Emergency department hyperoxia is associated with increased mortality in mechanically ventilated patients: a cohort study

This was a cohort study, using a database of patients that had mechanical ventilation initiated in the ED at a tertiary academic medical center (September 2009 to March 2016). The database was created as part of a clinical investigation that assessed outcomes associated with the implementation of ED lung-protective mechanical ventilation [24, 25]. All mechanically ventilated ED patients were screened for inclusion. The inclusion criteria were the following: (1) adult patients (age ≥18 years); (2) mechanical ventilation via an endotracheal tube; and (3) normoxia (partial pressure of arterial oxygen (P_aO₂) 60–120 mm Hg) on day 1 of ICU admission. The analysis was restricted to those patients with ICU normoxia, given the fact that (1) longer periods of exposure to hyperoxia in the ICU have been studied in the past; (2) this approach allowed us to better isolate a relatively brief hyperoxia exposure (i.e. in the ED) to test its association with outcome; and (3) the association between ED hyperoxia and outcome in mechanically ventilated patients had not been studied previously. Exclusion criteria were as follows: (1) death or discontinuation of mechanical ventilation within 24 hours of intubation; (2) chronic respiratory failure requiring mechanical ventilation; (3) presence of a tracheostomy; (4) transfer to another hospital; and (5) presence of acute respiratory distress syndrome (ARDS) while in the ED (defined by the Berlin criteria) [26]. This study was approved by the institutional review board under waiver of informed consent.

Demographic data, comorbidities, laboratory values, vital signs, illness severity, ED length of stay, and etiology of respiratory failure were collected. Data on treatments provided in the ED included the use of vasopressors and antibiotics and amount of intravenous fluid.

Mechanical ventilator settings provided in the ED were collected, along with gas exchange variables, plateau pressure, static compliance of the respiratory system, and driving pressure. Ventilator settings from the ICU were collected twice daily, for up to 2 weeks.

The definitions of comorbid conditions are provided in Additional file 1. Driving pressure (cm H₂O) was calculated as plateau pressure minus positive end-expiratory pressure (PEEP). Static compliance (mL/cm H₂O) of the respiratory system (C_RS) was calculated as:

The primary outcome was in-hospital mortality. Secondary outcomes were ventilator-free, ICU-free, and hospital-free days. Patients were followed until hospital discharge or death.

Patient characteristics were assessed with descriptive statistics, including mean (standard deviation (SD)), median (interquartile range (IQR)), and frequency distributions. Linear interpolation was used to deal with missing data on lactate values (n = 157 patients). The cohort was a priori categorized into three oxygen exposure groups based on P_aO₂ values obtained after intubation (i.e. in the ED, one arterial blood gas per patient). Hypoxia was defined as P_aO₂ < 60 mm Hg, normoxia as P_aO₂ 60–120 mm Hg, and hyperoxia as P_aO₂ > 120 mm Hg. Recognizing that there is no formal definition of hyperoxia, a P_aO₂ cutoff value of 120 mm Hg was used as it is congruent with the cutoff value used in other cohort studies that examined ICU hyperoxia exposure in a diverse cohort of mechanically ventilated patients (i.e. analysis not isolated to patients post cardiac arrest or those with stroke or traumatic brain injury) [8, 9, 27]. In a post hoc analysis, the hyperoxia group was further categorized into mild (P_aO₂ 121–200 mm Hg), moderate (P_aO₂ 201–300 mm Hg), and severe (P_aO₂ > 300 mm Hg) hyperoxia subgroups.

To assess predictors for the primary outcome of hospital mortality, categorical characteristics were compared using the chi-square test. Continuous characteristics were compared using analysis of variance (ANOVA) or the Kruskal-Wallis test depending on the distribution of the data. The Bonferroni correction was used to correct for multiple comparisons, and differences in P_aO₂ categories were considered statistically significant if P was < 0.017. Time (in days) to the primary outcome was assessed with the Kaplan-Meier survival estimate and log-rank test, comparing the normoxic and hyperoxic groups. A second Kaplan-Meier survival estimate was also calculated, which included the hyperoxic subgroups.

To determine independent predictors of mortality, a backward, stepwise, multivariable logistic regression model was used to evaluate death as a function of oxygen exposure group. Clinically relevant variables that were statistically significant in univariate analysis at P < 0.05 were candidates for model inclusion. The primary exposure of interest was ED P_aO₂. As fraction of inspired oxygen (F_iO₂) and PEEP are significantly linked to the exposure and statistically collinear with it, they were not entered into the multivariable model, neither was P_aO₂:F_iO₂ (expected differences based on the primary exposure group). Variables for inclusion or exclusion from the model were selected in sequential fashion based on the significance level of 0.10 for entry and 0.10 for removal. Normality, statistical interactions, and collinearity (i.e. variance inflation factor) were assessed, and the model used variables that were statistically independent. Model goodness of fit was assessed with the Hosmer-Lemeshow test and by examining residuals. Adjusted odds ratios (OR) and corresponding 95% confidence intervals (CI) are reported for the multivariable model, adjusted for all variables in the model. All tests were two-tailed, and a P value <0.05 was considered statistically significant.

The flow diagram of inclusions, exclusions, and the final study population are presented in Fig. 1. A total of 688 patients that were normoxic in the ICU were included in the final analysis.

Table 1 presents baseline characteristics of the study population related to the ED oxygenation group. The median (IQR) ED length of stay (hours) was 5.4 (3.5–7.9), with no difference in ED length of stay between the groups. There were no significant differences between the groups in comorbid conditions or indications for mechanical ventilation. Illness severity, as measured by Acute Physiology and Chronic Health Evaluation II (APACHE II) score, was higher in the ED hypoxia group. The most common reason for initiation of mechanical ventilation was sepsis.

Table 2 shows the ventilator variables in the ED and ICU. There were 300 patients (43.6%) who had exposure to ED hyperoxia, 350 (50.9%) who had exposure to ED normoxia, and 38 (5.5%) who had exposure to ED hypoxia.

Median (IQR) F_iO₂ was 80% (50–100) in patients in the ED hyperoxia group, which was significantly higher than the F_iO₂ in patients in the ED normoxia (68% (40–100)) and hypoxia (60% (40–80)) groups, P = 0.004. In the ED hyperoxia group, there was also a significantly higher median P_aO₂ (189 mm Hg (146–249)), and P_aO₂:F_iO₂ (270 (198–360)), P < 0.001.

There were no significant clinical differences between the groups with respect to day 1 ICU oxygenation and mechanical ventilation variables, though some statistical differences existed. Oxygenation and mechanical ventilation variables remained fairly static over the ICU stay, with little difference between day 1 variables and those calculated over the entire duration of mechanical ventilation (Additional file 2: Table S1).

In the entire cohort, hospital mortality occurred in 162 (23.5%) patients. Patients with ED hyperoxia had greater hospital mortality (29.7%), when compared to those with normoxia (19.4%) and hypoxia (13.2%). On Kaplan-Meier analysis, survival diverged significantly between the hyperoxia and normoxia groups (log-rank P = 0.021, Fig. 2).

The primary outcome analysis is shown in Table 3. The multivariable model was adjusted for age, gender, APACHE II, lactate, ED tidal volume, ED plateau pressure, ICU P_aO₂, and oxygen exposure group. After multivariable logistic regression analysis, ED hyperoxia was an independent predictor of hospital mortality (aOR 1.95 (1.34–2.85)). The complete multivariable model results are shown in Additional file 3: Table S2.

Secondary outcomes are also presented in Table 3. Compared to the normoxia group, there was a decrease in ventilator-free days (mean difference 3.7; 95% CI 2.0 to 5.4), ICU-free days (mean difference 3.5; 95% CI 1.9 to 5.1), and hospital-free days (mean difference 2.9; 95% CI 1.5 to 4.3) associated with ED hyperoxia, P < 0.001 for all.

The post hoc analysis examined mortality across hyperoxia subgroups. As the level of hyperoxia increased, hospital mortality was greater, (mild hyperoxia 28.0%, moderate hyperoxia 30.2%, severe hyperoxia 34.8%), though this was not statistically significantly different between the hyperoxia subgroups (Fig. 3). On Kaplan-Meier analysis, survival diverged significantly between the normoxia group and the hyperoxia subgroups (log-rank P < 0.001, Additional file 4: Figure S1).

This observational cohort study was conducted to examine the association between early hyperoxia exposure and clinical outcomes in mechanically ventilated patients with normoxia during their ICU stay. We found that the liberal use of oxygen in the ED was common, with a median (IQR) F_iO₂ of 70% (47–100), and a commensurate pre-ICU hyperoxia rate of 43.6%.

Pre-ICU exposure to hyperoxia in the ED was associated with a mortality rate of 29.7%, higher than in patients in both the hypoxia (13.2%) and normoxia (19.4%) groups. After controlling for confounders, including components of lung-protective ventilation (i.e. tidal volume, plateau pressure) and for baseline imbalances, hyperoxia remained an independent predictor of hospital mortality in multivariable analysis. Additionally, hospital mortality worsened across the hyperoxia subgroups. These data are congruent with prior studies of more prolonged exposure to hyperoxia in the ICU, yet to our knowledge this is the first study demonstrating an association between a comparatively brief early exposure to hyperoxia prior to ICU arrival, and worse clinical outcomes among mechanically ventilated patients [28].

Clinical guidelines recommend targeting oxygen saturations at 94–98% for most acutely ill patients, and recommend reducing oxygen therapy in patients with satisfactory oxygen saturation [1]. In an observational study of 101 mechanically ventilated patients in the ICU, patients spent > 70% of their total mechanical ventilation time with peripheral arterial oxygen saturation (S_pO₂) values of 96–100%, with mean P_aO₂ values of 144 mm Hg [29]. In 51 patients mechanically ventilated for > 48 hours, the majority of time was spent with S_pO₂ > 98% and 50% of all observations revealed hyperoxia [4]. A Dutch study of 126,778 arterial blood gas measurements from over 5000 mechanically ventilated ICU patients revealed a P_aO₂ > 120 mm Hg in 25% of the measurements, yet only 25% of the time was F_iO₂ decreased [2]. Finally, data from a single-center and multi-center study has shown that delivery of F_iO₂ > 90% is common and little titration of oxygen therapy occurs while patients are mechanically ventilated in the ED [21, 22]. Our current results, along with prior work in this area, further demonstrate that the liberal administration of oxygen is commonplace and hyperoxia is frequently tolerated, and extends these findings into the immediate post-intubation period in the ED.

With respect to clinical outcomes, there has been an increasing amount of data published in this domain over the last decade [6]. These data consist primarily of observational cohort studies with high heterogeneity among them in regards to methods, definition of hyperoxia, and the timing of the assessment of hyperoxia exposure [5, 6]. However, general themes show an association between hyperoxia and harm in the majority of studies. Hyperoxia seems to have a time-dependent and dose-dependent association with outcome, whereby early and severe hyperoxia in the ICU are particularly injurious, and with prolonged exposure [6, 9]. In addition to the production of free radicals, hyperoxia can cause vasoconstriction and a paradoxical decrease in oxygen delivery (within minutes) to prone regional areas (i.e. heart, brain, kidney) [30, 31]. These facts may help explain the findings in the current study, as the hyperoxia observed in the current investigation was in the most immediate post-intubation period, and much more pronounced (P_aO₂ 189 mm Hg (146–249) when compared to normoxic patients (P_aO₂ 88 mm Hg (76–101)). Additionally, the data suggest that our findings represent an effect of ED hyperoxia (and not carry-over into the ICU), as these patients were normoxic after ICU admission (Table 2), and throughout their ICU stay (Additional file 2: Table S1). This suggests that targeting normoxia from the initiation of mechanical ventilation may improve outcome, and makes complete physiologic sense.

A single-center randomized controlled trial (RCT) recently showed an absolute risk reduction in hospital mortality of 8.2% when normoxia was targeted in ICU patients; this is similar to the 8.3% difference in mortality between the normoxia and hyperoxia groups observed in this study, with survival curves of similar appearance [32]. However, another recent RCT failed to demonstrate any survival signal with a conservative oxygen target in mechanically ventilated patients, though this was a pilot feasibility study not powered for mortality [33]. Our study supports the notion that clinicians liberally administer oxygen in the immediate period of mechanical ventilation and routinely accept (and perhaps even target) hyperoxia. It also supports that this practice may be harmful. While no concrete recommendations can be given on the optimal P_aO₂ in mechanically ventilated ED patients, our data suggest little downside in immediately targeting normoxia, and provide data to suggest an opportunity to study this further in the ED.

This investigation has several limitations. It was a single-center study, and the oxygen administration practices and incidence of ED hyperoxia may not be externally valid or apply to other centers. Excess F_iO₂ administration in mechanically ventilated patients has been well-documented in the literature, suggesting these data are generalizable outside of our hospital [2, 29]. Given the study design, unmeasured confounders linked to hyperoxia could have accounted for the excess mortality in the hyperoxia group. For example, clinicians may err on the side of hyperoxia intentionally in conditions associated with lower oxygen delivery (i.e. anemia, low cardiac output), greater hypoperfusion, or higher illness severity. However, compared to the normoxia group, hyperoxic patients had similar hemoglobin levels, blood pressure, vasopressor use, lactate, illness severity, and fluid administration. Imbalances in baseline characteristics may have also influenced results, though these were adjusted for in our multivariable analysis. Nevertheless, the current results provide more evidence for the avoidance of excess oxygen, which is not providing any additional therapeutic benefit. While causation cannot be established with the design, the results are consistent with the majority of data on this topic. Furthermore, dose-response suggests causality, and greater mortality was observed across subgroups of increasing ED hyperoxia. The study also reflects real world practice in oxygen management, as it was conducted outside the auspice of a rigidly controlled randomized trial. There is potential for selection bias given the number of excluded patients in the study. The majority of exclusions were due to very early deaths or extubation within 24 hours; it is unlikely that acute hyperoxia would influence outcome in the acutely terminal patient or those stable enough to be extubated within 24 hours. Also, to better isolate an association between ED oxygen exposure and outcome, the analysis was restricted to those patients who were normoxic while in the ICU. Finally, the timing of arterial blood gas analyses was not obtained as part of a formal protocol therefore the exact duration of hyperoxia exposure is unknown. However, we observed marked differences between ED and day 1 ICU oxygenation data, suggesting that the primary exposure was driven by the ED. Furthermore, there were little differences between ICU day 1 data and data calculated over the entire time of mechanical ventilation, suggesting stability in oxygenation data over time (i.e. transient exposures in the ICU were less likely). However, without prospectively following all oxygenation and mechanical ventilation parameters closely in the ICU over time, there still exists the possibility that hyperoxia in the ICU affected some of our results, though our suspicion of this is low. In the future, the exact timing of all mechanical ventilator changes, oxygenation data, and arterial blood gas sampling should be documented to ensure that the exposure (and duration) has been reliably determined.