Hyperoxia for accidental hypothermia and increased mortality: a post-hoc analysis of a multicenter prospective observational study

This was a post-hoc analysis of a nationwide multicenter prospective observational study that was conducted by the Intensive Care with Extra Corporeal membrane oxygenation Rewarming in Accidentally Severe Hypothermia (ICE-CRASH) study group from December 2019 to March 2022 [17, 18]. The ICE-CRASH study included patients with accidental hypothermia at participating 36 tertiary care centers and was registered at the University Hospital Medical Information Network Clinical Trial Registry on April 1, 2019 (UMIN-CTR ID, UMIN000036132) prior to study initiation. The ICE-CRASH study was supported by the Japanese Association for Acute Medicine (approval no. 0005) and approved by the institutional review board for conducting research with human participants at the head institute of the ICE-CRASH study group (approval no. 18194 from Asahikawa Medical University). This study was conducted in accordance with the Helsinki Declaration, and written informed consent was waived due to the anonymity of the data.

The ICE-CRASH study enrolled consecutive patients aged 18 years or older with accidental hypothermia, including those with cardiac arrest, and hypothermia was defined as a core body temperature less than 32 °C measured at the emergency department (ED) on arrival. As there was no validated uniform rewarming strategy for accidental hypothermia during the study period, rewarming procedures were decided by an attending physician based on patient conditions such as hypothermia severity, vital signs, and the presence of cardiac arrest. Rewarming methods included blankets, warm parenteral fluids, warm baths, body cavity lavage, intravascular thermoregulated catheters, hemodialysis, and extracorporeal membrane oxygenation. Rewarming was typically initiated in the ED under cardiopulmonary monitoring and continued after intensive care unit (ICU) admission.

Data from the ICE-CRASH study (2019–2022) were reviewed retrospectively. Patients with accidental hypothermia were included if they were (1) 18 years old or older, (2) diagnosed with a body temperature less than 32 °C, and (3) had available arterial partial pressure of oxygen (PaO₂) data obtained at the ED. Patients who were in cardiac arrest when they arrived at the hospital were excluded because previous studies have suggested that hyperoxia can be harmful in post–cardiac arrest syndrome.

Patient data for the ICE-CRASH study were prospectively collected and entered into an online data collection portal at each hospital. Age, sex, Charlson comorbidity index, the activity of daily living (ADL), the etiology of hypothermia, the place of occurrence of hypothermia, transportation time from the scene to the hospital, vital signs on hospital arrival, presence of cardiac arrest on hospital arrival, rewarming methods, laboratory investigations and arterial blood gas assay that was obtained at the ED on arrival and at the ED or ICU after rewarming to 36 °C and was corrected by body temperature as appropriate at each institution, time from arrival to rewarming to 33 °C and 36 °C, Sequential Organ Failure Assessment (SOFA) score on admission, and mortality on the day of admission and during rewarming were all recorded. In addition, length of ICU and hospital stay, duration of ventilator use, cerebral performance category (CPC) at discharge, survival status 28 days after admission, and adverse events related to hypothermia/rewarming (ventricular fibrillation, hemorrhage, pneumonia, pancreatitis, and acute kidney injury) were all available.

According to previous research on hyperoxia in other diseases, hyperoxemia was defined as a PaO₂ level of 300 mmHg or higher [2, 4, 19]. Hyperoxia before the initiation of in-hospital rewarming was defined as hyperoxia at the ED on arrival. Trajectory of hyperoxia during rewarming was defined as average change in PaO₂ per hour until rewarming to 36 °C, that was calculated using PaO₂ before and after rewarming and the time duration of rewarming. Severe hypothermia was defined as a core body temperature of less than 28 °C [13, 20], and hemodynamic instability was defined as a systolic blood pressure (SBP) of less than 90 mmHg. The database lacked detailed indications for each rewarming procedure as well as hemodynamic status before, during, and after rewarming.

The primary outcome was 28-day mortality. Secondary outcomes included favorable neurological function at discharge (defined as a CPC of 2 or lower), ICU-, hospital-, and ventilator-free days to 28 days after admission, and the frequency of adverse events related to hypothermia/rewarming.

The primary outcome was compared between patients with and without hyperoxia using the Chi-square test as an unadjusted analysis. Then, inverse probability weighting (IPW) using propensity scores was conducted to adjust background characteristics between patients with and without hyperoxia [21, 22]. The propensity score for weighting was developed using a logistic regression model fitted with generalized estimating equations (GEE) to estimate the probability of hyperoxia exposure and account for within‐institution clustering [23]. Based on previous studies, relevant covariates were carefully selected from known or potential predictors for receiving supraphysiologic amounts of oxygen and predicting clinical outcomes in patients with accidental hypothermia (Additional file 1: Figure S1) [14, 20, 24‐26]. These covariates included age, sex, Charlson comorbidity index, ADL (independent/with limited help vs. considerably/completely dependent), etiology of hypothermia (intoxication, infection, and trauma), place of occurrence of hypothermia (indoor vs. outdoor), transportation time, vital signs on hospital arrival (Glasgow Coma Scale (GCS), SBP, heart rate, and respiratory rate), hypothermia severity, and arterial blood gas assay (lactate and base excess). Laboratory data on arrival, such as hematocrit, platelet count, prothrombin time, creatinine level, and glucose level, were also included as covariates because they are considered survival predictors in accidental hypothermia and are unaffected by PaO₂ on arrival [21, 26]. On the other hand, covariates related to rewarming information were not included in the model because hyperoxia exposure would affect such covariates [26]. Patients with missing covariates were excluded from the propensity score calculation. The discrimination ability of the propensity score was analyzed using the c-statistic [22]. The weight was calculated as the inverse of the propensity score of hyperoxia exposure. To avoid extreme weight based on propensity scores, patients with a propensity score of 0.05 or less or 0.95 or higher were excluded from the IPW analyses. The primary outcome was compared using the Chi-square test, and secondary outcomes were compared using odds ratios (ORs) or the Mann–Whitney U test [22].

Three sensitivity analyses were performed in order to validate the primary results. First, to validate results that were not dependent on the propensity score calculation, generalized estimating equation analysis with the logit link function was used to adjust patient backgrounds and differences in quality of care between participating hospitals [23]. Second, multivariate logistic regression was conducted with covariates for the propensity score calculation to confirm that the results were not dependent on the propensity score or within-institution clustering. Third, IPW was conducted with no restriction on the propensity score to validate that extreme weight truncation was appropriate [21, 22].

In addition, restricted cubic spline curves for estimating 28-day mortality by PaO₂ at the ED were created to identify any PaO₂ thresholds that affect the clinical outcomes of accidental hypothermia. For this, a generalized additive model was adopted. Then, to explore the ranges of hyperoxia that would affect outcomes, two different cut-offs were chosen from the spline curves and the hyperoxia was re-defined, with which the IPW analyses were repeated.

Moreover, as the effect of trajectory of hyperoxia would be expected to affect clinical outcomes, it was entered into a post-weight logistic regression model along with hyperoxia. In addition, another post-weight logistic model was analyzed, in which rewarming methods, time duration of rewarming, and the trajectory of hyperoxia were entered with hyperoxia.

Subgroup analyses were performed to investigate the relationships between hyperoxia, clinical characteristics, and 28-day mortality. Targeted subgroups were selected based on previous research into the clinical outcomes of patients with accidental hypothermia. The IPW analyses of the primary outcome were repeated in patient subgroups divided by age (< 65 vs. ≥ 65 years), presence of chronic cardiopulmonary diseases such as congestive heart failure and chronic lung disease, hemodynamic instability on hospital arrival (SBP ≥ 90 vs. < 90 mm Hg), and hypothermia severity (core body temperature < 28 °C vs. ≥ 28 °C). Subgroup analyses were also conducted in patients without hypoxia, defined as a PaO₂ level less than 60 mmHg.

Descriptive statistics are presented as a median (interquartile range (IQR)) or a number (percentage). The results were presented as a standardized difference and a 95% confidence interval (CI). The balance of covariates before and after weighting was evaluated with a standardized difference, in which less than 0.1 was considered insignificant [21]. The hypothesis was tested on the primary and secondary outcomes, with a two-sided α threshold of 0.05 considered significant. All statistical analyses were conducted using the IBM Statistical Package for the Social Sciences Statistics for Windows Version 28.0 (IBM Corp., Armonk, NY) and R Version 4.0.2 (R Foundation for Statistical Computing, Vienna, Austria).

Of the 499 patients with accidental hypothermia in the ICE-CRASH study, 338 adults had available PaO₂ data at the ED and arrived at participating hospitals without cardiac arrest; therefore, they were eligible for this study (Fig. 1).

In total, 65 patients (19.2%) had hyperoxia with PaO₂ levels of 300 mmHg or higher before initiating in-hospital rewarming. Patient characteristics are shown in Table 1. The median PaO₂ level was 366 mmHg in patients with hyperoxia and 135 mmHg in those without. Patients with hyperoxia had higher FiO₂, Carlson Comorbidity Index, and lactate on arrival, as well as lower GCS, SBP, body temperature, and base excess when compared to those without hyperoxia. Furthermore, a higher proportion of patients with hyperoxia had independent ADL and hypothermia indoors.

A propensity model for hyperoxia exposure was developed, and discrimination power was calculated, yielding a c-statistic of 0.699 (0.633–0.766). There were no patients excluded from IPW analyses due to missing covariates for propensity score calculation. Table 1 shows the patient characteristics after IPW with standardized differences, where differences in covariates such as patient demographics, comorbidities, hypothermia severity, and vital signs and laboratory data on arrival were successfully attenuated using the propensity score (standardized difference < 0.1). Propensity score distribution is also shown in Additional file 2: Figure S2.

Table 2 summarizes rewarming information after adjusting for patient backgrounds. The use of invasive rewarming devices, including thermoregulated catheters and hemodialysis, the time to rewarming to 33 °C and 36 °C, laboratory data after rewarming, the SOFA score, and mortality on the day of rewarming were comparable between patients with and without hyperoxia. The median PaO₂ and FiO₂ levels after rewarming were 90–110 mmHg and 0.2–0.3, respectively, and were comparable between patients with and without hyperoxia.

Patients with hyperoxia had significantly higher 28-day mortality than those without (25 (39.1%) vs. 51 (19.5%); OR 2.65 (95% CI 1.47–4.78); p < 0.001; Table 3), and similar results were obtained in the IPW analyses (34.0% vs. 23.8%; adjusted OR 1.65 (1.14–2.38); p = 0.008; Table 3). The three sensitivity analyses also showed a relationship between hyperoxia and increased 28-day mortality (Additional file 3: Table S1).

Furthermore, the restricted cubic spline curve of mortality prediction by PaO₂ revealed a convex downward curve of mortality odds as PaO₂ increased, with PaO₂ levels of approximately 60–250 mmHg having a lower risk of 28-day mortality among patients with accidental hypothermia (Fig. 2). In addition, based on the spline curve, 250 mmHg was chosen from the expected upper threshold of low risk of PaO₂ and 200 mmHg from in the middle of the range of low risk of PaO₂. Re-defined hyperoxia was associated with increased mortality only with PaO₂ ≥ 250 mmHg, not with ≥ 200 mmHg (Additional file 3: Table S1).

Moreover, in the post-weight logistic model, both hyperoxia and the trajectory of hyperoxia (average changes in PaO₂ per hour) were associated with increased 28-day mortality (adjusted OR 3.64 (1.92–6.89) for hyperoxia and 1.02 (1.01–1.03) for 1 mmHg/h decrease in the average change of PaO₂: slower normalization of PaO₂ from hyperoxia was associated with higher mortality). Another post-weight logistic model using in-hospital treatments, time duration of rewarming, and the trajectory of hyperoxia similarly revealed that hyperoxia was related to higher 28-day mortality (adjusted OR 2.54 (1.31–4.93)).

The secondary outcomes were summarized in Table 3. Hyperoxia was associated with fewer hospital-, ICU-, and ventilator-free days. In contrast, favorable neurological outcomes at hospital discharge and the frequency of adverse events related to hypothermia or rewarming were comparable between patients with and without hyperoxia, except for pancreatitis, which was more common in patients who experienced hyperoxia before rewarming than in those who did not (6.2% vs. 2.0%; OR 3.18 (1.24–8.12)).

In subgroup analyses (Table 4), a relationship between increased 28-day mortality and hyperoxia was observed in several subgroups: the elderly at 65 years of age or older, patients with chronic cardiopulmonary diseases, those without hemodynamic instability on hospital arrival, and those with severe hypothermia (OR 1.70 (1.15–2.50), 3.95 (1.45–10.74), 2.73 (1.64–4.56), and 1.69 (1.05–2.73), respectively).

In contrast, younger patients (< 65 years), those without chronic cardiopulmonary diseases, those with hemodynamic instability on hospital arrival, and those with non-severe hypothermia had comparable mortality regardless of hyperoxia exposure.

Furthermore, in the subgroup excluding patients with hypoxia (PaO₂ levels below 60 mmHg), hyperoxia was also associated with higher mortality (OR 1.81 (1.23–2.66)).