Background
Coronavirus disease 2019 (COVID-19) is caused by SARS-CoV-2 (severe acute respiratory syndrome coronavirus 2), which has been at the center of an ongoing global pandemic [
1]. Corticoid therapy is currently a key treatment regimen and recommended specifically for critically ill patients in the intensive care unit (ICU) [
2]. However, the effect of corticoid treatment for COVID-19 patients is still under evaluation. For example, one study showed that in hospitalized patients with COVID-19, the use of dexamethasone reduced 28-day mortality in both patients who received invasive mechanical ventilation or oxygen alone [
3]. Another systemic review and meta-analysis showed that dexamethasone reduced mortality and the use of mechanical ventilation in COVID-19 patients compared to those on standard care [
4]. A third study demonstrated that methylprednisolone administration reduced the risk of death in COVID-19 patients complicated with acute respiratory distress syndrome (ARDS), although there was no adjustment for immortal time bias and indication bias by time-varying confounding variables [
5]. In contrast, in a randomized trial from Brazil including 393 COVID-19 patients indicated no significance difference in 28-day mortality rates between administration of methylprednisolone and placebo [
6].
Due to uncertainty regarding efficacy of corticosteroid in the treatment of COVID-19, it is imperative to evaluate its role in the management of COVID-19. Immortal time bias and indication bias should be adjusted when evaluating the role of corticosteroid for COVID-19 patients in observational cohort studies. Immortal time bias refers to the requirement that patients survive long enough to receive the intervention of interest, leading to a potential incorrect overestimation of a positive treatment effect [
7]. Indication bias from time-varying confounding variables refers to having an association related to the indication of the intervention that evolves throughout the course of an illness [
8]. MSM allows proper adjustment for immortal time bias and indication bias. In this retrospective study, we investigated the therapeutic effect of corticosteroids for the patients with COVID-19 using multivariate analyses. Logistic regression, Cox proportional-hazards model (Cox) and marginal structural modeling (MSM) were utilized to explore the relationship between the initiation of corticosteroid therapy and 90-day mortality as well as SARS-CoV-2 RNA clearance. Importantly, baseline and time varying confounding factors were adjusted, with MSM allowing for proper adjustment of time-dependent confounding variables.
Methods
Study design and participants
We retrospectively studied all critically ill patients with COVID-19 admitted to 10 Intensive Care Units (ICUs) in Hubei province between December 30, 2019 and February 19, 2020. The data was censored on April 9, 2020. All patients were consecutive. Those receiving chronic corticosteroid therapy before the onset of critical illness were excluded. COVID-19 was diagnosed according to World Health Organization interim guidance. Critically ill patients were defined as those admitted to the ICU who required mechanical ventilation or had a fraction of inspired oxygen of at least 60%. This study was approved by the institutional ethics board of the participating medical centers (NO. 2020045).
Data collection
Data was extracted from the medical records of patients, reviewed and verified by the research team from the Department of Critical Care Medicine, Zhongnan Hospital of Wuhan University. Extracted data included epidemic data, demographic data, comorbidities, complications, treatments and time course of illness. Laboratory findings and respiratory mechanic variables on day 1 of ICU admission were also collected. Severity of illness was assessed using the Acute Physiology and Chronic Health Evaluation (APACHE) II score and organ dysfunction was assessed using the Sequential Organ Failure Assessment (SOFA) score, both of which were recorded on day 1 of ICU admission. The primary outcome was determined as 90-day mortality after ICU admission and secondary outcome as duration of SARS-CoV-2 RNA clearance.
ARDS was diagnosed according to the Berlin Definition and acute kidney injury (AKI) was diagnosed according to the Kidney Disease: Improving Global Outcomes (KDIGO) clinical practice guidelines [
9,
10]. Acute cardiac injury was defined as blood levels of hypersensitive troponin I above the upper reference limit (> 26.2 pg/ml). Duration of SARS-CoV-2 RNA clearance was defined as the time from ICU admission until the occurrence of two negative Nucleic-acid Amplification tests (NAAT) without a positive test afterward. Corticosteroid therapy was defined as the use of systemic corticosteroids when dosage equivalent to at least 60 mg of hydrocortisone per day [
11]. The type, daily dose, and duration of corticosteroid administration were collected. All preparations were converted to hydrocortisone-equivalent doses (methylprednisolone 1:5, dexamethasone 1:25, prednisolone 1:4).
Statistical analysis
Categorical variables were presented as frequency rates and percentages. Continuous variables were described using mean and standard deviations or medians and IQR values according to the distribution of the data. We compared characteristics of patients who received or not received corticosteroid therapy using chi-square test or Fisher’s exact test for categorical variables and student t-test or Mann–Whitney test for continuous variables, as appropriate. Subgroup analyses were performed by stratification to (1) high-dose corticosteroid therapy (≥ 300 mg of hydrocortisone equivalent per day) and low-dose corticosteroid therapy (< 300 mg of hydrocortisone equivalent per day) versus no treatment with corticosteroids, respectively; (2) corticosteroid therapy initiated in the first 3 days of ICU admission versus no corticosteroid therapy; (3) corticosteroid therapy started after Day 3 versus no corticosteroid therapy.
All statistical analyses were performed using R Project. A
p value less than 0.05 was considered statistically significant. We calculated the a priori sample size using online software (A-priori Sample Size Calculator for Multiple Regression:
https://www.danielsoper.com/statcalc/calculator.aspx?id=1). Totally, 14 variables were selected in our study. We assumed that anticipated effect size was 15%, desired statistical power was 80%, alpha level (type I error) was 0.05 [
12‐
14]. The calculated sample size was 135. In current study, 294 patients were included (survivor: 148, non-survivor: 146). We tested the associations of corticosteroid therapy and 90-day mortality or SARS-CoV-2 RNA clearance using three approaches: logistic regression, Cox and MSM.
Logistic regression
Logistic regression adjusted for baseline differences only. We included in the multivariable model a priori–decided baseline variables of clinical interest and all significant variables at the univariable level (p < 0.2).
Cox proportional hazards regression
Cox was used to adjust for baseline differences and immortal time bias. To examine outcomes as a time to event (i.e., death), we performed a Cox adjusting for the same above-mentioned baseline covariates with corticosteroid therapy as a time-varying covariate.
Marginal structural model
MSM was used to adjust for baseline differences and time varying confounders (indication bias and immortal time bias). Time-varying variables were determined as PaO2/FiO2, ventilation status and SOFA scores on the day of corticosteroid therapy initiation. Refining marginal structural models involves the calculation of two weights for each observation: treatment selection weight and censoring weight. The treatment selection weight at time k is a ratio of two weights. The numerator is the product of probabilities that a patient receives his observed treatment at time k, given the baseline variable. The denominator is calculated similarly by incorporating the time-varying variable (SOFA on the day, SOFA on the previous day, mode of ventilation on the day and mode of ventilation on the previous day). The weights are recalibrated until the first day of corticosteroid therapy and kept constant afterward. Since SOFA scores were recorded on Days 1, 3, 7, 14, and 28, we imputed missing values for the remaining days using the following rules: (a) If ICU length of stay (LOS) ≥ 28, we used linear interpolation for each of the intervals 4–6 days, 8–13 days and 15–27 days. (b) If 14 ≤ ICU LOS ≤ 27, we used linear interpolation for 4–6 days and 8–13 days and the Last-Observation-Carried-Forward method from day 15 onwards. (c) Missing values are managed similarly for ICU LOS less than 14 days. Ventilation status was recorded also on Days 1, 3, 7, 14, and 28. We imputed the mode of ventilation between these days using the Last-Observation-Carried-Forward method.
The same approach was used to calculate the censoring weight for early patient dropout. Patients were considered censored at hospital discharge or at Day 90, and the weights for censoring are calculated as the ratio of a subject’s probability of remaining uncensored up to day k. The final weight for each observation was obtained by multiplying the treatment selection weights and the censoring weights. We used weight truncation to manage outliers; weights greater than the 95th percentile value were fixed at the 95th percentile value, and weight less than the 5th percentile value were fixed at the 5th percentile value. This process continued until the mean of weight was close to 1. The time-varying intercept was modeled by a smoothing function of time, using restricted cubic splines with five knots for days from beginning of follow-up day. For the analysis of the associations of corticosteroid therapy and SARS-CoV-2 RNA clearance, patients were censored if they never cleared SARS-CoV-2 RNA or if they were discharged alive from the hospital before clearing SARS-CoV-2 RNA.
Discussion
Our study included 294 critically ill COVID-19 patients, with 62.2% patients having received corticosteroid therapy. Among those who received corticosteroid therapy, 96% were treated with methylprednisolone and 69.4% received corticosteroid treatment prior to ICU admission. After adjustment of baseline and time-dependent variables, an association between early corticosteroid treatment (≤ 3 days after ICU admission) and higher mortality was observed. Additionally, our study did not find an association between corticosteroid use and SARS-COV-2 RNA clearance.
Corticosteroids were widely used in the outbreaks of COVID-19 [
15]. Severe COVID-19 is associated with a dysregulated host inflammatory response. The increasing number of infected epithelial cells and cell debrides trigger a massive cytokine release—the so-called ‘cytokine storm’ [
16]. Histological examination of lungs from COVID-19 patients showed diffuse alveolar damage with cellular fibromyxoid exudates. Moreover, interstitial mononuclear inflammatory infiltrates, dominated by lymphocytes, were visible in lungs [
17]. These findings provided a rationale for the use of corticosteroids in order to reduce inflammation. Consequently, the most severely affected COVID-19 patients (those with severe hypoxemic and higher SOFA scores) in our study were administrated corticosteroids (particularly methylprednisolone).
We found no significant association between corticosteroid treatment and 90-day mortality when subgroup analysis was not performed. Similarly, in Siemieniuk et al., short course treatment with methylprednisolone in hospitalized patients with COVID-19 was not associated with a reduction in mortality [
4]. However, the RECOVERY trial by Huang et al. revealed that the use of dexamethasone resulted in lower 28-day mortality in patients randomized to receive invasive mechanical ventilation or oxygen alone, but not in patients who received no respiratory support [
1]. Dexamethasone was selected in RECOVERY study, but methylprednisolone was the most commonly-used corticosteroid in our study. Differences in treatment regimens may lead to different anti-inflammatory effects and may have varying outcomes [
18]. Additionally, the median duration of glucocorticoid therapy was 7 days in the RECOVERY trial and 9 days in our study. Long-term use of corticosteroids may cause adverse effects that obscure their effectiveness. Fadel et al. also showed that early short course of methylprednisolone reduced the requirement for mechanical ventilation and mortality [
19]. However, since this study was a quasi-experimental study, the effects of multiple independent variables and their interactions on dependent variables cannot be analyzed. Additionally, concerns regarding statistical power should be considered in that study.
Notably, association of early initiation of corticosteroid therapy (≤ 3 days after ICU admission) and higher mortality was revealed when subgroup analysis was performed in our study, suggesting that patient selection had an influence on the outcome in COVID-19 patients who were initiated on corticosteroid treatment. For example, Siemieniuk et al. showed that corticosteroid treatment reduced mortality in hospitalized COVID-19 above the age of 60, but had no beneficial effects less severe patients [
4]. Similarly, a benefit was not seen among less severe patients who did not require oxygen support in RECOVERY trial; there was no statistically significant upward trend in mortality [
1]. Previous studies determined that the median time of illness onset to ARDS was 12 days for COVID-19 patients [
20,
21]. In our study, the median time of illness onset to ICU admission was 11 days, with most corticosteroid therapy initiated prior to ICU admission. Therefore, we speculate that the adverse effects of corticosteroid therapy may be harmful for COVID-19 patients without ARDS. The anti-inflammatory role of steroids may be useful to mitigate the cytokine storm in COVID-19 patients with ARDS [
22,
23].
Our results did not reveal that corticosteroid use was associated with SARS-CoV-2 RNA clearance. Several observational studies have reported that corticosteroid therapy was linked to persistent viral RNA shedding in patients with avian influenza A (H7N9), MERS, and SARS [
24‐
26]. A retrospective study conducted by Xu et al. showed that prolonged viral RNA shedding in patients with COVID-19 was associated with corticosteroid treatment [
27]. Herein, after adjusting for both immortal bias and indication bias, we found that corticosteroid usage was not associated with prolonged viral RNA shedding time. Accordingly, a retrospective COVID-19 study by Shi et al. indicated that corticosteroid had little effect on the duration of viral excretion [
28]. The adverse effects of early initiation of corticosteroid therapy may be mediated by other side effects and require further investigation.
Our multi-center study analyzed the relationship between corticosteroid treatment and 90-day mortality of COVID-19 patients using logistic regression, Cox and MSM. Statistical power of our study is compelling and credible evidence for the effect of the corticosteroid treatment in COVID-19 patients was supplied. However, we also determined some limitations to our analyses. Firstly, the majority (127/183 or 69%) of the patients received corticosteroids prior to ICU admission, creating potential sampling bias, which may affect the evaluation of the role of corticosteroids. Secondly, we only focused on the effect of corticosteroids on mortality and RNA shedding. Arrhythmias, gastrointestinal bleeding, secondary infection and other corticosteroid-related side effects should be explored in future.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.