This study represents an analysis of a real-world dataset, the Medical Information Mart for Intensive Care-IV (MIMIC-IV), that was created by the Massachussets Institute of Technology (MIT) and provides critical care data for over 60,000 patients admitted to intensive care units at the Beth Israel Deaconess Medical Center (BIDMC) between 2008 and 2019 [25, 26]. This dataset provides granular information on demographics as well as many physiological variables, treatment received and mortality up to 1-year post-discharge. This study was conducted following the standards as defined by the Declaration of Helsinki. Since MIMIC-IV only includes anonymized information, patients’ consent to participate was waived at the local institution. The Research Ethics Board at Hospital Clinic in Barcelona did not require to undergo further protocol approval.

Patients were considered eligible if they had been admitted to the Medical, Medical/Surgical or Coronary ICUs and presented with acute hypoxemic respiratory failure, as defined by a ratio of oxygen saturation (SpO2) to inspired oxygen fraction (FiO2) ≤ 200 and a SpO2 ≤ 97% within 48 h of ICU admission and were not yet intubated. Patients could be receiving oxygen through facemask, high flow nasal cannula or non-invasive ventilation. We also wanted to exclude patients with immediate and major reason for endotracheal intubation. Therefore, exclusion criteria were a respiratory rate > 39 breaths per minute, a Glasgow Coma Scale ≤ 12 or a SpO2/FiO2 < 88 and the absence of a “Full Code”. These criteria were created to provide realistic limits to the inclusion of patients, since equipoise regarding withholding intubation would likely not hold in the latter subset.

To estimate the effect of immediately initiating invasive ventilation on survival in patients with hypoxemic respiratory failure without prior history of intubation during the ICU admission, we emulated a target trial comparing intubation within one hour versus delaying intubation. Patients were eligible for the target trial in the first hour that they met eligibility criteria and for every subsequent hour in which they also met eligibility criteria, up to 48 h (Additional file 1: Table S1). This arbitrary time point was chosen because most intubations occur during this period and to provide greater homogeneity between patients.

To emulate the target trial, we identified all subjects that fulfilled the inclusion criteria (and this was considered the time that eligibility had been first met, or hour 1). This procedure was repeated throughout hours 2–48 for all remaining eligible patients who had not received invasive ventilation previously. At each hour to still be considered eligible, patients had to remain non-intubated at the beginning of the interval and had to continue to fulfill the inclusion criteria as well as not to fulfil any of the exclusion criteria. Thus a patient who remained eligible and non-intubated could contribute up to 48 observations to the target trial emulation [27] (Fig. 1). This methodology was followed to aim at reproducing what often happens in the clinical setting where clinicians continuously reassess their patients regarding the decision for intubation.

Study flowchart. Patients could be included if they had been admitted in any of the following ICUs: Medical, Medical/Surgical or Coronary ICU, had not been intubated previously and did not present any exclusion criteria. Afterwards, if they presented with all the inclusion criteria, it was considered that they had met eligibility and they were included in target trial number 1. Each patient could later contribute to future observations in the following 48 h, provided he/she did not receive intubation in the current target trial and that he/she continued to present eligibility in the following hours. For example, 723 patients were excluded from target trial number 2 with 469 patients having received intubation and 254 patients not presenting with further eligibility (either because of any new exclusion criteria, not further inclusion criteria or both). A total of 38,272 patient-observations were included of which 747 corresponded to observations where intubation took place. SF: SpO2/FiO2, RR: respiratory rate, GCS: Glasgow Coma Scale

The main outcome evaluated on this study was one-year mortality while, 30-day mortality, ICU and hospital length of stay were defined as secondary outcomes.

When missing data was present at any given hour after first eligibility, last observation carried forward was used for physiological data; under the assumption that physiological data would not deviate significantly from a previous value unless there existed a new entry in patients’ charts (see Additional file 1).

At each evaluated time point throughout hour 1 to hour 48, patients’ probability to receive mechanical ventilation was estimated. To calculate this, a logistic regression with the receipt of mechanical ventilation as the dependent variable and variables supposed to play a role in the decision for intubation were used as independent variables. The variables included the time since fulfilling the inclusion criteria, age, comorbidities as measured by the Elixhauser comorbidity index, FiO2, SpO2/FiO2, respiratory rate, Glasgow Coma Scale, the use of any vasopressors and the admitting unit. After this propensity score had been estimated, stabilized inverse probability weights (IPW) were computed to adjust for confounding (see Additional file 1) [28, 29]. This approach resulted in a population that was weighted at each hour by their probability of receiving intubation [19, 30].

On this population, one-year mortality was later assessed in a time-to-event fashion with the use of a weighted Cox model. This model also accounted for systolic, median and diastolic blood pressure, temperature, creatinine and bilirubin levels as well as platelet count because these values may influence mortality independently from the decision to intubate patients. Hazard ratios (HR) are reported as an average of treatment effect over the study time and survival curves were constructed using a stratified Cox model [31] (see Additional file 1). 95% confidence intervals were calculated by estimating robust standard errors to account for the multiplicity of same-subject observations [27]. Unadjusted and adjusted mortalities were calculated using survival probabilities estimated with a non-parametrically bootstrapped Cox model with 1000 repetitions. ICU and hospital length of stay were assessed using weighted medians (and interquartile ranges) after bootstrapping and differences between groups with their 95% confidence intervals are presented. Reported p-values are two-sided and the level of significance was set at 0.05.

Several additional analyses were conducted in restricted populations or using different statistical methods for confounding adjustment. First, the inclusion criteria were tightened to include a population of patients that besides hypoxemia also presented with a ROX ≤ 4.88 at eligibility. This cut-off was had previously shown to predict intubation in patients with acute hypoxemic respiratory failure under high-flow oxygen therapy [32]. Second, the effect of time since eligibility was further evaluated considering nested target trials within 5 groups: first hour after first eligibility, 2nd to 6th hour, 7th to 12th, 13th to 24th and 25th to 48th hour. Third, we repeated the main analysis using two doubly robust approaches, one with augmented inverse probability weighting (AIPW) and a second one using targeted maximum likelihood (TMLE). Fourth, we carried out overlap IPW weighting to limit the analysis to subjects with a realistic probability of receiving either treatment under investigation. Fifth, we repeated the analysis by restricting to the Medical ICU only. Sixth, we conducted a complete-case analysis. This was done to check the robustness of our study findings (see Additional file 1).

To construct the dataset for this study Google BigQuery was connected to MIMIC-IV and the R software (R Foundation for Statistical Computing, Vienna, Austria) was used for statistical analysis. All the code is available at https://github.com/rmartigas/causal-inference-invasive-ventilation-MIMIC-IV.