Corticosteroid treatment and mortality in mechanically ventilated COVID-19-associated acute respiratory distress syndrome (ARDS) patients: a multicentre cohort study

Background Some unanswered questions persist regarding the effectiveness of corticosteroids for severe coronavirus disease 2019 (COVID-19) patients. We aimed to assess the clinical effect of corticosteroids on intensive care unit (ICU) mortality among mechanically ventilated COVID-19-associated acute respiratory distress syndrome (ARDS) patients. Methods This was a retrospective study of prospectively collected data conducted in 70 ICUs (68 Spanish, one Andorran, one Irish), including mechanically ventilated COVID-19-associated ARDS patients admitted between February 6 and September 20, 2020. Individuals who received corticosteroids for refractory shock were excluded. Patients exposed to corticosteroids at admission were matched with patients without corticosteroids through propensity score matching. Primary outcome was all-cause ICU mortality. Secondary outcomes were to compare in-hospital mortality, ventilator-free days at 28 days, respiratory superinfection and length of stay between patients with corticosteroids and those without corticosteroids. We performed survival analysis accounting for competing risks and subgroup sensitivity analysis. Results We included 1835 mechanically ventilated COVID-19-associated ARDS, of whom 1117 (60.9%) received corticosteroids. After propensity score matching, ICU mortality did not differ between patients treated with corticosteroids and untreated patients (33.8% vs. 30.9%; p = 0.28). In survival analysis, corticosteroid treatment at ICU admission was associated with short-term survival benefit (HR 0.53; 95% CI 0.39–0.72), although beyond the 17th day of admission, this effect switched and there was an increased ICU mortality (long-term HR 1.68; 95% CI 1.16–2.45). The sensitivity analysis reinforced the results. Subgroups of age < 60 years, severe ARDS and corticosteroids plus tocilizumab could have greatest benefit from corticosteroids as short-term decreased ICU mortality without long-term negative effects were observed. Larger length of stay was observed with corticosteroids among non-survivors both in the ICU and in hospital. There were no significant differences for the remaining secondary outcomes. Conclusions Our results suggest that corticosteroid treatment for mechanically ventilated COVID-19-associated ARDS had a biphasic time-dependent effect on ICU mortality. Specific subgroups showed clear effect on improving survival with corticosteroid use. Therefore, further research is required to identify treatment-responsive subgroups among the mechanically ventilated COVID-19-associated ARDS patients. Supplementary Information The online version contains supplementary material available at 10.1186/s13613-021-00951-0.


Data Collection and Validation
Data were collected using a paper CRF (case Report Form). CRF collect and record all protocol-required information, which is transcribed from patient source documents, such as hospital records and laboratory data during the patient's participation in the study. Before sending the CRF to the Study Coordinator (AR), these data were de-identified by removing the patient's name, medical record number, etc., and giving the patient a unique study number. We implemented a double data entry model for potential errors in realtime. Data was entered twice by two different Data Entry personnel based on the same set of data collected in the paper CRFs. All data were reviewed, and values that appeared incongruent or out of range were manually validated by confirming the accuracy of the data with the Study Coordinator (AR). The database was validated and cleaned before the statistical analysis and finally, the study database was locked to prevent any further changes, and to ensure data consistency and integrity for the statistical reporting and analysis.

Study definitions
The confirmation of case of SARS-CoV-2 infection was accomplished by positive reverse transcriptionpolymerase chain reaction, either at hospital or ICU admission, from specimens collected with nasopharyngeal and oropharyngeal swabs according to the WHO recommendations [1]. Lower respiratory tract specimens were collected when patients were under MV and high clinical suspicion was present, if previous tests were negatives. Shock at ICU admission was defined as the requirement of any dose of vasopressor therapy within the first six hours of admission to maintain appropriate main blood pressure, despite adequate fluid resuscitation targeted by dynamic hemodynamic parameters and/or echocardiography. Acute kidney injury was defined according the RIFLE criteria [2]. Community-acquired respiratory co-infection (CARC) or bacterial co-infection was considered in patients with confirmation of SARS-CoV-2 infection showing recurrence of fever, increase in cough and production of purulent sputum plus positive bacterial/fungal respiratory or blood cultures within the first two days of ICU admission [3].

Statistical analysis
Missing data were handled with multiple imputation by chained equations [4]. All patients considered in the study received Corticosteroids within the first day of ICU stay at the latest. Time zero of follow up was ICU admission, but we discarded all patients censored within 48h of ICU admission to avoid immortal time bias.
Genetic matching (GM) was used to reduce treatment selection bias and balance the covariance matrix for both groups. GM uses a genetic search algorithm to iteratively determine the weight of each of the covariates to find an optimal balance between matched groups [5]. The matching was one to one with replacement and ties (so that one treated unit could be matched to more than one untreated unit after weighting them appropriately) and without calipers. Variables included in the propensity score matching model were those baseline variables related to the outcome, as recommended elsewhere [6]. The matching weights were included as case weights in the Cox procedures (see Therneau, T., Grambsch, P., Modeling Survival Data: Extending the Cox Model. Springer-Verlag, 2000., page 161 for more information.).
Centre effect for ICU mortality was investigated by multilevel logistic regression analysis through a conditional random intercept model using inter-hospital variation as a random-effects variable. Regression coefficients were summarized as the variance with standard deviation (SD) and the interclass correlation coefficient (ICC). The ICC can be interpreted as "the proportion of the variance explained by the grouping structure in the population". This index goes from 0, if the grouping conveys no information, to 1, if all observations in a group are identical. When ICC is large, it means the between-class variance cannot be ignored and therefore a multilevel model is preferred. It has been suggested that if ICC > 0.1, one should consider the use of a multilevel model. When the ICC is not different from zero or negligible, one could consider running traditional one level regression analysis [7,8].
As a complementary analysis to the main outcome (ICU mortality), we conducted a survival analysis through a Cox regression to investigate whether survival times were related to covariates, and estimating the effect size of a corticosteroid treatment after adjusting for potential confounders. Survival analysis examines and models the time it takes for events to occur, then focuses on the distribution of survival times.
The Cox proportional-hazards model is essentially a regression model for investigating the association between the survival time of patients and one or more predictor variables and it permits to analyse how specified factors influence the rate (hazard rate) of a particular event happening (e.g., death) at a particular point in time. Therefore, the validity of the Cox regression model relies on the assumption of proportionality of the hazard rates of individuals which must remain constant over time with different covariates values [9]. Discharge alive from ICU has been identified as a competitive event for ICU mortality [10], then the survival analysis was performed using the Cox regression through a cause-specific hazard model [11,12]. This model enables the interpretation of etiological relationships and the estimation of effect between the exposure and both outcomes. An important assumption of survival analysis such as the Cox model is that censoring is 'independent'. This independent censoring assumption implies that patients who are censored at a certain time point should be representative for those still at risk (and thus in the risk set) at that point in time. However, this assumption cannot be made if, for example, the survival time of an individual is censored, being withdrawn as a result of a deterioration or an amelioration in his/her physical condition. This is probably the case in the ICU where patients are discharged alive, and thus withdrawn from the survival analysis, because they need no more intensive care, usually due to amelioration or deterioration of their vital conditions. Patients are therefore discharged alive (censored) because they have a lower risk or higher risk of hospital death than the average. These patients are therefore not the same patient population as those who stayed within the hospital. Resulting censoring is 'informative', meaning that censoring carries information about or depends on the survival time. In other words, informative censoring defined a competing risk, given that discharge from the hospital affects the probability of experiencing the event of interest (death before discharge). In this setting, standard survival methods are no longer valid, and specific methods need to be considered.
When the Cox model was made, the proportional hazards assumption was strongly violated for corticosteroids. Hence, the variable of interest (corticosteroids) was considered as a time-varying covariate, which occurs when a given covariate changes over time during the follow-up period, a common phenomenon in clinical research [13]. Proportional hazards may not hold over the entire time axis but may hold approximately over shorter time periods. The effect of a time-varying covariate (corticosteroid treatment) becomes stronger or weaker over time, which can be explored via stratification by time. Therefore, we carried out a time-dependent Cox regression using a step function to deal with nonproportional hazards [14]. The step function consisted in a partitioning of the time axis dividing the follow up into shorter time periods, hence the proportional hazard assumption held within each interval of the partition. We established to divide the study time frame in two intervals (at 17 th of follow up) when the proportional hazards assumption was met. The rationale behind 17 days-step term was based on the pronounced change of case-fatality rates according with both the life-tables and the survival curves in this timeframe. With this methodology, we modelled the effect of corticosteroids on mortality in two ranges: the short and long-term. The coefficients for these two time periods have been estimated separately so that each covariate can relate differently to survival during the two time periods. The benefit of partitioning the time axis is that we can model the effect of the corticosteroid treatment for each period of the study. With the time split at 17 days of follow up, permitted to observed that the hazard rates of individuals of each interval were constant over time, hence, the assumption of proportionality was met. In this way we model the effect of Corticosteroids on mortality in two ranges: the short and the long-term. The results of time-toevent data were expressed as hazard ratios (HR) and 95% CI." Prespecified subgroup sensitivity analysis with exploratory nature was performed with propensity score matching for each study subgroup to evaluate whether the observed effect of corticosteroids on ICU mortality was consistent across subgroups, and to assess the robustness of our findings. ICU mortality was investigated either by comparing proportions in the matched subsets and survival analysis with causespecific hazard model. Subgroups were based on previous research as well as clinical relevance and categorized as: age (< 60, ³ 60), severity of ARDS (mild, moderate and severe), and time since the symptom onset to the initiation of corticosteroids (< 7 days, ³ 7days). To account for multiplicity and avoid the potential inflation of the type I error rate as a result of multiple testing in the subgroup analysis, we used the Benjamini-Hochberg method for controlling the false discovery rate [15,16].
Ventilator-free days were defined as follows: 28 -"x" if successfully weaned from mechanical ventilation "x" days after ICU admission; for subjects who died within 28 days of mechanical ventilation or who were mechanically ventilated for more than 28 days, ventilator-free days were coded as zero. Differences between groups were assessed by Wilcoxon rank sum test and reported as means and standard deviations. We also evaluated ventilator-free days with time-to-event analysis censored at 28 days, with the event of interest the successful liberation from the mechanical ventilation and mortality as a competitive event. We    Table S4.

Respiratory characteristics of study groups (corticosteroids vs no corticosteroids) stratified by the severity of ARDS in COVID-19 ventilated patients within the first day of ICU admission.
Data are numbers (%) or medians (interquartile range). ARDS = Acute respiratory distress syndrome; PaO2/FiO2 = arterial oxygen partial pressure to fractional inspired oxygen ratio; PEEP = positive endexpiratory pressure; RR = respiratory rate; pCO2 = partial pressure of carbon dioxide; Vt = tidal volume; PBW = Predicted body weight; ECMO = extracorporeal membrane oxygenation; MV = mechanical ventilation.

Subgroup analysis with time-dependent Cox regression and step function. Summary of hazard ratios accounting for time (short and long-term).
Analysis was performed with Cause-specific hazard model for both events accounting for the timedependency effect of corticosteroid treatment with step function by partitioning time period of follow up on day 17 th of follow up, as in the primary analysis. To account for multiplicity, "p" values were adjusted by the Benjamini-Hochberg method for controlling the false discovery rate. *Post-hoc analysis. GAP corticosteroids mean the time since symptom onset to corticosteroid exposure. Models were adjusted for gender, age, body mass index, hospital GAP, ICU GAP, diagnosis GAP, shock, ACE inhibitors, ARBs, Comorbidity, asthma, COPD, chronic kidney disease, haematological disease, diabetes mellitus, neuromuscular disease, autoimmune disease, ischemic heart disease, hypertension, immunosuppression, dyslipidaemia, hypothyroidism, APACHE II, SOFA, pulmonary infiltrates, lactate dehydrogenase, white blood cells count, creatinine, urea, C-reactive protein, procalcitonin, Lactate, Ddimer, antibiotics, oseltamivir, lopinavir plus ritonavir, remdesivir, interferon, hydroxychloroquine, Tocilizumab, bacterial co-infection, ARDS severity, fractional of inspired oxygen (FiO2), positive endexpiratory pressure, tidal volume, partial pressure of carbon dioxide, pH, RIFLE criteria, myocardial dysfunction and corticosteroid treatment (short and long-term). GAP corticosteroids mean the time (in days) since onset of symptoms to corticosteroid initiation. ICU = intensive care unit; ACE = angiotensin converting enzyme; ARBs = angiotensin receptor blockers; COPD = chronic obstructive pulmonary disease; APACHE = Acute physiology and chronic health evaluation; SOFA = sequential organ failure assessment; ARDS = acute respiratory distress syndrome; CRP = C-reactive protein.   The fourteen remaining patients with corticosteroids received hydrocortisone (n=10) or combination of corticosteroids (n=4). Large number of patients were included during the first wave in Spain (March-April 2020).

Figure S3. Schoenfeld Residual Plot for corticosteroid treatment.
The solid line (βt) gives the estimated effect of the predictor through time (with confidence intervals depicted as dashed lines). Significant violation of the assumption of proportional hazards were found for the corticosteroid treatment. Figure S4. Love plot of the covariate balance of propensity score matching with unadjusted and adjusted standardized mean differences. Hospital gap is the time period from symptoms onset to hospital admission. ICU gap is the time period from hospital to ICU admission. ICU = Intensive care unit; COPD = Chronic obstructive pulmonary disease; APACHE = Acute Physiology And Chronic Health Evaluation; SOFA = Sequential Organ Failure Assessment; ARDS = Acute Respiratory Distress Syndrome; RIFLE = RIFLE criteria for acute kidney injury (Risk, Injury, Failure, Loss, End stage).

Figure S5. Forest plot of the Cause-specific hazard model for ICU discharged alive (competitive event of ICU mortality).
Corticosteroids were associated with lower probability of discharge alive in short-term, but exposure on admission in long-term was associated with higher probability of discharge alive. BMI: body mass index. Diagnosis gap means the time (days) from disease onset to the confirmation of the diagnosis of SARS-CoV-2 infection. Hospital gap (days) means the time from disease onset to hospital admission. ICU gap (days) means the time from hospital to ICU admission. ACE = Angiotensin-converting enzyme; ARBs = Angiotensin receptor blockers; COPD = Chronic obstructive pulmonary disease; APACHE = Acute Physiology And Chronic Health Evaluation; SOFA = Sequential Organ Failure Assessment; LDH = lactate dehydrogenase; WBC = white blood cells count; CRP = C-reactive protein; ARDS = Acute respiratory distress syndrome; FiO2: fraction of inspired oxygen; PEEP = positive endexpiratory pressure; Vt = tidal volume; pCO2 = partial pressure of carbon dioxide; RIFLE criteria: Risk, Injury, Failure, Loss, End stage. Figure S6. Pre-specified subgroups analysis. Survival plots of the Cause-specific hazard model for ICU mortality trough time-dependent Cox regression with step-function. Adjusted for gender, age, body mass index, hospital GAP, ICU GAP, diagnosis GAP, shock, ACE inhibitors, ARBs, Comorbidity, asthma, COPD, chronic kidney disease, haematological disease, diabetes mellitus, neuromuscular disease, autoimmune disease, ischemic heart disease, hypertension, immunosuppression, dyslipidaemia, hypothyroidism, APACHE II, SOFA, pulmonary infiltrates, lactate dehydrogenase, white blood cells count, creatinine, urea, C-reactive protein, procalcitonin, Lactate, Ddimer, antibiotics, oseltamivir, lopinavir plus ritonavir, remdesivir, interferon, hydroxychloroquine, Tocilizumab, bacterial co-infection, ARDS severity, fractional of inspired oxygen (FiO2), positive endexpiratory pressure, tidal volume, partial pressure of carbon dioxide, pH, RIFLE criteria, myocardial dysfunction and corticosteroid treatment (short and long-term). ICU: intensive care unit, ACE: angiotensin converting enzyme, ARBs: angiotensin receptor blockers, COPD: chronic obstructive pulmonary disease, APACHE: Acute physiology and chronic health evaluation, SOFA: sequential organ failure assessment, ARDS: acute respiratory distress syndrome.

Figure S7. Post-hoc subgroup sensitivity analysis according with corticosteroid treatment duration.
Plot A depicted the survival analysis with cause-specific hazard model for ICU mortality among patients with corticosteroid treatment duration less than seven days compared with none. In a shorter course of treatment, corticosteroids were not associated with short-term effects on survival, but significant negative long-term effects were observed. Plot B showed the survival analysis with cause-specific hazard model for ICU mortality among patients with corticosteroids treatment duration up to seven days or longer. In a longer course of treatment, corticosteroids presented the same time-dependent effect on survival as in the primary analysis.

Figure S8. Post-hoc sensitivity subgroup analysis according with tocilizumab (Yes/No).
Exploratory analysis showed that patients who received Corticosteroids plus Tocilizumab had significant association with short-term survival benefit without negative long-terms effects on mortality. In the No Tocilizumab subgroup, corticosteroids had no effects in short-term survival whereas significant negative long-term effects were found. Figure S9. Survival plot of the sub-distribution hazard model for ventilator-free days accounting for competitive event (ICU death) stratified according to corticosteroid treatment. sHR is the subdistribution hazard ratio expressing the association between corticosteroids and ventilator-free days.