Note from the National Guideline Clearinghouse (NGC): The National Institute for Health and Clinical Excellence (NICE) commissioned an independent academic centre to perform a systematic literature review on the technology considered in this appraisal and prepare an assessment report. The Evidence Review Group (ERG) report for this technology appraisal was prepared by Biomedical Journals Technology Assessment Group (BMJ-TAG) (see the "Availability of Companion Documents" field).
The manufacturer assessed the ARISTOTLE and AVERROES trials against criteria adapted from guidance for undertaking reviews in health care issued by the Centre for Reviews and Dissemination, as provided in the NICE template for manufacturer/sponsor submission of evidence to the single technology appraisal (STA) process. The ERG independently validated ARISTOTLE and AVERROES, and agrees with the manufacturer's assessments. The ERG considers both ARISTOTLE and AVERROES to be well-designed randomised controlled trials (RCTs). The ERG acknowledges that there was an imbalance in treatment discontinuations between the treatment groups in each of the RCTs (fewer discontinuations with apixaban in both RCTs) and notes that the manufacturer has used an intention to treat (ITT) analysis to report the results of each study. The ERG considers the manufacturer's approach to data analysis to be appropriate.
Summary and Critique of Submitted Clinical Effectiveness Evidence
The manufacturer presents data for two RCTs, ARISTOTLE and AVERROES, in the clinical effectiveness section of the manufacturer's submission (MS). The ERG considers that only ARISTOTLE met the inclusion criteria for this STA based on the final scope issued by NICE; only ARISTOTLE will be discussed in further detail below. Details and results of AVERROES are presented in Appendix 9.3 of the ERG report (see the "Availability of Companion Documents" field).
Description of ARISTOTLE Trial
ARISTOTLE was an international, multicentre, randomised double-blind phase III trial comparing the clinical efficacy and safety of apixaban with warfarin (vitamin K antagonist, VKA). The patient population of ARISTOTLE had atrial fibrillation (AF) and at least one additional risk factor for stroke. The primary objective of ARISTOTLE was to determine whether apixaban was non-inferior to warfarin (international normalised ratio [INR] target range 2.0–3.0) for the combined end point of stroke and systemic embolism (SE).
See Section 4 of the ERG report for more information on the ARISTOTLE trial; including statistical approaches, summary statement and summary of results.
Description and Critique of Network Meta-analysis
As a result of the absence of head-to-head trials, the manufacturer carried out two network meta-analyses (NMAs). The aim of these NMAs was to compare apixaban with dabigatran and rivaroxaban, as specified in the final scope issued by NICE. NMA 1 consisted of patients suitable for treatment with warfarin and compared apixaban, warfarin, dabigatran and rivaroxaban. NMA 2 was reported in the MS to be in a population of patients unsuitable for vitamin K antagonist (VKA) therapy and compared apixaban, dabigatran, rivaroxaban and aspirin. In the MS, the manufacturer stated that although aspirin was not included in the NICE final scope it remained a relevant comparator in VKA unsuitable patients. The manufacturer’s rationale for this was that aspirin is recommended in clinical guideline 36 (CG36) for patients at moderate to high risk of stroke or SE who are unsuitable for VKA therapy. In addition, the manufacturer stated that aspirin is widely used in clinical practice in England and Wales.
A total of three RCTs were identified as suitable for inclusion in NMA 1; these are detailed in Section 4.4.3 of the ERG report.
For each outcome of interest, event rates from each RCT were used in the NMA to calculate hazard ratios (HRs). The event rate was defined as the total number of events across all patients, divided by the total patient-years exposed. The ERG notes that if the event rate was not reported in published sources the manufacturer estimated event rates from the number of patients experiencing an outcome. The ERG is unable to provide any comment on the likely impact of this calculation as the true event rates are unknown. It is thus impossible to establish whether the calculated event rate could be over or under estimating the true event rate. However, the manufacturer reported that this method of calculating event rates accurately predicted event rates in the studies where data were available for both event rates and the number of patients with each event.
The manufacturer used a Bayesian Markov Chain Monte Carlo simulation in WinBUGS to conduct the NMA. In the MS, the manufacturer reported that both fixed and random effects models were fitted to the data; the model with the best fit was chosen for the reporting of the results. Model fit was determined using the deviance information criterion (DIC) and residual deviance for each outcome assessed. The manufacturer reported that there was little difference in model fit between the fixed and random effects models and all outcomes were reported using a fixed effects model. The manufacturer's rationale for the use of a fixed rather than random effects model was built around the small number of studies in the network (three studies). The manufacturer considered that as a result of the small number of included studies, a random effects model would produce poor estimates of the variation in between-study treatment effects. The ERG considers the manufacturer’s use of a fixed effects model to be a reasonable choice given the limited data set.
See Section 4 of the ERG report for more information on clinical effectiveness analysis.
The manufacturer constructed a Markov model to evaluate the long- and medium-term consequences of apixaban for the prevention of stroke or SE in AF patients. The model captured the clinical and economic consequences of treatment by modelling the movement of patients between discrete health states over a lifetime time horizon.
The Markov model submitted by the manufacturer consisted of 18 health states, including the absorbing state of death. Patients transitioned between health states in cycles of 6 weeks with only one clinical event permitted per cycle. The ERG notes that, given the influence of individual patient characteristics on outcomes in AF, a discreet event simulation rather than a Markov cohort modelling approach may have been more appropriate. However, the ERG acknowledges that a well-constructed Markov model may be sufficient to capture the mean differences in costs and consequences associated with prophylactic treatments in AF. Furthermore, the ERG notes that the use of Markov modelling techniques is consistent with previous novel oral anticoagulant (NOAC) HTA submissions. The model structure for patients on first-line therapy is displayed in Figure 3 of the ERG report. The model structure for patients on second-line therapy is shown in Figure 3 of the ERG report.
Model Validation and Face Validity Check
The manufacturer stated that the model was assessed for internal (verification) and external (validation) validity. Verification was carried out by two independent economists and used extreme value analysis to identify any flawed algorithms or irregularities. Validation was carried out by assessment of the face validity of the model with clinicians and comparison of the model results against published results.
In support of the apixaban submission, the manufacturer carried out probabilistic sensitivity analyses (PSAs), deterministic sensitivity analyses, and scenario analyses.
Probabilistic Sensitivity Analyses
The sensitivity of the model to parameter uncertainty while not accounted for in the base case model results has been explored in probabilistic sensitivity analyses. Parameters were assigned a probability distribution from which estimates were simultaneously sampled for 2,000 runs. Table 61 of the ERG report summarises the type of distribution used for each group of parameters considered within the sensitivity analyses; Figures 5 and 6 and Table 62 of the ERG report present the probabilistic results.
Within the deterministic sensitivity analyses, the manufacturer assessed the univariate sensitivity of the model to a total of 117 parameters. Each parameter was alternately assigned a low and high value estimated from the 95% confidence intervals associated with that parameter; where confidence intervals were not available or could not be derived, variation was assumed to be either 10% or 25% of the mean. Figures 7 to 10 of the ERG report present the deterministic sensitivity analysis results for the 13 most influential parameters in each comparison.
In total, the manufacturer carried out 19 scenario analyses (in the VKA suitable population) around various model assumptions. Table 63 of the ERG report gives the details of each scenario analysis carried out. The results of each analysis with respect to impact on the base case incremental cost-effectiveness ratio (ICER) are displayed in Figure 11 of the ERG report; impact is assessed by percentage change from the base case ICER (for apixaban versus each comparator). The majority of scenario analyses decrease the base case ICER (apixaban versus comparator), suggesting that the assumptions of the base case model were conservative (i.e., any bias was likely to be against apixaban).
See Sections 5 and 6 of the ERG report for more information on methods of cost-effectiveness analysis.