Background

The Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) has infected over 500 million people worldwide and resulted in more than 6 million deaths (https://covid19.who.int) [1, 2]. COVID-19, the disease caused by the SARS-CoV-2, is a multisystemic disease [3]. Its most severe presentation is acute respiratory distress syndrome (ARDS), secondary to pneumonia [4,5,6]. Most critically ill patients with COVID-19 receive advanced respiratory support, defined as high-flow nasal cannula (HFNC), non-invasive mechanical ventilation (NIV), or invasive mechanical ventilation (IMV) [3, 7, 8]. Up to 30% of hospitalised patients with COVID-19 are treated with one of these interventions [9, 10]; however, the use and need for support have changed over time depending on COVID-19 vaccination coverage, circulating viral variants, an evolving treatment evidence base and practice variation [11, 12].

Given the high demand for respiratory support and the insufficient capacity of intensive care units (ICU) and resources during the pandemic, especially in low- and middle-income countries (LMIC), the use of less invasive alternatives emerged as an alternative to provide advanced respiratory support [13, 14]. A global survey in 2020 found that HFNC (54%) and NIV (47%) were the most frequently used types of advanced respiratory support in patients with severe COVID-19 [15]. Up to 37% of patients who received NIV support ultimately required IMV [16], with high fatality ratios, especially in Latin America [17].

The objectives of this global study are to describe the clinical characteristics and outcomes of patients treated with HFNC, NIV, and IMV during the first two years of the pandemic, to determine risk factors associated with HFNC and NIV failure, and to estimate the association of later administration of IMV on clinical outcomes. We also compare the respiratory support types used in high-income countries (HICs) with those used in LMIC.

Methods

This is a prospective observational study of hospitalised patients from five continents. The study Consortium framework is provided by the International Severe Acute Respiratory and Emerging Infection (ISARIC)—World Health Organization (WHO) Clinical Characterisation Protocol for Severe Emerging Infections [18, 19]. The protocol, case report forms, consent forms, and study information are available on the ISARIC website (https://isaric.org). This standardised protocol uses tiered data collection tailored to a range of resource settings [19]. Investigators from 69 countries collected prospective data using the ISARIC case report form (CRF) built on Research Electronic Data Capture (REDCap, version 8.11.11, Vanderbilt University, Nashville, Tenn.) [20] hosted by the University of Oxford. Other investigators collected data using locally hosted systems and submitted it to the ISARIC dataset for centralised mapping. All investigators retain full rights to their data.

This observational study required no change to clinical management and encouraged patient enrolment in other research projects. The ISARIC-WHO Clinical Characterisation Protocol was approved by the World Health Organization Ethics Review Committee (RPC571 and RPC572). Also, local ethics approval was obtained for each participating country and site according to local requirements.

Study population

We included hospitalised patients with confirmed SARS-CoV-2 infection by reverse transcription-polymerase chain reaction (RT-PCR) in a respiratory sample treated with advanced respiratory support, defined as either HFNC, NIV, or IMV [3]. Patients with no recorded demographic data or vital signs within the first 24 h of hospital admission were excluded, as were patients whose 28-day vital status was unknown.

Variables and measurement

We recorded age, sex, income classification according to the World Bank (https://data.worldbank.org/country) of the country of recruitment, comorbidities, vital signs on admission, laboratory measurements during the first 24 h of hospital admission, treatment with advanced respiratory support at any point during hospitalisation, systemic complications, and treatments used during hospitalisation. The case report form completion guide is available online (https://isaric.org).

We identified the first wave of the pandemic for each participating country and composed a dichotomous variable to evaluate the impact of being admitted during the first wave on clinical outcomes.

We stratified patients in the cohort based on the first type of respiratory support received within the first 24 h of hospital admission. High-flow nasal cannula (HFNC) was defined as respiratory support continuously applied through large-bore nasal prongs using a heated and humid gas flow at an initial flow more significant than 20 L/min (or up to 80 L per minute) and a fraction of inspired oxygen of up to 1.0. Non-invasive mechanical ventilation (NIV) was defined as any type of positive pressure therapy delivered through a fitted mask and was preferred in patients with oxygen requirements over 6–15 L/min or laboured breathing. Continuous positive pressure (CPAP) or bi-positive pressure (BiPAP) may occur and be considered NIV. Invasive mechanical ventilation (IMV) is any mechanical ventilation administered to the patient after endotracheal intubation or tracheostomy. The decision to use this modality was left to the health care providers and not per study protocol.

Patients were considered to have failed the non-invasive respiratory strategy (i.e. HFNC or NIV) if they were subsequently treated with IMV during hospitalisation.

Outcomes

The primary outcome evaluated in this study was 28-day mortality. Secondary outcomes included the rate of and risk factors for failing non-invasive respiratory support (i.e. HFNC or NIV), the association of failure with clinical outcomes, and the frequency of respiratory strategies used in HIC versus LMICs.

Statistical methods

Continuous variables were expressed as median (interquartile range), and categorical variables as counts (percentages). For the primary outcome of 28-day mortality and secondary outcome of non-invasive respiratory failure, random forest (RF) models were used to identify the factors associated with these outcomes. The RF model uses multiple randomised individual decision trees that operate as an ensemble, where each decision tree gives a predicted class. The class obtained most frequently among the decision trees becomes the RF model prediction. A total of 500 estimators were used in this model. A more detailed explanation of the RF models is presented in the supplement.

To evaluate the performance of the RF model, the area under the model's receiver operating characteristics curve (AUROC) was used; for this, a tenfold cross-validation method was used, in which the data set was divided into ten subsets, and the validation was repeated ten times. Each time, one of the subsets was used as the test cohort, and the other nine subsets were used as training subsets, then the average AUROC was calculated and reported. When used for classification, RF models perform an implicit feature selection, a general indicator of each specific feature relevance, and can be computed as the Gini importance.

Then, we fitted two multivariable logistic regression models to estimate associations with the risk of 28-day fatality ratio or non-invasive respiratory failure, respectively. Variables identified as relevant by the RF model were included as explanatory variables. Odds ratios (ORs) were presented with forest plots.

A patient treated with respiratory support might receive different strategies during hospital admission. Thus, we developed alluvia diagrams to understand how patients were treated with other respiratory methods over time, stratified by the countries’ income classification. We constructed chord diagrams to provide a graphical representation of these patients' comorbid conditions and demographics differentiated by the income classification. A significance level of < 0.001 and a confidence level of 95% was chosen to determine statistical differences. This was selected as large datasets, such as the ISARIC COVID-19 dataset, might identify minor differences as significant even when the differences are not clinically relevant. Adjusting the rejection level of the null hypothesis could control this limitation inherent to large datasets and the possibility of incurring type one error. All data processing and statistical analysis were performed using Python version 3.8 with the following data packages: Pandas version 1.2.4, Tidyverse version 1.3.0, Bioconductor version 3.12.

Results

A total of 66,565 patients were included in this study (Fig. 1). Most patients were male (63.5% [42,256/66,565]) and treated in HICs (82.6% [55,004/66,565]). Specifically, 78.2% ([52,039/66,565]) of the cohort was hospitalised and treated in Europe. Regarding the age of the patients included in the cohort, 44.0% ([29,317/66,565]) of patients were between 60 and 80 years old. During the first 24 h of hospital admission, patients were most frequently treated with HFNC (42.5% [28,256/66,565]), followed by NIV (36.2% [24,112/66,565]) and IMV (21.3% [14,197/66,565]). Demographic characteristics, physiological variables and laboratories at hospital admission are shown in Fig. 2 and Tables 1 and 2.

Fig. 1
figure 1

Flow chart. This figure shows patients included in the analysis and cohort selection process

Full size image
Fig. 2
figure 2

Probability density of patients' basic demographics (age and sex), according to the first ventilation treatment received. A Complete cohort. B Patients from high-income countries. C Patients from low middle-income countries

Full size image
Table 1 Baseline characteristics of patients, stratified by the different advance ventilatory supports
Full size table
Table 2 Physiological parameters and laboratories of patients during the first 24-h hospital admission, stratified by the different advance ventilatory supports
Full size table

Patients’ characteristics, in-hospital treatments, and systemic complications

More than 85% of the patients had at least one comorbidity. Hypertension (41.3% [27,521/66,565]) and diabetes mellitus (30.3% [20,164/66,565]) were the most frequently reported comorbid conditions (Table 1). A total of 22.8% [15,190/66,565] of patients were current or past smokers. Complications were also common during the hospital admission (not at hospital presentation), 23.2% [15,470/66,565] developed ARDS, and 20.1% [13,353/66,565] were reported to have an acute renal injury (ARI).

During hospital admission, 61.3% (40,810/66,565) patients received corticosteroid treatment, and 54.6% [36,336/66,565] were admitted to the ICU. Vasopressor/inotrope therapy was used in a quarter of all patients (20.4% [13,592/66,565]), increasing in use according to ventilatory requirement (7.7% [2188/28,256] vs. 17.8% [4282/24,112] vs 50.2 [7122/14,197]). Approximately half of those treated with IMV received vasopressors/inotropes at some point during hospitalisation (50.2% [7122/14,197]). Almost one-quarter of the patients were placed in prone position (23.7% [15,778/66,565]), more commonly in those patients treated with IMV (12.0% [3384/28,256] vs. 27.5% [6628/24,112] vs 40.6% [5766/14,197]). A total of 15.5% [10,287/66,565] of patients failed HFNC/NIV. Moreover, 71.2% [7327/10,287] of the patients that failed HFNC/NIV were registered in HIC and 28.8% [2960/10,287] in LMIC. Finally, 28-day mortality was similar between the different advance ventilatory supports (33.9% [9581/28,256] vs. 38.1% [9177/24,112] vs. 37.1% [5263/14,197]).

Comparing respiratory support of patients admitted in HIC or LIMC

The cumulative frequency of advanced respiratory treatments was stratified by national income classification (Fig. 3). Patients admitted to the hospital in HICs were more frequently treated with HFNC (48.0% [26,399/55,004]), followed by NIV (38.6% [21,237/55,004]) and IMV (13.4% [7368/55,004]). In contrast, patients admitted in LMICs were less frequently treated with HFNC (16.1% [1857/11,561]), and the majority received IMV (59.1% [6829/11,561]) (Table 1; Fig. 3). We also found differences in distribution among the different types of respiratory support when stratified by income classification and respiratory support (Fig. 4).

Fig. 3
figure 3

Cumulative frequency (net number of patients) of ventilation treatment given to patients. A Complete cohort. B Patients from high-income countries. C Patients from low middle-income countries

Full size image
Fig. 4
figure 4

Chord graphic with demographics and comorbidities of patients according to the type of first ventilation treatment received. A Complete cohort. B Patients from high-income countries. C Patients from low middle-income countries

Full size image

Patients treated with IMV in HICs had fewer comorbid conditions and were more frequently between 40 and 70 years old. In sharp contrast, patients in LMIC who were younger than 40 years old often received IMV and were more frequently male. Also, they were mostly treated with IMV rather than non-invasive respiratory strategies (Fig. 4).

Changes in respiratory supports

Figure 5 presents the alluvia diagrams illustrating how patients progressed among respiratory support during hospital admission. Notably, patients who required more than one respiratory treatment had higher mortality than those treated with only one type of support, whether the first respiratory support was HFNC, NIV, or IMV (Fig. 5).

Fig. 5
figure 5

Alluvia diagram of the patients’ transitions between ventilation treatments and clinical outcomes. The width of the links is proportional to the number of patients. A Complete cohort. B Patients from high-income countries. C Patients from low middle-income countries

Full size image

Risk factors for failing HFNC or NIV as first respiratory support

The failure rate of HFNC or NIV was 15.5% [10,287/66,565]. According to the Gini importance, the variables most strongly associated with non-invasive ventilation failure (either HFNC or NIV) were age, lower platelets, and higher leukocyte count during the first 24 h of hospital admission (Fig. 6A). In the logistic regression analysis, we found that high leukocyte counts at hospital admission (OR [95% CI]; 5.86 [4.83–7.10]), treatment in an LMIC (OR [95% CI]; 2.04 [1.97–2.11]), and tachypnoea at hospital admission (OR [95% CI]; 1.16 [1.14–1.18]) were strongly associated factors with IMV treatment as rescue treatment (Fig. 6B, C).

Fig. 6
figure 6

An automatised model to determine risk factors associated with non-invasive ventilation failure. A variables more strongly related to non-invasive ventilation failure according to the Gini importance. B the contribution of the variables to the output; the red values indicate a high-value contribution of the variable, and the blue values a low-value contribution. The positive values in the plot indicate a high probability of 28-day fatality, and negative values indicate a low likelihood of 28-day fatality. Panel C presents a logistic regression model, showing variables more strongly associated with the 28-day fatality ratio. The most significant variables were leucocyte count, low-/middle-income country attention, higher respiratory rate, and higher systolic blood pressure

Full size image

Clinical outcomes and risk factors associated with 28-day fatality ratio

Almost half of the patients treated with HFNC [46.3%; 11,954/28,256] and 37.1% (5263/14,197) of patients treated with IMV died within 28 days. The variables identified as risk factors associated with the 28-day fatality ratio are shown in Fig. 7. Older age (OR [95% CI]; 2.42 [2.36–2.48]), cardiac arrest during hospitalisation (OR [95% CI]; 1.86 [1.81–1.92]), receiving treatment in an LMIC (OR [95% CI]; 1.56 [1.53–1.60]), and higher leukocyte counts at hospital admission (OR [95% CI]; 1.47 [1.39–1.55]) were the main adjusted risk factors associated with 28-day mortality. Notably, NIV/HFNC failure (OR [95% CI]; 1.27 [1.25–1.30]) was also highly associated with fatality. Other factors were acute kidney injury (OR [95% CI]; 1.23 [1.21–1.25]), ARDS (diagnosed during the hospital admission, not during the first 24 h) (OR [95% CI]; 1.12 [1.10–1.14]), increased heart rate at admission (OR [95% CI]; 1.15 [1.13–1.18]), increased respiratory rate at admission (OR [95% CI]; 1.15 [1.13–1.17]), chronic cardiac diseases (OR [95% CI]; 1.17 [1.14–1.19]), chronic pulmonary diseases (OR [95% CI]; 1.12 [1.10–1.14]), and diabetes mellitus (OR [95% CI]; 1.07 [1.05–1.09]). The model used to predict the 28-day fatality ratio had a good discriminatory capacity when evaluated by the AUROC (mean [SD] 0.78 [0.05], Fig. 7).

Fig. 7
figure 7

An automatised model to determine risk factors associated with the 28-day fatality ratio. A The contribution of the variables to the output; the red values indicate a high-value contribution of the variable, and the blue values a low-value contribution. The positive values in the plot indicate a high probability of 28 fatalities, and negative values indicate a low likelihood of 28-day fatality. B A logistic regression model, showing variables more strongly associated with the 28-day fatality ratio. C Each cross-validation trial's receiver operative curve (ROC) for the subset of the selected variables. The blue curve represents the average of the ROC curves of each test, and the average area under the ROC is also presented. The most significant variables associated with the 28-day fatality ratio were age, cardiac arrest, low-/middle-income country attention, and leucocyte count. Also, patients that fail the non-invasive or high-flow nasal cannula are independently associated with a higher 28-day fatality ratio

Full size image

Discussion

In this large, multinational, prospective cohort study, we found that patients with severe COVID-19 were mainly treated with non-invasive respiratory strategies (i.e. HNFC or NIV) in HICs; in contrast, patients with severe COVID-19 in LMICs were more frequently treated with IMV. We found that the 28-day fatality ratio was similar among patients treated with HFNC, NIV, or IMV worldwide. Notably, we found that patients treated with IMV as rescue therapy (i.e. failure of non-invasive treatments) had a higher 28-day fatality ratio than patients treated with IMV earlier in their disease course. The risk factors associated with failing the non-invasive respiratory strategies were high leukocyte counts at admission, increased heart rate at admission, and being treated in an LMIC. Notably, being admitted during the first pandemic wave did not impact clinical outcomes or respiratory treatments.

Early in the pandemic, healthcare workers identified that patients with hypoxemia could be treated with HFNC [21,22,23]. International guidelines also recommend non-invasive respiratory support as the first treatment, and many centres utilise HFNCs outside formal ICU settings [13]. Notably, the widespread usage of HFNC and NIV in patients with severe COVID-19 was recommended by experts and guidelines but not supported by high-quality data. Later, Ospina-Tascon et al. [12] carried out a multicentre, open randomised clinical trial and found that the early treatment with HFNC compared to conventional oxygen treatment was associated with a lower necessity of IMV (34.3 Vs 51.0, HR: 1.39; 95% CI 1.00–1.92; p = 0.04). Then, Perkins et al. [24] in the RECOVERY-RS trial found that NIV was associated with a lower requirement of tracheal intubation and lower 30-day mortality when compared to conventional oxygen therapy (absolute difference, − 8% [95% CI, − 15 to − 1%], p = 0.03). Our study found that HFNC, NIV, and IMV have similar 28-day fatality ratios, in concordance with prior literature. However, we found that HFNC was mainly used in HIC, which might be in relation to the capacity of these countries to acquire this new technology during the pandemic and the ability of these countries to expand their bed capacity to treat patients with HFNC outside of the ICU. Also, some patients or their families do not accept endotracheal intubation and prefer non-invasive strategies, though our study did not collect these data.

In contrast to HIC, the most common respiratory treatment in patients with severe COVID-19 utilised in LMIC has been IMV, as is evident in our data. Estenssoro et al. [17] described the results of a prospective observational cohort of patients admitted to 64 ICUs in Argentina. They included 1909 patients treated with IMV and found that lung-protective respiratory strategies were widely used but with a high fatality rate among patients included in the cohort (57.7%, 1101/1909). In another study in Brazil, Ranzani et al. [7] found that 23% (45,205/232,036) of patients admitted to the hospital were treated with IMV. They also found that the fatality rate among those receiving IMV was 80% during the first pandemic wave and 87% during the second wave [7, 25]. Notably, they found that 14% (5976/44,055) of the patients treated with IMV were treated outside of the ICU [25]. These results highlight that fatality rates and treatments changed during the pandemic and differed for each country. Moreover, these data align with our results, showing that IMV was frequently used in LMIC and that many patients with severe COVID-19 were treated outside of ICU [8, 26]. Notably, the impact of ICU admission on clinical outcomes was already explored in our cohort and published elsewhere [3]. We found that ICU admission was associated with better clinical outcomes independently of disease severity, treatments received, income classification, and system saturation (i.e. the number of new COVID-19 detected the day patients was admitted).

Even though non-invasive respiratory support has been proven effective in treating patients with severe hypoxemia during COVID-19, up to 30% of the patients were treated with IMV as a rescue treatment. Thus, it is essential to identify which patients might be at risk of failing under the non-invasive respiratory strategy and not to delay IMV in these patients. Rodriguez A. et al., in one of the largest prospective cohorts of patients admitted to the ICU due to severe flu infection, found that patients who failed NIV had a mortality rate three times higher than those who did not fail [27]. Also, they found that patients who failed NIV had higher mortality than those treated with IMV as initial treatment (38.4 vs 31.3, p = 0.18). In a multicentre COVID-19 study, Boscolo A. et al. found that 704 patients who failed non-invasive respiratory support had an accumulative fatality rate of 43% [28]. Our findings support that patients with severe COVID-19 who fail the initial respiratory support with non-invasive treatments have a higher mortality rate and were independently associated with 28-day fatality. Also, we found that patients with higher leukocyte counts at admission, higher respiratory rate at admission, and being in an LMIC were at higher risk of failing the non-invasive respiratory strategies. Thus, patients with these characteristics should be carefully evaluated to avoid delays in initiating IMV when appropriate.

Our study has strengths and limitations that are important to acknowledge. First, the respiratory support interventions were not according to a standardised protocol, leaving clinical teams to choose when to use HFNC, NIV, or IMV; thus, demographic or clinical characteristics may differ across the groups studied. However, we performed a robust statistical analysis using random forest analyses and logistic regression, adjusting for several confounders. This allowed us to evaluate linear and nonlinear relations in a supervised statistical approach. Second, most patients in our study were registered in Europe and HICs, which might constitute a significant selection bias. However, we had more than 11,000 patients in LMICs in Africa, South America, and Asia, including a large cohort of patients and contributing to our results' global generalisability. Third, we do not have complete data on specific respiratory parameters used during the support (i.e. peep, flows, FiO2, volumes, among many others), limiting our capacity to assess the rates of protective respiratory strategies, among other essential factors. Thus, these results cannot imply a causal association between respiratory support device treatments and clinical outcomes. Each patient should be evaluated carefully with decisions on the type of respiratory support based upon the evolving evidence base applied to their specific clinical condition and goals of care. Finally, throughout the COVID-19 pandemic, patients were treated with a large variety of medications and supportive clinical protocols; it is challenging to make conclusions about the factors associated with 28-day fatality using observational study methodologies in such a dynamic context.

Conclusions

Patients hospitalised with confirmed COVID-19 are often treated with advanced respiratory support. HFNC was the primary initial respiratory support used during the pandemic; however, this treatment was mainly used in HIC. In contrast, IMV was the primary respiratory treatment utilised in LMIC. Non-invasive respiratory treatments (i.e. HFNC and NIV) could be used as the first respiratory support in patients with severe COVID-19; however, it is crucial to identify patients at risk of failing because delaying IMV may be associated with worse clinical outcomes. Further studies are needed to confirm these associations.