Oxygen Sweet Spot for Brain-Injured ICU Patients Is Wider Than Guidelines Suggest

For decades, intensive care clinicians have hovered over arterial blood gas results in patients with devastating brain injuries, trying to decide exactly how much oxygen is enough — and how much is too much. A landmark analysis published in Neurocritical Care now offers the most precise answer yet, drawing on more than 13,000 mechanically ventilated patients across three large international datasets to define oxygen thresholds that minimize the risk of death in the intensive care unit. The findings, generated with a sophisticated Bayesian statistical framework, suggest that the safe zone for arterial oxygen tension is considerably wider than current European guidelines recommend, while confirming that extreme oxygen levels on either end of the spectrum carry real danger.

The study, led by Lavienraj Premraj and Christopher Camarda of Monash University and Monash Health in Australia, together with an international team of neurocritical care investigators, focused on patients with non-traumatic acute brain injury: intracerebral hemorrhage, aneurysmal subarachnoid hemorrhage, and acute ischemic stroke. These conditions strike more than a million people worldwide each year, and many of the most severely affected require invasive mechanical ventilation. In the ICU, the management of oxygenation is a daily balancing act. Too little oxygen risks starving vulnerable brain tissue — the penumbral zones surrounding a hemorrhage or an infarct — of the fuel it needs to survive. Too much oxygen, a condition known as hyperoxia, has been repeatedly linked in observational studies to oxidative stress, vasoconstriction, and worse outcomes.

Despite the stakes, the specific targets have remained stubbornly uncertain. The European Society of Intensive Care Medicine consensus panel currently recommends maintaining the partial pressure of arterial oxygen, or PaO2, between 80 and 120 millimeters of mercury in ventilated patients with brain injury. But that recommendation rests on limited prospective evidence. Major randomized trials, including ICU-ROX, UK-ROX, and most recently the LOGICAL trial in cardiac arrest patients, have generally failed to show that conservative oxygen strategies improve survival or neurological recovery, leaving clinicians without a clear mandate. None of those trials monitored oxygen directly inside brain tissue, and none was designed to pinpoint the PaO2 level associated with the lowest mortality in brain-injured patients specifically.

To break this impasse, the research team turned to Bayesian inference, a statistical approach that formally combines prior evidence with new data rather than treating each dataset in isolation. The final model was fitted to the Australian and New Zealand Intensive Care Society Adult Patient Dataset, a binational registry capturing 99 percent of Australian ICU admissions, which yielded 11,264 eligible ventilated patients: 4,846 with intracerebral hemorrhage, 4,319 with aneurysmal subarachnoid hemorrhage, and 2,099 with acute ischemic stroke. Crucially, the priors — the starting assumptions of the Bayesian machinery — were derived from two independent cohorts: the MIMIC-IV critical care database from the United States, contributing 1,789 patients, and the European ENIO observational study, contributing 544. The model was fitted sequentially, with posterior estimates from each cohort updating the priors for the next, allowing effect estimates to accumulate transparently across populations without assuming that any single dataset tells the whole story.

The analysis adjusted for a parsimonious set of clinically relevant confounders, including age, sex, Glasgow Coma Scale score on admission, and chronic cardiovascular or respiratory comorbidity. In the final model, estimates were further adjusted using the Australia and New Zealand Risk of Death score, with the oxygen-related components of the score deliberately removed to avoid statistical circularity, and random intercepts for hospital site and admission year accounted for institutional and temporal clustering. Thresholds were then identified using a region of practical equivalence: the widest continuous interval of PaO2 values within 4 percent of the minimum predicted mortality, a criterion that had to hold in 99 percent of the posterior samples. This stringent definition was designed to ensure that the resulting ranges were robust rather than artifacts of statistical noise.

The results revealed a striking heterogeneity across brain injury subtypes. For subarachnoid hemorrhage, the relationship between early PaO2 and ICU mortality retained a pronounced U-shape even after adjustment, with mortality rising steeply at both low and high oxygen tensions. The optimal range for these patients was 100 to 190 millimeters of mercury. For acute ischemic stroke, the curve was flatter but the optimal band was narrower still, at 100 to 160 millimeters of mercury. Intracerebral hemorrhage proved the most forgiving, with a broad safe zone stretching from 80 to 230 millimeters of mercury. In every case, the derived ranges were substantially wider than the 80 to 120 millimeter window endorsed by the ESICM, yet patients managed within these broader ranges showed essentially the same probability of ICU mortality as those falling within the narrower guideline-defined normoxia.

The danger zones, by contrast, were consistent and sobering. Extreme hyperoxia, defined as a PaO2 above 299 millimeters of mercury, was associated with increased ICU mortality across all three brain injury subtypes, with posterior probabilities ranging from 84 percent for intracerebral hemorrhage to 99 percent for subarachnoid hemorrhage. Hypoxia below the lower limit of the safe range also carried elevated risk, particularly in subarachnoid hemorrhage patients, where the odds of ICU death rose by roughly 28 percent. Even mild hyperoxia — oxygen levels above the upper boundary of the derived safe range but below the extreme threshold — was very likely harmful in patients with subarachnoid hemorrhage and ischemic stroke. Exploratory analyses using day-three blood gas values from the MIMIC-IV and ENIO cohorts reinforced these patterns, and showed that patients who developed or persisted in hyperoxia between day one and day three fared worse than those who avoided it entirely.

The mechanistic story behind these numbers is biologically plausible. In subarachnoid hemorrhage, elevated arterial oxygen may amplify an already substantial oxidative burden, promoting cellular and endothelial injury that increases the risk of vasospasm and delayed cerebral ischemia, both of which are strongly linked to death. In ischemic stroke patients treated with mechanical thrombectomy, high oxidative stress is already present, and excess oxygen may worsen reperfusion injury rather than rescue the penumbra. Notably, the study’s finding that higher PaO2 does not reliably translate into better brain tissue oxygenation echoes mechanistic work showing that arterial oxygen content is a poor proxy for oxygen delivery to ischemic tissue, particularly when cerebral autoregulation is disrupted.

The authors are careful to spell out the limitations. This was an observational study, so causation cannot be proven, and residual confounding may persist despite robust adjustment. The analysis relied on a single representative PaO2 value from the first 24 hours rather than continuous oxygen exposure, and none of the datasets included direct measurements of brain tissue oxygen tension. Functional neurological outcomes — arguably the endpoint that matters most to patients and families — could not be assessed. The choice of the 4 percent mortality band for defining the safe zone is, as the authors acknowledge, ultimately arbitrary, though they imposed a high standard of certainty by requiring consistency across 99 percent of posterior samples. Randomized trials, including the ongoing Mega-ROX Brains program, will be needed to confirm whether these liberalized ranges are truly safe.

Even with those caveats, the study represents a genuine advance in a field where practice has long outpaced evidence. By triangulating three international datasets and nearly 14,000 patients, it provides the first empirical evaluation of existing PaO2 recommendations in ventilated brain-injured patients, and it does so with a statistical framework that honors uncertainty rather than hiding it. For intensivists, the practical message is twofold: the anxiety-inducing narrowness of the current guideline window may be unnecessary, particularly for patients with intracerebral hemorrhage, but the upper limits are not negotiable — pushing arterial oxygen above roughly 300 millimeters of mercury is very likely to harm, whatever the underlying brain pathology. As neurocritical care moves toward personalized, physiology-guided ventilation, this work offers a data-driven map of where the true safe harbor lies.

Subject of Research: Optimal arterial oxygen tension thresholds in mechanically ventilated patients with non-traumatic acute brain injury

Article Title: Early Arterial Oxygen Tension Thresholds for Mechanically Ventilated Patients with Non-traumatic Acute Brain Injury: A Bayesian Analysis of Three Large Neurocritical Care Datasets

Article References: Early Arterial Oxygen Tension Thresholds for Mechanically Ventilated Patients with Non-traumatic Acute Brain Injury: A Bayesian Analysis of Three Large Neurocritical Care Datasets. (n.d.). https://doi.org/10.1007/s12028-026-02659-0

Image Credits: AI Generated

DOI: 10.1007/s12028-026-02659-0

Keywords: arterial oxygen tension, PaO2 thresholds, acute brain injury, mechanical ventilation, neurocritical care, subarachnoid hemorrhage, intracerebral hemorrhage, ischemic stroke, Bayesian analysis, hyperoxia, ICU mortality, ESICM guidelines