Data source

This retrospective cohort study used data from HealthVerity Real-time Insights and Evidence (Philadelphia, PA, USA), comprising hospital chargemaster data linked to closed medical and pharmacy claims (from commercial, Medicare, and Medicaid health plans) for inpatient and outpatient encounters across 50 states and includes > 25 million US patients. The study was approved and patient informed consent was waived under an applicable exemption for deidentified data by the WCG Institutional Review Board; the study followed the 2005 Guidelines for Good Epidemiologic Practice, Best Practices for Conducting and Reporting Pharmacoepidemiologic Safety Studies Using Electronic Healthcare Data Sets, and the 2015 International Society of Pharmacoepidemiology Good Pharmacoepidemiology Practices.

Patients

The HealthVerity database was queried for patients ≥ 12 years of age admitted for hospitalization for ≥ 2 days with a diagnosis of COVID-19 (ICD-10 code U07.1) in any position (ie, nonprimary diagnoses were included) during the inpatient encounter between May 1, 2020 (when remdesivir received EUA from the US Food and Drug Administration), and September 30, 2021. Only the first hospitalization fulfilling all inclusion criteria was considered. Included patients had ≥ 365 days of continuous insurance enrollment prior to hospital admission through Day 2 of hospitalization (with a ≤ 30-day allowable gap) and survived the first 2 days of hospitalization. Exclusion criteria included diagnosis of Long COVID (ICD-10 code U09.9) prior to admission, pregnancy-related diagnoses during the hospitalization, and prior remdesivir use. Because manifestations of Long COVID could include exacerbation of ongoing disease, those experiencing symptoms consistent with Long COVID prior to admission were not excluded, and preexisting Long COVID–like symptoms were accounted for in propensity score weighting.

Additionally, 2 subgroups were identified: patients with a potentially immunocompromising condition (identified by ≥ 1 diagnosis code for HIV, hematologic or solid malignancy, organ transplantation, rheumatologic disease, or other immunosuppressive conditions during the ≤ 365 days before hospital admission)²⁹ and patients with ≥ 1 moderately-to-severely immunocompromising therapy or condition (identified via an algorithm based on National Institutes of Health criteria; definitions for subgroups are summarized in Supplementary Table 4)³⁰.

Study design

The day of hospital admission was defined as Day 0; the index date was defined as Day 2. The remdesivir-exposed group included patients who received ≥ 1 dose of remdesivir on Day 0 or Day 1 based on chargemaster drug descriptions or procedure codes; the unexposed comparator group included patients with no evidence of remdesivir use during the first 2 days of hospitalization. Covariates for adjustment were measured in the baseline period from Day−365 to Day−1 (Supplementary Fig. 2). In an attempt to balance for preexisting disease states that could be mistaken for Long COVID in the primary assessment window postinfection and confound interpretation of Long COVID outcomes being measured, the baseline preinfection ICD-10 codes were assessed from Day –365 through Day –15. This baseline window concluded at Day –15 rather than Day –1 because some Long COVID outcomes overlap with acute infection symptoms and conditions.

For each Long COVID outcome, the outcome assessment period began 90 days after admission (consistent with the WHO definition of Long COVID⁶) until the earliest of the following: occurrence of the specific outcome of interest, inpatient death, disenrollment of insurance, data cutoff (April 30, 2022), or 270 days after admission. When a Long COVID outcome occurred, that patient was censored for that particular Long COVID outcome. However, follow-up continued for other potential Long COVID outcomes, and thus a single patient could contribute to ≥ 1 Long COVID outcome. For patients still hospitalized at Day 90, outcomes were assessed from the day after discharge through the end of follow-up.

Potential Long COVID outcomes were identified and defined based on physician coauthor input and a previously published systematic literature review and meta-analysis (including 52 studies) assessing Long COVID outcomes from 28 days to 1 year following COVID-19 hospital discharge⁴. The primary composite outcome of Long COVID occurred if patients had ≥ 1 ICD-10 diagnosis code (Supplementary Table 5) or ICD-10 code U09.9 (“post-COVID condition, unspecified”) in any position in claims or chargemaster data during the study assessment period. Secondary outcomes included each of the 16 Long COVID outcomes (neuropsychiatric features, dyspnea/breathlessness, fatigue, joint pain/arthralgia, cognitive dysfunction, chest pain, cerebrovascular disease, thromboembolic disease, cough, ischemic heart disease, diarrhea, headache, muscle pain/myalgia, dysautonomia, taste disturbance/dysgeusia/ageusia, or smell disturbance/anosmia). Outcomes occurring in the assessment period could be new onset (incident), ongoing from acute infection (persistent), or present in both the baseline and outcomes assessment time periods (prevalent). These outcome states are consistent with interpretations of WHO definitions of “new onset,” “persistent,” or “fluctuating” and emerging definitions from the RECOVER consortium focusing on prevalent symptoms after SARS-CoV-2 infection irrespective of their presence at baseline^6,39.

Statistical analysis

Descriptive analyses of baseline patient characteristics were performed across the full study population and compared between the 2 groups using absolute SMDs. To prevent bias that can arise from patients followed for different lengths of times, outcome event rates per 100 person-years were calculated as the number of patients experiencing the outcome divided by the total person-time contributed over the assessment period. The 95% CIs for the rates were calculated using the normal approximation.

In the primary analysis, a pseudo-population of patients who were remdesivir-exposed or unexposed, balanced for key characteristics, was created using weights derived from inverse probability weighting based on a propensity score for baseline characteristics, censoring, and competing risks (Supplementary Methods and Supplementary Table 6).

Log-binomial models were fit to estimate relative risks, 95% CIs (using the normal approximation using weighted standard errors), and P values for the composite outcome of any Long COVID outcome and for each of the 16 individual outcomes. Each outcome model only included patients who experienced the outcome of interest or had a complete 270 days of follow-up. Patients who were censored or had evidence of inpatient mortality contributed to weighting models but were not included in the outcome models; thus, the number of patients contributing to each model varied. The outcome models were weighted by the product of the 3 individual weights (treatment assignment, censoring, and competing risks)^40,41. The 95% CIs and P values reported here were not adjusted for multiple comparisons; however, because of multiple comparisons, statistical significance was determined using the Holm-Bonferroni method (setting the significance threshold according to the number of tests performed) to reduce the likelihood of incorrectly rejecting the null hypothesis (type 1 error) rather than comparing the P values to an α of 0.05. The weighting and modeling process was conducted within each subgroup.

An as-treated sensitivity analysis (ie, considered the treatment received, irrespective of assignment) was performed by censoring patients who were unexposed by the index date and later received remdesivir at the point of crossover; crossover was incorporated into the weighting as a censoring event.