High-Flow Nasal Oxygen in Patients With COVID-19-Associated Acute Respiratory Failure

Ricard Mellado-Artigas; Bruno L. Ferreyro; Federico Angriman; Maria Hernandez-Sanz; Egoitz Arruti; Antoni Torres; Jesus Villar; Laurent Brochard; Carlos Ferrando


Crit Care. 2021;25(58) 

In This Article


Study Design and Setting

We conducted a prospective, multicentre, cohort study of consecutive patients with COVID-19 associated acute respiratory failure admitted to 36 hospitals from Spain and Andorra (see Supplementary file).[16] The study was approved by the referral Ethics Committee of Hospital Clínic, Barcelona, Spain (#HCB/2020/0399), and conducted according to the amended Declaration of Helsinki. This report follows the "Strengthening the Reporting of Observational Studies in Epidemiology (STROBE)" guidelines for observational cohort studies.[25] Gathering of data is ongoing and as of August 13, 2020, a total of 1,129 patients have been included. A preliminary communication was presented as an abstract at the annual European Respiratory Society conference in September 2020.[26]

Study Population

We included adult patients (≥ 18 years old) admitted to the ICU between March 12 and August 13, 2020. Patients were included if they had positive confirmatory nasopharyngeal or pulmonary tract sample and received support with either HFNO or intubation on the first day of ICU admission. Main exclusion criteria were intubation outside the ICU, a PaO2/FiO2 ratio > 300 mmHg, a respiratory rate on day 1 > 35 breaths/min, a Glasgow Coma Score < 13, and pH < 7.25.[18,27] The rationale for the aforementioned eligibility criteria was based on a population that (with equipoise) could theoretically be randomized to a strategy of early intubation or HFNO in the first 24 h of critical illness, under the framework of a target randomized trial (Additional file 1: e-Table 1).[28] The final analytical cohort was obtained by propensity score matching based on potential confounders measured at baseline.

Data Collection

Patients' characteristics were collected prospectively according to a previously standardized common protocol. Each investigator had a personal username/password and entered data into a specifically pre-designed online data acquisition system (CoVid19.ubikare.io) endorsed and validated by the Spanish Society of Anesthesiology and Critical Care (SEDAR).[29] Patient confidentiality was protected by assigning a de-identified code. Recorded data included patients' demographics [age, gender, body mass index (BMI)], comorbidities, time from onset of symptoms and from hospital admission to initiation of respiratory support and vital signs (temperature, mean arterial pressure, heart rate), laboratory parameters (complete blood count, coagulation tests, electrolytes, creatinine) and severity assessment scales such as the Sequential Organ Failure Assessment (SOFA)[30] and Acute Physiology and Chronic Health Evaluation II (APACHE II) scores.[31] Site investigators collected what they considered to be the most representative data of each day from ICU admission to ICU discharge. Patients were followed up until hospital discharge to assess for in-hospital mortality.

Study Exposure and Outcomes

The main exposure was the use of HFNO as the initial oxygenation strategy in the first 24 h (conservative group), and the comparison was the use of invasive mechanical ventilation in the first 24 h (early intubation group).[6] Because data were collected once per day and the duration of HFNO use was not recorded, patients that were switched from HFNO to invasive-mechanical ventilation on day 1 were considered as part of the early intubation group.[6] We considered that, in these patients, the HFNO use may have been too short to have a meaningful effect in a patient's outcome.[6] The decision to intubate was left at the discretion of the treating physicians at each participating centre. The primary outcome of interest was VFDs at day 28, calculated as 28 minus the days that a particular patient remained mechanically ventilated.[32] To account for the competing risk of death, deceased patients were considered to have 0 VFDs. Secondary outcomes included ICU length of stay, intubation rate and all-cause in-hospital mortality (and up to 60 days). A subgroup analysis considering patients intubated early (on day 1) versus those intubated late (from day 2 and onwards) was performed.

Statistical Analysis

Demographics, comorbidities, vital signs and laboratory markers at ICU admission were compared between both treatment groups using standardized mean differences. To account for potential confounding of the effects of HFNO on all outcomes of interest, we performed a propensity-score matched analysis.[33] Specifically, we built a multivariable logistic regression model to estimate the log-odds of receiving HFNO on the first day of ICU. The criteria to include variables in this model were based on those potentially affecting the likelihood of outcome occurrence and receipt of study treatments[34] and were performed based on subject matter knowledge with the help of a direct acyclic graph (DAG) (e-Figure 1 in Additional file 1).[35,36] Selected variables included gender, APACHE II, SOFA, Glasgow Coma Scale, systolic blood pressure, pH, respiratory rate, arterial partial pressure of carbon dioxide (PaCO2), body mass index (BMI), creatinine, bilirubin, platelet, leucocyte and lymphocyte count, lactate, immunosuppression and hospital group (divided into quartiles based on the proportion of patients receiving intubation from the total). The matching procedure was conducted on a 1:1 fashion without replacement and with the calliper of the logit (propensity score) set at 0.2.[33] Proper adjustment was assessed with standardized mean differences (SMD) in the matched population, and covariate imbalance defined using a SMD > 0.2 threshold.[33] Missing data on important confounders were handled using multiple imputations with a Monte Carlo Markov chain method (details in Additional file 1).[37]

Once the matched cohort was constructed and after balance assessment, we used simple linear regression to assess mean differences in VFDs at 28 days and ICU length of stay (in days) between treatment groups. For all-cause in-hospital mortality, we used generalized linear models (with identity link and binomial distribution) to estimate risk differences and a crude logistic regression model to estimate odds ratios. For all models, 95% confidence intervals (CI) were constructed based on robust standard errors to account for the matching procedure.

Sensitivity Analyses

Several sensitivity analyses were performed to assess the robustness of our findings for the study outcomes. First, we performed a complete case analysis, excluding patients that had any missing data on the selected variables to construct the propensity score. Second, we repeated our primary analysis for the treatment effect by adjusting for those baseline variables that were not balanced (i.e., SMD > 0.2) by our matching procedure.[38] Third, given that treatment assignment was not random, both residual and unmeasured confounding remain possible. Hence, we estimated the E-value as a way to determine the association between an unmeasured confounder with both the exposure (HFNO) and outcome that would fully explain the estimated effect (see details in Additional file 1). Fourth, we changed our exposure classification, keeping patients who initially received HFNO and switched within the 24-h window to invasive mechanical ventilation as part of the conservative strategy (HFNO). This was done to evaluate whether the initial classification yielded optimistic estimates by assigning sicker patients with early HFNO failure to the early intubation group. Finally, we assessed the modification of the effect of HFNO on the primary outcome of interest according to baseline severity measured by the PaO2/FiO2 ratio. For subgroup analysis, we used Wilcoxon rank-sum test and Fisher's test as appropriate.

We used a threshold of 0.05 for statistical significance. All reported tests are two-sided. The R software (R Foundation for Statistical Computing, Vienna, Austria; packages mice, lme4 and sjstats packages) and STATA v.14.2 were used for all analysis. The E-value was computed using a freely available online calculator (www.evalue-calculator.com). Graphs were constructed using BioRender.com.