Comparative Effectiveness of Dexamethasone in Hospitalized COVID-19 Patients in the United States

Study Design

This retrospective cohort study utilized the Premier Health Database (PHD) [16] to identify hospitalized adult patients (≥ 18 years) with COVID-19 from July 1, 2020, to January 31, 2021. This study period was selected because dexamethasone was considered the standard of care for hospitalized COVID-19 patients from July 2020 onwards. The index date was the date when patients first initiated any oxygen therapy (Supplementary Fig. S1). The baseline period starts from admission to the index date.

Data Source

The PHD is one of the most comprehensive inpatient electronic health databases in the US and represents approximately 25% of annual inpatient admissions [16]. With more than a thousand contributing hospitals or healthcare systems, PHD contains information on inpatient discharges from geographically diverse non-profit, non-governmental, community, teaching hospitals, and health systems from rural, and urban areas. The data include demographics, early diagnosis, admission and discharge diagnoses, and information on billed services, including costs at the departmental level. The PHD contains de-identified, Health Insurance Portability and Accountability Act compliant data and is exempted from Institutional Review Board oversight [16].

Compliance with Ethics Guidelines

This is an observational study that uses previously collected data and does not impose any form of intervention and was performed in accordance with the Helsinki Declaration of 1964 and its later amendments. Data have been deidentified to protect subject privacy and to be fully compliant with the US patient confidentiality requirements, including the Health Insurance Portability and Accountability Act of 1996, and did not require institutional review board waiver or approval.

Patient Population

Patients were included in this study if they had at least one inpatient diagnosis of COVID-19 and initiated oxygen therapy during the study period. For the dexamethasone group, patients with the following were included: (1) ≥ 1 report of a COVID-19 diagnosis (ICD-10 CM: U07.1) from an inpatient hospital stay during the study period; (2) gender entry not missing; (3) presence of oxygen use (as defined in the Adaptive COVID-19 Treatment Trial Ordinal Scale [ACTT OS] 5, 6, or 7) [17, 18]; (4) who initiated dexamethasone between 1-day pre- to 1-day post-index period. Similar criteria were used for including patients in the comparator group except for the use of dexamethasone between 1-day pre- to 1-day post-index period. The differences in steroid usage between the dexamethasone group and the comparator group were balanced using all the baseline characteristics available in the database. Patients with death recorded before or on index date were excluded from both the groups. For the comparator group, patients with dexamethasone use within 7 days prior to index were also excluded (Fig. 1).

Fig. 1figure 1

Flow chart for patient selection. Index date was the date when patients first initiated any oxygen therapy

Analyses of Primary Endpoint

The primary endpoint was the binary outcome of in-hospital mortality (includes patients that were discharged and deceased at the hospice) for patients receiving dexamethasone versus those who did not receive dexamethasone at the time of receiving oxygen therapy. The effect of dexamethasone on in-hospital mortality was estimated using multiple methods through a frequentist model averaging framework (FMA). FMA is an ensemble approach in which multiple pre-specified candidate methods/models are entered and cross-validation is used to select or upweight the methods/models that perform best. FMA addresses model uncertainty and has been shown to produce more robust estimates of treatment effects than using a single pre-selected method [19]. Specifically, the mean squared prediction error (MSPE) for each candidate method is computed via (five fold) cross validation and weights are derived for each candidate method with higher weights given to candidate methods with smaller MSPE. For binary and continuous outcomes (including the primary outcome of in-hospital mortality) final treatment estimates are obtained by either selecting the method with the smallest MSPE (averaged across bootstrap samples) or a weighted average of the treatment effect estimates across all methods is computed. However, as there is no guidance in the literature on FMA weighting for time to event outcomes, such analyses were ranked by model fit, e.g., Akaike information criterion or Bayesian information criterion followed by the lowest average absolute standardized difference of means.

Table 1 provides the details of the pre-specified individual methods used in the FMA analysis. A total of 23 pre-specified individual models/methods were entered into the FMA analysis, including analyses based on weighting, matching, stratification, and regression. For those models that used propensity score (PS), the variables used in balancing the baseline covariates are presented in Supplementary Table S1. In brief, the following variables were used to adjust for confounding: any underlying comorbidities during in-hospital stay, baseline demographics, serious infection, background treatments (remdesivir, enoxaparin, and corticosteroid), ACTT OS scale, and hospital admission status.

Table 1 Summary of models

Prior to conducting the analysis of treatment effects, the feasibility of the analysis was confirmed by first assessing the positivity assumption through quantifying the overlap in population characteristics between the two treatment groups. This was done both graphically through overlapping histogram plots and quantitatively through statistics summarizing differences in populations (standardized mean difference, Tipton’s index). Second, the ability of PS to produce balance in baseline patient characteristics between the two treatment groups was assessed using standardized mean differences, variance ratios, and graphical assessment of the distribution of each covariate between groups. This feasibility step was completed prior to conducting outcome analyses.

The validity of the treatment effect estimates depends on the statistical assumption of no unmeasured confounding. To quantify the strength of the findings relative to potential unmeasured confounding variables, the E-value was estimated [20]. The E-value is the minimum strength of association on the risk ratio scale that an unmeasured confounder would need to have with both the treatment and outcome to fully explain the observed outcome, conditional on the observed covariates. To aid in the interpretation of the E-value, we compared the E-value with the strength of associations of select observed covariates.

Finally, sensitivity analyses were performed to check whether the primary endpoint analysis varied according to the various choices on the patient population and modeling strategies described in Table 1. We conducted nine sensitivity analyses, categorized in five ways—(1) restricting the analysis to patients who were classified as ACTT OS 5–7 and ACTT OS 5 or 6 at index date, (2) FMA including additional models based on alternative PS models, (3) excluding patients who received dexamethasone prior to the index date or for the comparator cohort that received dexamethasone post-index, (4) including patients with an index date from April 2020, and (5) including patients with an index date from April 2020, but not including the index date as part of the PS.

All analyses were conducted using SAS 9.4 and R version 4.0.3 (R Foundation for Statistical Computing, Vienna, Austria).

Analyses of Secondary Endpoints

Secondary endpoints included 28-day mortality, time to in-hospital mortality, progress to invasive mechanical ventilation or death for patients who were not receiving invasive mechanical ventilation on the index date, time to discharge, proportion discharged alive by day 28, and length of hospital stay. The effect of dexamethasone on all the secondary endpoints was compared between patients who received dexamethasone versus those who did not, within 1 day of index, adjusting for confounders, using the same statistical approach used for the primary outcome measure (see Table 1).

留言 (0)

沒有登入
gif