Hyperoxia

Categorical variables were compared using the chi-square test. Continuous variables were compared using student t-test or Wilcoxon rank-sum test, based on the distribution of the data. We used Spearman’s correlation coefficient (r) to assess the relationship between 1- and 6-hour PaO₂, and the corresponding FiO₂ and SaO₂. We used multivariable logistic regression analysis to identify what patient and management characteristics were associated with hyperoxia (see Supplemental Methods).
For the primary outcome, we calculated relative risk (RR) using multivariable generalized linear regression with a log link^{33 (link)} to test if exposure to hyperoxia during the initial six hours after ROSC was an independent predictor of poor neurologic function at hospital discharge. We a priori defined hyperoxia as PaO₂ > 300 mmHg on one or more ABG analyses, based on previously described definition for hyperoxia.^{6 (link), 7 (link), 10 (link), 13 (link)}A priori, we selected the following candidate variables for the regression model on the grounds that they were previously demonstrated to predict outcome in post–cardiac arrest patients: (1) age (decile); (2) initial cardiac rhythm [asystole or pulseless electrical activity (PEA) vs. ventricular fibrillation/pulseless ventricular tachycardia (VF/VT)];^{34 (link)} (3) metabolic acidosis (defined as one or more recorded base deficit ≤ -6 during the initial six hours after ROSC, based on previously published literature);^{35 (link)} (4) arterial hypotension (mean arterial pressure < 70 mmHg during the initial six hours after ROSC);^{21 (link)} (5) pre-arrest comorbidities (i.e., Charlson comorbidities index);^{36 (link)} (6) prolonged duration of CPR (CPR duration > 20 min);^{37 (link)} and (7) location of cardiac arrest (in- vs. out-of-hospital).^{38 (link)–41 (link)} Backward elimination with a criterion of p < 0.05 for retention in the model was used. Statistical interactions and collinearity were assessed. Goodness of fit of the model was evaluated with the deviance test. This analysis was repeated for both secondary outcomes. For the main analyses listwise deletion was used for missing co-variables. We also report results using multiple imputation for missing co-variables. These models used robust standard error and took into account the random effects at the institution (i.e. site of enrollment) level.
We performed several additional pre-planned sensitivity analyses for the primary outcome. First, we entered additional covariates beyond those pre-specified into a multivariable generalized linear regression model with a log link. Second, we assessed whether cardiac arrest location (pre-hospital or in-hospital) had different results. Finally, we performed a sensitivity analysis limited to only patients who survived to hospital discharge (detailed description of sensitivity analyses is discussed in Supplemental Methods).
We also examined the association between PaO₂ and outcome across different thresholds to define hyperoxia (i.e. PaO₂ > 100, 150, 200, 250, 300, 350, and 400 mmHg on one or more ABG analyses). We entered each threshold into a multivariable generalized linear regression model with a log link and calculated relative risks with 95% confidence intervals (CI) for poor neurological outcome adjusting for candidate variables retained in the original model. We graphed the relative risks with 95% CI and inspected the graph to assess if there was a threshold signal for neurological outcome over increasing PaO₂ cut points.
To reflect the duration of hyperoxia exposure during the initial six hours after ROSC, we used the first PaO₂ measurement to represent the PaO₂ exposure during the time from ROSC to the first ABG measurement. We then calculated the time intervals between ABG measurements and inferred that the PaO₂ remained constant at the level observed in the earlier measurement until the time point of the subsequent measurement (i.e. last value carried forward). Similar methodology to estimate PaO₂ exposure has been used previously.^{16 (link)} We then added up the total time patients had exposure to hyperoxia during the initial six hours after ROSC. To test the impact of duration of hyperoxia exposure, we entered duration of exposure as a continuous variable (calibrated for one hour) into a multivariable generalized linear regression model with a log link adjusting for the candidate variables retained in the original model. Given some subjects had ABG analyses ordered by treating physicians in addition to the protocol, we adjusted the model for the total number of ABG analyses obtained during the initial six hours after ROSC, as well as time to first ABG.