Dabigatran, Rivaroxaban, or Apixaban versus Warfarin in Patients with Nonvalvular Atrial Fibrillation: A Systematic Review and Meta-Analysis of Subgroups

Background. New oral anticoagulants (NOAC; rivaroxaban, dabigatran, apixaban) have become available as an alternative to warfarin anticoagulation in non-valvular atrial fibrillation (NVAF). Methods. MEDLINE and CENTRAL, regulatory agencies websites, clinical trials registers and conference proceedings were searched to identify randomised controlled trials of NOAC versus warfarin in NVAF. Two investigators reviewed all studies and extracted data on patient and study characteristics along with cardiovascular outcomes. Relative risks (RR) and 95% confidence intervals (CI) were estimated using a random effect meta-analysis. Results. Three clinical trials in 50,578 patients were included. The risk of non-hemorrhagic stroke and systemic embolic events (SEE) was similar with the NOAC and warfarin (RR = 0.93; 95% CI = 0.83–1.04), while the risk of intracranial bleeding (ICB) with the NOAC was lower than with warfarin (RR = 0.46; 95% CI = 0.33–0.65). We found differences in the effect size on all strokes and SEE depending on geographic region as well as on non-hemorrhagic stroke, SEE, bleeding and mortality depending on time in therapeutic range. Conclusion. The NOAC seem no more effective than warfarin for prevention of nonhemorrhagic stroke and SEE in the overall NVAF population, but are generally associated with a lower risk of ICB than warfarin.


Introduction
Atrial fibrillation (AF) represents the most common sustained cardiac arrhythmia, affecting more than 6 million people in Europe [1,2]. AF, particularly when it is persistent/permanent, predisposes patients to the development of atrial thrombi, which may embolize to the systemic circulation, being associated with a 4-to 5-fold increase in the risk of ischemic stroke [3].
Vitamin K antagonists (VKA; coumarins, like warfarin and acenocoumarol) have been the only oral anticoagulants available over the last 60 years [4]. These agents are effective to prevent stroke in patients with AF [5], but their management remains problematic due to their narrow therapeutic index and variability in drug exposure, necessitating routine coagulation monitoring (international normalised ratio (INR)), clinical surveillance, and continuous patient education [6]. As a result, approximately only half of eligible patients with AF receive oral anticoagulation with VKA [7].
Dabigatran etexilate (Pradaxa, Boehringer Ingelheim) [8], rivaroxaban (Xarelto, Bayer HealthCare) [9], and apixaban (Eliquis, Bristol Myers Squibb) [10] are new oral anticoagulants (NOAC) available in Europe and other countries. Unlike VKA, these new compounds exhibit a predictable dose response and do not require routine coagulation monitoring, but their anticoagulant effect declines quickly in case of 2 Thrombosis poor compliance, and not coagulation monitoring tests or specific antidotes are currently available [4]. Among other indications, these new compounds have been tested for prophylaxis of stroke and systemic embolic events (SEE) in patients with nonvalvular atrial fibrillation (NVAF) using a combined primary outcome of all strokes (nonhemorrhagic and haemorrhagic) and SEE. The pivotal studies conducted in that indication are usually large clinical trials recruiting heterogeneous populations in different geographic regions [11]. The potential influence of differences in clinical, demographic, and geographic factors across studies on the relative efficacy and safety of the NOAC, as well as the clinical relevance of these differences, has not been thoroughly analysed. For that purpose, we systematically reviewed the data from randomised controlled trials with the NOAC for prevention of stroke and SEE in patients with NVAF. To maximize the real-world applicability of our results to relevant subgroups, we considered randomised phase III controlled trials with available subgroup analyses comparing any of the approved NOAC (i.e., rivaroxaban, dabigatran, and apixaban) with warfarin for prevention of stroke and SEE in patients with NVAF. At least one of the daily doses tested in the experimental arms had to correspond to the approved regime for the NOAC: dabigatran 150 or 110 mg twice daily (BID), rivaroxaban 20 mg once daily (OD), and apixaban 5 mg BID. At least one of the drug control groups had to correspond to warfarin, dose adjusted to achieve a target INR of 2 to 3.

Data Extraction. Two investigators (Antonio Gómez-
Outes and Ana Isabel Terleira-Fernández) independently and separately assessed trials for eligibility and extracted data. If a trial was covered in more than one report we used a hierarchy of data sources: public reports from regulatory authorities (US Food and Drug Administration, European Medicines Agency), peer-reviewed articles, reports from the web based repository for results of clinical studies, and other sources.
The following study characteristics were collected: number of patients; dosage of the experimental and control groups; trial phase and design; inclusion and exclusion criteria; main efficacy and safety outcomes; main populations and period of analyses; definition of noninferiority; adjudication committees of clinical events; median length followup; and time in therapeutic range (TTR). We assessed study quality using the Jadad scale [12]. Additionally, we used the Cochrane Collaboration's tool for assessing risk of bias in randomised studies [13].
The prespecified primary efficacy outcome was nonhemorrhagic (ischemic/undefined) stroke and SEE [14]. The main safety outcome was intracranial bleeding (ICB; composite of hemorrhagic stroke, subdural, subarachnoid, or epidural hemorrhage) [14]. A net clinical outcome was included as the composite of all strokes and SEE, which was consistent with the primary efficacy outcome in individual trials and in previous publications assessing the net clinical benefit of anticoagulant therapy [14]. A subanalysis of all strokes and SEE occurring after study drug discontinuation (i.e., temporary interruptions, permanent discontinuation, and after end of study) was also conducted. Other secondary outcomes included the components of the main efficacy and safety endpoints as well as major bleeding, major gastrointestinal bleeding, and mortality.
We collected outcome data on the overall trial populations as well as on the following 12 relevant subgroups: geographic region (Europe or other regions), prior stroke or transient ischemic attack (yes or no), quality of warfarin therapy (TTR ≥ 65% or <65%), degree of thromboembolic risk by applying the CHADS 2 score (≥2 or <2 score points), age (≥75 yrs or <75 yrs), gender (male or female), diabetes (yes or no), heart failure (yes or no), renal function (creatinine clearance ≥50 mL/min or <50 mL/min), type of AF (permanent/persistent or paroxysmal), prior use of VKA (yes or no), and concomitant use of acetylsalicylic acid (yes or no). Finally, we contacted sponsors or the main investigators for missing outcome data of interest.

Statistical Analysis.
We carried out direct comparisons between the NOAC and warfarin on an intention-to-treat basis, according to PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) recommendations [15].
For the meta-analysis we calculated relative and absolute risks and their respective 95% confidence intervals (CI) for each study and for the pooled studies. Heterogeneity was assessed using the Cochran test [16] and the Higgins 2 test [17]. A Cochran's < 0.10 and 2 > 50% were considered to show significant heterogeneity between studies or subgroups [17]. We used the random effects model described by DerSimonian and Laird for the main analysis [18]. We carried out subgroup analyses according to relevant clinical characteristics (see "Outcome measures"). Pooled absolute risk estimates (obtained using risk difference meta-analysis) were annualized taking into account median followup during the studies. We also calculated the number needed to treat to benefit (NNTB) or to harm (NNTH) (as the inverse of the absolute risk difference) and corresponding confidence intervals for main outcomes [19]. We performed sensitivity analyses using the fixed effects method described by Mantel and Haenszel [20] and including only studies at low risk of bias. Baseline characteristics were compared across studies using Chi-square test for categorical variables and one-way analysis of variance (ANOVA) for continuous variables. Statistical calculations were done using the RevMan statistical software, version 5.1 (Nordic Cochrane Center) [21] 2 pharmacokinetic/pharmacodynamic studies 2 phase II trials 3 subanalyses not relevant to the meta-analysis 2 publications of a phase III trial that did not include a warfarin control group

Descriptive Analysis.
The literature search identified 1561 articles, 27 of which related to clinical trials or protocols with rivaroxaban, dabigatran, or apixaban in AF ( Figure 1) and were selected for checking as full text. Three articles of phase III clinical trials with dabigatran (RE-LY study) [22], rivaroxaban (ROCKET-AF study) [23], and apixaban (ARISTOTLE study) [24] and their corresponding protocols [25][26][27] were eligible for inclusion. We also included 11 subanalyses of these trials that were considered relevant for the meta-analysis [28][29][30][31][32][33][34][35][36][37][38] and 1 article corresponding to update of events [39] reported in the RE-LY study [22]. The remaining 9 articles did not meet inclusion criteria and were excluded [40][41][42][43][44][45][46][47][48]. We also identified five public reports from the US Food and Drug Administration website [49][50][51][52][53] that included supplementary data of the RE-LY, ROCKET-AF, and ARISTOTLE studies, as well as one additional analysis of posttreatment events in the ROCKET-AF study [54] and one additional subanalysis of the ARISTOTLE study [55]. Table 1 shows the characteristics of the trials and treatments. The 3 studies comprised 50,578 patients and compared dabigatran ( = 12,091) [22], rivaroxaban ( = 7131) [23], or apixaban ( = 9120) [24] with warfarin ( = 22,236) [22][23][24]. In RE-LY, dabigatran was administered at fixed twice-daily doses of 150 mg or 110 mg [22]. On the contrary, in ROCKET-AF, the rivaroxaban 20 mg OD dose had to be down-titrated to 15 mg in patients with moderate renal impairment [23], while in ARISTOTLE, the apixaban 5 mg BID dose had to be down-titrated to 2.5 mg BID for patients with two or more of the following criteria: age ≥ 80 years, body weight ≤ 60 kg, and/or serum creatinine level ≥ 1.5 mg per decilitre [24]. RE-LY scored 3 points on the Jadad scale and was considered to be at moderate risk of bias because it was an open study [22]. ROCKET-AF and ARISTOTLE were double-blind trials [23,24], scoring 5 points on the Jadad scale, and were judged to be at low risk of bias. The risk-of-bias assessment following Cochrane recommendations [13] showed similar results, with RE-LY [22] considered to be at unclear risk of bias and ROCKET-AF [23] and ARISTOTLE [24] judged to be at low risk of bias (see supplementary Table A1). Table 2 shows the characteristics of patients and events rates in the warfarin control group in each study. Median age ranged between 70 and 73 years across trials, with a predominance of male sex. However, other characteristics, like percentage of patients with CHADS ≥ 2, prior stroke/transient ischemic attack, congestive heart failure, age > 75 years, type of AF, prior use of VKA or acetylsalicylic acid, and percent population enrolled in Europe, were widely heterogeneous. Demographic data were also compared for patients with prior stroke or transient ischemic attack (see supplementary Table A2) and a similar heterogeneity was found.

Primary Efficacy
Outcome. The NOAC were not more effective than warfarin in preventing nonhemorrhagic stroke and SEE in the overall study populations (RR = 0.93; CI 0.83 to 1.04) ( Figure 2). However, subgroup analyses suggest a trend towards superiority of the NOAC in centres with TTR <65% (Figure 2). Approximately 92% of the events were nonhemorrhagic strokes and only 8% were SEE. The separate 4 Thrombosis Table 1: Characteristics of the studies and treatments.

All Strokes and Systemic Embolic Events (Intention-to-Treat).
In the overall study populations, the NOAC afforded a lower relative risk of events than warfarin (RR = 0.82; CI 0.74 to 0.91) ( Figure 4). However, there were significant geographic differences. Within Europe, the NOAC did not significantly reduce the rates of events compared with warfarin, while outside of Europe the NOAC appeared superior in this regard ( Figure 4). The full subgroup analyses of all strokes or SEE are included in supplementary Figure A3.

All Strokes and Systemic Embolic Events after Treatment.
There was no statistically significant difference between the NOAC and warfarin in the risk of stroke and SEE after temporary interruptions or permanent discontinuations ( Figure 5). However, after end of study, patients who were transitioned from the NOAC to warfarin experienced significantly more events within the first 30 days of transition than patients on the warfarin group (RR = 3.87; CI 2.00 to 7.51) ( Figure 5). The vast majority of events corresponded to ischemic/undefined stroke.

Major Bleeding and Deaths.
There was a trend towards lower rates of major bleeding with the NOAC in comparison with warfarin in RE-LY and ARISTOTLE and neutral in ROCKET-AF ( Figure 6). The trend towards reduction in major bleeding rates was homogeneous in patients with previous stroke or transient ischemic attack (RR = 0.87; CI 0.76 to 1.00) and in patients recruited in Europe (RR = 0.86; CI 0.75 to 0.99) ( Figure 6). Between-subgroups significant  heterogeneity was shown depending on TTR ( = 0.09; 2 = 66%), with better results of the NOAC on major bleeding in centres with TTR < 65% than in centres with TTR ≥ 65% ( Figure 6). The full subgroup analyses of major bleeding are included in supplementary Figure A4.
There was a trend towards higher rates of major gastrointestinal bleeding with the NOAC in comparison with warfarin in RE-LY and ROCKET-AF and no increase in ARISTOTLE (Figure 7). Analysing the site of bleeding, rivaroxaban showed a trend towards increased risk of major upper gastrointestinal bleeding and major lower gastrointestinal bleeding, while dabigatran increased the risk of major lower gastrointestinal bleeding (Figure 7).

Absolute Difference in
The numbers of all strokes and SEE avoided per 1000 patients per year were highly dependent on the geographic region and quality of warfarin therapy, with no significant differences in European patients or in centres with TTR ≥65%. The absolute risk reduction in major bleedings was mainly observed in centres with TTR <65% (−7.1; CI −12.7 to −1.6) ( Table 4).
The absolute risk difference in all-cause death per 1000 patients treated per year (−3.90; CI −6.34 to −1.40) was highly dependent on quality of warfarin therapy. No significant differences between the NOAC and warfarin were seen in centres with TTR ≥65% (−0.9; −4.8 to 2.9), while a significant absolute risk reduction was achieved in centres with TTR <65% (−6.9; CI −11.2 to −2.6).
3.6. Sensitivity Analyses. Sensitivity analyses using the fixed effects model and including only studies at low risk of bias Test for subgroup differences: P = 0.34; I 2 = 0% Test for subgroup differences: P = 0.74; I 2 = 0%   were consistent with the main analysis (see supplementary  Table A3).

Role of Funding.
All studies were sponsored by pharmaceutical companies. In all cases, the sponsor was involved in the study design and oversight with the collaboration of a research institute and a scientific committee. In RE-LY [22], the Population Health Research Institute (Hamilton, ON, Canada) independently managed the database and performed the primary data analyses. In ROCKET-AF [23], the Duke Clinical Research Institute (Durham, NC, USA) coordinated the trial, managed the database, and performed the primary analyses independently of the sponsors. In ARISTOTLE [24], the primary analyses were performed both at Bristol-Myers Squibb and at the Duke Clinical Research Institute. At least one of the authors of the publications was employee of the sponsor.

Discussion
This systematic review, comprising more than fifty thousand patients enrolled in 3 randomised clinical trials, is to our knowledge the first systematic attempt to assess separately the efficacy of the NOAC in preventing thromboembolic events (nonhemorrhagic stroke and SEE) and major prohemorrhagic effects (ICB) [14] in NVAF. The data indicate that the NOAC have a generally similar efficacy than warfarin in the prevention of nonhemorrhagic stroke and SEE. This efficacy endpoint, which does not include hemorrhagic stroke, differs from the main outcome chosen for pivotal trials with NOAC in NVAF, which has been a net clinical endpoint including all strokes (ischemic, hemorrhagic, or undefined/unknown type) and SEE [22][23][24]. Our results are not inconsistent with the primary efficacy analyses of the respective studies but clearly suggest that, in the overall population, the weight of the effect tends to rely on the reduction of ICB, rather than on the antithromboembolic effect. Two relevant meta-analyses of the NOAC in NVAF have been recently published [56,57]. These meta-analyses showed an overall clinical benefit of the NOAC versus warfarin in NVAF, which is consistent with the results of our metaanalysis regarding the net clinical endpoint of all strokes and SEE. With respect to the assessment of efficacy, there are some methodological differences between our meta-analysis and those conducted by Dentali et al. [56] and Miller et al. [57]. We analysed the composite of ischemic/undefined strokes and SEE, while the other meta-analyses only included ischemic strokes [56] or ischemic/undefined strokes [57]. Notwithstanding, the meta-analysis by Dentali  Test for subgroup differences: P = 0.99; I 2 = 0% Test for subgroup differences: P = 0.09; I 2 = 66.0% Heterogeneity: 2 = 0.05; 2 = 18.34, df = 2 (P = 0.0001); I 2 = 89%   ischemic stroke (RR = 0.92; CI 0.81 to 1.04) [56], which is consistent with the similar efficacy in preventing nonhaemorrhagic stroke and SEE found in our meta-analysis. On the other hand, the meta-analysis by Miller et al. [57] included the dabigatran 150 mg BID dose but excluded the dabigatran 110 mg BID dose, as the meta-analysis was conducted from a US perspective and the dabigatran 110 BID dose is not currently approved in the US for use in NVAF. This issue could result in an overestimation of the efficacy of the NOAC versus warfarin in preventing ischemic/undefined strokes in their meta-analysis (RR = 0.87; CI 0.77 to 0.99) [57]. On the contrary, we included both dabigatran doses (150 mg BID and 110 mg BID), because both are already approved in Europe and many other regions for use in NVAF. The long-term extension of the RE-LY study has shown no differences in efficacy between the high and low dabigatran dose in the long term [58], which further supports our decision of including both dabigatran doses in the meta-analysis.
To the best of our knowledge, our systematic review is the first one that analyses specific subgroups and gives absolute risks estimates, thus providing a clear picture about the absolute benefit in efficacy or safety that may be expected with the NOAC in the heterogeneous population of patients with AF.
Although this review shows that the overall net clinical benefit of the NOAC versus warfarin is favourable, the magnitude of such benefit may be however influenced by a number of factors, as suggested by subgroup analyses. In RE-LY and ARISTOTLE, superiority in the composite of all strokes and SEE was mainly gained at expenses of events that occurred in non-European countries (e.g., South America, Asia, and Africa), while all the NOAC were consistently not superior to warfarin in Europe. In the ROCKET study, with a higher proportion of European patients, these differences were not apparent. It is hard to believe that geography itself influences treatment effect, but it may influence the way patients are managed in clinical practice [59,60]. Potential interaction factors accounting for geographic differences may comprise the quality of oral anticoagulation and control of associated risk factors for thrombosis (e.g., hypertension, diabetes, and heart failure). The benefit of oral anticoagulation is largely dependent on the quality of INR control achieved by centers and countries as measured by TTR [61,62]. The use of centerbased TTRs as a proxy for individual-level INR control is a matter of controversy, but it may be considered a reasonable approach in clinical trials comparing the NOAC and warfarin in AF [28]. Individual-level comparisons between the NOAC and warfarin would increase the relevance of the results to decision-making, but these comparisons are very difficult to conduct given that there were no comparable INRs in the treatment arms with the NOAC. While the understanding of the determinants of individual TTR remains incomplete, it is clear that the providers of care, and the systems within which they work, have a profound effect on the quality of anticoagulation [63]. Beyond statistical significance, subgroup analyses suggest that the net benefit of the NOAC seems better than

Total (95% CI)
Test for subgroup differences: P = 0.81; I 2 = 0% that of warfarin in situations in which quality of oral anticoagulation is poor, given that thromboembolic complications, major bleeding, and mortality may be decreased, as well as in patients with prior stroke or transient ischemic attack, particularly if they have concomitant predictive factors for developing ICB (e.g., leukoaraiosis) [64], given that the absolute risk reduction in ICB may be significant. Switching between anticoagulants may be sometimes clinically indicated and have deleterious consequences if not properly planned. In ROCKET-AF and ARISTOTLE, an excess in (mainly ischemic) strokes occurred in the rivaroxaban and apixaban arms upon discontinuation at the end of the trial. This resumption of events was probably related to inadequate control of anticoagulation, but induction of a hypercoagulable state by long-term treatment with the NOAC has not been ruled out [51].
The present systematic review has limitations. The main efficacy outcome in our study (nonhemorrhagic stroke and SEE) was a part of the main net clinical outcome (all strokes and SEE) in individual studies. In addition, the 14 Thrombosis  main safety outcome in our study (ICB) was a secondary safety endpoint in individual studies. On the other hand, we conducted subgroup analyses, which have well-known limitations [65]. Testing multiple subgroups, even though prespecified, creates the possibility of false-positive findings. However, when subgroups are described in the protocol of the original trials along with a stated hypothesis, these secondary analyses may be used to illustrate applicability across patient subgroups [65,66]. RE-LY, ROCKET-AF, and ARISTOTLE trials included a heterogeneous population (Table 2), and in such a situation, subgroup analyses are reasonable. In addition, most subgroups included a significant number of patients (e.g., 14,527 had prior stroke/TIA and 21,695 were recruited in European centers; Table 2) thus having sufficient statistical power to detect clinically meaningful differences between treatments. At the time of translating the results from these clinical trials into practise, some additional considerations are necessary. On the one hand, thrombotic and haemorrhagic events in the real-world anticoagulated AF population are higher than those reported in clinical trials [67,68], probably due to the strict selection of population and close followup applied in clinical trials. In particular, patients aged 75 years or older were underrepresented in clinical trials (range: 31% to 43% of patients) compared with real-world AF cohorts (range: 47% to 64% of patients) [69]. This issue may have important implications in bleeding risk, as renal function declines with age and all NOAC undergo renal elimination to a greater or lesser extent. Postmarketing reports of serious bleedings have frequently involved patients generally not qualified for the NOAC (i.e., severe renal insufficiency) [70]. This finding, accompanied by the current unavailability of specific antidotes, emphasizes the need for their appropriate use according to product labelling in order to minimise bleeding risk [8][9][10].
Finally, there is a need for strategies that could optimize anticoagulation quality and improve clinical outcomes in AF [71]. Beyond the use of NOAC in selected patients, these strategies may include systems facilitating algorithm-based warfarin dosing in the anticoagulation clinics [72], as well as the use of home-monitoring and self-management of anticoagulation with VKA in suitable candidates [73].

Conclusion
The NOAC seem no more effective than warfarin in preventing nonhemorrhagic stroke and SEE in NVAF. However, they are generally associated with a lower risk of ICB than warfarin. The net benefit of the NOAC seems better than that of warfarin in situations in which quality of oral anticoagulation is poor, given that thromboembolic complications, major bleeding, and mortality may be decreased, as well as patients with prior stroke or transient ischemic attack, as the absolute risk reduction in ICB may be particularly significant. However, the absolute benefit of the NOAC tends to be of a lesser magnitude in Europe than in other regions, which might be due to regional differences in quality of oral anticoagulation and overall management of associated risk factors for thrombosis. These findings would deserve further investigation.