Lower or higher oxygenation targets in the intensive care unit: an individual patient data meta-analysis

This IPDMA of the HOT-ICU and HOT-COVID trials was conducted in accordance with the protocol published before inclusion of the last patient in the HOT-COVID trial [20]. Any deviations from the protocol are explicitly stated in the electronic supplementary material (ESM) 2. The results are presented in accordance with the Preferred Reporting Items for a Systematic Review and Meta-Analysis of Individual Participant Data (PRISMA-IPD) [21]. The PRISMA-IPD checklist is presented in the ESM 1 (eTable 1).

The HOT-ICU and HOT-COVID trials were investigator-initiated, multi-centre, parallel-group, randomised clinical trials. Both utilised centralised randomisation with computer-generated concealed assignment sequences in permuted blocks of varying sizes with the HOT-ICU trail being stratified for trial site, chronic obstructive pulmonary disease (COPD), and active haematological cancer, and the HOT-COVID trial being stratified for trial site. For both, detailed protocols and statistical analysis plans were published prior to trial completion [17, 18, 22].

The HOT-ICU trial was conducted from June 20, 2017, to August 3, 2020, and randomised 2928 patients [5]. The HOT-COVID trial was conducted from August 25, 2020, to March 8, 2023, and randomised 726 of a planned 780 patients; it was stopped early, prior to assessment of trial results [16], due to slow recruitment. No patients were randomised into both trials. Both trials were approved by the Danish Health and Medicines Agency, the Health Research Ethics Committee in the North Denmark Region, the Danish Data Protection Agency, and by all required authorities in participating countries, and were prospectively registered at ClinicalTrials.gov (NCT03174002 and NCT04425031).

Both trials included adult patients (≥ 18 years of age) admitted to an ICU with acute hypoxemic respiratory failure. At randomisation, patients received supplemental oxygen either through an open system with a flow rate of at least 10 L per minute, or through a closed system with HOT-ICU requiring a minimum fraction of inspired oxygen (FiO₂) of 0.50, while HOT-COVID allowed any FiO₂. Patients were expected to receive supplemental oxygen for at least 24 h in the ICU and were required to have a functioning arterial line for frequent PaO₂ monitoring. Arterial lines are considered standard of care for patients admitted to the ICU with acute hypoxemic respiratory failure in all participating countries. The two trials differed only in HOT-COVID mandating a verified Severe Acute Respiratory Syndrome COronaVirus 2 (SARS-CoV-2) infection at randomisation [17, 18]. Full lists of inclusion and exclusion criteria for both trials are provided in the ESM 1.

In both trials, patients were randomised 1:1 to receive supplemental oxygen targeting a PaO₂ of either 8 kPa (Lower Oxygenation Group) or 12 kPa (Higher Oxygenation Group). The intervention was administered throughout the entire ICU stay, including any readmissions, for up to 90 days after randomisation. PaO₂ levels were achieved via titration of administered FiO₂ based on intermittent arterial blood gas analyses correlated to the continuous measurement of peripheral oxygen saturation.

The primary outcome was 90-day all-cause mortality. The secondary outcome was the absolute number of days alive without life support in 90 days from randomisation. Life support was defined as any use of either mechanical ventilation, vasopressor or inotropic support, or renal replacement therapy. Details are provided in the ESM 1.

The intervention effects on the primary and secondary outcome were assessed in a total of 14 subgroups according to the follow baseline characteristics; age, sex, admission type (medical, elective surgical, or emergency surgical), ventilatory support [invasive mechanical ventilation, non-invasive ventilation or continuous positive airway pressure (CPAP), or open systems], COPD, pneumonia, intracranial pathology (traumatic brain injury, haemorrhagic or ischaemic stroke, and patients resuscitated from cardiac arrest), heart disease (ischaemic heart disease, chronic heart failure, and acute myocardial infarction), chronic dialysis, intestinal ischaemia, cancer (active haematological cancer and metastatic cancer), shock (plasma lactate > 2 mmol/L and use of continuous vasopressor or inotropes), COVID-19, and in patients receiving mechanical ventilation according to their PaO₂/FiO₂ ratio [20]. The expected direction of the intervention effect is presented both in the protocol [20] and in ESM 1 (eTable 2). The credibility of the effect modifications in the subgroup analyses was estimated using the Instrument to assess the Credibility of Effect Modification Analyses (ICEMAN) tool [23]. A total of eight mandatory domains and one optional were considered, and the overall credibility of the proposed effect modification was ranked as ‘very low’, ‘low’, ‘moderate’, or ‘high’. Subgroups with a p value for interaction below 0.10 were graded with ICEMAN in accordance with the guidelines for this instrument requiring that all outcomes with a potential interaction are assessed [23]; nevertheless, only subgroup interactions with a p value below 0.05 were considered statistically significant to provide a conservative interpretation of subgroup results given the risk of chance findings [20]. Details on baseline characteristics, outcome definitions, and ICEMAN are provided in the ESM 1.

All analyses were conducted in the intention-to-treat population, being all randomised patients, except for those where follow-up data could not be obtained due to withdrawn or unobtainable consent according to national regulations. All tests for statistical significance were two-sided, with alpha levels set at 0.05. Subgroup analyses are reported with 95% confidence intervals (CIs) and p values for the test of interaction. No adjustments for multiplicity were conducted in the primary analyses [20]. Given the risk of multiplicity, we post hoc included sensitivity analyses in which we adjusted the subgroup analyses for multiplicity ad modum Jakobsen et al. [24].

Individual patient data from both trials were appended in a one-step approach using a Generalised Linear Mixed Model (GLMM) allowing for fixed slopes and random intercepts and robust variance estimation [25, 26]. Data were organised into clusters according to trial and site levels with the intercepts being allowed to vary in all clusters. The fixed slopes assume a uniform effect of the oxygenation targets across all clusters, assuming a consistent physiological impact of the intervention irrespective of trial or site locations. Random intercepts permit variations in baseline outcome levels among clusters, thus accounting for differences in patient demographics or other factors unrelated to the intervention. We treated sites participating in both trials as separate clusters for each trial [17, 18, 20]. Trial-specific stratification variables were retained (eFigure 1).

The primary outcome was analysed using a GLMM Poisson family with a log-link, and the secondary outcome using a GLMM Gaussian family with an identity link. Sensitivity analyses with adjustment for the trial-specific stratification variables COPD and active haematological cancer, and the clinically important baseline factors age, metastatic cancer, admission type, and sequential organ failure assessment (SOFA) score were performed for both the primary and the secondary outcome. We supplemented the analysis of the primary outcome with Kaplan–Meier plots. Subgroup heterogeneity of treatment effects (HTE) analyses for the primary and the secondary outcomes were performed using a generalised linear model with fixed slopes, and robust cluster variance estimation, while retaining data clustering as in the primary analyses. Further, we assessed HTE between the two trials for the primary and the secondary outcomes in supplemental post hoc interaction analyses. These tests used the trial (HOT-ICU or HOT-COVID) and the intervention group as interaction terms. No imputations for missing data were conducted since the percentage of missing data for all variables and outcomes in any analysis was less than the pre-specified 5% cut-off level necessitating this (ESM 1) [20]. All analyses were conducted using Stata statistical software release 18 (StataNordic).

A total of 3654 patients from 35 ICUs across Denmark, Switzerland, Finland, The Netherlands, Norway, the United Kingdom, and Iceland were enrolled in two trials: 2928 patients (80.1%) in the HOT-ICU trial and 726 (19.9%) in the HOT-COVID trial, with a total of 1827 patients in the Lower Oxygenation Group and 1827 patients in the Higher Oxygenation Group. For the primary outcome of 90-day mortality, data were available for 3585 patients (98.1%): 2888 from the HOT-ICU and 697 from the HOT-COVID trial, with 1792 patients in the Lower Oxygenation Group and 1793 patients in the Higher Oxygenation Group. For the secondary outcome of days alive without life support in 90 days, data were available for 3575 patients (97.4%): 2878 from the HOT-ICU and 697 from the HOT-COVID trial, with 1786 patients in the Lower Oxygenation Group and 1789 patients in the Higher Oxygenation Group.

Baseline characteristics were similar between the two oxygenation groups, except for the presence of cardiac arrest and the end-expiratory pressure in invasively mechanically ventilated patients (Table 1). When comparing the trials, included patients differed in all baseline characteristics except for positive end-expiratory pressure (PEEP) in patients receiving non-invasive ventilation or CPAP, and in the proportion on long-term dialysis. Complete data are available in the ESM 1 (eTable 3).

The median recorded PaO₂, with corresponding arterial oxygen saturation (SaO₂) and FiO₂, in the ICU were all lower in the Lower Oxygenation Group than in the Higher Oxygenation Group throughout the 90-day intervention period (Fig. 1). Details on ICU treatments, including use of mechanical ventilation, prone positioning, inhaled vasodilators, extracorporeal membrane oxygenation, circulatory support, renal replacement therapy, blood transfusions, and ventilator settings (positive end-expiratory pressure, peak inspiratory pressure, and tidal volume) are provided in the ESM 1 (eTable 4). In the HOT-COVID trial, more patients were treated with proning, inhaled vasodilators, and extracorporeal membrane oxygenation, but less received vasopressors or inotropes, renal replacement therapy, and red blood cell transfusions compared to patients in the HOT-ICU trial (ESM 1, eTable 5).

At 90 days after randomisation, a total of 724 of 1792 patients (40.4%) in the Lower Oxygenation Group and 733 of 1793 patients (40.9%) in the Higher Oxygenation Group had died (adjusted risk ratio, 0.99; 95% CI, 0.92–1.07; P = 0.80) (Figs. 2 and 3). Additional adjustment for baseline factors in the sensitivity analysis produced similar results (Fig. 2). The post hoc analysis of HTE between the trials showed no significant interaction with a p value of 0.12.

The strengths of the current individual patient data meta-analysis include the publication of the protocol prior to inclusion of the last patient in the HOT-COVID trial, and methodology consistent with previously conducted patient-level meta-analyses in critically ill patients [46‐48].

In addition, the inclusion of a large ICU population with acute hypoxemic respiratory failure, identical trial protocols for oxygenation targeting, recruitment from a sizeable number of ICUs in several countries, and the use of an endorsed tool for the evaluation of credibility in effect modification analyses [23, 49], all add to the external validity of the findings.

Limitations of the study include the lack of multiplicity adjustments, which especially in the subgroup analyses increases the risk of false positive findings. The results should therefore be interpreted cautiously and only as hypothesis-generating, a fact which is underlined by the results of the post hoc multiplicity-adjusted subgroup analyses. The decision not to multiplicity-adjust the primary analyses was made to avoid deflation of the study-wide alpha level, thus addressing a general null hypothesis rather than the specific null hypotheses for each subgroup analysis. Moreover, such adjustments assume independence between analyses, which is improbable given that patients are likely to share relevant characteristics despite being analysed in different subgroups. Also, the findings of this study may not apply to ICU patients without hypoxemic respiratory failure or those less ill. Differences in baseline characteristics between the two trials were evident. These differences likely reflect different patient populations: HOT-COVID patients were more homogeneous, primarily suffering from COVID-19 respiratory failure, whereas HOT-ICU patients were more heterogeneous and sicker, with multiple organ failure. Additionally, treatment algorithms for COVID-19 evolved during the pandemic. Initially, high-flow systems were avoided due to fears of transmission, but later, these systems were widely used to prevent intubation. These differences are unlikely due to variations in the interpretation of the common inclusion criteria, as both trials were conducted at the same sites by the same clinicians. The post hoc between-trial HTE analysis suggests that baseline or treatment differences may have influenced the intervention effects on the secondary outcome of days alive without life support, highlighting a limitation of our study. However, the between-trial HTE analysis for the primary outcome of 90-day mortality revealed no statistically significant heterogeneity.

Finally, we specifically included data from the HOT-ICU and HOT-COVID trials to reduce bias stemming from indirectness, thus enhancing the precision of overall effect estimates and allowing for investigation of subgroup effects. This selection was deliberate as both trials employed identical interventions and were conducted in similar settings, with randomised patients receiving supplemental oxygen via both open and closed systems. Unlike other large randomised controlled trials (RCTs) [6‐10], which included a broader patient population, these two trials focused solely on patients with severe acute hypoxemic respiratory failure at randomisation and maintained the intervention throughout the entirety of ICU admission for up to 90 days. Although the trials were compatible, we recognise that their data may not be entirely independent and apply a hierarchical mixed model with clustering at both the trial and site levels.