Two reviewers independently screened all relevant articles published on MEDLINE, the Cochrane Library, EMBASE, and the Web of Science databases from inception to March 28, 2021. The search terms used were “COVID-19,” “2019 novel coronavirus infection,” “coronavirus disease-2019,” “2019-nCoV disease,” “2019 novel coronavirus disease,” “severe acute respiratory syndrome coronavirus 2,” “Wuhan coronavirus,” “Wuhan seafood market pneumonia virus,” “SARS-CoV-2,” “SARS2,” “anticoagulant,” “anticoagulation,” “heparin,” “unfractionated heparin,” “UFH,” “fondaparinux,” “enoxaparin,” “low-molecular-weight heparin,” “heparin, low molecular weight,” “LMWH,” “thromboprophylaxis,” “antithrombotic,” and “anti-thrombosis” (see Supplementary Table 2 for the detailed search strategy). The titles and abstracts of the resulting articles were examined to exclude irrelevant studies. The full texts of the remaining articles were read to determine if these articles meet our eligibility criteria. Randomized controlled trials (RCTs) or observational studies that compared the effect of anticoagulation vs. non-anticoagulation on the mortality and/or bleeding events of hospitalized COVID-19 patients were included. The exclusion criteria were as follows: (1) absence of original data; (2) enrollment of non-hospitalized COVID-19 patients; (3) case series/report; (4) absence of a comparator group; and (5) non-human studies.

This meta-analysis mainly investigated the impact of anticoagulation on the in-hospital mortality and bleeding risk of inpatients with COVID-19. The Newcastle–Ottawa Scale was used to assess the methodological quality of observational studies (case–control or cohort studies), which has eight criteria and yields scores ranging from 0 to 9. Studies with scores ≥7 were regarded as high quality (15). The Grading of Recommendations Assessment, Development and Evaluation (GRADE) framework was applied for rating the quality of evidence for each outcome in the pooled analysis, and the quality of evidence was rated as high, moderate, low, and very low quality (16). The credibility of results from subgroup analyses was assessed by specific criteria (17, 18).

The primary outcome was all-cause in-hospital mortality during hospitalization for COVID-19. The secondary outcome was the incidence of bleeding events during hospitalization for COVID-19. Definition of bleeding events was according to definitions in the individual studies, mainly including (1) an active source of bleeding; (2) suspected bleeding without confirmation of an active bleeding source; (3) confirmatory imaging or other evidence (neuroimaging for intracranial bleed); and (4) bleeding necessitating a transfusion of packed red blood cells.

Dichotomous variables such as mortality or bleeding events were expressed as risk ratio (RR) and 95% confidence interval (CI). Heterogeneity was evaluated using the I² statistic and Cochrane Q test to assess the degree of inter-study variation. I² values of 0–24.9%, 25–49.9%, 50–74.9%, and 75–100% were considered as having no, mild, moderate, and significant thresholds for statistical heterogeneity, respectively (19, 20). A random-effects model using restricted maximum likelihood (21), which is thought to be better than the conventional DerSimonian–Laird method (22), was performed to pool the data due to potential clinical heterogeneity. Potential publication bias was assessed by inspection of funnel plots when the total number of included studies surpasses 10, with Begg's rank correlation test performed subsequently. Predefined sensitivity analyses were conducted using a random-effects model and a leave-one-out approach. We also conducted two exploratory sensitivity analyses by excluding the clinical trials performed in the USA and by excluding those performed outside the USA in the meta-analysis of mortality, respectively. All statistical analyses were performed using Stata version 12.0 and Review Manager version 5.3.

The study selection process is presented in Figure 1. The literature search yielded 4,421 potentially relevant records. We retained 2,107 relevant studies after removing duplicate studies. After we evaluated the abstract and title of each record, 2,060 studies were excluded because they did not meet the inclusion criteria or they met the exclusion criteria. Of the remaining 47 studies, 36 were excluded after reviewing the full-text manuscript (10 used direct oral anticoagulants as a control group, eight did not compare the mortality of COVID patients receiving anticoagulation with those not receiving anticoagulation directly, 12 enrolled outpatients, and six did not have insufficient outcome data or baseline data for further meta-analysis). A total of 11 observational studies reporting the effect of anticoagulants on mortality (9–11, 23–30) (n = 11) and bleeding events (9, 25, 27–29) (n = 5) in hospitalized patients with COVID-19 were included in the review.

This meta-analysis is based on a pooled sample of 20,748 hospitalized COVID-19 patients with reported information related to mortality and a pooled sample of 11,302 patients with reported information related to bleeding events. The sample sizes ranged widely across studies from 81 to 4,389. All studies were retrospective in study design. Among the included studies, seven were conducted in the USA (9, 10, 25, 27–30), two studies were conducted in China (11, 23), one study was conducted in Spain (26), and one study was conducted in Italy (24). Three studies were conducted exclusively in the ICU settings (9, 11, 24). The mean age in most studies was over 60 years. The proportion of male subjects ranged from 47 to 69%. Heparin and enoxaparin were the most used anticoagulants in the included studies. Table 1 further summarizes the detailed characteristics of each study, and anticoagulation doses and routes are listed in Supplementary Table 3.

The Newcastle–Ottawa Scale score for the included studies ranged from 6 to 9, while 10 studies scored above 7 and were of high quality. A full assessment is shown in Table 2.

Eleven studies reported the overall mortality of COVID-19 patients receiving prophylactic/therapeutic dose anticoagulation therapy or not receiving anticoagulation, which ranged widely across studies from 8.1 to 27.3% for anticoagulation group and 16.4–29.6% for non-anticoagulation group in mild-to-moderate COVID-19 patients, while it ranged from 29.8 to 40% vs. 62.7 to 78.9% in severe to critical COVID-19 patients, respectively (9–11, 23–30). Of 14,854 hospitalized COVID-19 patients receiving anticoagulant administration, 2,430 patients died during hospitalization, whereas 1,459 of 5,894 patients who did not receive anticoagulant administration died. Figure 2 shows that anticoagulation therapy was significantly associated with reduced in-hospital mortality in inpatients with COVID-19 (RR = 0.70, 95% CI 0.52–0.93, p = 0.01, I² = 94%). Using the GRADE framework, we rated the quality of evidence in mortality as high (Table 3).

Three studies further reported the mortality in COVID-19 patients admitted to ICU (9, 11, 24). There was also a significant decrease in mortality risk in critically ill COVID patients receiving anticoagulant therapy compared with those not receiving anticoagulant therapy (RR = 0.59, 95% CI 0.43–0.83, p = 0.002, I² = 76%, Figure 3). The quality of evidence for mortality in hospitalized patients admitted to ICU was rated as moderate because plausible confounding is not considered to reduce or increase demonstrated effect (Table 3).

There were five studies that specified bleeding events during the anticoagulant therapy for COVID-19 patients (25, 27–29, 31). The meta-analysis with a total of 11,302 patients found that the pooled risk ratio of bleeding risk did not favor either of the two groups (RR = 0.93, 95% CI 0.71–1.23; p = 0.63, I² = 0%, Figure 4). The quality of evidence for bleeding events was rated as moderate because plausible confounding is not considered to reduce or increase demonstrated effect (Table 3).

As shown in Figure 5, visual assessment of the funnel plot did show a little substantial asymmetry, but Begg's rank correlation test did not indicate the evidence of publication bias across studies of anticoagulant therapy in COVID-19 patients and mortality (p = 0.213).

In order to assess the stability of the results of the current meta-analysis, we first performed a one-study-removed sensitivity analysis for mortality. Statistically, similar results were obtained after omitting each of the studies (Table 4), indicating the stability of this meta-analysis. Given that the mortality of in-hospital COVID-19 patients differs significantly among countries due to multiple factors such as reporting bias, transparency, and definition of COVID-19 mortality, we further performed two exploratory sensitivity analyses to test the stability of the results in this meta-analysis. As shown in Table 5, the RR value of anticoagulation therapy was 0.75 (0.63, 0.89) when clinical studies performed in the USA were excluded. However, an insignificantly reduced in-hospital mortality risk was observed when trials conducted outside of the USA were excluded (RR = 0.67, 95% CI 0.43–1.02, p = 0.06, Table 6).