The Association Between Arterial Oxygen Level and Outcome in Neurocritically Ill Patients is not Affected by Blood Pressure

We retrospectively screened the Finnish Intensive Care Consortium (FICC) database for adult (age ≥ 18) patients admitted to five university-affiliated tertiary referral hospitals’ ICUs in Finland in the years 2003–2013 with various types of brain injury [13]. FICC coordinates an intensive care benchmarking program and a multicenter database that collects data from several Finnish ICUs; TietoEvry (Kuopio, Finland) maintains the central database. FICC was established in 1994, and the initial aim was to improve intensive care quality in Finland. We included patients with different types of brain injury: the admission diagnoses included TBI, subarachnoid hemorrhage (SAH), intracranial hemorrhage (ICH), acute ischemic stroke (AIS), and resuscitated CA. We defined the diagnoses of TBI, ICH, SAH, AIS, and CA according to the Acute Physiology and Chronic Health Evaluation III (APACHE III) and the International Classification of Diseases and Related Health Problems, 10th Revision [14, 15]. CA patients included both in-hospital and out-of-hospital arrests admitted to the ICU. We grouped ICH and AIS patients together due to the small number of participants in each group. We only included patients on mechanical ventilation support during the initial 24 h in the ICU. We excluded readmissions. All participating ICUs followed internationally recognized treatment guidelines pertaining at the time. The study included no interventions. Data about admission diagnosis, premorbid physical performance, arterial oxygen value (PaO₂) and other physiological data, and Simplified Acute Physiology Score II (SAPS II) components and scores were collected from the FICC database [16, 17]. Data were extracted in September 2016. The study data have previously been used in two studies reporting healthcare-associated costs after various types of brain injury, and data have partially been used in a study reporting associations between hyperoxemia and mortality after TBI [9, 18, 19]. The ethics committee of the Operative Division of Helsinki University Hospital, the Finnish National Institute for Health and Welfare, and the Social Insurance Institution of Finland (Kela) approved the study.

Categorical data are presented as counts and percentages and continuous data as medians with interquartile ranges (IQR). For mortality prediction, we constructed a multivariable logistic regression model, including age, admission diagnosis, independent premorbid physical performance (fully capable in self-care/dependent on help), intracranial pressure (ICP) monitored (yes/no), any vasoactive used (yes/no), and severity of disease (a modified SAPS II score excluding points for age, type of admission, oxygenation, and systolic blood pressure). In this model, we included categorized PaO₂, MAP tertiles, and their interaction term PaO₂*MAP. We chose the confounders to the model by clinical relevance and used the variance inflation factor to control for multicollinearity. First, we analyzed all the diagnosis cohorts together in one multivariable regression model, and, second, we conducted sensitivity analyses with diagnosis cohorts separately (TBI, CA, SAH, and ICH + AIH). In the sensitivity analyses, the TBI model was further adjusted for the Marshall classification of TBI based on the results of the initial computer tomography scans [21]. The results from the regression analyses are presented as odds ratios (OR) with 95% confidence intervals (95% CI) and p values shown in tables. Additionally, we visualized 1-year mortality with Kaplan–Meier curves. We inspected functional outcome with multivariable logistic regression model similar as with mortality prediction but without the variable premorbid physical performance (analysis only included pre-admission independent patients). Last, we wanted to demonstrate the independent relationship between PaO₂ and predicted probability of 1-year mortality in MAP tertiles in different diagnosis cohorts. We calculated the predicted probability of mortality with multivariable logistic regression analysis separately in every diagnosis cohort with the following variables: age, premorbid physical performance, ICP monitoring use, vasoactive use, modified SAPS II score, and TBI Marshall classification when applicable. We used locally weighted scatterplot smoothing (Loess) curves to visualize the relationship between PaO₂ and the predicted probability of 1-year mortality. In the scatterplot, all PaO₂ values > 40 kPa were coded as 40 kPa, to avoid few high values’ strong impact to the curve shape. We tested the model accuracy with all regression analyses with receiver operating characteristic (ROC) curves and determined the area under the curve (AUC). We used the Hosmer–Lemeshow Ĉ test to evaluate the model fit. Before proceeding with any statistical method, we ensured that the assumptions of the method in question were appropriately met. We performed all analyses with SPSS Statistics 25.0 for Mac OS (IBM Corp, Armonk, NY, USA).

The FICC database included between 2003 and 2013 14,196 adult ICU-treated patients with traumatic or hypoxic ischemic brain injury. We excluded 2138 patients with CA during intensive care from the study. After excluding patients with no ventilation assistance and patients with missing data, the study sample included 8290 patients (Fig. 1). Frequencies by admission diagnoses were 1974 (24%) with TBI, 3446 (41%) with CA, 1135 (14%) with SAH, and 1735 (21%) with AIS or ICH. The median age was 61, and 2571 (31%) of the study population were female. Altogether, 3912 (47%) patients died within a year of the insult. Regarding survivors, 2704 (62%) survived as independent and 1674 (38%) remained permanently disabled. According to our threshold for hyperoxemia (PaO₂ > 18.3 kPa), 836 (10%) patients were hyperoxemic, and 5340 (64%) received some vasoactive medication. The baseline characteristics for all patients and patients by diagnosis groups are presented in Table 1 (found at the end of the manuscript). Additionally, we determined PaO₂ and MAP values medians and interquartile ranges by admission year to confirm comparability through study period (2003 to 2013), see results from Additional Table 1.

In the multivariable regression analyses, PaO₂ was not an independent predictor of 1-year mortality: the OR for hyperoxemia versus normoxemia was 1.16 (95% CI 0.85–1.59) and for hypoxemia versus normoxemia 1.24 (95% CI 0.96–1.61). Higher MAP was associated with lower mortality, OR for MAP 60–68 mmHg versus MAP < 60 mmHg was 0.73 (95% CI 0.64–0.84) and OR for MAP > 68 mmHg versus MAP < 60 mmHg 0.80 (95% CI 0.69–0.92). Additionally, higher age, dependent premorbid physical performance, and more severe disease (higher modified SAPS II score) were independent predictors of higher mortality (Table 2]. There were no significant interactions between groups of PaO₂ and MAP (the interaction term PaO₂*MAP was nonsignificant). See Table 2 for all multivariable analysis results of mortality. In the sensitivity analyses, when testing possible subgroup effect in the different diagnosis cohorts, the interaction PaO₂*MAP was not statistically significant in any of the diagnosis cohorts. Further PaO₂ was not an independent predictor of mortality in any of the diagnosis cohorts. MAP < 60 mmHg was associated with higher mortality compared to MAP 60–68 mmHg in all other admission diagnosis cohorts except for ICH + AIS. More severe disease and higher age predicted higher mortality in every diagnosis cohort. All results for the sensitivity analyses in disease cohorts are found in Additional Table 2. Results with 90-day mortality were consistent with 1-year mortality (Additional Table 3).

The Kaplan–Meier analysis was used to compare mortality in different oxygen groups between MAP tertiles. One-year mortality was highest in the hypoxemia group in all MAP tertiles: 294 (70%) of the hypoxemic cases with MAP < 60 mmHg, 136 (50%) cases with MAP 60–68 mmHg, and 89 (50%) cases with MAP > 68 mmHg were dead at 1 year after the insult. Mortality was roughly equal between normoxemia and hyperoxemia groups in all MAP tertiles, although patients with MAP < 60 mmHg had overall higher mortality compared to MAP 60–68 and > 68 mmHg. The Kaplan–Meier results were statistically significant: the Logrank p value was < 0.001 for the analysis in all hemodynamic conditions (Fig. 2).

After exclusion of all patients being disabled prior to admission, 7386 patients remained in the analysis with the functional outcome. Again, higher age and more severe disease predicted lower probability of independent functional outcome. PaO₂ was not independently associated with the functional outcome. The middle MAP (60–68 mmHg) predicted higher probability of independent functional outcome compared to MAP < 60 mmHg. The interaction term PaO2*MAP was not statistically significant. Results for multivariable analysis with the functional outcome are found from Additional Table 4.

The relationships between PaO₂ and the predicted probability of 1-year mortality were visualized with Loess curves in every diagnosis cohort separately for different MAP tertiles (Fig. 3). Sixteen PaO₂ values > 40 kPa were coded as 40 kPa in the analyses. There were no dramatic differences in the relationships between PaO₂ and predicted mortality between MAP tertiles (similar curve shapes) in any diagnosis cohorts. Thus, the association between PaO₂ and mortality is constant regardless of MAP. Yet, within all diagnosis cohorts, there was a minor trend toward higher mortality with lower PaO₂ (PaO₂ about < 10 kPa). The effect was most visible in the lowest MAP (< 60 mmHg). Overall in SAH and ICH + AIS cohorts the predicted mortality was slightly higher with hyperoxemia compared to normoxemia but the small amount of cases with hyperoxemic PaO₂ values makes the association weak.

In this large retrospective cohort analysis of patients with various type of brain injury, we found no association between the presence of hyperoxemia and improved long-term outcome in patients with low MAP during the first 24 h. Low MAP (< 60 mmHg) was an independent predictor of mortality, but the effect was not alleviated in patients with higher oxygen. Our additional analysis focusing on long-term functional outcome as a surrogate marker of neurological recovery resulted in similar findings. The current analysis, acknowledging its retrospective design and clear limitations to show causality, nonetheless, does not support the hypothesis that higher blood oxygen content in neurocritically ill patients could compensate for the harmful effects of impaired systemic circulation and low MAP.

Multiple studies have been conducted on the associations between oxygen and outcome in patients with different types of brain injury. In TBI patients, hyperoxemia has been associated with worse survival with extreme and better survival with moderate hyperoxemia [3, 9, 22, 23]. However, in a recent meta-analysis of hyperoxemia after TBI, including six epidemiological studies, hyperoxemia did not influence mortality [24]. Extreme hyperoxemia after SAH has been found to increase mortality and is claimed to induce or exacerbate delayed cerebral ischemia [25‐27]. ICH and AIS are regularly studied together: one study found an association between hyperoxemia and higher mortality, whereas another did not [6, 28]. One meta-analysis studied hyperoxemia and mortality after SAH, ICH, or AIS and concluded that when all three cohorts were inspected together, no association existed between high oxygen and outcome. Additionally, another meta-analysis inspected SAH and AIS in separate cohorts and hyperoxemia was associated with higher mortality in AIS but not in SAH patients [24, 29]. In a large meta-analysis of 25 randomized controlled trials with 16 037 acutely ill heterogenous ICU-treated patients, Chu et al. compared conservative versus liberal oxygen therapy [30]. In individual studies, the results were heterogenic, but the combined effect suggested benefit with conservative compared to liberal oxygen therapy. None of the previous studies have in any detailed fashion studied the differential association between the oxygen and MAP with outcome. Guidelines on oxygenation and blood pressure after acute neurologic injury are debatable, and lack high-quality evidence base. For TBI specific PaO₂ or SpO₂ targets are not given, only monitoring of jugular venous saturation and PbO₂ is recommended without specific targets [31]. For ischemic stroke SpO₂ > 94% and for CA SpO₂ 94–98% are recommended and routine supplemental oxygen advised not to be used if SpO₂ -target is met [32, 33]. Loose blood pressure targets are specified after TBI (systolic blood pressure, SBP > 100 mmHg), CA (MAP > 65 mmHg and SBP > 90 mmHg), and SAH (< 160 mmHg) [33‐35].

It is well recognized that aerobic metabolism is reduced in injured brain areas and oxygen therapy has been claimed to restore the brain aerobic function [7, 36, 37]. An increase in PaO₂ enhances oxygen diffusion and leads to higher brain tissue oxygen (PbO₂) even in the setting of normal cerebral perfusion pressure and MAP [12]. High PaO₂ has also been shown to cause vasoconstriction which can reduce brain edema and enhance brain perfusion [36, 38‐41]. Increasing PaO₂ could be of theoretical benefit especially if adequate cerebral perfusion pressure is not achieved and, indeed, the use of 100% oxygen in the severely shocked patient (i.e., during severe hemorrhage) is common practice, even though conclusive evidence in humans is lacking. Early onset NBO therapy after AIS has been claimed to preserve the hypoperfused core infarction area and penumbra [42, 43]. In a large randomized clinical trial, routine low-dose supplemental oxygen (2–3 l/min) administered in the ICU continuously or at night only for the initial 72 h after AIS did not improve outcome, but NBO therapy was concluded to be safe [11]. It is, however, largely unclear if the injured brain can utilize supranormal oxygen levels [12, 44]. Furthermore, by far the known hyperoxemia mediated changes to cerebral vascular tension, perfusion, and tissue are partly opposing and true net change in oxygen delivery to the brain site in need can remain near zero [36, 38, 40, 45].

We did not find any clear association between oxygen level and MAP during the first 24 h and outcome in patients with various types of brain injury. The “optimum” PaO₂ during the initial ICU treatment after brain injury, despite the hemodynamic conditions, seems to be near the physiological level or slightly but not greatly above it, which supports the findings of several studies indicating better outcomes of moderate but not extreme hyperoxemia exposure after brain injury [4, 6, 9, 10, 22]. Some clinical studies have been conducted in patients treated following cardiac arrest, but thus far the evidence is conflicting and larger trials are needed in neurocritically ill patients.

We acknowledge several strengths of our study. Our sample of brain injury patients is large and includes patients from several diagnosis cohorts with differences in initial disease pathophysiology but similarities in injury responses after the insult [46]. In addition, patients were treated at multiple centers in Finland, increasing validity. We used data from the Social Insurance Institution of Finland (Kela) database, including data on the disability allowance of study participants 1 year after admission, as a surrogate marker for long-term functional outcome. All Finnish residents belong to the social insurance system and data on possible dates of death were available from Finnish population register for all patients; hence, loss of follow-up in this study is nonexistent.

However, we also recognize limitations. Due to limitations in the used dataset, we only had a single PaO₂ value for our analyses (the value recorded in connection with the lowest PaO₂/FiO₂ ratio during the first 24 h in the ICU) and could not inspect actual oxygen burden over time. However, we have confirmed in a previous study that in mechanically ventilated TBI patients the PaO₂ used in the current study correlates well with the mean PaO₂ of the first 24 h in the ICU [9]. Further on, this same strategy of categorizing oxygenation status with just one PaO₂ value has also been used in scientific studies focusing on CA patients [47]. In addition, our threshold for hyperoxemia (> 18.3 kPa) was set relatively low compared to some previous studies which may mask some of the effects of severe hyperoxemia [4, 22, 23]. However, the used PaO₂ value is likely the lowest measured during the first 24 h of ICU care, indicating that other PaO₂ measurements during the 24-h period were higher. For some patients, this study can misleadingly group a patient as hypoxemic based on only one temporal low PaO₂ value. Also, we grouped ICH and AIS patients together due to the small number of participants in each group; a similar grouping is been undertaken in other studies [6]. We only had the lowest MAP value of the first 24 h in the ICU and therefore cannot draw conclusions about the duration of hypotension and its effects. Hence, used MAP levels should not be seen as absolute indicators for actual blood pressure levels over time but as surrogate markers of the hemodynamic status over the first 24 h. There is, however, evidence that the lowest MAP measured during the initial hours in the ICU is associated with outcome in patients after CA as well as in patients with circulatory shock related to sepsis and acute pancreatitis [48‐50]. Last, at the time of the data collection, arterial CO₂ was not included in the FICC database, and we could not consider it as a confounding factor in this study although it to a great degree impacts on cerebral perfusion pressure.