This is a registered meta-analysis on PROSPERO (https://www.crd.york.ac.uk/prospero/). The registration number is CRD42020214713.

The participants, intervention, comparison, outcome, and study design approach (PCIOS) were used to select clinical studies (Table 1). Reviews, meta-analyses, non-human studies, case reports, and conferences were excluded. Studies that did not compare the clinical outcomes between UFH and bivalirudin were excluded as well. Two authors (S. Liang and J. Zhu) independently searched the PubMed, EMBASE, and The Cochrane Library databases for articles published from inception until 1 June 2021, using the heading terms “heparin,” “unfractionated heparin,” “bivalirudin,” “extracorporeal membrane oxygenation,” “ECMO,” “ECMO treatment,” “ECLS,” or “ECLS treatment”. No language restrictions were used. The references of relevant literature were also searched to look for more eligible studies.

Data were extracted by the same two independent readers (S. Liang and J. Zhu) who performed the literature search and study selection; the researchers were not blinded to the authors and institutions of included studies. Disagreements were solved by a third reader (M. Ma). Y. He supervised the whole process. This meta-analysis followed the guidelines for preferred reporting items for systematic reviews and meta-analyses (PRISMA) (Shamseer et al., 2015). The two reviewers extracted the following information independently: the first author, published year, study design (prospective/retrospective), study duration, total patients and number of patients in the bivalirudin and UFH groups, the doses in the bivalirudin and UFH group, and the incidence of thromboses, major bleeding, and mortality (per-patient).

For the observational studies, the Newcastle-Ottawa Scale (NOS) was used to assess the risk of bias. The NOS ranges from 0 (lowest) to 9 (highest), and studies with scores ≥6 are considered high quality. For RCTs, the modified Jadad quality scoring scale is used for the quality assessment, which includes the generation of random sequences, distribution methods, randomized concealment, double-blinding, withdrawals and dropouts. The Jadad score among four to seven is considered as good quality.

Sensitivity analysis of the included studies was conducted via a one-by-one elimination method to evaluate the meta-analysis’s stability. A Galbraith plot was used to find the cause of heterogeneity. Egger’s test was used to test the publication bias via Stata version 15.1 Software (The StataCorp LP, Texas City, United States).

For studies describing the results via median and interquartile range (IQR), Standard deviations (SDs) of the mean differences (MDs) were obtained as described by former researches (Wan et al., 2014; Luo et al., 2018). The RevMan version 5.3.5 Software (The Cochrane Collaboration, Copenhagen, Denmark) and Stata version 15.1 Software (The StataCorp LP, Texas City, United States) was used for all statistical analyses. Statistical heterogeneity was assessed by using the Cochrane Q and the I square statistics. Heterogeneity was interpreted as absent (I ² :0–25%), low (I ² : 25.1–50%), moderate (I ² : 50.1–75%), or high (I ² : 75.1–100%) (Higgins et al., 2003). The use of a random-effects model was also considered when the number of studies was relatively small, and a random-effects model was applied to estimate the continuous outcome data for data with a p-value ≤0.1 and an I ² -value >50%, which indicated statistical heterogeneity (Higgins et al., 2003). Otherwise, a fixed-effects model was used. The overall log with its 95% CI was used as the summary of the overall effect size. A p-value <0.05 was considered statistically significant.

The literature search produced 125 total findings (101 on EMBASE, 4 on The Cochrane Library, and 20 on PubMed); 80 full texts were retrieved after duplicates were removed. The titles and abstracts of studies were screened, after which 53 articles were excluded due to the following reasons: systematic reviews (n = 2), reviews, letters and editorials (n = 34), case reports (n = 17). A total of 27 full-text articles were reviewed, and 17 were excluded later because they lacked the comparison between UFH and bivalirudin (n = 15) or they are only with abstracts (n = 2). Finally, ten unique retrospective observational studies (Ranucci et al., 2011; Pieri et al., 2013; Ljajikj et al., 2017; Berei et al., 2018; Macielak et al., 2019; Brown et al., 2020; Hamzah et al., 2020; Kaseer et al., 2020; Schill et al., 2021; Seelhammer et al., 2021) with 847 patients were included for the quantitative analysis. All articles were published before 1 June 2021. The literature screening process is presented in Figure 1.

Table 2 shows basic information from the included studies; Table 3 shows group definition of the bivalirudin group and clinical outcomes. Generally, these included studies met most NOS quality indicators. However, the control group of all the studies did not meet the standard of “community controls” and “no history of diseases” as the controls was from a hospital. Moreover, as the included studies were all case-control retrospective studies, they were not blinded to the case/control status. According to the NOS, all the included studies were considered as high quality (Supplemental Table S1).

Nine studies (Ranucci et al., 2011; Pieri et al., 2013; Ljajikj et al., 2017; Berei et al., 2018; Macielak et al., 2019; Brown et al., 2020; Hamzah et al., 2020; Kaseer et al., 2020; Schill et al., 2021) reported the incidence of major bleeding. Two studies (Macielak et al., 2019; Brown et al., 2020) were not included due to the different ways of expression (per ECMO day). The incidence rate of major bleeding is 0.223 and 0.139 per ECMO day in the UFH and bivalirudin group in Macielak et al. (Macielak et al., 2019)’s study, 0.308 and 0.062 per ECMO day in Brown et al. (Brown et al., 2020)’s study, respectively. Regarding Ljajikj et al. (Ljajikj et al., 2017)’s study, we considered both delayed chest closure and intracranial bleeding as major bleeding because the authors thought that delayed chest closure might also be the result of diffuse persisting bleeding.

Seven studies (Ranucci et al., 2011; Pieri et al., 2013; Ljajikj et al., 2017; Berei et al., 2018; Hamzah et al., 2020; Kaseer et al., 2020; Schill et al., 2021) were included in the meta-analysis, and moderate heterogeneity was observed (I ² = 59%, p = 0.02), therefore a subgroup analysis was conducted. Five studies (Ranucci et al., 2011; Pieri et al., 2013; Ljajikj et al., 2017; Berei et al., 2018; Kaseer et al., 2020) were included in the adults group, whilst two studies (Hamzah et al., 2020; Schill et al., 2021) were include in the children group (Figure 2). Low heterogeneity was observed in both adults and children group (I ² = 48% and 48%, p = 0.10 and 0.16, respectively), therefore a fixed-effects model was used. The results showed that the difference of pooled incidence of major bleeding was significantly reduced in children (I ² = 48%, p = 0.01, odd ratio (OR) = 0.17, 95% confidence interval (CI): 0.04–0.66) in the bivalirudin group, but not in adults (I ² = 48%, p = 0.65, OR = 0.87, 95% CI: 0.46–1.62). The heterogeneity decreased after subgroup analysis, which indicates that the age maybe one of the sources of heterogeneity.

The sensitivity analysis of the incidence of major bleeding of the included studies showed that all studies’ estimate was within 95% CI of the total effect except for Berei et al. (Berei et al., 2018)’s study, which means the analytical stability may be affected (Supplementary Figure S1). After removing the study the difference of pooled incidence of major bleeding was still not significantly reduced in the adult group (I ² = 0%, p = 0.05, OR = 0.40, 95% CI: 0.16–0.99).

Ten studies (Ranucci et al., 2011; Pieri et al., 2013; Ljajikj et al., 2017; Berei et al., 2018; Macielak et al., 2019; Brown et al., 2020; Hamzah et al., 2020; Kaseer et al., 2020; Schill et al., 2021; Seelhammer et al., 2021) reported the incidence of thrombosis. Two studies (Macielak et al., 2019; Brown et al., 2020) were not included due to the different ways of expression (per ECMO day). The incidence rate of thrombosis is 0.207 and 0.089 per ECMO day in the UFH and bivalirudin group in Macielak et al. (Macielak et al., 2019)’s study, 0.043 and 0 per ECMO day in Brown et al. (Brown et al., 2020)’s study, respectively.

Thrombosis can be divided into patient thrombosis and in-circuit thrombosis/interventions. Eight studies (Ranucci et al., 2011; Pieri et al., 2013; Ljajikj et al., 2017; Berei et al., 2018; Hamzah et al., 2020; Kaseer et al., 2020; Schill et al., 2021; Seelhammer et al., 2021) reported the incidence of patient thrombosis and six studies (Ranucci et al., 2011; Pieri et al., 2013; Berei et al., 2018; Hamzah et al., 2020; Kaseer et al., 2020; Seelhammer et al., 2021) reported the incidence of in-circuit thrombosis/interventions group (Figure 3). Low heterogeneity was observed in both group (I ² = 0%, p = 0.96 and 0.81, respectively), therefore a fixed-effects model was used. The results showed that both the difference of pooled incidence of patient thrombosis (I ² = 0%, p = 0.02, OR = 0.58, 95% CI: 0.37–0.93) and in-circuit thrombosis/interventions (I ² = 0%, p = 0.0005, OR = 0.40, 95% CI: 0.24–0.68) was significantly reduced in the bivalirudin group. The sensitivity analysis showed that all studies’ estimate was within 95% CI of the total effect, which means the analytical stability was not affected (Supplementary Figure S2)

Seven studies (Ranucci et al., 2011; Pieri et al., 2013; Ljajikj et al., 2017; Berei et al., 2018; Hamzah et al., 2020; Kaseer et al., 2020; Schill et al., 2021; Seelhammer et al., 2021) reported the incidence of mortality. Mortality can be divided into in-hospital mortality, 30-day mortality, and 1-year mortality. Seven studies (Ranucci et al., 2011; Pieri et al., 2013; Berei et al., 2018; Hamzah et al., 2020; Kaseer et al., 2020; Schill et al., 2021; Seelhammer et al., 2021) were included in the in-hospital mortality group, three studies (Ljajikj et al., 2017; Berei et al., 2018; Kaseer et al., 2020) were included in the 30-day mortality group, and only one study (Ljajikj et al., 2017) was included in the 1-year mortality group (Figure 4). Low heterogeneity was observed in both in-hospital mortality and 30-day mortality group (I ² = 0%, p = 0.58 and 0.53, respectively), therefore a fixed-effects model was used. The results showed that the difference of pooled incidence of in-hospital mortality was significantly reduced in the bivalirudin group (I ² = 0%, p = 0.007, OR = 0.64, 95% CI: 0.46–0.88), but the difference of pooled incidence of 30-day mortality was not significant (I ² = 0%, p = 0.61, OR = 0.83, 95% CI: 0.41–1.68). The sensitivity analysis showed that all studies’ estimate was within 95% CI of the total effect, which means the analytical stability was not affected (Supplementary Figure S3)

Eight studies (Ranucci et al., 2011; Pieri et al., 2013; Berei et al., 2018; Macielak et al., 2019; Hamzah et al., 2020; Kaseer et al., 2020; Schill et al., 2021; Seelhammer et al., 2021) reported the ECMO duration, five of which (Pieri et al., 2013; Hamzah et al., 2020; Kaseer et al., 2020; Schill et al., 2021; Seelhammer et al., 2021) described the ECMO duration between the bivalirudin group and the heparin group via IQR (Figure 5). Six studies (Ranucci et al., 2011; Pieri et al., 2013; Berei et al., 2018; Macielak et al., 2019; Kaseer et al., 2020; Seelhammer et al., 2021) were included in the adult group and three studies (Hamzah et al., 2020; Schill et al., 2021; Seelhammer et al., 2021) were included in the children group. Low heterogeneity was observed in the adult group (I ² = 41%, p = 0.13), therefore a fixed-effects model was used. The results showed that the MD of pooled ECMO duration was not significant between the two groups (I ² = 41%, p = 0.75, MD = -3.19, 95% CI: -23.01–16.63).

High heterogeneity was observed in the children group (I ² = 76%, p = 0.02), therefore a random-effects model was used. The results showed that the MD of pooled ECMO duration was not significant between the two groups (I ² = 76%, p = 0.65, MD = 40.33, 95% CI: -135.45–216.12).

After removing the study conducted by Schill et al. (2021), the heterogeneity of this outcome decreased significantly (I ² = 0%, p = 0.68), which indicated the main source of heterogeneity. Therefore, a fixed-effects model was used. The results showed that the MD of pooled ECMO duration was not significant between the two groups (I ² = 0%, p = 0.25, MD = -53.30, 95% CI: -143.16–36.56). However, a directional change occurred after removing Schill et al. (2021), indicating that the results of this meta-analysis maybe not that stable, more studies should be included. For adults’ ECMO duration, the sensitivity analysis showed that all studies’ estimate was within 95% CI of the total effect, which means the analytical stability was not affected (Supplementary Figure 4A). For children’s ECMO duration, the sensitivity analysis showed that two studies’ (Schill et al., 2021; Seelhammer et al., 2021) estimate was not within 95% CI of the total effect, which means the analytical stability maybe affected (Supplementary Figure 4B).

Egger’s and Begg’s tests suggested no significant publication bias of the incidence of major bleeding (Egger p = 0.093 and Begg p = 0.368), patient thrombosis (Egger p = 0.116 and Begg p = 0.035), circuit thrombosis (Egger p = 0.503 and Begg p = 0.452), in-hospital mortality (Egger p = 0.551 and Begg p = 0.764), 30-day mortality (Egger p = 0.757 and Begg p = 1), ECMO duration in adults (Egger p = 0.156 and Begg p = 0.133) and children (Egger p = 0.282 and Begg p = 0.296).

Compared with the former systematic review (Sanfilippo et al., 2017), our study introduced new clinical studies and expanded the sample size, and we conducted the first meta-analysis. However, there are still some limitations in our study. Initially, the studies included were retrospective small-size studies, which means the argumentation intensity may not be strong enough, and only a hypothesis can be generated. We hope that there will be more large-scale RCTs in the future. Secondly, though sensitivity and subgroup analyses were performed, the patients’ variable character may lead to heterogeneity, which may affect the stability of the results. Future research is essential to ensure the homogeneity of the population as much as possible. Finally, the lack of specific results or available research data may restrict our subgroup analysis, such as the use of bivalirudin versus UFH in a different type of ECMOs (VV or VA), and different indications (STEMI, ARDS, heart transplantation, and so forth); these can be further investigated in future studies and enrolled in the sensitivity analysis. Despite these limitations, our meta-analysis provides valuable insight into the use of bivalirudin in the anticoagulation therapy of ECMO.