5.1. Summary of Available Evidence

In this systematic review, five unique RCTs evaluating non-inferiority of the three NOACs versus adjusted-dose warfarin were identified. Three large trials were assessed as high quality through critical appraisal (ROCKET-AF, RE-LY, ARISTOTLE), and were included in the network meta-analysis and used to support the economic evaluation. Two RCTs (PETRO, ARISTOTLE-J) were excluded from the analysis because of small sample size and lack of data for the outcomes specified a priori. No non-randomized evidence was located in the published literature.

5.2. Interpretation of Results

5.3. Knowledge Gaps

No head-to-head studies have been conducted comparing NOACs. As a consequence, there was insufficient evidence to draw definitive conclusions regarding the relative safety and efficacy between the various oral anticoagulants. Rigorously conducted, longer-term studies with larger sample sizes will be required to determine if any of the oral anticoagulants are superior to one another with regards to effectiveness and safety.

Evidence for apixaban has not been reviewed by the FDA (and not approved by Health Canada at the time of this therapeutic review). As a result, detailed data from the FDA Clinical Summary Report was not available for apixaban; only one published subgroup analyses was available.

There was also limited comparative data or subgroups are not consistently defined across the trials for: weight, impaired renal function, prior history of GI bleed, concurrent use of NSAIDs. In addition, detailed subgroup analysis data was only available for two major end points: stroke/SE and major bleeding. Studies in subpopulations, considering a variety of end points, are especially pertinent given the large budget impact of these NOACs. Patient-level network meta-analyses could be helpful to compare risk-benefit profiles of NOACs and warfarin across subpopulations.

5.4. Strengths and Limitations

This is the first systematic review to simultaneously assess the relative safety and cost-effectiveness of warfarin, dabigatran, apixaban, and rivaroxaban in patients with atrial fibrillation. The review was conducted according to Cochrane Collaboration guidelines using standardized, reproducible methods for the identification of evidence, data abstraction, quality assessment, and analysis. Both direct and indirect evidence was synthesized using standard evidence synthesis methodologies, and considering both a Bayesian and frequentist approach to the models. A comprehensive economic evaluation was conducted using available cost data and the results of the network meta-analyses. Other strengths of this analysis were its comprehensiveness regarding the oral anticoagulants considered, the number of outcomes assessed, and the similarity of trials included in analysis regarding year of publication.

Despite the aforementioned strengths, a number of limitations related to the available evidence warrant discussion. Network meta-analysis involves pooling of trials. To avoid the introduction of bias, it is imperative that clinical and methodological variation across studies is minimized. If variability does exist, the assessment of its effects on network meta-analysis results is required. We observed variability in study and subject characteristics that may be important predictors of treatment effects. In particular, ROCKET-AF included higher-risk patients, with a CHADS₂ score of at least 2. This limits the comparability of treatment effects to lower-risk patients and to results from RE-LY and ARISTOTLE, which included more patients at low risk. Patients enrolled in ROCKET-AF were also older in age and had more congestive heart failure, more diabetes mellitus, more hypertension, and more prior stroke or transient ischemic attack. ROCKET-AF also reported results using the per-protocol population for many outcomes, whereas RE-LY and ARISTOTLE used ITT.

While some may argue against conducting a network meta-analysis in light of the aforementioned clinical and methodological heterogeneity, the reality is that decisions in the clinical setting and at the policy level have to be made despite these issues. In the absence of a formal indirect comparison or network meta-analysis, health professionals and decisions-makers may make informal comparisons between oral anticoagulants without understanding or acknowledgment of clinical and methodological variation across studies. That being said, a formal network meta-analysis where clinical and methodological variation across studies was well-described, and where formal adjustments were made attempting to resolve some of the heterogeneity across studies, was justified. To address clinical and methodological heterogeneity, we reported results for ROCKET-AF for both intentions to treat, when available. However, this information was not consistently reported in ROCKET-AF. We also performed subgroup analyses where we stratified results across studies by CHADS₂, age, and centre TTR. While these subgroup analyses address some of the limitations, issues regerading heterogeneity are only partly resolved. We do not have information regarding the similarity of patient populations for individual subgroups considered, as randomization is broken. Consequently, whereas some subpopulations (e.g., CHADS₂ ≥ 2) may be more similar across studies in subgroup analyses relative to reference case analyses, heterogeneity may still remain. Furthermore, we did not have data to conduct multiway subgroup analyses (e.g., CHADS₂ ≥ 2 and age ≥ 75 and cTTR ≥ 65.5%). In light of the data, therefore, it is difficult to ascertain subpopulations where warfarin use may be appreciably more or less effective, although data from individual one-way subgroup analyses allude to a less favourable risk-benefit profile for newer anticoagulants in patients who are older and well-controlled on warfarin. Given the budget impact of NOACs, patient-level network meta-regression analyses are desperately needed to compare the risk-benefit profiles of NOACs and dose-adjusted warfarin across subpopulations.

While ROCKET-AF and ARISTOTLE were designed as double-blind trials, treatment was not blinded in RE-LY, except for the dose of dabigatran, which might be a potential source of bias such as performance bias (i.e., systematic differences between groups in the care provided or exposure to factors other than the interventions of interest). However, we were unable to adjust for differences in blinding due to the small number of studies identified. Consequently, it is unclear what impact lack of blinding in RE-LY may have on effect estimates, although the effect may not be substantive given the nature of the end points considered in RE-LY (e.g., stroke/SE, major bleeding).

The small number of trials limited our ability to adjust for heterogeneity using other techniques such as meta-regression or sensitivity analysis (i.e., removing results from individual studies). While it was possible to expand the evidence network to include data for other treatments such as aspirin, or aspirin plus clopidogrel, and ultimately facilitate the use of more sophisticated methods to address heterogeneity, this was beyond the scope of this review. It is, however, noteworthy that expanding the evidence network to include additional information (and potentially attempt to address heterogeneity) would involve the potential introduction of heterogeneity — studies comparing aspirin with warfarin are older and may have involved different patients and treatment patterns than those enrolled in newer studies of anticoagulants.

There is a potential asymmetry of information that may introduce bias. Both RE-LY and ROCKET-AF have been reviewed by the FDA, while apixaban is currently under review. Further, several subgroup publications were available for RE-LY and ROCKET-AF, but only one publication and a conference abstract were available for ARISTOTLE. Therefore, data for dabigatran and rivaroxaban may have been more heavily scrutinized than apixaban, potentially biasing results in favour of apixaban. If data emerges over the course of FDA or Health Canada reviews suggesting that the benefits of apixaban are less favourable, then the network meta-analysis and cost-effectiveness will have to be revisited.

While our analysis considered both fixed and random effects models, again the small number of trials limited the analyses. The applicability of random effects models was compromised because vague or non-informative prior distributions exerted a large degree of influence on any inference. While we could have considered the use of more informative priors, this was beyond the purview of this review. Similarly, the random effects interaction term trial-by-treatment had to be excluded from the GLMM model because of the small number of trials. Without including random effects terms, the full variability in the data cannot be included with the results appearing more precise. Due to these limitations, the fixed effects model was associated with a better fit based on the assessment of the deviance information criterion and comparison of residual deviance to the number of unconstrained data points. However, results from the fixed effects model, particularly those for reference case analyses, do not account for heterogeneity across studies and should be interpreted in light of this issue. As a result, results from fixed effects network meta-analysis and those for study-level results are very similar — a finding attributable to the few studies included in the evidence network and lack of a “meta-analysis” for individual pairwise comparisons within the evidence network (i.e., only one study for each pairwise contrast).

The absolute risk reduction was calculated by multiplying the relative risk (and corresponding 95% confidence interval) by the point estimate of the event rate in the warfarin arm. This approach does not take into account the uncertainty around the event rate. Future studies could characterize uncertainty to a greater degree using a probability distribution of the event rate. Nevertheless, such an approach would likely not produce substantially different results given the large sample sizes in the studies and subgroups.

Not all outcomes of interest were reported, in particular for various predefined subgroups. In particular, life-threatening bleeding and the composite ischemic/uncertain stroke or SE were often not reported in studies, and CV-mortality was reported within vascular mortality. All-cause mortality, intracranial bleeding, MI, and major GI bleeding were not reported by subgroups of interest re TTR, age, and CHADS₂ score. Therefore, the complete risk-benefit profile of newer drugs versus warfarin for subpopulations is not known. Given the budget impact of newer anticoagulants, patient-level network meta-analyses are needed to compare the risk-benefit profiles of NOACs and warfarin across subpopulations.

As all major studies were performed as multinational or multicentre trials, generalizability to the Canadian health care system may be limited. Treatment of comorbidities and management of patients who are candidates for warfarin may differ between various countries. Further, TTR of warfarin treatment showed substantial differences between the trials, and was also affected by geography. In RE-LY and ARISTOTLE, INR control rates were substantially better than in ROCKET-AF. Skillful warfarin use as a predictor of treatment success might play a role in transferring trial results to the Canadian setting. In ROCKET-AF, the non-inferiority of rivaroxaban was achieved for the primary outcome stroke and SE in the intention to treat (ITT) population, mostly due to a rather high number of events in the transition phase after stopping the study drug. Thus, the magnitude of treatment effect was smaller in the ITT population. The implication of this observation for drug use in clinical routine is unclear but does indicate differences in methodology between the trials.

While outcome definitions for efficacy end points are similar throughout the included trials, definitions of bleeding events differed substantially, in particular for minor (RE-LY) or non-major clinically-relevant bleeding (ROCKET-AF and ARISTOTLE). Thus, these bleeding rates were markedly higher in RE-LY and ROCKET-AF compared to ARISTOTLE, limiting the comparability of results in network meta-analysis.

Variability in patient risk and VKA experience at baseline, notably in ROCKET-AF, could impact the trial mean TTR, as could the variations in study population assessment (ITT versus as-treated per protocol). Patients who are older or have higher rates of comorbidities could have reduced access to TTR testing. Cross-trial assessment of TTR variations is made difficult by study design, as an open-label design could allow clinicians to make adjustments in warfarin arms more frequently.

The inclusion and exclusion criteria identifying the patients’ eligibility for the studies on which the adverse events data are based may be different than the patients seen in clinical practice, leading to altered asdverse events profiles.

The limited follow-up from the clinical trials and the sensitivity of results to the duration of treatment effect leads to uncertainty regarding whether the new anticoagulants will be cost-effective in the long term. Aspirin was not included in the economic evaluation and would have been a relevant comparator for the analysis of patients with a CHADS₂ score of 0 or 1.