1. Introduction

During the first year of the coronavirus disease 2019 (COVID-19) pandemic, efforts were undertaken to identify effective pharmacological interventions for COVID-19 which were readily available in hospitals for repurposing. Early in the pandemic, the RECOVERY Collaborative Group demonstrated that use of dexamethasone resulted in lower 28-day mortality rates for patients on respiratory support [1]. Dexamethasone is a commonly used glucocorticoid. Early findings, which supported the proposed mechanism of reducing the systemic hyper-inflammatory immune response and thus reducing lung and end-organ damage, made corticosteroids a promising intervention [2–4]. Further evidence, including multiple meta-analyses, has supported the benefits of dexamethasone and other corticosteroids in hospitalized COVID-19 patients for reduction of mortality [5–7]. A small randomized clinical trial in Brazil from September 2020 also supported the benefit of dexamethasone for number of days alive and free from mechanical ventilation for hospitalized patients [8].

Observational studies have covered broader patient populations and have resulted in mixed evidence. In a systematic review and meta-analysis, 73 observational studies showed benefit for corticosteroids. However, an additional 32 studies affording comparison had high patient and between-study heterogeneity and suggested a detrimental effect of corticosteroids on mortality, although both the high heterogeneity and selection bias likely had a critical and confounding role in this result [6]. Furthermore, an observational study of dexamethasone in hospitalized COVID-19 patients not receiving respiratory support showed evidence of potential increased risk of 90-day mortality [9], and these findings appear to have been recapitulated in a more recent meta-analysis of hospitalized patients not receiving oxygen therapy [10]. Comparatively, a cohort study of the effectiveness of combined dexamethasone with remdesivir compared to standard-of-care treatment suggests reduction of mortality even for moderate patients; additionally, progression to mechanical ventilation was reduced [11].

Dexamethasone remains a popular corticosteroid for management of hospitalized COVID-19 patients; by one estimate based on 137,870 hospitalized adult COVID-19 patients in the United States (U.S.) from 1/1/2020 to 2/28/2021, 39.1% received dexamethasone during their hospitalization [12]. In the first year of the pandemic, one study also showed increasing numbers of non-hospitalized patients with COVID-19 being prescribed dexamethasone, despite NIH recommendations [13]. The World Health Organization (WHO) currently recommends the use of corticosteroids for patients with severe or critical COVID-19 only [14].

Many questions remain concerning the use of dexamethasone as a front-line clinical therapy for hospitalized COVID-19 patients, especially in real-world settings with a high degree of heterogeneity [15]. Such real-world evidence is possible at scale through use of large population-based data repositories, curated to the point of common measures across sources, and sufficiently large sample sizes to support meaningful comparisons. In response to this need during the pandemic, the informatics community developed the National COVID Cohort Collaborative (N3C) Data Enclave in the United States (U.S.), enabling secondary analysis of electronic health record (EHR) data related to COVID-19 [16].

In this work, we aim to leverage data from the N3C to estimate the real-world effectiveness of treatment with dexamethasone in hospitalized COVID-19 patients in the U.S. during the first year of the pandemic, with analyses stratified by remdesivir treatment. In addition to the methodological rigor realized through use of a centralized, harmonized, and highly granular EHR data repository, we demonstrate the usage of data imputation to handle missing biomarkers and a matched propensity score approach to handle biomarkers of disease severity with N3C data to generate high quality evidence. Use of the varied comorbidity and laboratory test data available in this resource allows for improved estimation of patient severity and enables an observational study of the efficacy of dexamethasone while minimizing bias. These methods could be applied to similar questions regarding real-world treatment effectiveness of pharmacological interventions within the N3C and beyond, and methods demonstration is an important contribution of our study as well.

2. Methods

This is a comparative effectiveness study of dexamethasone treatment in hospitalized COVID-19 patients in the U.S. from the first year of the pandemic (1/1/2020 to 2/23/2021), prior to broad vaccine availability and evolution of major viral variants. It uses a retrospective cohort design based on a curated, national, data repository derived from the electronic health record (EHR). In what follows, we first describe the data and patient inclusion criteria. Then, we describe the imputation of missing data. Lastly, we describe propensity score matching of treated and control groups to control for confounders, and estimation of treatment of effect. The study was determined to be “not human subjects research” by the University of Texas at Austin Institutional Review Board. The overall project is conducted under a Data Use Agreement (DUA) between N3C and the University of Texas (UT) at Austin (PI: PJR) and an N3C-approved Data Use Request (DUR) for this specific study.

3. Results

3.1 Patient characteristics

Fig 1 describes the application of inclusion criteria to develop the study cohort (dexamethasone- and non-dexamethasone-treated patients) from all available patients in the N3C data enclave. Of 4,937 COVID-19 positive patients treated with dexamethasone, 3,645 (82.9%) were inpatient hospitalizations two or more days long where the patient did not also receive prednisone, methylprednisolone, or hydrocortisone. Of these patients, 2,469 received dexamethasone on the first or second day of the visit. After removal of the two sites that had a data quality issue, 2,457 patients remained and were used in imputation and matched to non-dexamethasone treated controls. Of 272,290 non-dexamethasone-treated COVID-19 positive patients, 28,874 (10.6%) were inpatient hospitalizations of length two or more days where the patient did not also receive prednisone, methylprednisolone, or hydrocortisone. After removal of the two sites that had a data quality issue, 28,439 patients remained and were used in imputation and as potential controls during PS matching.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Development of cohort from All N3C patients.

Application of inclusion criteria to patients in the N3C and development of matched pairs for comparing the effect of dexamethasone among patients who had and had not received remdesivir, independently groups. Note that within the remdesivir analysis, 390 dexamethasone-treated patients were dropped between the PS-matching and logistic regression stages because there were not enough remdesivir controls with propensity scores similar enough to that of the dexamethasone group.

https://doi.org/10.1371/journal.pone.0294892.g001

After application of the inclusion criteria, and before imputation and PS matching, there were 2,457 dexamethasone-treated patients, and 28,349 potential non-dexamethasone-treated controls. The dexamethasone-treated group was slightly older (61±16 years) compared to the non-dexamethasone-treated group (56±19 years). Notably, 1,194 patients from the dexamethasone-treated group received remdesivir as part of their treatment, while 1,014 patients from the non-dexamethasone-treated group received remdesivir as part of their treatment. The remaining demographics, laboratory measurements, and comorbidities of these two groups with percentage missing, if applicable, are presented in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Characteristics of dexamethasone-treated patients and potential controls at baseline before imputation and PS matching.

https://doi.org/10.1371/journal.pone.0294892.t001

3.2 Modeling and outcomes

3.2.1 Imputation and propensity score matching.

After imputation, the distribution of imputed and observed values for continuous laboratory values was comparable (S1 Fig).

Propensity score matching of non-dexamethasone-treated controls to dexamethasone-treated patients in both the remdesivir and non-remdesivir groups improved balance. The absolute standardized mean difference between the treated and control group of all continuous covariates and all levels of categorical covariates included in assigning propensity scores was reduced to <0.1 (S2A and S2B Fig). Achieving balance between groups required exclusion of treated units within the dexamethasone and remdesivir treated group which could not be successfully matched to controls. After matching, 2,067 total dexamethasone-treated patients remained. Eight-hundred-four patients were also treated with remdesivir while 1,263 patients were treated with dexamethasone only. Three-hundred-ninety remdesivir and dexamethasone treated patients were dropped between the logistic regression and PS-matching stages because there were not enough controls with propensity scores similar enough to that of the remdesivir and dexamethasone group. In the dexamethasone only group, all units were successfully matched to three controls. The characteristics of the matched groups are summarized in Table 2, with further details in S1 Table.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Dexamethasone treatment and matched control group summary.

For (A) the non-remdesivir stratum, characteristics of dexamethasone-treated patients and 3:1 PS matched non-dexamethasone-treated controls, after imputation. For (B) the remdesivir stratum, characteristics of dexamethasone-treated patients and 1:1 PS matched non-dexamethasone-treated controls, after imputation. Corresponds with S1 Table. For each laboratory value, the percentage which was imputed is included in S2 Table.

https://doi.org/10.1371/journal.pone.0294892.t002

3.2.2 Effect of dexamethasone.

Rates of in-hospital death or referral to hospice were lower in the dexamethasone-treated group compared to the non-dexamethasone-treated matched control group for those receiving remdesivir (OR = 0.74, 95% CI: 0.53 to 1.02), those not receiving remdesivir (OR = 0.77, 95% CI: 0.62 to 0.95), and both remdesivir and non-remdesivir treated groups combined (OR = 0.77, 95% CI: 0.64 to 0.91), although the effect in the smaller remdesivir stratum was only borderline significant.

The use of dexamethasone also was associated, albeit at a weaker level, with a lower incidence of a combined severe outcome or in-hospital death/hospice referral for those receiving remdesivir (OR = 0.83, 95% CI: 0.64 to 1.09), those not receiving remdesivir (OR = 0.84, 95% CI: 0.71 to 1.00), and both remdesivir and non-remdesivir treated strata combined (OR = 0.82, 95% CI: 0.71 to 0.94) (Table 3).

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Outcomes for dexamethasone-treated patients and matched controls.

Outcomes and crude odds ratios (95% CI) comparing dexamethasone recipients to PS-matched non-recipients in N3C hospitalized COVID-19 patients. Non-recipients of remdesivir are matched 3:1 (non-dexamethasone to dexamethasone), while remdesivir recipients are matched 1:1.

https://doi.org/10.1371/journal.pone.0294892.t003

In sensitivity analyses, the effect of dexamethasone on reduction of in-hospital death/hospice referral (OR = 0.75, 95% CI: 0.59 to 0.95, p = 0.017) and combined severe outcome or in-hospital death/hospice referral (OR = 0.82, 95% CI: 0.68 to 0.98, p = 0.028) in the non-remdesivir stratum remained similar and significant after adjusting for the effect of age, Q-score, AST, creatinine, platelet count, white blood cell count, and PS via stratum indicators for the 4 PS quartiles (Table 4A). Within the strata defined by quartiles of propensity score, the effect of dexamethasone on reduction of both study outcomes in the non-remdesivir stratum after adjustment was stronger in the fourth quartile of propensity scores (presumably patients who presented as most severe) (OR = 0.62, 95% CI: 0.42 to 0.9, p = 0.014 mortality outcome; OR = 0.61, 95% CI: 0.44 to 0.82, p = 0.002, combined severe/mortality outcome) than in the aggregate cohort. In the first through third quartiles, the effect was non-significant and somewhat heterogeneous (S3A Table).

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Effect of dexamethasone in logistic regression models.

Prediction of in-hospital death/hospice referral and combined in-hospital death/hospice referral and severe outcome by receipt of dexamethasone with logistic regression models for (A) non-remdesivir stratum and (B) remdesivir stratum.

https://doi.org/10.1371/journal.pone.0294892.t004

In the remdesivir-treated group, the use of dexamethasone showed a borderline statistically significant benefit in reducing in-hospital death or referral to hospice (OR = 0.72; 95% CI: 0.51 to 1.00, p = 0.054); however, the effect of dexamethasone on the combined severe or in-hospital death/referral to hospice outcome was non-significant (Table 4B). Within strata, in the remdesivir-treated group, effects were largely non-significant (although p = 0.048 in the second quartile) and heterogeneous, likely owing to the small sample sizes (S3b Table).

The expected rationale of the fourth quartile of PS being the most severe was further characterized by assessing the laboratory values by PS. We see trends across laboratory values demonstrating severity differences between quartile of PS for both the remdesivir and non-remdesivir treated groups (S4 Table).

The same benefit of dexamethasone for both the remdesivir and non-remdesivir groups was observed in further sensitivity analyses where the logistic regression models above with linear adjusters were extended to investigate quadratic effects for the laboratory variables included, though the effect of dexamethasone on the combined outcome was again not significant in the remdesivir group (S5 Table).

4. Discussion

Our analysis of multi-site EHR data in the first year of the pandemic confirmed existing clinical trial findings that dexamethasone shows in-hospital mortality benefit. We also found that use of dexamethasone results in a reduction in the secondary combination outcome of in-hospital mortality or severe outcome, defined by the use of ECMO or mechanical ventilation. The effect of dexamethasone on reducing the combined severe and in-hospital death/hospice referral outcome was not as strong as the effect on in-hospital death/hospice referral alone. Similar outcomes were observed in the remdesivir-treated and non-remdesivir-treated groups. Although we did not directly assess co-administration of remdesivir with dexamethasone, our analysis suggests that remdesivir did not appear to impact the benefit of dexamethasone in reduction of either mortality or the secondary outcome.

The treatment effects observed in the cohorts were subjected to sensitivity analyses. Treatment effect size was the same when adjusting for both linear and quadratic effects of included covariates; however, we found that the odds ratios for the secondary outcome in the remdesivir groups were non-significant. It is likely that more data are needed in the remdesivir groups to confirm the effect.

Our results are mostly in agreement with existing evidence, but it is important for us to expand on the context of current NIH guidelines and more recent evidence since the first year of the pandemic. Current guidelines strongly recommend dexamethasone as the primary immunomodulator of choice for therapeutic management of hospitalized COVID-19 patients requiring any type of oxygen. Furthermore, a similar study published in April 2023 which also evaluated dexamethasone for inpatients using a national, retrospective, propensity score–weighted cohort study during the first year of the pandemic found significantly reduced odds of mortality or discharge to hospice in those requiring supplemental oxygen or MV and/or ECMO, corroborating our results [27]. Even though our study does not specifically look at benefits in those requiring supplemental oxygen, the findings do not contradict the recommendations. Our cohort is broad, and the study assesses all people who received dexamethasone in this early pre-vaccine period. Our goal was to assess the average effect of treatment among the treated, but we recognize there may be heterogeneity of treatment effects across subpopulations. This is captured in our assessment of treatment effect by quartile of propensity score. In both remdesivir and non-remdesivir groups, there was significant variability of the odds ratios across strata; however, one notable trend was a statistically significant treatment benefit for both study outcomes in the fourth quartile of the non-remdesivir analysis group. As the propensity score is the probability of being assigned treatment given a certain set of covariates, and dexamethasone is more likely to be administered to severe patients, those who have covariates in the PS stratum which suggest a more severe condition at baseline may lead a provider to be more likely to administer dexamethasone. This suggests those in the fourth quartile of PS benefited most from dexamethasone in comparison to other patients receiving dexamethasone, consistent with current practice of use of dexamethasone primarily in more severe patients receiving oxygen support.

Specifically regarding combined remdesivir/dexamethasone therapy, for patients requiring conventional oxygen, there is a moderate recommendation for co-administration of remdesivir with dexamethasone; for patients on high-flow nasal cannula or non-invasive ventilation, there is a weak recommendation for co-administration of remdesivir. No statement is made about remdesivir co-administration for patients already on mechanical ventilation or ECMO [14, 28]. Indeed, evidence remains mixed regarding remdesivir plus dexamethasone treatment. A prospective controlled non-randomized study found a significant reduction in mortality for remdesivir/dexamethasone treatment, compared to dexamethasone alone [29]; however, a retrospective, multicenter cohort study found addition of remdesivir was not associated with shorter hospitalization or in-hospital mortality [30]. Again, our study did not directly assess remdesivir/dexamethasone, or remdesivir, treatment but suggests potentially no reduction of dexamethasone benefit with addition of remdesivir. More specific subgroup analyses directly addressing this issue are likely needed.

The landscape of the pandemic has also changed since the first year of the pandemic, and this includes the emergence of variants of concern (VOCs) and vaccines. We expect efficacy to be generally similar across different variants, because the damaging inflammation (such as cytokine storm, chest/lung inflammation) which is characteristic of the severe immune response triggered by COVID-19 infection is the target of corticosteroid treatment. Furthermore, the emergence of the alpha variant (B.1.1.7) in January 2021 is unlikely to affect our analysis, since while the transmissibility of the alpha variant was found to be higher than the original strain and studies have suggested an increase in mortality in the population due to B.1.1.7 overall [31], a cohort study found no association between SARS-CoV-2 lineage and death for hospitalized patients [32].

It is still unclear how vaccination status is associated with severe/mortality outcomes for hospitalized patients. However, findings have shown that vaccinations are clearly effective for preventing severe COVID-19 infection, that is, preventing a hospitalization in the first place (3.5 to 17.5 times higher hospitalization rates for non-vaccinated) [19].

Regarding limitations, we addressed several statistical issues in the course of our study. First, as a result of treatment guidelines at the time, the number of patients who had received remdesivir in the absence of dexamethasone was not large. The two strata do in principle allow for comparison of dexamethasone treatment effects between those also receiving and those not receiving remdesivir. We did not detect strong evidence of such differences, although our sample size was limited to be able to detect such interaction terms.

Second, regarding concerns about the considerable amount of missing data, we note that other N3C investigators have also struggled with this issue and have nevertheless been able to draw robust and useful conclusions using the N3C platform, and missing data is a fact of life in research based on the electronic health record. A strength of our analysis is that we matched on the number of missing lab values (before imputation) as a predictor of dexamethasone treatment (versus no such treatment). Indeed, before matching, those treated with dexamethasone had far more complete laboratory profiles, most likely reflecting severity of disease. This component of matching may be one of the more powerful variables to achieve balance on severity (which is not directly observed). Also, regarding the use of single imputation, we note that this may result in standard errors being slightly underestimated but should not affect the imputed values themselves.

There are also some other general limitations associated with the use of electronic health record data that have been harmonized across many sites. Key data such as detailed ventilator flow settings, ICU admission, oxygen saturation (SpO₂), and supplemental oxygen that would have enabled us to answer additional questions about the effects of dexamethasone were not available in the N3C. Specific details about the administration of dexamethasone including delivery route (IV vs oral) and dosage were absent in a portion of patients who received dexamethasone, limiting the potential to study how these factors specifically affect outcomes.

Overall, the large number of sites and patient-level variables available in the N3C were critical in enabling our study and may enable future research in related directions. We also demonstrated the value of the N3C as a resource with the use of methods described in this paper to conduct robust secondary analyses of EHR data producing high quality evidence evaluating the effectiveness of interventions to manage COVID-19. This provides a framework to equip the field to respond quickly to generate evidence to guide management interventions when facing the next emerging, rapidly evolving pandemic when there is not sufficient time to conduct robust prospective clinical trials.

In conclusion, dexamethasone reduced in-hospital mortality and severe outcomes in hospitalized COVID-19 patients; dexamethasone co-administered with remdesivir resulted in similar outcomes. Furthermore, the most severe patients at baseline may benefit most from dexamethasone treatment, as there is likely heterogeneity of treatment effects within the treated group.

Supporting information

Acknowledgments

The analyses described in this publication were conducted with data or tools accessed through the NCATS N3C Data Enclave (https://covid.cd2h.org) and N3C Attribution & Publication Policy v1.2-2020-08-25b supported by NCATS U24 TR002306. This research was possible because of the patients whose information is included within the data and the organizations (https://ncats.nih.gov/n3c/resources/data-contribution/data-transfer-agreement-signatories) and scientists who have contributed to the on-going development of this community resource.

Disclaimer: The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health or the N3C program.

We gratefully acknowledge the following core contributors to N3C: Adam B. Wilcox, Adam M. Lee, Alexis Graves, Alfred (Jerrod) Anzalone, Amin Manna, Amit Saha, Amy Olex, Andrea Zhou, Andrew E. Williams, Andrew Southerland, Andrew T. Girvin, Anita Walden, Anjali A. Sharathkumar, Benjamin Amor, Benjamin Bates, Brian Hendricks, Brijesh Patel, Caleb Alexander, Carolyn Bramante, Cavin Ward-Caviness, Charisse Madlock-Brown, Christine Suver, Christopher Chute, Christopher Dillon, Chunlei Wu, Clare Schmitt, Cliff Takemoto, Dan Housman, Davera Gabriel, David A. Eichmann, Diego Mazzotti, Don Brown, Eilis Boudreau, Elaine Hill, Elizabeth Zampino, Emily Carlson Marti, Emily R. Pfaff, Evan French, Farrukh M Koraishy, Federico Mariona, Fred Prior, George Sokos, Greg Martin, Harold Lehmann, Heidi Spratt, Hemalkumar Mehta, Hongfang Liu, Hythem Sidky, J.W. Awori Hayanga, Jami Pincavitch, Jaylyn Clark, Jeremy Richard Harper, Jessica Islam, Jin Ge, Joel Gagnier, Joel H. Saltz, Joel Saltz, Johanna Loomba, John Buse, Jomol Mathew, Joni L. Rutter, Julie A. McMurry, Justin Guinney, Justin Starren, Karen Crowley, Katie Rebecca Bradwell, Kellie M. Walters, Ken Wilkins, Kenneth R. Gersing, Kenrick Dwain Cato, Kimberly Murray, Kristin Kostka, Lavance Northington, Lee Allan Pyles, Leonie Misquitta, Lesley Cottrell, Lili Portilla, Mariam Deacy, Mark M. Bissell, Marshall Clark, Mary Emmett, Mary Morrison Saltz, Matvey B. Palchuk, Melissa A. Haendel, Meredith Adams, Meredith Temple-O’Connor, Michael G. Kurilla, Michele Morris, Nabeel Qureshi, Nasia Safdar, Nicole Garbarini, Noha Sharafeldin, Ofer Sadan, Patricia A. Francis, Penny Wung Burgoon, Peter Robinson, Philip R.O. Payne, Rafael Fuentes, Randeep Jawa, Rebecca Erwin-Cohen, Rena Patel, Richard A. Moffitt, Richard L. Zhu, Rishi Kamaleswaran, Robert Hurley, Robert T. Miller, Saiju Pyarajan, Sam G. Michael, Samuel Bozzette, Sandeep Mallipattu, Satyanarayana Vedula, Scott Chapman, Shawn T. O’Neil, Soko Setoguchi, Stephanie S. Hong, Steve Johnson, Tellen D. Bennett, Tiffany Callahan, Umit Topaloglu, Usman Sheikh, Valery Gordon, Vignesh Subbian, Warren A. Kibbe, Wenndy Hernandez, Will Beasley, Will Cooper, William Hillegass, Xiaohan Tanner Zhang. Details of contributions available at covid.cd2h.org/core-contributors.