Study design and setting
This study is reported as per the Strengthening the Reporting of Observational Studies in Epidemiology guideline (Additional file 1). The current study was a retrospective analysis of data from the Japanese Association for Acute Medicine Out-of-Hospital Cardiac Arrest (JAAM-OHCA) registry that were collected between June 2014 and December 2017. This registry provides for the nationwide, multicenter, prospectively focused collection of pre-hospital and in-hospital data from patients with OHCA in Japan [12]. The registry included all OHCA patients who were transported to participating institutions. Pre-hospital data were obtained from the All-Japan Utstein Registry of the Fire and Disaster Management Agency, as previously reported [13]. In-hospital data were collected via an internet-based system by physicians or medical staff at each institution. The JAAM-OHCA registry committee integrated pre- and in-hospital data, as previously described [14].
Participants
In this study, we included adult patients (age ≥ 18 years) in the registry who were introduced to ECMO during cardiac arrest in the emergency room. Patients for whom no PaO2 data were available after ECMO initiation were excluded. In addition, we excluded patients who experienced hypoxia (initial measurement of PaO2 < 60 mm Hg) after the start of ECMO.
Data Collection
Patient demographics and pre-hospital factors were extracted from the JAAM-OHCA registry. The data was segmented as follows: age, sex, witness status (emergency medical service personnel or others), presence of a bystander who performed cardiopulmonary resuscitation, etiology of cardiac arrest (cardiac or non-cardiac), initial cardiac rhythm, pre-hospital adrenaline administration, pre-hospital airway management, pre-hospital shock delivery, and response time (time from call to scene arrival, time from scene to hospital arrival). In addition, in-hospital factors and outcomes were extracted as follows: cardiac rhythm on arrival, in-hospital shock delivery, in-hospital adrenaline administration, antiarrhythmic drug administration, transient return of spontaneous circulation before ECMO pump-on, time from hospital arrival to ECMO pump-on, time from hospital arrival to initial blood gas analysis after ECMO pump-on, initial blood gas analysis data (pH, PaO2, partial pressure of arterial carbon dioxide [PaCO2], bicarbonate ion concentration, lactate level) after ECMO pump-on, intra-aortic balloon pump use, percutaneous coronary intervention, targeted temperature management, and cerebral performance category 30 days after cardiac arrest.
Exposure and Definition
We divided the eligible patients into three groups according to their initial PaO2 levels after ECMO pump-on. The three groups were as follows: normoxia group, 60 mm Hg ≤ PaO2 ≤ 200 mm Hg; moderate hyperoxia group, 200 mm Hg < PaO2 ≤ 400 mm Hg; and extreme hyperoxia group, PaO2 > 400 mm Hg.
Outcome Measures
Outcomes were assessed by emergency physicians at participating hospitals 30 days after cardiac arrest. The primary outcome was a 30-day neurologically favorable outcome after cardiac arrest. A neurologically favorable outcome was defined as a cerebral performance category of 1 or 2. The cerebral performance categories included the following five outcomes: 1) good cerebral recovery, 2) moderate cerebral disability, 3) severe cerebral disability, 4) coma or vegetative state, and 5) death or brain death [15]. The secondary outcome was 30-day survival after cardiac arrest.
Statistical Analyses
Descriptive statistics were calculated for all variables of interest. Continuous variables are reported as medians and interquartile ranges (IQRs), while categorical variables are summarized using counts and percentages. Categorical variables in the three groups were analyzed using the Chi-square test, and continuous variables were analyzed using the Kruskal–Wallis test. Cubic splines were used to examine the potential nonlinear effects of PaO2 levels on 30-day neurologically favorable outcomes.
We used multiple imputations to compensate for missing data, and ten imputed datasets were generated [16]. Univariate logistic regression analysis was performed to calculate the crude odds ratio (OR) of the PaO2 level group for 30-day favorable neurological outcomes or 30-day survival after cardiac arrest.
Subsequently, we performed multiple propensity score analysis in the multivariate analysis in order to adjust and control for multiple independent variables [17]. A multiple propensity score is a conditional probability of patients being categorized into three or more groups given baseline covariates. Multiple propensity score analysis was applied to compare three or more groups [18]. First, we performed a multinomial logistic regression analysis by setting one of the three PaO2 groups as the dependent variable. The following covariates were used to calculate the multiple propensity scores: age, sex, witness, bystander administration of cardiopulmonary resuscitation, initial cardiac rhythm, pre-hospital shock delivery, pre-hospital adrenaline administration, pre-hospital advanced airway management, time from call to scene, time from scene to hospital arrival, etiology, cardiac rhythm on arrival, transient return of spontaneous circulation before ECMO pump-on, time from hospital arrival to ECMO pump-on, and time from hospital arrival to blood gas analysis.
Second, we performed a binomial logistic regression analysis to determine the adjusted ORs of the PaO2 level group for 30-day favorable neurological outcomes or 30-day survival after cardiac arrest, adjusting for multiple propensity scores and in-hospital variables, including PaCO2, bicarbonate ion concentration, lactate level, intra-aortic balloon pump use, percutaneous coronary intervention, and targeted temperature management.
For robustness, we performed the sensitivity analysis (multivariate logistic regression analysis for 1-month neurological favorable outcomes and 1-month survival after cardiac arrest) using data of the cases with all characteristics recorded (complete case analysis).
ORs and 95% confidence intervals (CIs) were calculated. All statistical tests were two-sided, and a p-value of < 0.05 was considered significant. All statistical analyses were conducted using R 3.3.1 (R Foundation for Statistical Computing, Vienna, Austria) and SPSS 24.0 for Mac (IBM Corp., Armonk, NY, USA).