Summary of main results

Deaths attributed to vascular causes and total ischaemic events (including ischaemic stroke/transient ischaemic attacks (TIAs), non‐fatal systemic embolic events (SEE) and non‐fatal myocardial infarction (MI)) were not significantly different between direct thrombin inhibitors (DTIs) and warfarin. Importantly, the analysis of individual doses showed that dabigatran 150 mg was superior to warfarin for this same endpoint.

Bleeding events ‐ namely fatal and non‐fatal haemorrhages including haemorrhagic strokes ‐ were significantly lower with the DTIs. The overall estimate was mainly influenced by ximelagatran and by the dose of dabigatran 110 mg twice daily.

Other outcomes evaluated including adverse events that led to discontinuation of treatment were significantly more frequent in the DTI group. Deaths from all causes were comparable between the DTIs and warfarin.

Overall completeness and applicability of evidence

Several aspects of this review attest to its completeness: (1) we conducted a thorough search and included data from published and unpublished studies in the analysis; (2) RE‐LY 2009 was complemented with its update publication by Conolly 2010 (included in RE‐LY 2009 references) and by the briefing document prepared by Boehringer Ingelheim Pharmaceuticals for the FDA (Boehringer Ingelheim 2010 ‐ included in RE‐LY 2009 references), allowing us to break down composite endpoints accurately with minimal assumptions.

The assessment of the external validity of our results should take into consideration the following.

1. The eight studies that met the eligibility criteria were considered adequate to address the main objectives of the review. However, studies varied in important aspects: (a) design; (b) number of randomised participants; (c) duration of study, and (d) definitions of primary and secondary outcomes.

2. All studies were conducted at multiple sites and included participants from the US and Canada, Europe, Asia and Australia. However, fewer black people and Hispanics were enrolled compared to white people and Asians.

3. Populations were broadly similar for important participant characteristics relevant to the treatment in evaluation (age, stroke severity, type of atrial fibrillation (AF), exclusion of people with marked renal impairment) but they varied in other characteristics such as percentage of vitamin K antagonist (VKA)‐naïve participants included or concomitant antiplatelet therapy.

4. Time in the therapeutic range (TTR) varied across studies and across different sites within a given study. However, mean TTR was compatible with the one seen in sites with good control of warfarin therapy in everyday clinical practice (Baker 2009), thus avoiding underestimation of the benefits of the standard therapy.

Although challenging for combined analysis, the clinical heterogeneity across studies better supports the external validity of the results as this variation in participant populations is more likely to reflect practice in the real world.

Quality of the evidence

We included eight RCTs involving 27,623 participants and three DTIs. Studies were heterogeneous in their individual quality assessment; see Figure 4 and Figure 5. They also varied in design (phase II and phase III), primary and secondary analyses (intention‐to‐treat (ITT) versus per protocol (PP)) and displayed some diversity in the reporting of outcomes. Nonetheless, all eight studies were judged to be adequate to address the main objectives of the review.

SPORTIF V 2005 and RE‐LY 2009 analysed all randomised participants according to the ITT principle for primary outcomes. SPORTIF II 2003 and SPORTIF III 2003 describe an ITT analysis but three participants from each study were randomised and not analysed. Overall, 16 participants were randomised but not analysed. The remaining four studies performed a PP analysis only including participants who took at least one dose of study treatment. Secondary outcomes from SPORTIF III 2003 and SPORTIF V 2005 included in the safety analysis were based on an on‐treatment (OT) population that allowed a maximum continuous interruption of 30 days (or up to 60 days total) without study medication. Primary safety analyses in RE‐LY 2009 used the randomised set (ITT), and sensitivity analyses using the safety dataset (OT) were performed.

Potential biases in the review process

The risk of having introduced significant bias into the review process is small owing to the fact that all three review authors independently performed study selection, data abstraction and assessment of risk of bias, and cross‐checked them for reproducibility. Added to this, we carried out a comprehensive electronic search complemented with various other sources to minimise overall assumptions.

A limitation of our review is the lack of sufficient data to perform head‐to‐head comparisons between different DTIs. Combining all of them may introduce bias to both the efficacy and safety analyses, as we observed important differences between drugs that were not explained by drug class, such as the hepatotoxicity seen with ximelagatran that was not seen with the other DTIs, and some unexplained variations in bleeding effects across studies.

Another limitation is the relatively short evaluation periods. Anticoagulation in the AF population is intended to be long‐term, and there is therefore a chance that additional adverse events arise with longer periods of treatment. Moreover, people with AF tend to be older and to have more comorbidities and polypharmacy, which increases their risk of developing adverse events.

Importantly, a single study (RE‐LY 2009) represented 75% of the population analysed in our primary results. The dominant effect that this large study has on the overall estimates should be considered in the interpretation of results. Important differences could be present but undetected, given the smaller samples of the other studies.

Finally, we performed several post hoc sensitivity analyses, which may appear to introduce bias. However, the rationale for each analysis is extensively discussed and the findings clearly help in the interpretation of the protocol‐driven composite endpoints.

Publication bias must be considered due to the asymmetry of funnel plots and to the fact that no studies were found published in a language other than English.

Agreements and disagreements with other studies or reviews

DTIs represent a major advance in stroke prevention in people with non‐valvular AF, as results from our analyses indicate a trend towards similar ‐ and perhaps superior ‐ effectiveness and increased safety. These results are in agreement with the findings from other recently published meta‐analyses: Capodanno 2012, Lip 2012 and Miller 2012. However, these meta‐analyses included drug classes other than DTIs, such as Factor Xa inhibitors, which are beyond the scope of our review.

When comparing our results with those from individual trials, it is important to consider that our primary outcomes were constructed in a different way from those from the original studies. The most relevant difference is that we excluded haemorrhagic stroke from our primary efficacy outcome. The rationale behind this exclusion is that AF does not cause haemorrhagic stroke itself, but anticoagulants do. Moreover, haemorrhagic stroke is not prevented by anticoagulants but rather is associated with their use. Therefore, haemorrhagic stroke belongs more appropriately to the primary safety outcome. The introduction of haemorrhagic stroke in the primary safety endpoint rather than in the primary efficacy endpoint does not undermine its critical clinical relevance. It merely places it where it more appropriately belongs.

For our composite primary efficacy endpoint of vascular deaths and ischaemic events, we found no difference between DTIs and warfarin. However, the sensitivity analysis for individual doses revealed that dabigatran 150 mg twice daily was superior to warfarin for this same endpoint. This is a concordance with the results from RE‐LY 2009 where the higher dose of dabigatran was found to be superior to warfarin for their primary efficacy outcome of stroke and SEE as well as for their outcome of vascular deaths, particularly those due to strokes. Importantly, our analysis shows that the benefits of dabigatran over warfarin are maintained even after removing the influence of including haemorrhagic strokes in the comparison.

The inclusion of vascular deaths in our primary efficacy endpoint is also clinically relevant because death may act as a competing risk factor for stroke/SEE (Schatzkin 1989). In a geriatric population with considerable comorbidities, the competing risk of death is especially high (Berry 2010). The inclusion of death in a composite endpoint is one way of addressing competing risks.

The efficacy endpoint in our review also included MI. The inclusion of this outcome in the composite endpoint can be controversial as it can be argued that the pathophysiology of MI differs from that of stroke and other embolic events. However, we deemed this inclusion appropriate based on its clinical relevance and the fact that a recent meta‐analysis of seven trials that evaluated dabigatran for various indications found a higher risk of MI compared with warfarin (Uchino 2012). Also, in the RE‐LY 2009 trial, the frequency of MI was greater in participants receiving dabigatran than in those treated with warfarin. We performed a sensitivity analysis excluding MI to assess the impact of having included this outcome in the comparison. As expected, the results tended to favour the DTIs but without reaching statistical significance.

We also considered pulmonary embolism (PE) and deep vein thrombosis (DVT) in our primary analysis as systemic embolic events. The consideration was based on the fact that systemic embolism was defined in the largest included study as an acute non‐intracerebral and non‐coronary vascular event. Although the pathophysiology of arterial and venous thrombosis differs in many aspects, the inclusion of these events in the composite outcome seems reasonable per study definitions. RE‐LY 2009 showed slightly higher rates of PE events in both dabigatran groups compared with warfarin but the overall number of these events was low. Both MI and PE were considered as efficacy outcomes in RE‐LY 2009 but the authors did not include either of them in the primary efficacy analysis. This is important because the efficacy of the DTI was comparable to that of warfarin in our analysis even after the introduction of MI and PE in the composite endpoint.

Our safety analysis was designed to evaluate fatal and non‐fatal haemorraghic events referring to major bleeds rather than overall bleeding events, because major bleeds carry important clinical consequences relevant to health professionals when deciding whether or not to use an anticoagulant. A reduction in major bleeding is of particular clinical relevance, as there is increasing evidence to support a significant association between a major haemorrhagic event and adverse prognosis as indicated by an increased risk of death in people with AF (Sharma 2012). In our safety analysis, DTIs were found to be superior to warfarin with fewer fatal and non‐fatal haemorraghic events including haemorrhagic strokes. The sensitivity analyses showed that this favourable result was mainly influenced by ximelagatran and dabigatran at the dose of 110 mg twice daily. Interestingly, although numerically fewer major bleeding events were seen in all three SPORTIF trials, none of the individual studies found a statistically significant difference between the DTI and warfarin. In contrast, RE‐LY 2009 did report a significant reduction in the occurrence of major bleeding events with both doses of dabigatran compared with warfarin.

In the evaluation of new drugs, the assessment of adverse events is fundamental. However, the possibility of researcher bias needs to be considered in this analysis. Researcher bias refers to investigators having more stringent discontinuation criteria for the new treatment in comparison with a standard drug with which they have more experience and feel more comfortable. Since warfarin was given in an open‐label fashion in most studies, this bias could be important. Our analysis included only events that led to treatment discontinuation because participant compliance with long‐term treatment is of critical importance in dealing with chronic conditions like stroke prevention in AF. Our analysis showed that the adverse events leading to treatment discontinuation occurred more frequently with the DTIs compared with warfarin. The most frequent adverse events were gastrointestinal complaints. Interestingly, the analysis of individual drugs showed that adverse events leading to treatment discontinuation were significantly more frequent with dabigatran than with AZD0837.

Finally, we found all‐cause mortality to be comparable between DTIs and VKAs. The results are in agreement with RE‐LY 2009, where fewer overall deaths were observed in both dabigatran groups compared with the warfarin group, although this difference did not reach statistical significance.