Comparative Effectiveness and Safety of Direct Oral Anticoagulants in Patients with Atrial Fibrillation: A Systematic Review and Meta-Analysis of Observational Studies

There are no head-to-head randomized controlled trials comparing different direct oral anticoagulants (DOACs). Thus, we systematically reviewed and meta-analyzed observational studies assessing the comparative effectiveness and safety of DOACs for stroke prevention in patients with atrial fibrillation (AF). We systematically searched MEDLINE and EMBASE up to February 2019 for observational studies comparing different DOACs head-to-head in patients with AF. Two independent reviewers identified studies, extracted data, and assessed the risk of bias using the Risk Of Bias In Non-randomized Studies of Interventions (ROBINS-I) tool. Random-effects models were used to meta-analyze data across higher-quality studies. We identified 25 cohort studies including 1,079,565 patients with AF treated with DOACs. Meta-analysis of the 19 studies at moderate risk of bias yielded a similar risk of ischemic stroke for rivaroxaban versus dabigatran (six studies; hazard ratio [HR] 0.93; 95% confidence interval [CI] 0.83–1.04; I2: 0%), apixaban versus dabigatran (five studies; HR 0.94; 95% CI 0.82–1.09; I2: 0%), and apixaban versus rivaroxaban (four studies; HR 1.07; 95% CI 0.93–1.23; I2: 0%). Regarding major bleeding, there was an increased risk for rivaroxaban versus dabigatran (six studies; HR 1.33; 95% CI 1.20–1.47; I2: 22%) and decreased risks for apixaban versus either dabigatran (eight studies; HR 0.71; 95% CI 0.64–0.78; I2: 0%) or rivaroxaban (eight studies; HR 0.56; 95% CI 0.48–0.65; I2: 69%). As head-to-head trials comparing different DOACs do not exist, available evidence derives exclusively from observational studies. These data suggest that while dabigatran, rivaroxaban, and apixaban have a similar effect on the risk of ischemic stroke, apixaban may be associated with a decreased risk of major bleeding compared with either dabigatran or rivaroxaban.


INTRODUCTION
Atrial fibrillation (AF) is a common cardiac arrhythmia that increases the risk of ischemic stroke five-fold.[1] While vitamin K antagonists (VKAs) have long been the primary oral anticoagulants for stroke prevention in AF, they are prone to drug-drug interactions and need frequent monitoring.[2] Direct oral anticoagulants (DOACs), including the thrombin inhibitor dabigatran and the Factor Xa inhibitors rivaroxaban, apixaban, and edoxaban, expanded recently our pharmacologic arsenal.They were found to be either non-inferior or superior to the VKA warfarin for stroke prevention in large randomized controlled trials and have several advantages over VKAs, including more rapid onset of anticoagulation and decreased need for monitoring.[3] Consequently, treatment guidelines now recommend DOACs as first-line oral anticoagulation among patients with AF. [4][5][6] To date, there are no large, head-to-head trials comparing different DOACs in patients with AF.Moreover, there is a need to assess the comparative effectiveness and safety of DOACs in real-world settings.While four publications to date have systematically reviewed and metaanalyzed available real-world data, [7][8][9][10] one used outdated tools for the assessment of the risk of bias, [7] while others omitted bias assessment altogether.[8,9] Moreover, numerous studies reporting head-to-head comparisons among DOACs that were recently published were not included in these earlier works.[11][12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27] Thus, the objective of this systematic review and meta-analysis of observational studies was to provide an up-to-date synthesis of the available real-world evidence on DOAC comparative effectiveness and safety in patients with AF, while thoroughly assessing the risk of bias of the included studies.

METHODS
This systematic review and meta-analysis was conducted according to a pre-specified protocol and is reported following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses [28] and Meta-Analysis of Observational Studies in Epidemiology.[29] 2.1 Search strategy MEDLINE and EMBASE were systematically searched from inception to February 28, 2019 for observational studies published in English in the peer-reviewed literature and comparing DOACs to each other in patients with AF.The search strategy was tailored to each database and included index terms (MeSH and Emtree) and text words related to AF and DOACs (see Electronic Supplementary Material eTable 1).We also scanned the bibliographies of the included articles and relevant reviews for further references.

Inclusion and exclusion criteria
Randomized controlled trials, cross-sectional studies, letters to the editor, commentaries/editorials, and previous reviews and meta-analyses were excluded.Conference abstracts were also excluded, as their results are often preliminary, and they contain insufficient information to adequately assess risk of bias.To minimize the potential effects of publication bias, we excluded studies with less than 1000 DOAC users.Studies looking at DOAC use in AF patients undergoing ablation were also excluded, as their results are not generalizable to AF patients in general.
Studies eligible for inclusion were cohort or case-control studies comparing DOACs (apixaban, dabigatran, rivaroxaban, or edoxaban) to each other in patients with AF.The primary effectiveness outcome was ischemic stroke, while the primary safety outcome was major bleeding.
Secondary effectiveness outcomes were all-cause mortality, myocardial infarction, and systemic embolism.Secondary safety outcomes included intracranial hemorrhage, hemorrhagic stroke, gastrointestinal bleeding, and other bleeding events.

Study selection
Two independent reviewers (either CMD/SY or AD/SY) performed study selection.Titles and abstracts were screened to identify potentially relevant studies and duplicates; all studies identified as potentially relevant by either reviewer proceeded to full-text review.Full-text review established the final set of included studies, with discrepancies resolved by consensus.

Data extraction
Two independent reviewers (either CMD/SY or AD/SY) extracted data using a pilot-tested female sex] score [31] or their components, and HAS-BLED [hypertension, abnormal renal/liver function, prior stroke, bleeding history or predisposition, labile international normalized ratio, age >65 years, drugs] score [32] and its components), and study outcomes.Other items extracted to describe the methodological approach and assess risk of bias included use of a new-user design, exposure definition (e.g., intention-to-treat, as-treated, time-dependent, etc.), and handling of treatment switch or discontinuation.The main summary measures of interest were hazard ratios (HR) or odds ratios (OR) with 95% confidence intervals (CIs).Effect estimates were presented for the comparisons rivaroxaban versus dabigatran, apixaban versus dabigatran, and apixaban versus rivaroxaban.For articles reporting effect estimates with a different DOAC as comparator (e.g., dabigatran versus rivaroxaban), comparator was changed, and reciprocal results were calculated.

Assessment of risk of bias
Two independent reviewers (AD/SY) assessed the risk of bias using the Risk Of Bias In Non-randomized Studies of Interventions (ROBINS-I) tool.[33] Seven domains were assessed: bias due to confounding; bias in the selection of study participants; bias in the classification of interventions; bias due to departure from intended interventions; bias due to missing data; bias in the measurement of outcomes; and bias in the selection of the reported results.Based on the assessment of each domain, an overall risk of bias was assigned as low, moderate, serious, or critical, with the overall risk determined by the highest risk assigned in any individual domain.[33] Given the potential for confounding inherent in observational studies, the highest quality studies were those with an overall moderate risk of bias.A moderate risk of confounding bias was ascribed to studies considering at least the following covariates in their design or analysis: age, sex, prior use of warfarin, use of antiplatelets, previous stroke (for stroke outcomes), CHADS2 or CHA2DS2-VASC score or their components (for stroke outcomes), previous bleeding (for bleeding outcomes), and HAS-BLED score or its components (for bleeding outcomes).

Study characteristics
All twenty-five included studies were cohort studies published between 2016 and 2019.They included a total of 1,079,565 patients (380,682 treated with dabigatran, 452,611 with rivaroxaban, and 246,272 with apixaban).The follow-up durations ranged from 89 to 422 days (Table 1).

Assessment of risk of bias
Based on ROBINS-I, nineteen studies were assigned a moderate risk of bias, [11, 12, 16-27, 34, 37, 38, 40, 41] four were assigned a serious risk of bias, [13,14,35,39] and two were assigned a critical risk of bias [15,36] (see Electronic Supplementary Material eTable 4).As one of the studies at moderate risk of bias reported only absolute risk differences, [18] its results are presented in the tables but not included in qualitative or quantitative data synthesis.One domain leading to a major increase in the risk of bias was 'risk of bias due to confounding', resulting from confounding by indication, contraindication and/or severity associated with previous use of VKAs, [15,35,36,39] time-varying confounding due VKA use during follow-up, [14] or from residual confounding due to failure to adjust for important confounders.[13] Eighteen studies used propensity score-based approaches in their analyses to control for confounding.[11, 13, 16, 17, 19-This is a post-peer-review, pre-copyedit version of an article published in 'Drug Safety'.The final authenticated version is available online at: https://doi.org/10.1007/s40264-019-00842-1.The following terms of use apply: https://www.springer.com/gp/ope11 26, 34, 37-41] A propensity score is defined as the probability of getting exposed to a medication, given a set of covariates.[42].As this score summarizes all patient characteristics into a single covariate, it reduces the potential for overfitting.However, the possibility of confounding due to unmeasured covariates cannot be excluded.
Another domain responsible for an increased risk of bias was 'bias in selection of participants into the study', resulting from the inclusion of previous users of VKAs [35,39] or DOACs, [15,36] as well as from potential informative censoring in the setting of an as-treated exposure definition.[11-13, 16, 18, 19, 21-23, 25-27, 34, 38-41] Of note, no study using an as-treated definition included statistical approaches to address informative censoring (e.g., inverse probability of censoring weights).However, three studies using both as-treated and intention-totreat definitions (in sensitivity analyses) while not having other sources of selection bias were ascribed a low risk in this respect given the complementary nature of these analyses.[20,24,37] Moreover, considering the short follow-up of the included studies (<1 year) and the resulting low risk of exposure misclassification, studies using an intention-to-treat approach were ascribed a low risk of "bias in classification of interventions".Finally, 'bias in selection of reported results' due to the absence of a prespecified study protocol also affected the quality of most of the included studies (see Electronic Supplementary Material eTable 4).
The results on DOAC comparative effectiveness and safety did not considerably change when comparing low-dose regimes (see Electronic Supplementary Material eTable 14) or using alternative exposure definitions (see Electronic Supplementary Material eTable 15).

DOAC effectiveness and safety in higher quality studies
When considering only the nineteen studies at moderate risk of bias and only outcomes assessed by more than one study, qualitative data synthesis remained inconclusive regarding the risk of ischemic stroke (HR range for rivaroxaban versus dabigatran: 0.73 to 1.12; HR range for apixaban versus dabigatran: 0.40 to 1.22; HR range for apixaban versus rivaroxaban: 0.67 to 1.27).
Data suggested an increased risk of major bleeding for rivaroxaban versus dabigatran (HR range: 1.05 to 1.69), and decreased risks for apixaban versus either dabigatran (HR range: 0.50 to 0.94) or rivaroxaban (HR range: 0.39 to 0.88).
Regarding all-cause mortality, we found a trend towards an increased risk for rivaroxaban versus dabigatran (HR range: 0.99 to 1.52), a similar risk for apixaban versus dabigatran (range: HR 0.91 to 1.14), and a trend towards a decreased risk for apixaban versus rivaroxaban (HR range: 0.81 to 0.94).There was also a similar risk of myocardial infarction for rivaroxaban versus dabigatran (HR range: 0.88 to 1.11).Moreover, data suggested an increased risk of systemic embolism for rivaroxaban versus dabigatran (HR range: 1.09 to 1.39), and a trend towards decreased risks for apixaban versus either dabigatran (HR range: 0.37 to 0.76) or rivaroxaban (HR range: 0.49 to 0.56), albeit all studies had wide 95% CIs.
Regarding intracranial hemorrhage, data suggested an increased risk for rivaroxaban versus dabigatran (HR range: 1.05 to 1.81), but data on apixaban were heterogenous (HR range versus dabigatran: 0.65 to 1.75; HR range versus rivaroxaban: 0.51 to 1.39).There was also a trend towards an increased risk of hemorrhagic stroke for rivaroxaban versus dabigatran (HR range: 1.70 to 4.55), a similar risk for apixaban versus dabigatran (HR range: 0.72 to 1.08), and heterogenous results for apixaban versus rivaroxaban (HR range: 0.32 to 1.49).Finally, regarding gastrointestinal bleeding, the results were heterogeneous for rivaroxaban versus dabigatran (HR range: 0.85 to 1.52) but suggested decreased risks for apixaban versus either dabigatran (HR range: 0.39 to 0.86) or rivaroxaban (HR range: 0.33 to 0.94).

DISCUSSION
The objective of our study was to synthesize the available real-world evidence on the comparative effectiveness and safety of DOACs.Overall, we identified twenty-five studies meeting our inclusion criteria.Considering only nineteen higher-quality studies, our meta-analyses suggest a similar risk of ischemic stroke for rivaroxaban versus dabigatran (HR, 0.93; 95% CI, 0.83 to 1.04), apixaban versus dabigatran (HR, 0.94; 95% CI, 0.82 to 1.09), and apixaban versus rivaroxaban (HR, 1.07; 95% CI, 0.93 to 1.23).Moreover, we observed an increased risk of major bleeding for rivaroxaban versus dabigatran (HR, 1.33; 95% CI, 1.20 to 1.47) and decreased risks for apixaban versus either dabigatran (HR, 0.71; 95% CI, 0.64 to 0.78) or rivaroxaban (HR, 0.56; 95% CI, 0.48 to 0.65).Some studies included in this systematic review had several limitations that warrant consideration.Using the ROBINS-I tool, we found that nineteen studies were assigned a moderate risk of bias, [11, 12, 16-27, 34, 37, 38, 40, 41] while six studies were assigned a serious or critical risk of bias.[13-15, 35, 36, 39] A potential limitation observed in all studies with a serious or critical risk of bias was confounding by indication, contraindication, and/or severity related to previous use of VKAs.The remaining studies considered previous VKA use in their design, either by matching on propensity scores that included previous VKA use as a variable or by excluding previous VKA users.While the first approach does not eliminate the possibility of residual confounding since aspects such as duration of previous VKA use are not taken into consideration, the second approach may yield findings of decreased generalizability as many DOAC users are previous VKA users.[43] The prevalent new-user study design, a newly-developed approach incorporating both new users and switchers from previous medications that considers the duration of previous treatment could offer an alternative in this setting.[44] Another major limitation was the indiscriminate inclusion of prevalent users, [15,36] which may result in under-ascertainment of early adverse events, depletion of susceptibles, and adjusting for covariates in the causal pathway.[45,46] Our findings of a similar risk of ischemic stroke among DOACs well as the decreased risk of major bleeding with apixaban compared with either rivaroxaban or dabigatran are congruent with those of a recent systematic review of network meta-analyses of randomized controlled trials.[47] Moreover, our findings that rivaroxaban could be associated with an increased risk of major bleeding and all-cause mortality compared with dabigatran are congruent with those of the meta-analysis by Bai et al. [7] However, while Bai et al. reported no differences between rivaroxaban and dabigatran regarding intracranial hemorrhage (HR, 1.22; 95% CI, 0.85 to 1.59), [7] our pooled estimate suggested a 71% increased risk for rivaroxaban.A possible explanation for this discrepancy is that Bai et al. also included two studies assigned a serious of bias in our quality assessment that suggested a decreased risk for rivaroxaban.[35,39] The higher risks for different types of bleeding observed with rivaroxaban compared with dabigatran or apixaban could be a result of the dosing regimens.Indeed, while DOACs have similar plasma half-lives, [48] rivaroxaban is given once daily as opposed to dabigatran and apixaban that are given twice daily.It is conceivable that once-daily regimens could lead to higher peak levels and to increased risk of bleeding.However, to our knowledge, a correlation between rivaroxaban plasma levels and bleeding events has yet to be shown.
Our study has several strengths.First, it provides an up-to-date synthesis of the available literature in a dynamically evolving field, including several recent studies not captured in previous systematic reviews and considering overall almost half a million DOAC users.Second, this study presents robust data on the comparative effectiveness and safety of apixaban, a relatively recently approved DOAC.Finally, we used ROBINS-I to evaluate the quality of the included studies, a tool that enables a robust assessment of the risk of different biases such as confounding or selection bias, and restricted meta-analysis to higher quality studies.
Our study also has some limitations.First, our review is affected by the limitations of the included studies such as residual confounding due to clinical data not typically captured by administrative databases (e.g., smoking, diet).Second, while the exclusion of studies with <1000 DOAC users provides an objective, pre-specified threshold based on underlying event rates, there is a possibility that some underpowered but potentially eligible studies could have been excluded.Finally, as the included studies were conducted using computerized healthcare databases from different jurisdictions, confounding due to jurisdiction-specific factors such as formulary restrictions cannot be excluded.

CONCLUSIONS
Our systematic review and meta-analysis suggest no major differences in the risk of ischemic stroke, all-cause mortality, myocardial infarction, or systemic embolism between dabigatran, rivaroxaban, and apixaban in patients with AF.However, rivaroxaban is associated with an increased risk of bleeding compared with dabigatran, while apixaban is associated with a decreased risk of bleeding compared with either dabigatran or rivaroxaban.Thus, current observational evidence supports the notion that while differences among DOACs regarding effectiveness appear to be small, apixaban should be preferred in AF patients at higher risk of bleeding.

Figure 1 .
Figure 1.Forest plot demonstrating individual and pooled relative risks of ischemic stroke

Figure 2 .
Figure 2. Forest plots demonstrating individual and pooled relative risks of major bleeding

Supplementary Material eTable 2). Study
characteristics included study design, location, data source, study period, sample size (overall and by exposure group), follow-up duration, patient characteristics (age, sex, CHADS2 [congestive heart failure, hypertension, age ≥75 years, diabetes mellitus, prior stroke or transient ischemic attack] score[30] or CHA2DS2-VASc [congestive heart failure, hypertension, age ≥75 years, diabetes mellitus, prior stroke or transient ischemic attack, vascular disease, age 65-74 years, form, with discrepancies resolved by consensus (see Electronic

Table 1 . Characteristics of observational studies on effectiveness and safety of DOACs among patients with AF
This is a post-peer-review, pre-copyedit version of an article published in 'Drug Safety'.The final authenticated version is available online at: https://doi.org/10.1007/s40264-019-00842-1.The following terms of use apply: https://www.springer.com/gp/ope31

Table 2 . Results of meta-analyses for the comparative effectiveness and safety of DOACs among patients with AF
Abbreviations: DOACs, direct oral anticoagulants; AF, atrial fibrillation; HR, hazard ratio; CI, confidence interval Figure 2