Admission Oxygen Saturation and Mortality in Acute Pulmonary Embolism Patients: Observational Data From Large ICU Databases

The relationship between blood oxygenation and clinical outcomes of acute pulmonary embolism (APE) patients in intensive care unit (ICU) is unclear, which could be nonlinear. The study aimed to determine the association between admission pulse oximetry-derived oxygen saturation (SpO2) levels and mortality, and to determine the optimal range with real-world data. Patients diagnosed with APE on admission and staying in ICU for at least 24 hours in the Medical Information Mart for Intensive Care III (MIMIC-III) database and the eICU Collaborative Research Database (eICU-CRD) were included. Logistic regression and restricted cubic spline (RCS) models were applied to determine the nonlinear relationship between mean SpO2 levels within the rst 24 hours after ICU admission and in-hospital mortality, from which we derived an optimal range of SpO2. Subgroup analyses were based on demographics, treatment information, scoring system and comorbidities. oximetry-derived oxygen the studies in Epidemiology

mismatch between ventilation and perfusion, which contributes to hypoxaemia and threaten the life of patients. Therefore, supplemental oxygen including conventional oxygen supplementation, high-ow oxygen or mechanical ventilation is needed in treatment of APE patients with arterial oxygen saturation (SaO2) < 90% [3]. However, animal and human studies have shown that excessive oxygen can promote vasoconstriction, in ammation, and oxidative stress on pulmonary and cardiovascular systems, and may increase mortality risk in critically ill adults [4][5][6]. There are also reports that hyperoxaemia is common in intensive care unit (ICU), which was associated with increased in-hospital and in-ICU mortality [7][8][9][10]. In addition, a retrospective study has shown that the in-hospital mortality of APE patients in ICU requiring mechanical ventilation reached 41.0% compared with 8.0% in the non-ventilated group (P < 0.0001) [11]. Thus, blood oxygenation and mortality of APE patients in ICU may be nonlinear and have a U-shaped relation, which means that too low or too high blood oxygenation may correlate with higher mortality.
However, few studies support this hypothesis. Considering the poor prognosis of APE patients and the prevalent use of blood oxygenation in many risk factor studies, it is necessary to determine the association between blood oxygenation and mortality, which may help predict the mortality risk accurately and optimize oxygen therapy support of APE patients.
In critically ill patients with cardiorespiratory compromise, the blood oxygen level is commonly measured continuously by using peripheral pulse oximetry (SpO2), which helps detecting hypoxaemia and guiding oxygen therapy [12,13]. Interestingly, one study involving APE patients showed oxygen saturation in air < 88% at admission was an independent predictor of PE-related death with incremental prognostic value to the 2014 European Society of Cardiology (ESC) risk model [14]. However, the study de ned oxygen in arbitrarily de ned categories rather than as a continuous variable and were unable to de ne an optimal target range. A meta-analysis study indicated oxygen supplementation with higher versus lower fractions or oxygenation targets may increase mortality of patients admitted to ICU [15]. Furthermore, there is still a lack of available recommendations on the optimal range of SpO2 in the current APE guidelines [3,16].
Since hypoxaemia and hyperoxaemia are both related to adverse outcomes, it is necessary to determine the optimal range of SpO2, which helps minimize the competing risks of hypoxaemia and hyperoxaemia of APE patients in ICU, especially those who need oxygen therapy. A large-scale multicenter study was therefore required to elucidate oxygen saturation targets of APE patients to optimize clinical treatment and guide future research. This study is now feasible with the application of large data sources such as the Medical Information Mart for Intensive Care III database (MIMIC-III) and the eICU Collaborative Research Database (eICU-CRD) which are open-access databases of ICU patients [17,18]. In this study, we aimed to determine the nonlinear relationship between admission SpO2 levels and in-hospital mortality of patients with APE in ICU, and to determine the optimal range of oxygen saturation for clinical practice and future researches.

Data Description
Data were collected from the MIMIC-III (Version: 1.4) and the eICU-CRD (Version: 2.0) in accordance with the ethical standards of the institutional review board of the Massachusetts Institute of Technology (authorization code: 35655780) and with the 1964 Declaration of Helsinki and its later amendments [19]. Requirement for individual patient consent was waived because the project did not impact clinical care  and all protected health information was deidenti ed, which was approved by the institutional review   boards of Beth Israel Deaconess Medical Center. MIMIC-III covered 61,532 ICU admissions between 2001  and 2012  Eligible patients had to have typical symptoms suggestive of APE (de ned as dyspnoea, chest pain, presyncope or syncope, haemoptysis, atrial arrhythmias, electrocardiographic changes indicative of right ventricular strain-such as inversion of T waves in leads V1-V4, a QR pattern in V1, a S1Q3T3 pattern, and incomplete or complete right bundle branch block) [3,21]. Patients were excluded meeting the following criteria: repeat ICU stays; patients aged < 18 years; stayed in the ICU less than 24 hours; had no value for hospital mortality; had no value for SpO2 and other important medical data records. Patients staying less than 24 hours were typically admitted after elective surgery with very low mortality. These cases were removed because this would lead to prognostic de-enrichment while not providing a large enough exposure window for the effects of hyperoxaemia or hypoxaemia to become apparent.
The primary outcome was hospital mortality, with ICU mortality as a secondary outcome. The primary independent variable was SpO2, which were measured multiple times within the rst 24 hours after ICU admission. The mean values of SpO2 were used in our analysis as a measure of the central tendency of patients' conditions. SpO2 is usually measured hourly in MIMIC-III and eICU-CRD, which were veri ed and entered into a chart by a nurse.
The data were extracted from the databases with PostgreSQL (Version:10.14). The codes that support the MIMIC-III and eICU-CRD documentation and website are publicly available, and contributions from the community of users are encouraged (https://github.com/MIT-LCP/mimic-code, https://github.com/MIT-LCP/eicu-code). The variables in the study contained demographics, vital signs, comorbidities, scoring system, treatment information. Only the records of the rst 24 hours after ICU admission were included. For treatment information, the oxygen therapy subgroup using mechanical ventilation was included. The vasopressor subgroup (de ned as using any of dopamine, epinephrine, milrinone, norepinephrine, phenylephrine, vasopressin) were included. As extensive missing data could lead to bias, variables with over 20% missing values were excluded in the subsequent analyses. Correspondingly, single imputation was conducted for variables with less than 20% missing values based on the complete conditional speci cation and predictive mean matching method [22,23]. Each incomplete variable was imputed with by an independent model, which ensured the validity of the results.

Statistical analysis
Baseline characteristics of enrolled patients were presented and compared among groups of SpO2 < 95%, SpO2 within 95-98%, 98%<SpO2 by using either Kruskal-Wallis rank test or Pearson's chi-square test as appropriate. Continuous variables were characterized as median (interquartile range [IQR]), while categorical data were presented as count and proportion.
Both hypoxaemia and hyperoxaemia are associated with adverse outcomes suggest a nonlinear relationship between SpO2 and mortality. Therefore, restricted cubic splines (RCS) based on logistic regression models [24], were used to identify the nonlinear relationship between ICU admission SpO2 and mortality. According to the plot results, an optimal oxygen saturation was determined, and the total cohort was then divided into several subgroups with different SpO2 levels for the subsequent analyses.
Univariate logistic regression analysis was used to con rm the associations between SpO2 levels and mortality among the total cohort and different subgroups. For the calculation of estimated odds ratio (OR) in each group, 95-98% SpO2 was de ned as non-exposed group, while SpO2%<95%, 98%<SpO2 were de ned as exposed group. Results were presented as a forest plot with group sizes, estimated OR (95% con dence interval [CI]), P values for the independent variable. Furthermore, multivariate logistic regression analyses including the full logistic regression analysis and stepwise logistic regression analysis with backward elimination of the total cohort were conducted to identify the relationship between SpO2 levels and mortality. The grouping strategy was consistent with univariate logistic regression analysis. Results were presented as a table with OR (95% CI) and P values.
A two-tailed P value of less than 0.05 was set up to be statistically signi cant. All statistical analyses were conducted using R software (Version: 4.0.4).

Results
After application of the inclusion and exclusion criteria, the nal study cohort consisted of 1109 APE patients, of whom 129 (12%) patients died during hospitalization and 80 (7.2%) died in ICU. The detailed information on the inclusion and selection process was summarized in Fig. 1. In total, 32 variables were extracted from the MIMIC-III and eICU-CRD databases, among which 11 variables had missing values (Additional le 1: Table S1). The baseline characteristics of APE patients grouped by different SpO2 levels (SpO2 < 95%, SpO2 within 95-98%, 98%<SpO2) were summarized in Table 1. The total cohort with a median age of 64 (IQR: 51-75) comprised 542 (49%) male patients. Notably, patients with SpO2 < 95% and 98%<SpO2 both had higher hospital mortality than those of patients with SpO2 within 95-98% (SpO2 < 95% versus 95-98% SpO2 versus 98%<SpO2: 16% versus 10% versus 14%; P = 0.056). The relationship between admission SpO2 level and all-cause in-hospital mortality of APE patients was nonlinear, and the smooth curve tted between them was a U-shaped curve (P overall =0.0002, P nonlinear =0.0001), as shown in Fig. 2. Although low SpO2 levels correlated more strongly with mortality, high SpO2 levels were also associated with increased mortality. Based on the attest part of the U-shape in Fig. 2, we determined the range of SpO2 within 95-98% as the optimal oxygen saturation range. Furthermore, the total cohort was divided into three groups with different SpO2 levels: the groups with SpO2 < 95%, or SpO2 within 95-98%, or 98%<SpO2. For subgroup analyses of demographics, treatment information and scoring system in Fig. 3, there were 8 out of 15 subgroups that had a non-linear Ushaped relationship between admission SpO2 level and in-hospital mortality: age (including < 70 years, ≥ 70 years), female, medical ICU, other ICU, high risk of simpli ed Pulmonary Embolism Severity Index (sPESI), high risk of Pulmonary Embolism Severity Index (PESI), mild and moderate of Glasgow Coma Scale (GCS) (All P overall and P nonlinear <0.05). For subgroup analyses of comorbidities in Fig. 4, there were 4 of 9 comorbidity subgroups that had a non-linear U-shaped association between SpO2 level and inhospital mortality: infectious disease, cancer, congestive heart failure, hypertension (All P overall and P nonlinear <0.05). Except sepsis subgroup, the bottom of the curves for all subgroups appeared within the range of 95-98% SpO2.
Univariate logistic regression was performed to determine the association between different SpO2 groups and hospital mortality among the total cohort and subgroups (Fig. 5). Since there was no statistical signi cance about mortality between the reference group (95-98% SpO2) and the group with 98%<SpO2 from the results of univariate logistic analysis for the total cohort (Additional le 1:

Discussion
In this study, our analyses demonstrated a U-shaped relationship between ICU admission SpO2 levels and all-cause in-hospital mortality of patients with APE. Furthermore, our study also presented the lowest mortality for the mean of SpO2 within 95-98%, which could be the optimal oxygen saturation targets and bene t oxygen therapy for patients with APE. These results were consistent for demographics, treatment information, scoring system, comorbidity subgroups, and when ICU mortality was used in place of hospital mortality.
Pulse oximetry is universally used for monitoring respiratory status of patients in ICU, which assess SaO2 based on Beer-Lambert law, and the correlation between SaO2 and SpO2 is generally high which makes SpO2 signi cant in the continuous monitoring and titration of oxygen supplement in ICU [25,26]. In addition, arterial oxygen tension (PaO2) is commonly used to guide oxygen therapy in ICU as well.
Compared to PaO2, SpO2 has the advantages of low cost, noninvasive, and repetitive measurement of oxygenation. Furthermore, SpO2 is also clinically more relevant, for which the ventilator and inspired oxygen fraction settings are based on SpO2 changes rather than on intermittent arterial blood gas assays [27]. Additionally, using SpO2 to titrate supplemental oxygen is better than xed inspired oxygen fractions, which risk over-oxygenation in patients with narrow alveolar-arterial oxygen gradients and under-oxygenation in those with wide gradients. Because of the sigmoid shape of the oxyhemoglobin dissociation curve, oximetry may not detect hypoxaemia in patients with high PaO2 levels [28]. However, for the SpO2 range of 95-98%, the correlation between SpO2 and PaO2 would be fair, with little risk of underestimation of either hypoxaemia or hyperoxaemia [29].
Risk strati cation of patients with APE is mandatory for determining the appropriate therapeutic management approach. Blood oxygen saturation has been used as one of the important prognostic indicators of several APE risk scores such as PESI and sPESI [30,31]. However, few studies have explored the prognostic value of SpO2 level as a continuous variable of APE patients. In our study, the assessment of SpO2 level within the rst day after ICU admission could serve as a preliminary prognostic marker for short-term mortality among all APE patients, which could help stratify the risk of APE patients and personalize their treatment. Similar to our results, Kline et al. found that a room-air pulse oximetry reading < 95% at diagnosis was associated with a signi cantly higher probability of in-hospital complications (death, cardiogenic shock, or respiratory failure) from pulmonary embolism [32]. However, this study was composed of emergency department patients, the proportion of patients with SpO2 within 98-100% was less, which may not fully re ect the harm of hyperoxaemia in APE patients. Considering the U-shaped relationship between SpO2 and mortality, perhaps at least three groups of SpO2 level groups are required to explore its impact on the prognosis of APE patients.
Although oxygen therapy is recommended in patients with APE and SaO2 < 90%, there is still lack of strong evidence on the optimal oxygenation target for guiding the oxygen therapy in the current APE guidelines [3,16]. Our study presented the lowest mortality at a SpO2 within 95-98%. However, for the group with 98%<SpO2, the univariate logistic regression showed 100% SpO2 was associated with higher increased risk of all-cause in-hospital mortality compared to 95%-98% SpO2 (OR: 2.455; P = 0.01), but there was no signi cant difference between 99% SpO2 and 95%-98% SpO2 (OR: 1.033; P > 0.05), and then multivariate logistic regression analyses further presented that there was no signi cant difference between 99% SpO2 and the reference group (OR: 0.611; P > 0.05). This indicates that 99% SpO2 is not a risk indicator for APE patients which could serve as a buffer range for oxygen targets. Additionally, the SpO2 range with the lowest mortality in the mechanical ventilation subgroup was 97-99%, which also indicated 99% SpO2 was less risk and was consistent with the ndings by Zhou et al. that time spent oxygen saturation 95-99% is associated with reduced mortality in critically ill patients with mechanical ventilation [33]. However, considering that patients with 99% SpO2 still have a relatively high risk of hyperoxaemia, we cautiously set the upper limit of the optimal SpO2 target at 98% [29,34]. Besides, an upper level of 98% may reduce excessive use of high concentration oxygen therapy and avoid the potential risks of hyperoxia.
Notably, the British Thoracic Society (BTS) guideline recommends oxygen should be prescribed to achieve a target saturation of 94%-98% in patients with pulmonary embolism except for in those at risk of hypercapnic respiratory failure [5]. However, the recommendations about the use of oxygen in pulmonary embolism are based on expert opinion in BTS guideline and the evidence level is relatively low, lacking strong evidence support. Our ndings are consistent with the recommendations of BTS guideline and could provide it with a higher level of evidence support. Additionally, the Oxygen-ICU trial showed that patients who received conservative oxygen therapy with SpO2 levels between 94% and 98% had lower inhospital mortality and in-ICU mortality than those received conventional therapy with SpO2 levels between 97% and 100% (hospital mortality, 24.2% vs. 33.9%, P = 0.03; ICU mortality, 11.6% vs. 20.2%, P = 0.01), which also supports our optimal range of 95-98% for APE patients although this trial may overestimate the true treatment effect of conservative group [35,36]. Furthermore, a retrospective study involving ICU patients with at least 48 hours of oxygen therapy showed that the optimal range of SpO2 was 94-98%, which is also consistent with our ndings [37]. Our results are potentially impactful and targeting SpO2 between 95% and 98% might optimize survival for all APE patients in ICU. Since pulse oximetry is widespread and affordable, implementation of the 95-98% target would be feasible, even in resource-limited environments.
Several studies showed that overuse of oxygen therapy in ICU is common and is associated with adverse outcomes, including higher mortality, longer duration of mechanical ventilation and longer hospitalization [10,38,39]. Furthermore, from a dose-response relationship between arterial oxygen levels and outcomes, severe hyperoxia was more consistently associated with poor outcomes (hospital and ICU mortality) than mild hyperoxia [9]. On the other side, the hypoxaemia could be induced by APE, which contributes to the increase in pulmonary artery pressure and in severe cases it aggravates the hemodynamic impairment [40]. Thus, optimal oxygen targets are required to minimize and balance the competing risks of hypoxaemia and hyperoxaemia in APE patients. From a physiological standpoint, it is necessary to avoid hypoxaemia to maintain su cient oxygen delivery to important tissues, while avoiding hyperoxaemia is also necessary to avoid the formation and toxic effects of reactive oxygen species (ROS) [5,41]. Since our study included all APE patients, the results were applicable to APE patients with or without oxygen therapy. Our ndings could provide a more solid foundation for the selection of SpO2 targets in the treatment groups for future treatment and research on APE patients.
Several limitations of our study should be acknowledged. Firstly, the data were from the United States, and the results may not be fully applicable to other ICUs with different practices or resources. Second, even though we performed multiple subgroup and multivariate regression analyses, residual confounding could exist. Particularly, confounding by comorbidities would mean overestimation of the association between hyperoxaemia or hypoxaemia and mortality. Finally, the study is a retrospective study based on large ICU databases and shared the same aws as other observational studies. Thus, prospective cohorts were needed for further validation.

Conclusion
Among patients with APE, the relationship between admission SpO2 levels and in-hospital mortality followed a U-shaped curve. The lowest mortality was observed for an SpO2 range of 95-98%. Future randomized trial could adopt an SpO2 range of 95-98% as the reference target.   Restricted cubic spline plot of the relationship between mean SpO2 levels and the odds ratio for all-cause in-hospital mortality in APE patients. Solid red line represents the smooth curve t between variables. The shaded area around the spline curve represents the 95% con dence interval. The reference for the odd ratio was set at 1. The P value for overall association and P value for non-linearity are shown in the subtitle position. APE = Acute pulmonary embolism; CI = Con dence interval; OR = Odds ratio; SpO2 = Pulse oximetry-derived oxygen saturation. Restricted cubic spline plots of the relationship between mean SpO2 levels and the odds ratio for inhospital mortality in APE patients from different subgroups. The subgroup analyses were based on demographics, treatment information, scoring system. Solid red line represents the smooth curve t between variables. The shaded area around the splines curve represent the 95% con dence interval. The reference for the odd ratio was set at 1. The P value for overall association and P value for non-linearity are shown in the subtitle position. APE = Acute pulmonary embolism; CI = Con dence interval; GCS = Glasgow Coma Scale; ICU = Intensive care unit; OR = Odds ratio; PESI = Pulmonary Embolism Severity Index; sPESI = Simpli ed Pulmonary Embolism Severity Index; SpO2 = Pulse oximetry-derived oxygen saturation.

Figure 4
Restricted cubic spline plots of the relationship between mean SpO2 levels and the odds ratio for inhospital mortality in APE patients with different comorbidities subgroups. Solid red line represents the smooth curve t between variables. The shaded area around the splines curve represent the 95% con dence interval. The reference for the odd ratio was set at 1. The P value for overall association and P value for non-linearity are shown in the subtitle position. APE = Acute pulmonary embolism; CI =