Search strategy

We followed a comprehensive search strategy to identify any relevant articles on the topic, mainly from 4 medical databases including PubMed, Cochrane Library, Web of Science, and Embase. We also searched relevant papers using the Google search engine and major preprint platforms including Medrix, bioRxiv, and SSRN. Tailored search terms featured “2019‑nCoV,” “COVID‑19,” “Coronavirus,” “SARS‑CoV‑2,” and “Wuhan Coronavirus” (Supplementary material, Table S1). No language restriction or publication status criteria were set. Reference lists of relevant articles were also screened for eligible studies. The last search was performed on April 7, 2020.

Study selection

Two investigators (LP and LH) independently screened the manuscripts of the potentially eligible studies. Another investigator (SH) checked the results, and disagreement was resolved by consensus. Inclusion criteria were as follows: 1) randomized controlled trials (RCTs), cohort studies, case control studies, and cross‑sectional studies; 2) study settings and patient characteristics were provided; and 3) detailed data on drug interventions and outcomes were available. Drug interventions included the use of antiviral agents, glucocorticoids, antibiotics, and intravenous immunoglobulin. Outcomes referred to the number of survivors and nonsurvivors at the end of the follow‑up of each study. Exclusion criteria were: 1) duplicate reports; 2) preliminary studies that included patient groups overlapping with those presented in most recent reports.

Quality assessment

The Newcastle–Ottawa quality assessment scale was used to assess study quality and risk of bias for retrospective studies.² The scale consists of 3 elements (selection, comparability, and exposure) and is covered by 8 items. According to this scale, the number of stars was used to evaluate study quality. A total of 4 stars can be awarded for selection, 2 for comparability, and 3 for exposure. Studies with 1 to 3 stars were considered as those of low quality; studies with 4 to 6 stars, of moderate quality; and studies with 7 to 9 stars, of high quality. The modified Jadad score (7 points) was used to assess the quality of RCTs, with classification criteria of high quality (6–7 points), moderate quality (4–5 points), and low quality (1–3 points).³ Two investigators (LP and LH) independently performed quality assessment, and the third investigator (WL) checked the results and resolved any disagreement.

Definition of interventions and outcomes

All‑cause mortality at the end of follow‑up of each study was regarded as the primary outcome. A pharmacological intervention was defined as a situation in which patients received a specific drug of interest (including antiviral agents, glucocorticoids, antibiotics, and intravenous immunoglobulin). Pooled analyses were performed to evaluate the association of intervention effects and patient outcomes according to the study definitions.

Data extraction

We used a standard strategy to extract the following data from each study: study characteristics (authors, date of publication, study design, duration of follow‑up, and sample size), participants (age and sex), patients with COVID‑19 who received pharmacological interventions (antiviral agents, glucocorticoids, antibiotics, and intravenous immunoglobulin) or not, and outcomes (number of nonsurvivors and survivors). Data were independently extracted by 2 investigators (XG and WJ) and checked by the third investigator (DC). Our protocol was not published or registered owing to the rapid emergence of this infectious disease.

Statistical analysis

Statistical analyses were performed using the STATA software, version 14 (StataCorp, College Station, Texas, United States). As most studies were retrospective and expected to be heterogenous, we chose the random effects model for data synthesis.⁴ For retrospective studies, we used odds ratios (ORs) and 95% CIs as effect measures. For RCTs, we pooled results using relative risk (RR) and 95% CIs. All the ORs and RR with corresponding 95% CIs were graphically visualized on forest plots. Heterogeneity across studies was evaluated using the Cochrane Q test and the I² test (I² = 100% [(Q – df) / Q]). The I² value of 0% to 49%, 50% to 74%, and higher than 75% indicated low, moderate, and high heterogeneity, respectively.⁵

Subgroup analyses and sensitivity analysis plans were proposed based on the quality of studies, study design, participants, and types of drugs, as appropriate. Publication bias was assessed by funnel plots if more than 10 studies were included. A 2‑sided P value less than 0.05 was considered significant.

Search results and study selection

The flowchart of study selection is presented in Figure 1. We identified 3421 references by the initial database query and manual search. Among them, 523 were removed as duplicates and 2857 were excluded after title and abstract screening. Eventually, 41 articles were eligible for full‑text review. Thirty‑five studies were excluded due to the following reasons (Supplementary material, Table S2): 21 did not include relevant grouping variables; 6 did not report relevant data on pharmacotherapy; 4 were review articles; 2 were case reports; 1 did not include a control group; and 1 was a correspondence. Six studies presenting patients’ pharmacotherapy data and outcomes were included for systematic review and meta‑analysis.

Characteristics of the included studies

The main characteristics of the included studies are shown in Table 1. A total of 1142 patients were included. Geographically, all studies originated from China, with varied sample sizes ranging from 52 to 274 patients. Of these, 5 were retrospective and observational^6-10 and there was a single RCT.¹¹ Of the 6 included studies, all except the study by Yang et al⁷ received funding. Patients enrolled in these studies were at a median age of 40 to 69 years and predominantly male (55% to 67%).