Efficacy and safety of corticosteroid regimens for the treatment of hospitalized COVID-19 patients: a meta-analysis

In March 2020 the WHO designated COVID-19, which is caused by SARS-CoV-2, as a global pandemic [1–4]. The spread of SARS-CoV-2 has been considered to be a severe public health crisis due to its high transmission, hospitalization and mortality rates [5–8]. One of the largest factors attributed to the high mortality rate associated with COVID-19 was the depletion of hospital resources as a result of surges in cases [9–11]. The care of critically ill COVID-19 patients require valuable and scarce supplies such as ICU beds, ventilators, personal protective equipment and hospital staff [12]; thus, physicians have sometimes been forced to triage critically ill COVID-19 patients [13] and divert resources from other specialties such as oncology, leading to the rescheduling of appointments for non-COVID patients [14].

Recent advancements in SARS-CoV-2 vaccination programs, including the US FDA-approved Comirnaty vaccine produced by Pfizer-BioNTech [15], aims to curb the transmission of the virus, induce herd immunity and relieve the pressure of the pandemic on hospital resources. However, new viral variants of interest (VOIs), such as Lambda and Delta [16], threaten to introduce new COVID-19 surges in the coming months with their increased infectivity [17], immune resistance [17,18] and reduced sensitivity to antibody neutralization [19]. These newly discovered variants, in addition to widespread vaccine hesitancy movements [20] and poor vaccine accessibility in low- to middle-income countries (LMICs) [21] are likely to extend the pandemic well into 2022 [22]. Therefore, further investigations are still required to determine the optimal treatment regimen for the management of COVID-19 surges in the coming months.

Currently, most of the proposed COVID-19 treatment regimens consist of repurposed pharmacological compounds that were initially chosen based on their antiviral efficacy against SARS-CoV-2 in vitro or in small observational studies. However, many repurposed regimens such as hydroxychloroquine and lopinavir/ritonavir have been consistently shown in randomized controlled trials (RCTs) and meta-analyses to confer no efficacy against SARS-CoV-2 in clinical settings [23–26]. In early 2020, a series of observational and prognostic studies brought a new category of repurposed therapies, corticosteroids, into the limelight [27]. These early findings are corroborated by encouraging results from the RECOVERY trial, which is the world's largest clinical trial for COVID-19 treatments to date [28]. The RECOVERY trial concluded that dexamethasone may confer survival benefits among patients receiving mechanical ventilation. Unfortunately, no other authoritative RCTs, including the REMAP-CAP [29] and CoDEX trial [30], were able to identify a significant beneficial effect associated with corticosteroids. This was partially due to insufficient sample sizes as a result of premature trial terminations following the publication of results from the RECOVERY trial [31]. Additionally, pre-pandemic studies of corticosteroids in other viral pneumonias, including the original 2003 SARS-CoV [32], the Middle East respiratory syndrome (MERS) [33] and influenza [34], have associated the use of corticosteroids with delayed viral clearance and increased mortality. These conflicting findings call into question the efficacy of corticosteroids, prompting further investigations into this category of treatments.

Biologically, the rationale for the use of corticosteroids in the treatment of COVID-19 is to reduce excessive host inflammatory responses in the lungs through the inflammatory phase of the disease, during which immunopathological factors play a bigger role in disease progression compared with viral replication [35,36]. For some patients, impaired interferon responses [37,38] and the activation of intracellular antiviral pathways triggered by viral RNA may lead to symptoms that mimic macrophage activation syndrome (MAS) during the inflammatory phase [39]. MAS is associated with impaired cytolytic capabilities of CD8⁺ T cells and natural killer (NK) cells [40,41], which are normally tasked with lysing infected cells to prevent viral replication and the excessive secretion of inflammatory cytokines [42]. This impairment is driven by high levels of pro-inflammatory cytokine IL-6, which forms a vicious cycle of cytokine-driven cytokine secretion [43] leading to cytokine storms [39]. IL-6 is also implicated in the excessive differentiation of Th2 cells and Th17 cells while inhibiting the differentiation of antiviral Th1 cells [40,44–46]. This action prevents the host immune system from mounting an effective antiviral response while inducing a nonproductive, eosinophilic immune response [47]. COVID-19 patients with elevated IL-6 levels and eosinophilic immune responses have been associated with the development of severe lung injuries and acute respiratory distress syndrome (ARDS) [48,49].

As one of the most commonly used therapies for cytokine storms and MAS in patients with autoimmune and inflammatory diseases, corticosteroids appear to be the ideal regimen for treating COVID-related hyperinflammatory disorders [50]. With appropriate administration times and dosages, it is proposed that corticosteroids can reduce the production of proinflammatory cytokines and inhibit overactive immune cells, thus preventing lung damage and ARDS [51–53]. On the other hand, corticosteroids' immunosuppressive properties may impair viral clearance and increase the rate of dangerous nosocomial infections [54]. It is evident that the role of corticosteroids requires further clarification to examine whether its immunomodulatory efficacy outweighs its adverse effects. Thus, the current meta-analysis assessed the efficacy and safety of corticosteroids compared with standard of care, and studied the impact of factors such as dosage and time of administration on patient-important outcomes.

Methods

We conducted this systematic review and meta-analysis in accordance with the PRISMA 2020 guidelines [55]. The completed PRISMA 2020 checklist is provided in Supplementary Table 1. This review was prospectively registered on PROSPERO (CRD42021233108), the international prospective register of systematic reviews [56].

Study identification

To identify relevant studies for inclusion in this review, we created an English search strategy based on database-specific COVID-19 search strings provided by the Rudolph Matas Library of the Health Sciences, Tulane University (LA, USA) [57] in combination with treatment-relevant keywords such as ‘dexamethasone’, ‘prednisone’, ‘methylprednisolone’, ‘corticosteroid’, ‘glucocorticoid’ and ‘steroid’ etc. Using this strategy, we searched the following databases from 1 January 2020 to 12 July 2021: Medical Literature Analysis and Retrieval System Online (MEDLINE), Excerpta Medica Database (EMBASE), and PubMed. Additionally, we utilized a Chinese search strategy equivalent to our English strategy, and systematically searched the following Chinese literature databases from 1 January 2020 to 12 July 2021: Wanfang Data, Wanfang Med Online, SinoMed, China National Knowledge Infrastructure (CNKI) and Chongqing VIP Information (CQVIP). The search strategies used for the database searches are appended in Supplementary Tables 2–9. Last, we hand searched the reference sections of six previous systematic reviews [34,58–62] to identify relevant studies that were not identified by our database searches.

Eligibility criteria

We included both randomized and non-randomized comparative studies that met the following inclusion criteria: compared any corticosteroid therapy with standard of care, or compared any corticosteroid therapy with an adjuvant therapy to adjuvant therapy alone, and included hospitalized COVID-19 patients. We only included studies using adjuvant therapies if the adjuvant therapy between the control and experimental arms are the same in order to minimize the effect of the adjuvant therapies on treatment outcomes.

Outcome measures

Our efficacy outcomes included: time to negative conversion of SARS-CoV-2 tests, overall length of hospitalization, length of ICU hospitalization, incidence of progression to ICU hospitalization, incidence of mortality, incidence of progression to mechanical ventilation and length of mechanical ventilation. The safety outcomes included: incidence of all-cause adverse events, incidence of investigator-defined severe adverse events, incidence of hyperglycemia or corticosteroid-induced diabetes mellitus and incidence of nosocomial bacterial or fungal infections.

Study selection

All title/abstract entries identified from our literature searches were entered into an independent and in duplicate title/abstract screening process via Rayyan (https://www.rayyan.ai/) [63]. Abstract entries that were deemed eligible were then entered into a subsequent in duplicate full-text screening process. During the selection process, all disagreements were referred to a senior author for resolution. The PRISMA flowchart [64] depicting our study selection process is shown in Figure 1.

Data extraction

We performed data extraction independently and in duplicate using an extraction sheet developed a priori. Categories of extracted information included: patient demographics and baseline characteristics, descriptions of study methodology, treatment descriptions and outcome measures. The full list of extracted items are listed on our PROSPERO registration page. For observational studies reporting data for both matched and unmatched cohorts, we only extracted and analyzed the data from matched cohorts. All disagreements were resolved by recruiting a senior author to review the data.

Missing data

For studies with missing data required for meta-analysis, including mean values for continuous outcomes and measures of variance, we made attempts to contact the corresponding authors to obtain unpublished data. For studies that presented median and interquartile range (IQR) for continuous outcomes, we used methods recommended by Luo et al. [65] and Wan et al. [66] to estimate the mean and standard deviation (SD) for meta-analysis, given that the data is normally distributed as examined using methods proposed by Shi et al. [67]. We did not include skewed median and IQR values in the meta-analysis and elected to describe these results narratively.

Risk of bias assessment

The risk of bias in our included RCTs were assessed independently and in duplicate using version 2 of the Cochrane risk-of-bias tool for randomized trials (RoB2) [68]. The risk of bias of non-randomized comparative studies was assessed in duplicate using the risk of bias in non-randomized studies of interventions (ROBINS-I) tool [69]. All disagreements were referred to a senior author for resolution.

Quality of evidence

We assessed the quality of evidence for our outcomes using the Grading of Recommendations, Assessment, Development and Evaluations (GRADE) framework [70,71]. A summary of the meta-analysis outcomes and GRADE ratings were presented in a GRADE summary of findings table [72] generated using GRADEpro (https://gradepro.org/) [73].

Statistical analysis

All statistical analyses were conducted using R v4.1.1 (https://www.r-project.org/) [74]. For outcomes with more than two studies presenting eligible data, we performed random-effects meta-analyses using the meta v4.19 library (https://cran.r-project.org/web/packages/meta/) [75]. For dichotomous/incidence outcomes, we expressed the treatment effect as odds ratios (ORs) and 95% CIs for data pooling. The treatment effects of continuous outcomes were expressed as mean differences (MDs) and 95% CIs. Results from individual studies and data synthesis were displayed on forest plots.

Zero event studies

For studies reporting zero events in one or both of its treatment arms, we applied treatment arm continuity correction (TACC) [76] to complete the meta-analysis.

Heterogeneity assessment

The presence of heterogeneity was examined using the Cochran's Q test [77] with a significance level of p < 0.10 as recommended by the Cochrane Handbook [78]. Heterogeneity was subsequently quantified using I² statistics [77,79]. We interpreted 30% <I² <75% as moderate heterogeneity and I² ≥75% as serious heterogeneity, as per recommendations from the Cochrane Handbook [78].

Publication bias & small study effects

We created funnel plots for outcomes with 10 or more included studies to identify small study effects within the meta-analysis as an indication for publication bias [78]. Egger's regression test was used to quantitatively evaluate the presence of asymmetry within the funnel plots [80]. We did not draw funnel plots nor conduct Egger's regression test for outcomes with fewer than 10 included studies as Egger's test does not have sufficient power in these circumstances [81].

Meta-regression & subgroup analyses

We performed meta-regression analyses on the proportion of patients with severe disease (defined as per individual investigators' criteria), mean baseline sequential organ failure assessment (SOFA) scores, and follow-up duration for mortality.

Furthermore, we performed the following subgroup analyses: corticosteroid regimens, corticosteroid dose, with methylprednisolone equivalent dose of ≤1.5 mg/kg/day or ≤100 mg/day defined as low dose and methylprednisolone equivalent dose of >1.5 mg/kg/day or >100 mg/day defined as high dose, similar to the threshold set in previous studies [82,83], time from hospitalization to corticosteroid administration, with a time to administration ≤3 days defined as early/immediate administration, and a time to administration >3 days defined as late/delayed administration. Additionally, we performed subgroup analyses by study design (randomized studies vs observational studies) and risk of bias (low/some concerns/moderate risk of bias vs high/serious/critical risk of bias) to examine the impact of different study methodologies.

Results

Included studies

We identified and screened 9865 potentially eligible titles and abstracts (following deduplication using Endnote 20 [https://endnote.com/]). A total of 250 full-text articles were subsequently retrieved and screened. After screening, 10 RCTs [28–30,84–90] and 85 observational studies [83,91–174] were included in this systematic review and meta-analysis (Figure 1), with a total of 42,205 hospitalized COVID-19 patients. In the article selection and data extraction process, we contacted the corresponding authors from seven studies [95,147,149,155,175–177] with unclear or missing data and two authors [149,155] responded to our inquiries. Of the remaining studies, two [95,147] were included with a subset of the reported outcomes without response from authors, and three studies [175–177] were excluded due to a lack of relevant outcomes/insufficient data.

Across all included studies, 37 studies (38.9%) [84,86,89,90,93,94,99,101,104,106,107,112,115,119,121,125,126,130,136,138,142,143,146,147,152–155,161,162,165,167–170,172,174] included treatment arms that assessed methylprednisolone, 10 studies (10.5%) [28,30,88,100,105,114,117,125,141,171] included treatment arms that assessed dexamethasone, 3 studies (3.2%) [29,85,87] administered hydrocortisone, 1 study (1.1%) [92] administered prednisone and 1 study (1.1%) [145] administered prednisolone. The study by Ko et al. [125] had two corticosteroid arms, with one arm receiving dexamethasone and the other arm receiving methylprednisolone. The remaining studies did not specify the exact corticosteroids used and/or only provided equivalent dosages. In terms of adjuvant therapies, 4 studies [93,113,150,160] included tocilizumab as adjuvants, while 1 study [135] used interferon-α2b as an adjuvant.

Forty-six studies (48.4%) [28,83,84,86,90,91,96,98,100,101,104,107–110,112,115–117,120–122,124,130,132,135,137,139,141,143–145,150,151,154,156–158,164,165,168–170,172,173] included all hospitalized COVID-19 patients, 20 studies (21.1%) [87,93,97,99,102,103,126,131,133,134,136,138,142,146,147,152,155,160,163,174] only included hospitalized patients with severe disease, while 6 studies (6.3%) [88,92,106,111,119,128] included hospitalized patients without severe disease. Last, 22 studies (23.2%) [29,30,85,89,94,105,113,114,118,123,125,127,129,140,148,149,153,159,161,162,167,171] only included ICU hospitalized patients, and 1 study (1.1%) [166] only included hospitalized patients receiving mechanical ventilation. Characteristics of the included studies are tabulated in Supplementary Table 12.

Risk of bias

The risk of bias in the included RCTs was evaluated using RoB2 (Figure 2A & Supplementary Table 10). Seven RCTs (70.0%) [29,30,85,86,88–90] were rated as having some concerns in regards to their risk of bias, mainly due to the use of open-label study designs. The remaining 3 RCTs [28,84,87] were rated as having a low risk of bias. No RCT was rated as having a high risk of bias.

The risk of bias in our included observational studies were assessed using ROBINS-I (Figure 2B & Supplementary Table 11). Thirty-one studies (36.5%) [92,95,96,98,99,103,105–107,109,112,115,118,120,121,126,133,135,139,143,149,150,152,156,160,165,166,170,172–174] were rated as having critical risk of bias, and 33 studies (38.8%) [91,94,97,100,101,108,110,114,117,119,124,125,128–132,134,136,138,140,141,144–147,158,161,163,164,167–169] were rated as having serious risk of bias. Generally, studies were rated as having a critical or serious risk of bias due to confounding factors and poor descriptions of the regimens/interventions used. Fourteen studies (16.5%) [93,102,104,111,113,116,122,137,142,151,154,155,157,162] were rated as having moderate risk of bias, with no studies receiving a rating of low risk of bias. The remaining 7 studies (8.2%) [83,123,127,148,153,159,171] were rated as ‘no information’ due to insufficient reporting of study methodology, treatment descriptions and baseline information.

Treatment outcomes

Time to negative test conversion

Nineteen studies [86,102,106,112,119,121,128,129,131–133,136–138,164,165,169,170,172] with 2785 patients evaluated the effect of corticosteroid administration on time to negative test conversion and were included in the meta-analysis (Figure 3). The overall pooled MD was 1.80 (95% CI: 0.48–3.11) with serious and significant heterogeneity (I² = 78%; P_Q <0.01). There was no noticeable asymmetry on the funnel plot (Supplementary Figure 1), and Egger's regression test showed no significant small-study effects (P_Egger = 0.82). There were no significant subgroup differences between the pooled results from observational studies versus RCTs (p = 0.26; Supplementary Figure 2). Similarly, there were no significant subgroup differences between pooled results from high risk of bias studies compared with low risk of bias studies (p = 0.77; Supplementary Figure 3).

Ten studies [86,106,112,119,121,136,165,169,170,172] assessing 879 patients compared methylprednisolone versus standard of care, while the remaining studies did not disclose the exact corticosteroid regimens used (Supplementary Figure 4). Including only studies that administered methylprednisolone, the pooled MD was 1.81 (95% CI: 0.33–3.29) with substantially decreased heterogeneity compared with the overall analysis (I² = 32%; P_Q = 0.15). There were no significant subgroup differences between subgroups of studies assessing low fixed-dose regimens versus low weight-adjusted dose regimens (p = 0.53; Supplementary Figure 5), nor were there significant subgroup differences between subgroups of studies providing early administration of corticosteroids versus late administration of corticosteroids (p = 0.95; Supplementary Figure 6). Meta-regression analyses did not demonstrate significant correlations between the treatment effect and the proportion of severe patients (p = 0.70; Supplementary Figure 7) or baseline mean SOFA scores (p = 0.53; Supplementary Figure 8).

Three observational studies [120,122,143] reported median time to negative test conversion with IQR, however, these data could not be imputed due to a lack of normality. Based on the Kruskal–Wallis H test, the Mann-Whitney U test and the Student's t-test, all three studies reported no significant differences in the duration to negative test conversion in patients administered corticosteroids versus control.

Length of stay

Thirty studies [86,88,90,92–94,98,102–104,107,115,116,122,126,128,134,135,138,144,147,151,162,164–167,170,172,173] assessing 10,454 patients compared the effect of corticosteroids versus control on length of stay and were included in the meta-analysis (Figure 4). The overall pooled MD was 1.60 (95% CI: -0.20–3.40) with serious and significant heterogeneity (I² = 91%; P_Q <0.01). There was no funnel plot asymmetry based on visual inspection (Supplementary Figure 9), and Egger's regression test showed no significant small-study effects (P_Egger = 0.36).

Subgroup analysis by different study design identified significant between-group differences between the pooled treatment effect from observational studies compared with RCTs (p < 0.05; Supplementary Figure 10). The pooled MD based on 27 observational studies [92–94,98,102–104,107,115,116,122,126,128,134,135,138,144,147,151,162,164–167,170,172,173] was 1.38 (95% CI: -0.62–3.38) with no reduction in heterogeneity (I² = 92%; P_Q <0.01), while the pooled MD based on 3 RCTs [86,88,90] was 3.62 (95% CI: 1.10–6.15) with no heterogeneity (I² = 0%; P_Q = 0.75). There were no significant subgroup differences between pooled results from high risk of bias studies compared with low risk of bias studies (p = 0.13; Supplementary Figure 11).

In addition, there were significant between-group differences between studies using different types of corticosteroids (p < 0.01; Supplementary Figure 12). One study [88] administered dexamethasone and reported the highest MD in length of stay (MD 4.64 [95% CI: 1.20–8.09]). Fourteen studies [86,90,94,104,107,115,126,147,151,162,165,167,170,172] administered methylprednisolone and reported a pooled MD of 0.28 (95% CI: -2.65–3.22). Last, one study [92] administered prednisone and reported the lowest MD of -5.93 (95% CI: -8.23 to -3.63). The remaining studies did not report the exact corticosteroid regimen used in the study.

In terms of subgroup analysis by dosage, there were no significant between-group differences in the pooled MD from studies administering high fixed-dose, high weight-adjusted dose, low fixed-dose, or low weight-adjusted doses of corticosteroids (p = 0.54; Supplementary Figure 13). There were also no significant differences between the pooled MD from studies with early corticosteroids administration versus studies with late corticosteroid administration (p = 0.14; Supplementary Figure 14). Meta-regression analyses did not demonstrate significant correlations between the treatment effect and the proportion of patients with severe disease (p = 0.08; Supplementary Figure 15) or baseline mean SOFA scores (p = 0.28; Supplementary Figure 16).

ICU length of stay

Seven studies [88,98,113,125,140,162,167] assessing 3233 ICU hospitalized patients compared the effect of corticosteroids on ICU length of stay compared with control and were included in the meta-analysis (Figure 5A). The overall pooled MD was -0.01 (95% CI: -3.50–3.47) with serious and significant heterogeneity (I² = 93%; P_Q <0.01). We did not draw the funnel plot and conduct Egger's regression test as there were less than 10 included studies.

Subgroup analysis by study design revealed significant differences between the overall MD from observational studies versus RCTs (p < 0.05; Supplementary Figure 17). Across the 6 observational studies [98,113,125,140,162,167], the pooled MD was -0.66 (95% CI: -4.48–3.16) with no reduction in heterogeneity (I² = 92%; P_Q <0.01). Only one study [88] was included in the RCT subgroup, which reported a MD of 4.00 (95% CI: 1.65–6.35). There were no significant subgroup differences between pooled results from high risk of bias studies compared with low risk of bias studies (p = 0.15; Supplementary Figure 18).

There were no significant differences between subgroup results from the subgroup analyses by corticosteroid type (p = 0.14; Supplementary Figure 19) or dose (high fixed-dose vs low weight-adjusted dose, p = 0.75; Supplementary Figure 20). However, there were significant subgroup differences between studies with early administration of corticosteroids versus studies with late administration of corticosteroids (p < 0.01; Supplementary Figure 21). Two studies administering early steroids yielded a pooled MD of 2.00 (-1.63–5.63) with a substantial reduction in heterogeneity compared with the overall meta-analysis (I² = 0%; P_Q = 0.75), while one study administering late corticosteroid yielded an MD of 10.50 (95% CI: 7.57–13.43). Last, meta-regression analyses did not reveal any significant correlations between the treatment effect and the proportion of patients with severe disease (p = 0.82; Supplementary Figure 22) or baseline mean SOFA scores (p = 0.87; Supplementary Figure 23).

Five studies [94,105,114,129,134] reported median length of ICU stay with IQR. However, as the data were significantly skewed, they could not be imputed and thus were not included in the meta-analysis. Based on the Mann-Whitney U test, two studies [94,129] did not find a significant difference between the median length of ICU stay in patients taking corticosteroids versus control, two studies [105,134] found that the length of ICU stay is significantly longer in patients taking corticosteroids versus control, and one study [114] found that the length of ICU stay is significantly longer in control patients versus patients taking corticosteroids.

ICU admission

Seventeen studies [86,90,91,93,97,98,102,104,107,110,116,131,134,139,144,155,157] with 10,391 patients compared the effect of corticosteroids versus control on incidence of ICU admission and were included in the meta-analysis (Figure 5B). The overall pooled OR was 0.87 (95% CI: 0.52–1.47) with serious and significant heterogeneity (I² = 87%; P_Q <0.01). Visual examination of the funnel plot (Supplementary Figure 24) and Egger's regression test showed no evidence of small-study effects (P_Egger = 0.24). There were no significant between-group differences in the subgroup analysis by study design (p = 0.42; Supplementary Figure 25) and by risk of bias (p = 0.61; Supplementary Figure 26).

Five studies [86,90,104,107,155] with 552 patients assessed the efficacy of methylprednisolone versus standard of care. The remaining studies did not disclose the exact type of corticosteroid used. The pooled OR for studies using methylprednisolone was 0.53 (95% CI: 0.29–0.97) with no heterogeneity (I² = 0%; P_Q = 0.71; Supplementary Figure 27).

There were no significant differences between the subgroup results from studies using different dosages of corticosteroids (p = 0.36; Supplementary Figure 28), nor were there any differences between the subgroup results from studies administering early corticosteroids versus late corticosteroids (p = 0.43; Supplementary Figure 29). Last, meta-regression analyses did not reveal significant correlations between the treatment effect and the proportion of patients with severe disease (p = 0.52; Supplementary Figure 30) and baseline mean SOFA score (p = 0.29; Supplementary Figure 31).

Mortality

Eighty-six studies [28–30,83–94,96–99,101–108,110–125,127–129,131–135,137–142,144–164,166–169,171–174] with 40,623 patients compared the effect of corticosteroids versus control in terms of mortality incidence and were included in the meta-analysis (Figure 6). The overall pooled OR was 1.00 (95% CI: 0.84–1.20) with serious and significant heterogeneity (I² = 76%; P_Q <0.01). Visual inspection of the funnel plot (Supplementary Figure 32) and Egger's regression test revealed no evidence of small-study effects (P_Egger = 0.44). Subgroup analyses showed that there were no significant differences in the pooled OR between studies of different designs (p = 0.27; Supplementary Figure 33) or between studies rated as high risk of bias versus studies with a low risk of bias (p = 0.17; Supplementary Figure 34).

Subgroup analyses by different corticosteroid regimens revealed significant between-group differences (p < 0.05; Supplementary Figure 35). Eight studies [28,30,88,105,114,117,141,171] using dexamethasone yielded a pooled OR of 0.94 (95% CI: 0.78–1.13) with no heterogeneity (I² = 0%; P_Q = 0.72). Three studies [29,85,87] assessed the efficacy of hydrocortisone and yielded a pooled OR of 0.88 (95% CI: 0.11–6.84) with moderate heterogeneity (I² = 55%; P_Q = 0.11). Thirty studies [84,86,89,90,94,99,101,104,106,107,112,115,119,121,138,142,146,147,151–155,161,162,167–169,172,174] assessed the efficacy of methylprednisolone and yielded a pooled OR of 0.62 (95% CI: 0.45–0.85) with moderate heterogeneity (I² = 36%; P_Q <0.05). Last, one study [145] assessed the efficacy of prednisolone with an OR of 0.38 (95% CI: 0.01–13.17), and another [92] assessed the efficacy of prednisone with an OR of 0.06 (95% CI: 0.01–0.59).

There were no significant differences between the subgroup results from studies using different dosages of corticosteroids (p = 0.06; Supplementary Figure 36), nor were there any differences between the subgroup results from studies administering early corticosteroids versus late corticosteroids (p = 0.28; Supplementary Figure 37). Last, meta-regression analyses did not reveal significant correlations between the treatment effect and the proportion of patients with severe disease (p = 0.56; Supplementary Figure 38), baseline mean SOFA score (p = 0.48; Supplementary Figure 39), and follow-up duration (p = 0.14; Supplementary Figure 40).

Incidence of mechanical ventilation

Thirty-seven studies [28,86,88,89,95–98,100–105,107,113,116,122,124,130,131,134,136,138,139,141,144–146,148,155–158,162,164,167] with 20,453 patients compared the effect of corticosteroids versus control in terms of incidence of mechanical ventilation and were included in the meta-analysis (Figure 7A). The overall pooled OR was 1.19 (95% CI: 0.81–1.76) with serious and significant heterogeneity (I² = 84%; P_Q <0.01). Visual inspection of the funnel plot (Supplementary Figure 41) and Egger's regression test showed no evidence of small-study effects (P_Egger = 0.90), and there were no significant between-group differences among studies using different study designs (p = 0.73; Supplementary Figure 42) or studies with high risks of bias versus studies with low risks of bias (p = 0.45; Supplementary Figure 43).

There were no significant differences between studies using different corticosteroid regimens (p = 0.95; Supplementary Figure 44), studies using different corticosteroid doses (p = 0.56; Supplementary Figure 45), or between studies administering early corticosteroids versus late corticosteroids (p = 0.89; Supplementary Figure 46). Last, meta-regression analyses revealed no significant correlations between the treatment effect and proportion of patients with severe diseases (p = 0.63; Supplementary Figure 47) and baseline mean SOFA score (p = 0.88; Supplementary Figure 48).

Length of mechanical ventilation

Seven studies [30,94,98,126,138,141,159] with 977 patients receiving mechanical ventilation compared the effect of corticosteroids versus control on the length of mechanical ventilation and were included in the meta-analysis (Figure 7B). Overall, the pooled MD was -1.71 (95% CI: -2.54 to -0.88) with no heterogeneity (I² = 0%; P_Q = 0.88). Because there were less than 10 included studies, we did not draw funnel plots nor did we conduct Egger's regression test. The subgroup analysis by study design (Supplementary Figure 49) and risk of bias (Supplementary Figure 50) showed no significant between-group differences (p = 0.67 for both subgroup analyses).

The subgroup analysis by different corticosteroid regimens (Supplementary Figure 51) and corticosteroid dosage (Supplementary Figure 52) also did not show significant between-group differences (p = 0.91 and p = 0.99 for regimen and dosage subgroups, respectively). Last, meta-regression analyses by the proportion of severe patients (Supplementary Figure 53) and baseline mean SOFA score (Supplementary Figure 54) did not reveal any significant correlations (p = 0.30 and p = 0.39 for the severe proportions and SOFA score regressions, respectively). We did not conduct subgroup analyses by time to administration due to a lack of data.

Five studies [87,114,122,129,155] reported median length of mechanical ventilation with IQR, however, they were not imputed and included in the meta-analysis due to skewed data. One study [87] did not perform any statistical testing, three studies [114,129,155] used the Mann-Whitney U test, and one study [122] used Student's t-test. Out of the four studies with statistical tests, one study [122] (using the t-test) did not find significant differences between the corticosteroid and control groups. Two studies [114,155] found a significantly shorter median length of mechanical ventilation in patients taking corticosteroids, and one study [129] found a significantly shorter median length of mechanical ventilation in the control patients. The study that did not conduct statistical tests [87] reported a median duration of mechanical ventilation of 15 h (IQR: 13–21 h) in the control group, compared with a median duration of 6 h (IQR: 4–20 h) in the corticosteroids group.

Adverse events

Nine studies [30,96,102,111,133,140,155,158,172] with 3409 patients reported incidence of adverse events in the corticosteroid and control groups (Figure 8A). The overall pooled OR was 1.89 (95% CI: 0.98–3.63) with moderate and significant heterogeneity (I² = 71%; P_Q <0.01). There were significant differences between the treatment effect from observational studies versus RCTs (p < 0.01; Supplementary Figure 55), where 8 observational studies [96,102,111,133,140,155,158,172] yielded a pooled OR of 2.17 (95% CI: 1.24–3.81) with similar heterogeneity compared with the overall analysis, and 1 RCT [30] reported an OR of 0.43 (95% CI: 0.14–1.26). In terms of risk of bias, there were no significant between-group differences in the subgroup analysis (p = 0.46; Supplementary Figure 56). We did not draw funnel plots nor conducted Egger's regression tests because less than 10 studies were included in the analysis.

There were no significant between-group differences by corticosteroid regimen (p = 0.08; Supplementary Figure 57), corticosteroid dosage (p = 0.05; Supplementary Figure 58) and time to administration (p = 0.79; Supplementary Figure 59). Meta-regression analysis showed no significant correlation associated with baseline mean SOFA score (p = 0.74; Supplementary Figure 60). We did not conduct meta-regression analysis by the proportion of patients with severe disease due to a lack of data.

Severe adverse events

Six studies [29,30,87,89,104,120] with 1142 patients reported incidence of severe adverse events (Figure 8B). The pooled OR was 0.87 (95% CI: 0.47–1.62) with no heterogeneity (I² = 0%; P_Q = 0.77). There were no significant between-group differences in the subgroup analysis by study design (p = 0.74; Supplementary Figure 61) or risk of bias (p = 0.96; Supplementary Figure 62). We did not draw funnel plots or conduct Egger's regression test as less than 10 studies were included.

Subgroup analyses by corticosteroid regimen (Supplementary Figure 63) and corticosteroid dosage (Supplementary Figure 64) revealed significant between-group differences (p < 0.01 for both subgroup analyses). Two studies [89,104] used high fixed-dose methylprednisolone and yielded an OR of 0.83 (95% CI: 0.77–0.89) with no heterogeneity (I² = 0%; P_Q = 0.99), while two studies used low fixed-dose hydrocortisone [29,87] and yielded an OR of 3.02 (95% CI: 2.73–3.34) with no heterogeneity (I² = 0%; P_Q = 0.99). One study [120] used a low weight-adjusted dose of unspecified corticosteroid and reported an OR of 1.00 (95% CI: 0.01–72.91), and one study [30] compared dexamethasone versus standard of care and reported an OR of 0.53 (95% CI: 0.17–1.62). There were no between-group differences for subgroup analysis by time to administration (p = 0.73; Supplementary Figure 65), and we did not perform any meta-regression analysis due to a lack of data.

Hyperglycemia

Eleven studies [30,86,90,104,111,120,136,155,158,164,168] with 2898 patients reported incidence of hyperglycemia (Figure 8C) with a pooled OR of 2.45 (95% CI: 1.08–5.56). There was significant and moderate heterogeneity (I² = 53%; P_Q = 0.02). Visual inspection of the funnel plot (Supplementary Figure 66) and Egger's regression test showed no significant small study effects (P_Egger = 0.18). No subgroup analysis showed significant between-group differences (study design [p = 0.60; Supplementary Figure 67], risk of bias [p = 0.67; Supplementary Figure 68], corticosteroid regimen [p = 0.16; Supplementary Figure 69], corticosteroid dosage [p = 0.69; Supplementary Figure 70] and time to administration [p = 0.69; Supplementary Figure 71]), and meta-regression analyses showed no correlation between the treatment effect and proportion of severe patients (p = 0.42; Supplementary Figure 72) and baseline mean SOFA score (p = 0.95; Supplementary Figure 73).

Nosocomial infection

Twenty-seven studies [30,85,89,90,96,104,105,107,111,113,114,116,120,129,130,134,140,142,144,149,150,153,155,156,158,168,170] with 9483 patients reported incidence of nosocomial infection (Figure 8D) with a pooled OR of 1.06 (95% CI: 0.82–1.37). There was significant and moderate heterogeneity (I² = 47%; P_Q <0.01). Visual inspection of the funnel plot (Supplementary Figure 74) and Egger's regression test showed no significant small study effects (P_Egger = 0.99). No subgroup analysis showed significant between-group differences (study design [p = 0.89; Supplementary Figure 75], risk of bias [p = 0.27; Supplementary Figure 76], corticosteroid regimen [p = 0.70; Supplementary Figure 77], corticosteroid dosage [p = 0.25; Supplementary Figure 78] and time to administration [p = 0.06; Supplementary Figure 79]), and meta-regression analyses showed no correlation between the treatment effect and proportion of severe patients (p = 0.72; Supplementary Figure 80) and baseline mean SOFA score (p = 0.24; Supplementary Figure 81).

Quality of evidence

The summary of findings for primary outcomes is tabulated in Table 1.

Discussions

Main findings

The current systematic review and meta-analysis included 10 RCTs and 85 observational studies involving 42,205 hospitalized COVID-19 patients, making it the most comprehensive systematic review available on the use of corticosteroid for treating COVID-19. We found that the use of corticosteroid regimens was not significantly associated with changes in length of ICU hospitalization, nor incidence of mechanical ventilation.

While the overall meta-analysis showed that corticosteroids had no impact on length of stay, the pooled MD from RCT studies showed that corticosteroids may significantly increase the length of stay by 3.62 days, while the 95% CI for the pooled MD from observational studies narrowly crossed the line of no effect. The overall meta-analysis also showed that corticosteroids had no impact on mortality nor incidence of ICU admission, but subgroup analysis by corticosteroid type showed that methylprednisolone may confer a significant survival benefit with a 38% reduction in the odds of death compared with standard of care, while other regimens such as dexamethasone and hydrocortisone conferred no significant survival benefit. Additionally, methylprednisolone may be associated with a significant 47% reduction in the incidence of ICU admission. Across mechanically ventilated patients, we found that corticosteroids significantly reduced length of mechanical ventilation. However, corticosteroids may also significantly increase time to negative conversion of SARS-CoV-2 tests, which could suggest that corticosteroid administration hinders viral clearance.

In terms of safety, the use of corticosteroids was not associated with a significant increase in the odds of adverse events or severe adverse events. However, the pooled OR from observational studies showed that corticosteroids may increase the odds of adverse events by 117%. We also found a significant difference in the odds of serious adverse events between studies using high fixed-dose methylprednisolone versus low fixed-dose hydrocortisone, with methylprednisolone being associated with a significant, 17% reduction in the odds of serious adverse events while hydrocortisone was associated with a significant 202% increase in the odds of serious adverse events. While assessing the odds of hyperglycemia and nosocomial infection – both common adverse events associated with corticosteroids – we found that corticosteroids significantly increased the odds of hyperglycemia by 145%, however, there were no significant increases in nosocomial infection associated with corticosteroids.

Clinical implications for corticosteroid efficacy

In this review, we found that corticosteroids had no protective effect against the incidence and duration of ICU hospitalization nor incidence of progression to mechanical ventilation. These findings are consistent with previous studies assessing the use of corticosteroids in other coronavirus infections such as SARS-CoV and MERS-CoV [178], demonstrating that corticosteroids may not be effective for preventing progression from mild-and-moderate COVID-19 to severe disease, nor is it effective for managing ICU-hospitalized patients. In fact, corticosteroids may significantly increase the length of stay for hospitalized patients (according to pooling of RCT studies) and increase time to viral clearance, potentially placing more strain on limited hospital resources and increasing the risk of viral transmission. As previously speculated, using immunosuppressive therapies indiscriminately for viral infections such as COVID-19 may complicate disease courses and perpetuate illness [179].

Interestingly, while we found that corticosteroids had no significant impact on the incidence of mortality and ICU admission, methylprednisolone appears to have an exceptional efficacy for reducing the odds of death compared with other corticosteroids such as dexamethasone and hydrocortisone. Previous studies, including a triple-blinded RCT [180] and an observational cohort study [125], also found that methylprednisolone may be more efficacious at reducing mortality compared with dexamethasone. In addition, when the meta-analysis is restricted to studies using methylprednisolone only, it was found that methylprednisolone may also significantly reduce incidence of ICU admissions. It was hypothesized that these observations were due to methylprednisolone having greater lung tissue penetration compared with other corticosteroids [181,182], making methylprednisolone a more optimal corticosteroid for treating hyperinflammatory reactions to SARS-CoV-2 and respiratory complications. However, we suspect that over-precision associated with the large quantity of observational studies in the methylprednisolone subgroup may have contributed to our finding as well [183]. Thus we cannot definitively conclude that methylprednisolone is superior compared with other corticosteroids for reducing mortality or ICU admission. Currently, most published RCTs focus on dexamethasone and hydrocortisone [28,29], but less frequently on methylprednisolone. Further investigations using large scale, high-quality RCTs are required to establish the potential impact of methylprednisolone on mortality and ICU admission.

A main finding of our review is that corticosteroids significantly reduced the length of mechanical ventilation in critically ill patients experiencing respiratory failure, while it was not effective in decreasing the odds of mechanical ventilation incidence. Mechanical ventilation is often required in COVID patients experiencing ARDS, defined as a reduction in the area of normoventilated lung due to inflammatory pulmonary edema [184]. While pulmonary edema is one of the main drivers of COVID-related mortality [185,186], prolonged mechanical ventilation is also associated with a higher risk of developing complications such as ventilator-associated pneumonia [187], death, increased hospital and ICU lengths of stay, and increased cost [187,188]. Thus, the use of corticosteroids may substantially improve the outcome of critically ill COVID-19 patients who require mechanical ventilation, as well as conserving valuable ventilator assets during COVID surges.

Compared with previous meta-analyses, the current review did not identify a robust association between the use of corticosteroids and mortality, unlike smaller systematic reviews conducted earlier during the pandemic [58,59,62]. However, our findings regarding prolonged duration of viral clearance and reduced length of mechanical ventilation is strongly supported by the findings of previous meta-analyses [60,61]. In summary, our results suggest that the role of corticosteroids for treating COVID-19 may be limited to severe and critically ill patients receiving mechanical ventilation for reducing length of mechanical ventilation. It may not be effective for reducing mortality or improving other efficacy outcomes for other patient groups. The effect of methylprednisolone on the odds of mortality requires further investigation and clarification. Last, while corticosteroids may be effective in certain patient groups, it is currently unknown whether it is related to SARS-CoV-2, its treatment, or both. Further research may be warranted.

Clinical implications for corticosteroid safety

Similar to previous meta-analyses [58,62], we did not identify any significant associations between the use of corticosteroids and the odds of all-cause adverse events and all-cause severe adverse events. However, excluding the CoDEX trial [30] from the meta-analysis resulted in a significantly increased odds of adverse events based on the treatment effects of the remaining observational studies. Given that the patient characteristics are similar between the CoDEX trial and patients in the observational study, especially as both the CoDEX trial and the observational studies largely excluded mild and moderate patients, a possible explanation of this finding is due to the use of corticosteroids in a subset of control patients in the CoDEX trial, as well as the open-label design which may lead to biased reporting of adverse events in the experimental group. The CoDEX investigators stated that these factors may have shifted the results toward null [30]; consequently, the inclusion of the CoDEX trial would have shifted our pooled results toward null as well.

Additionally, we found that the pooled results from studies using high fixed-dose methylprednisolone yielded a significantly decreased odds of severe adverse events, while studies using low fixed-dose hydrocortisone yielded a significantly increased odds of severe adverse events. As there was no heterogeneity within each of the aforementioned subgroups, and the studies included in the subgroups (including 3 RCTs and 1 propensity-matched cohort study) were unlikely to be substantially biased; we speculate that the differences in corticosteroid doses may explain this observation. It is possible that the high-dose methylprednisolone therapy was more efficacious at controlling COVID symptoms and undesirable effects from adjuvant medications than low-dose hydrocortisone, thus resulting in less reported adverse events. Overall, a majority of the adverse events reported by the subgroup studies were deemed to be unrelated to the intervention [29,87,89,104]. Nevertheless, the sample size supporting this observation is low, and further investigations are needed to confirm this finding.

In this review, we assessed the odds of hyperglycemia and nosocomial infections, two of the most common types of adverse events in patients undergoing systemic corticosteroid treatments [189]. Our review found that the use of corticosteroids is associated with a significant increase in the odds of hyperglycemia compared with control, indicating a need for glycemic monitoring for patients receiving corticosteroid therapy [190]. Interestingly, we did not identify a significant correlation between corticosteroid use and the odds of nosocomial infection, contrary to expectations due to corticosteroids' immunosuppressive properties. This is most likely due to the use of concomitant antimicrobials in COVID-19 patients to prevent superinfections. As the reporting of antibiotic and antifungal use among our included studies were extremely inconsistent, we could not properly evaluate the impact of corticosteroids on the odds of nosocomial infections.

Strengths & limitations

This systematic review and meta-analysis has several strengths. First, we conducted several subgroup analyses to assess sources of heterogeneity and to study the impact of study methodologies on the treatment effect, including comparing the pooled effect from randomized to non-randomized studies and studies with low-moderate risk of bias versus studies with high risk of bias. Secondly, we followed the most up-to-date systematic review guidelines and methods, such as PRISMA 2020, RoB2, and GRADE. Moreover, the large sample size in our analyses improved the power and precision of our study compared with previous systematic reviews.

The major limitation of this study is the inclusion of observational studies. Similar to our previous systematic reviews [25,26], a majority of our included observational studies were rated as having a critical or serious risk of bias according to ROBINS-I, highlighting the poor publication quality associated with pandemic-related literature [191]. The inclusion of observational studies in meta-analyses was also associated with large treatment effects [192,193] and overly precise estimates [194,195], potentially introducing biases into our significant findings. However, it must be noted that our subgroup analysis by study design showed that the findings from observational studies were largely in agreement with findings from RCTs for a majority of the outcomes. In addition, while we conducted numerous subgroup analyses and meta-regressions to explore the sources of heterogeneity and possible factors that may affect the efficacy of corticosteroids, we could not provide insight on the effect of virus variants, the general care provided, vaccination status and underlying conditions on the efficacy of corticosteroids over the course of the pandemic due to a paucity of data. With rapidly changing COVID-19 treatment strategies and the spread of the Omicron variant, future studies may wish to consider the potential impact of these factors.

Conclusion

This comprehensive systematic review and meta-analysis showed that the use of corticosteroids was not significantly associated with changes in length of ICU hospitalization, incidence of ICU admission, incidence of mortality, nor incidence of mechanical ventilation. Additionally, corticosteroids may increase length of stay (based on RCT data only) and increase the duration of viral clearance. However, corticosteroids were significantly associated with reduction in the length of mechanical ventilation; thus, it may confer beneficial effects for critically ill patients requiring mechanical ventilation. There was also evidence suggesting that methylprednisolone may be efficacious in reducing mortality and incidence of ICU admission; however, these findings should be further assessed using RCTs.

In terms of safety, corticosteroids were not associated with increased odds of adverse events or severe adverse events, although these findings need to be interpreted with caution due to potential confounding factors such as corticosteroid dosage. Corticosteroids were also not associated with increased odds of nosocomial infections, albeit the use of adjuvant antimicrobials complicates the interpretation of this finding. Last, corticosteroids use was associated with significantly increased odds of hyperglycemia, indicating a need for glycemic monitoring in COVID-19 patients receiving corticosteroid treatments.