury (TBI) [84]. The association of hyperoxia with outcome is even more controversial. After TBI, both hypoxaemia and hyperoxia were [85] or were not [86] independently associated with worse outcome. In two retrospective studies, including a mixed population of brain-injured patients, hyperoxaemia, defined as PaO₂ > 300 mmHg [87] or > 120 mmHg [88], was associated with increased in-hospital mortality and poor neurological outcome, even after adjusting for confounders. Patients with subarachnoid haemorrhage exposed to higher PaO₂ levels were also more likely to develop cerebral vasospasm [89, 90]; however, a retrospective analysis of patients needing mechanically ventilation did not find any relation between time-weighted PaO₂ and outcome [91]. Studies in acute ischaemic stroke in general [92], and in a sub-group needing mechanical ventilation [93], found no association between outcome and PaO₂ within the first 24 h. Even early hyperoxaemia (PaO₂ > 300 mmHg) did not affect mortality in mechanically ventilated TBI patients, notwithstanding severity on admission [94, 95]. Finally, PaO₂≈150-250 mmHg within the first 24 h post-TBI was associated with better long-term functional outcome after TBI [96]; however, the study excluded patient who died. Normobaric hyperoxia combined with intravenous thrombolysis was associated with more favourable neurological outcome than thrombolysis alone after ischaemic stroke [97].

Prospective studies have evaluated the effects of targeted hyperoxia after acute brain injury: Small studies in patients with acute ischaemic stroke not eligible for thrombolysis found either transient clinical improvement and smaller infarct size with high-flow O₂ [98, 99] or no effect of normobaric hyperoxia [100]. In a small RCT in mechanically ventilated TBI patients, F_IO₂ = 0.8 (vs. 0.5) improved 6-month neurological outcome [101], but conclusions should be cautioned due to methodological concerns. Exposure to F_IO₂ = 0.7 or 0.4 for up to 14 days after TBI influenced neither markers of oxidative stress or inflammation nor neurological outcome [102]. Finally, the Normobaric-Oxygen-Therapy-in-Acute-Ischemic–Stroke-Trial (NCT00414726) was prematurely halted after inclusion of 85/240 patients because of higher mortality in the high-flow O₂ group, although most deaths occurred following early withdrawal of life-support.

It remains open in acute brain injury, whether normoxaemia vs. targeted hyperoxaemia influences brain function and neurological recovery. Optimal PaO₂ targets, study populations, and specific forms of brain injury are currently unknown.

Supplemental O₂ is used because increasing the amount of physically dissolved O₂ during blood loss-related reductions in O₂ transport is thought to faster repay a tissue O₂ debt [103]. Despite its vasoconstrictor properties [6‐8], ventilation with 100% O₂ during experimental haemorrhage improved tissue PO₂ [104] and attenuated organ dysfunction [105, 106].

However, PaO₂ > 100 mmHg may enhance ROS formation [9‐11], especially during I/R and/or hypoxia/re-oxygenation, e.g. resuscitation from trauma-and-haemorrhage [6‐8].

A recent retrospective study in patients with prehospital emergency anaesthesia demonstrated that hyperoxia was present in most patients upon arrival in the hospital, however without relation to outcome [107]. Clinical data on the impact of hyperoxia on morbidity and mortality remain equivocal. No association was seen between mortality and PaO₂ in the first 24 h (median Injury Severity Score ISS = 29) [108]. Another observational study noted that 44.5% of patients mechanically ventilated in the emergency department had hyperoxaemia, this cohort having a higher Day-28 mortality [109]. From a French trauma registry (median ISS = 16), univariate analysis showed that admission PaO₂ > 150 mmHg coincided with a higher mortality, however, propensity score matching yielded the opposite result, namely supra-physiological PaO₂ levels were associated with significantly lower mortality [110]. Lower Day-28 mortality and less nosocomial pneumonia were seen in patients early after blunt chest trauma [111]. An analysis of 864,340 trauma patients (median ISS = 9) investigated the impact of supplemental O₂ in the ED; in all three patient categories predefined according to incremental SpO₂, supplemental O₂ was associated with a significantly higher ARDS incidence and mortality [112]. A retrospective analysis of patients with ISS ≥ 16 studied the impact of PaO₂ ≥ 300 mmHg during resuscitation [113]; while prolonged ICU stay was seen in patients not intubated in the ED, no effect was seen in the sicker cohort of mechanically ventilated patients. Finally, a retrospective multicentre study of trauma patients found SpO₂ > 96% over the first seven days was common place; the adjusted mortality risk was higher with greater F_IO₂ [114]. The currently recruiting “Strategy-to-Avoid-Excessive-Oxygen-for-Critically-Ill-Trauma-Patients (SAVE-O2)” (NCT04534959) will address any causality between hyperoxia and outcome.

Despite O₂ supplementation being common practice in patients with pronounced blood loss, no optimal target for PaO₂ is available.

During cardiac arrest, PbtO₂ drops rapidly to levels close to zero [115]. With cardiopulmonary resuscitation (CPR) PbtO₂ increases slowly, driven by the achieved cerebral perfusion pressure [116]. Guidelines recommend ventilation with 100%O₂ even though no clinical study has compared this against lower F_IO₂ [117]. Observational data suggest an association between higher PaO₂ during CPR and a higher likelihood of return of spontaneous circulation (ROSC), survival, and neurological outcome [118, 119]. After ROSC blood and brain PO₂ levels increase; mostly, this appears inevitable as F_IO₂ titration is impossible during CPR [120]. Given the connection between hyperoxia and ROS formation, there has been great interest in assessing whether avoidance of hyperoxaemia in the post-arrest phase could mitigate brain injury. Results are conflicting, either showing an association between hyperoxia and poor outcome [68, 121‐124], or not [125‐129]. Smaller randomised trials and sub-group analysis from larger trials have also been performed [130]. Overall, the evidence suggests that lower rather than higher O₂ targets are beneficial, even though any sweet spot for optimal PaO₂ is unknown [131]. The COMACARE pilot trial compared different PaO₂ targets and found no difference in two brain injury biomarkers [132, 133]. A sub-group analysis of the ICU-ROX study showed improved outcomes in restrictive compared to liberal O₂ treated patients at risk of hypoxic brain injury [134]. Opposite findings were seen in a sub-group of the HOT-ICU trial [64]. Current guidelines recommend targeting strict normox(aem)ia. The evidence suggests a signal to harm and, importantly, no indication of benefit from extreme hyperox(aem)ia; thus, this should be avoided [135].

Supplemental O₂ use has been standard practice for decades in acute myocardial infarction (AMI) [136]. Studies have nonetheless suggested side effects including coronary artery vasoconstriction [137]. Several large studies have shown either harm or lack of benefit from supplemental O₂ use in patients without hypoxaemia [138, 139]. A large cluster randomised controlled trial of > 40,000 patients with acute coronary syndrome (including patients with AMI) found no benefit with supplemental O₂ use overall, but evidence was inconclusive in patients with ST-elevation AMI [140]. Importantly, these trials included patients without hypoxaemia [138, 140]. Despite the lack of high-quality evidence, it appears prudent to avoid hypoxaemia (SaO₂ < 90%) in AMI patients.

Trials of intraoperative hyperoxia have mainly been performed in elective surgery to prevent surgical wound infection through increased tissue oxygenation [141, 142]. Initial enthusiasm was followed by larger trials with similar wound complication frequencies with F_iO₂ = 0.80 vs. 0.30 perioperatively [143, 144]. Concerns have been raised by shorter cancer-free survival in patients given F_iO₂ = 0.80 [145, 146]. A higher F_IO₂ is used to ensure adequate or, in some cases, supranormal end-organ oxygenation, although there is sparse evidence of benefit.

Both preoxygenation and high intraoperative F_IO₂ can cause resorption atelectasis [147], especially in patients with pulmonary comorbidity, as general anaesthesia itself reduces functional residual capacity and causes airway closure [148]. As ventilation-perfusion mismatch and shunt contribute to impaired oxygenation, use of F_IO₂ = 0.30–0.35 is therefore considered normal during general anaesthesia [149, 150]. F_IO₂ ≥ 0.80 caused significant atelectasis during preoxygenation, but this can be eliminated with a recruitment manoeuvre followed by 5-10cmH₂O PEEP [14], which clearly is not common practice. Failure to correct such iatrogenic atelectasis may trigger the use of excessive perioperative F_IO₂. In a large observational study [151], high intraoperative F_IO₂ was dose-dependently associated with major pulmonary complications and mortality after adjustment for all relevant risk factors. This association has not yet been confirmed in RCTs [152].

Based on a sub-group analysis in a systematic review, the World Health Organization proposed using F_IO₂ = 0.80 in all intubated patients to prevent postoperative wound infections [153]. This engendered controversial discussion [154, 155]. Most of the evidence for risks and benefits of hyperoxia during emergency surgery arise from RCTs of 385 laparotomy procedures and 210 open appendicectomies [143, 156]. While wound infections were significantly reduced with F_IO₂ = 0.80 in the appendicectomy study, the frequencies of surgical site infections, serious adverse events and mortality did not differ in the laparotomy trial [156, 157].

Acute perioperative patients should be carefully treated with respect to their ongoing medical conditions; most current evidence suggests greatest safety with O₂ titration to normoxaemia.

Current evidence suggests that PaO₂ > 300mmmHg should be avoided in most ICU patients. It remains uncertain whether there is a “sweet spot” PaO₂ target, which may vary for given clinical conditions. Systematic reviews using trial sequential analysis to take into account high vs. low bias risk found no effect (including all patients [39]) or increased mortality (including only ICU patients [157]) from higher oxygenation targets. Certainty evidence was low with futility for a 15% relative mortality risk increase. The currently recruiting “Mega-Randomised-Registry-Trial-Comparing-Conservative-vs.-Liberal-Oxygenation (Mega-ROX trial)” (CTG1920-01) in 40,000 patients should provide any “ideal target PaO₂”: The trial tests the hypothesis that conservative vs. liberal O₂ targets reduce mortality by 1.5% points in mechanically ventilated, adult ICU patients, i.e. 1,500 lives saved for every 100,000 patients treated. Since both conservative and liberal O₂ therapy may be best for certain patients, several parallel trials will evaluate pre-specified hypotheses in specific patient cohort patients accompanied by separate power calculations. For example, anticipating heterogeneity of treatment response, in septic patients or patients with acute brain pathologies (other than hypoxic brain injuries), the trial will test the opposite hypothesis that liberal (rather than conservative) O₂ will reduce mortality. Finally, the trial design cannot exclude that for some patient sub-groups, a different window of O₂ exposure is most suited.

So far, it appears prudent to target PaO₂ values within the normal range, i.e. carefully titrating PaO₂ to avoid both hypoxaemia and excess hyperoxaemia [158], particular as no clinically useful biomarker of O₂ toxicity is available, and data on the effects of hyperoxia on markers of oxidative stress are equivocal [10, 159‐161].