Apixaban for preventing stroke and systemic embolism in people with non-valvular atrial fibrillation

The main clinical effectiveness evidence for apixaban came from 2 international, multicentre, double-blind, double-dummy, placebo-controlled, randomised controlled trials, which had investigated apixaban. ARISTOTLE (n=18,201) compared apixaban (5 mg twice daily; 2.5 mg twice daily in selected patients) with warfarin (in patients with an international normalised ratio [INR] target range of 2.0 to 3.0). AVERROES (n=5,598) compared apixaban (5 mg twice daily; 2.5 mg twice daily in selected patients) with aspirin (81 mg to 324 mg once daily) in people 50 years or older with atrial fibrillation and at least 1 additional risk factor for stroke for whom treatment with warfarin had failed, or for whom warfarin was unsuitable or who were unwilling to take warfarin.

The primary objective of ARISTOTLE was to determine if apixaban was non-inferior to warfarin for the combined end point of stroke and systemic embolism. Stroke included both ischaemic stroke, caused by embolism from the heart, and haemorrhagic stroke, which can be a complication of anticoagulant treatment (although it may also occur spontaneously or as a result of secondary haemorrhage into an ischaemic stroke). ARISTOTLE included adults with atrial fibrillation or atrial flutter not resulting from a reversible cause and at least 1 additional risk factor for stroke (assessed by CHADS₂ criteria). It enrolled patients from 39 countries; 40% of participants were from Europe and this included patients from 41 sites in the UK. The average age was 69 years and 65% of the population were male. The mean time in therapeutic range for patients in the warfarin arm was 62.2%, and the median time in therapeutic range was 66%. Approximately 4% of the study population received 2.5 mg apixaban (those who had 2 or more of the following criteria: 80 years or older, a body weight of 60 kg or less, or a serum creatinine level of 1.5 mg/100 ml [133 micromole/l] or more). The mean CHADS₂ score at baseline was 2.1 and approximately 65% of patients had a CHADS₂ score of 2 or more.

In the intention-to-treat population, apixaban met non-inferiority criteria using a non-inferiority margin of 1.38, over a median follow-up of 1.8 years. Apixaban was associated with a significantly lower rate of stroke and systemic embolism than warfarin (hazard ratio [HR] 0.79, 95% confidence interval [CI] 0.66 to 0.95, p=0.01). The rate of fatal or disabling stroke was significantly lower in the apixaban group than the warfarin group (HR 0.71, 95% CI 0.54 to 0.94). When the outcomes included in the composite primary outcome (ischaemic or uncertain type, haemorrhagic stroke and systemic embolism) were analysed separately, apixaban was associated with a significant reduction in haemorrhagic stroke compared with warfarin (HR 0.51, 95% CI 0.35 to 0.75), but the decrease for apixaban compared with warfarin in ischaemic or uncertain type stroke or systemic embolism was not statistically significant (ischaemic or uncertain type stroke HR 0.92, 95% CI 0.74 to 1.13, p=0.42; systemic embolism HR 0.87, 95% CI 0.44 to 1.75, p=0.70). The rates of myocardial infarction, and pulmonary embolism or deep vein thrombosis, were lower with apixaban than warfarin, but were not statistically significant (HR 0.88, 95% CI 0.66 to 1.17, p=0.37, and HR 0.78, 95% CI 0.29 to 2.10, p=0.63 respectively). Apixaban was associated with fewer all-cause deaths than warfarin, which was of borderline statistical significance (3.52% and 3.94% respectively [HR 0.89, 95% CI 0.80 to 0.99, p=0.047]).

The manufacturer presented results for the primary efficacy outcomes for 21 pre-specified subgroups in ARISTOTLE including subgroups broken down by baseline risk of stroke or systemic embolism (grouped by CHADS₂ scores of 1 or less; 2; and 3 or more). ARISTOTLE was not statistically powered to demonstrate superiority in subgroup analyses. The hazard ratios for apixaban relative to warfarin for stroke and systemic embolism in the 3 stroke risk subgroups were consistently less than 1, but the confidence intervals of the CHADS₂ ≤1 and 2 groups crossed 1, meaning that that the difference between apixaban and warfarin was not statistically significant for these groups. The hazard ratios for stroke and systemic embolism in the groups of patients who received 5 mg and 2.5 mg apixaban were also both below 1 (the hazard ratios for CHADS₂ score subgroups and for the groups of patients who received 5 mg and 2.5 mg apixaban are commercial-in-confidence). The manufacturer also presented data for subgroups based on INR (international normalised ratio) control using quartiles of centre time in therapeutic range (less than 58.0%, 58.0% to 65.7%, 65.7% to 72.2% and more than 72.2%). A centre's time in therapeutic range was calculated as the median of individual time in therapeutic ranges among the centre's patients on warfarin. The manufacturer reported that the benefits of apixaban over warfarin in preventing stroke or systemic embolism were consistent (HR <1) regardless of INR control (centre time in therapeutic range <58.0% [HR 0.77, 95% CI 0.56 to 1.06], centre time in therapeutic range 58.0% to 65.7% [HR 0.80, 95% CI 0.56 to 1.15], centre time in therapeutic range 65.7% to 72.2% [HR 0.79, 95% CI 0.54 to 1.13], centre time in therapeutic range greater than 72.2% [HR 0.81, 95% CI 0.52 to 1.26]).

The adverse events and safety analyses were reported for the on-treatment population in ARISTOTLE (all patients who received at least 1 dose of study medication). Apixaban was superior to warfarin for the primary safety outcome of time from first dose of study drug to first occurrence of confirmed International Society on Thrombosis and Haemostasis (ISTH) major bleeding (HR 0.69, 95% CI 0.60 to 0.80; p<0.001). Apixaban resulted in significantly fewer bleeding events than warfarin for all of the major bleed types (intracranial major bleeding HR 0.42, 95% CI 0.30 to 0.58; other location major bleeding HR 0.79, 95% CI 0.68 to 0.93) and clinically relevant non-major bleeding events reported by the manufacturer apart from major gastrointestinal bleeding, for which the difference between apixaban and warfarin was not statistically significant (HR 0.89, 95% CI 0.70 to 1.15, p=0.37). There were similar proportions of patients who experienced adverse events with apixaban (81.5%) and warfarin (83.1%) and a lower proportion of patients who experienced bleeding adverse events with apixaban (25.2%) compared with warfarin (32.7%). Serious adverse events occurred in 35.0% of patients treated with apixaban and 36.5% of patients treated with warfarin. Fewer patients stopped the study drug in the apixaban group than the warfarin group (25.3% compared with 27.5% respectively, p=0.001); 7.6% of patients in the apixaban arm and 8.4% of patients in the warfarin arm stopped treatment because of an adverse event. The safety of apixaban was maintained across patients at different levels of stroke risk, regardless of warfarin control (time in therapeutic range) and in patients who needed dose reduction.

The primary objective of AVERROES was to determine if apixaban was superior to aspirin for preventing the composite outcome of stroke or systemic embolism in adults with at least 1 risk factor for stroke in whom vitamin K antagonists were unsuitable. In the intention-to-treat population apixaban reduced the rate of stroke and systemic embolism compared with aspirin over a mean follow-up of 1.1 years (HR 0.45, 95% CI 0.32 to 0.62, p<0.001). The rates of disabling or fatal stroke were also lower in patients who received apixaban compared with patients who received aspirin (HR 0.43, 95% CI 0.28 to 0.65). When considered as a separate outcome apixaban reduced the rates of ischaemic stroke compared with aspirin (HR 0.37, 95% CI 0.25 to 0.55) but did not statistically significantly reduce the rates of haemorrhagic stroke (HR 0.67, 95% CI 0.24 to 1.88, p=0.45). Apixaban was associated with a higher rate of all bleeding than aspirin (HR 1.30, 95% CI 1.10 to 1.53) and of major or clinically relevant non-major bleeding (HR 1.38, 95% CI 1.07 to 1.78). Although apixaban was associated with higher rates of major bleeding than aspirin, this was not statistically significant (HR 1.54, 95% CI 0.96 to 2.45, p=0.07).

No head-to-head data were available for apixaban compared with dabigatran etexilate (hereafter referred to as dabigatran) or rivaroxaban. The manufacturer used a Bayesian Markov chain Monte Carlo stimulation in WinBUGS to conduct 2 network meta-analyses using a fixed-effect model. The first meta-analysis included patients for whom vitamin K antagonist treatment was suitable and it compared apixaban, warfarin, dabigatran and rivaroxaban. The second meta-analysis was intended to assess a population of patients for whom vitamin K antagonists were unsuitable, comparing apixaban, dabigatran, rivaroxaban and aspirin.

The first meta-analysis included ARISTOTLE, RE-LY and ROCKET-AF trials. RE-LY compared dabigatran (150 mg and 110 mg twice daily) with warfarin. ROCKET-AF compared rivaroxaban (20 mg once daily) with warfarin. There were differences between the trials of apixaban, dabigatran and rivaroxaban: ARISTOTLE and ROCKET-AF were double-blind, double-dummy trials, whereas RE-LY was an open-label trial; the population in ROCKET-AF had a higher stroke or systemic embolism risk at baseline (baseline CHADS₂ of 3.6 [ROCKET-AF], 2.1 [ARISTOTLE], 2.1 [RE-LY]) and the mean percentage time in therapeutic range was lower in ROCKET-AF (55%) than in ARISTOTLE (62%) and RE-LY (64%). Where possible, the manufacturer used intention-to-treat data from each trial. However, the manufacturer highlighted that there was an absence of published intention-to-treat outcome data for some secondary outcomes from ROCKET-AF including fatal stroke, disabling stroke and non-disabling stroke. Therefore, data from the on-treatment population were also used. The second meta-analysis included ARISTOTLE, RE-LY, ROCKET-AF and AVERROES.

The manufacturer did not present any statistical analysis of heterogeneity but commented that potential sources of clinical heterogeneity between the trials were the differences in baseline stroke risk scores, study blinding, and whether the intention-to-treat or on-treatment populations had been used to assess efficacy and safety outcomes. Additionally, the manufacturer highlighted a statistically significant difference in myocardial infarction at baseline between treatment groups in ROCKET-AF.

The base-case results of the first meta-analysis indicated that there were no statistically significant differences between apixaban and rivaroxaban or dabigatran in the incidence of stroke, systemic embolism and all-cause mortality. The results did however suggest that apixaban was associated with a significantly lower incidence of myocardial infarction compared with dabigatran (150 mg or 110 mg twice daily). Apixaban was associated with a significantly lower incidence of all bleeding outcomes compared with rivaroxaban (intracranial haemorrhage, major bleeding, gastrointestinal bleeding, other major bleeding, clinically relevant non-major bleeding, any bleeding). Apixaban had a significantly lower incidence of all bleeding events except intracranial haemorrhage and clinically relevant non-major bleeding (which was not measured in RE-LY) than dabigatran 150 mg. Apixaban had a significantly lower incidence of any bleeding than dabigatran 110 mg. In addition, apixaban was associated with significantly fewer discontinuations compared with dabigatran 150 mg, dabigatran 110 mg and rivaroxaban. All of the hazard ratios from the first network meta-analysis are academic-in-confidence. The manufacturer reported that the results for apixaban compared with warfarin generated by the first meta-analysis were consistent with the pair-wise comparisons between warfarin and apixaban in ARISTOTLE (see sections 3.3 and 3.5).

The manufacturer conducted 2 sensitivity analyses of its first network meta-analysis. The first used data from a later publication of RE-LY (Connolly et al. 2010) rather than the RE-LY 2009 data. The results for the first sensitivity analysis were generally consistent with the base case; however, the reduction in myocardial infarction with apixaban compared with both doses of dabigatran was no longer statistically significantly different. The second sensitivity analysis used the safety on-treatment dataset from ROCKET-AF rather than the intention-to-treat data from this trial, which was also generally consistent with the base case. The hazard ratios from the sensitivity analysis of the manufacturer's first network meta-analysis are academic-in-confidence.

The manufacturer commented that there were no data for rivaroxaban or dabigatran in the population for whom warfarin was unsuitable, so data from ROCKET-AF (which assessed rivaroxaban compared with warfarin) and RE-LY (which assessed dabigatran compared with warfarin) were included, alongside ARISTOTLE and AVERROES. This meant that the second meta-analysis represented a mix of patients for whom warfarin was suitable and unsuitable ('warfarin-suitable' and '-unsuitable' populations).

The manufacturer also used data from ARISTOTLE, RE-LY and ROCKET-AF to estimate the distribution of stroke severity and bleed type associated with apixaban, warfarin, rivaroxaban and dabigatran. Mild, moderate, severe and fatal stroke were classified by modified Rankin scores, with scores of 0 to 2 classed as a mild stroke, scores of 3 to 4 classed as a moderate stroke, a score of 5 classed as a severe stroke and a score of 6 classed as a fatal stroke. Data corresponding to these modified Rankin scores were available for apixaban and warfarin from ARISTOTLE, but ROCKET-AF and RE-LY grouped stroke severity scores differently. The manufacturer therefore estimated the proportion of patients that would be expected to have scores of 3 to 4 or 5 in the group of patients reported as having a stroke with a modified Rankin score of 3 to 5 with rivaroxaban or dabigatran in ROCKET-AF and RE-LY. The manufacturer based this estimate on the relative proportions of patients treated with apixaban who had these scores in ARISTOTLE. The distribution of stroke severity across treatments in the population for whom a VKA antagonist was suitable is academic-in-confidence.

The manufacturer constructed a Markov model to evaluate the long- and medium-term consequences of apixaban for preventing stroke and systemic embolism in people with atrial fibrillation. The model considered warfarin-suitable and -unsuitable populations separately. The baseline characteristics of both populations were considered to be equivalent to the characteristics of a cohort of patients with a diagnosis of atrial fibrillation from a UK GP-based survey (Gallagher et al. 2011). Data from both network meta-analyses were used to inform the clinical effectiveness of treatments in the warfarin-suitable and -unsuitable populations respectively and to derive the transition probabilities used in the model. The risk of stroke was adjusted for baseline CHADS₂ score distribution. The risks of stroke, intracranial haemorrhage, myocardial infarction, other major bleeds and clinically relevant non-major bleeds were adjusted for age. The model had a lifetime time horizon. The intervention and comparators were implemented in the model according to their marketing authorisations. For dabigatran 150 mg it was assumed that patients would switch to the 110 mg dose when they reached 80 years in line with the marketing authorisation. The average dosage of warfarin in the warfarin-suitable population was assumed to be 4.5 mg once daily. The evaluation was undertaken from the perspective of the NHS and Personal Social Services in England and Wales, and costs and benefits were discounted at 3.5% per year after the first year.

The model had 18 health states, including death. Both event-related mortality and other-cause mortality were incorporated in the model. Hypothetical patients transitioned between health states in cycles of 6 weeks with only 1 clinical event permitted per cycle. Patients entered the model in the non-valvular atrial fibrillation ('NVAF') health state, and stayed in this state until they died or experienced 1 of the following permanent events: ischaemic stroke (mild, moderate, severe or fatal); haemorrhagic stroke (mild, moderate, severe or fatal); systemic embolism or myocardial infarction; or 1 of the following temporary events: other intracranial haemorrhage (that is not a haemorrhagic stroke); other major bleeds (gastrointestinal bleeds or other bleeds besides intracranial haemorrhage and gastrointestinal-related bleeds); clinically relevant non-major bleeds; or other cardiovascular hospitalisations (that is, cardiovascular hospitalisations unrelated to stroke or myocardial infarction). The model allowed a maximum of 2 lines of therapy. After a switch to second-line therapy, patients transitioned into the 'NVAF without original anticoagulant' health state and were at risk of the same events as patients in the 'NVAF' health state (with the exception of the switch to second-line therapy).

The manufacturer classified the events as permanent or temporary. Patients who experienced a permanent event accrued both acute and long-term maintenance costs and were not assumed to recover to their previous level of health. After a permanent event, patients in the model were not exposed to the risks of all events: patients who had systemic embolism or myocardial infarction stayed in those health states until they died; patients who had a non-fatal stroke could remain in that health state, have 1 recurrent stroke or die. Recurrent strokes were assumed to be of the same type as the initial event (ischaemic or haemorrhagic) but could be of different severity. The resource use and disutility associated with the second stroke was assumed to be equal to that of the most severe stroke experienced. After a temporary event, all patients were assumed to recover to their previous health status.

A switch from first-line to second-line therapy was permitted after discontinuation because of a clinical event (intracranial haemorrhage or other major bleed) or after discontinuation because of other causes. Patients could switch to aspirin or have no treatment. In the base case, anyone who discontinued treatment was assumed to receive aspirin as second-line treatment. Only a switch from first-line anticoagulation therapy to second-line therapy with aspirin altered patients' risk of subsequent clinical events. Patients who experienced certain permanent events also switched treatment: patients who had a myocardial infarction or haemorrhagic stroke were assumed to stop treatment, and patients receiving aspirin as second-line therapy switched to warfarin if they had an ischaemic stroke or systemic embolism. However, all other patients who had ischaemic stroke or systemic embolism were assumed to remain on their original treatment in the base case. The risk of subsequent events for patients after a permanent event was assumed to be independent of treatment received, so switching did not affect their risk profile.

The manufacturer conducted a systematic review of health-state utility value studies relevant to the health states considered in the model, focusing on studies that reported EQ-5D values. Values from 21 studies that presented EQ-5D data in a population with atrial fibrillation and 3 studies that reported EQ-5D values for a variety of chronic conditions after controlling for comorbidities were included. As there were some health states for which a utility value had not been identified, studies of a population with atrial fibrillation that reported utilities elicited by methods other than the EQ-5D were screened, and data from a further 8 studies were included. One further study was identified from the reference list from the submissions for:

• NICE's technology appraisal guidance on dabigatran etexilate for the prevention of stroke and systemic embolism in atrial fibrillation and

• NICE's technology appraisal guidance on rivaroxaban for the prevention of stroke and systemic embolism in people with atrial fibrillation.

The manufacturer used unit costs taken from NHS reference costs 2010 to 2011 where possible. If available, Healthcare Resource Group codes specified in the costing report for atrial fibrillation from NICE's previous guideline on atrial fibrillation (now replaced by NICE's guideline on atrial fibrillation) were used. The average daily drug acquisition costs were £2.20 for apixaban, £2.20 for dabigatran (either dose), £2.10 for rivaroxaban and £0.12 for warfarin (4.5 mg average daily dose). The manufacturer's model included intervention costs such as an annual INR monitoring cost of £248, which was an inflated estimate of the ERG's calculation in NICE's technology appraisal guidance on dabigatran etexilate, and a £3 renal monitoring cost for 19.4% of patients treated with dabigatran. The manufacturer used NHS reference costs for acute costs per episode for the temporary health states. The acute and long-term costs for systemic embolism and stroke were taken from a UK population-based assessment. Dyspepsia was the only adverse event that was not explicitly modelled as a health state, and a yearly cost of £27.60 was applied to all patients who had dyspepsia.

The manufacturer presented a deterministic base case for the warfarin-suitable and -unsuitable populations. In the population for whom warfarin was suitable, the ICER for apixaban compared with warfarin was £11,008 per QALY gained. This represented a gain of 0.164 QALYs for an incremental cost of £1,795. Dabigatran 110 mg twice daily was strictly dominated (was more costly and less effective) than the dabigatran blend (dabigatran used as per its marketing authorisation, that is, people who are younger than 80 years receive a 150 mg twice daily dose and people 80 years or older receive a 110 mg twice daily dose). Apixaban extendedly dominated rivaroxaban and the dabigatran blend (resulted in a lower ICER compared with warfarin despite having higher total QALYs and total costs than rivaroxaban and the dabigatran blend).

Although aspirin was not included as a comparator in the scope, the manufacturer compared apixaban with aspirin in a population for whom warfarin was unsuitable. In this population apixaban was associated with an ICER of £2,903 per QALY gained compared with aspirin.

The manufacturer assessed the univariate sensitivity of the model to 117 parameters using deterministic sensitivity analyses. In the warfarin-suitable population, parameters that had the most influential effect on the ICER for apixaban compared with warfarin were disutility associated with warfarin use, the hazard ratios for intracranial haemorrhage, ischaemic stroke or other-cause mortality during the trial, the cost of INR monitoring visit and the discount rate applied to QALYs. For apixaban compared with rivaroxaban or dabigatran, the most influential parameters were the hazard ratios associated with stroke, intracranial haemorrhage and other-cause mortality during the trial for these comparators compared with apixaban, the absolute stroke risk for apixaban, and the second-line stroke risk for aspirin. All of the ICERs calculated in the manufacturer's deterministic sensitivity analysis for apixaban compared with the comparator drugs were below £20,000 per QALY gained. In addition, the manufacturer carried out 19 scenario analyses. The majority of the scenario analyses decreased the base-case ICER (for apixaban compared with comparator). The probabilistic sensitivity analysis indicated that the probability that apixaban was cost effective at £20,000 and £30,000 per QALY gained was 80% and 87% respectively. For the dabigatran blend, rivaroxaban and warfarin the probabilities of being cost effective at £20,000 were 10%, 9% and 1% respectively. At £30,000 these were 5%, 7% and 0% respectively.

The ERG considered that, of the 2 trials of apixaban, only ARISTOTLE met the inclusion criteria for this technology appraisal, although it did acknowledge that aspirin is sometimes used in clinical practice in the UK. With respect to the network meta-analyses, the ERG did not consider the second analysis to be appropriate to determine the relative effectiveness of aspirin, apixaban, rivaroxaban and dabigatran in a population for whom vitamin K antagonists were unsuitable because the majority of trials in the second network meta-analysis included patients for whom warfarin was suitable. The ERG therefore focused its critique on the ARISTOTLE trial and the first network meta-analysis which compared the safety and efficacy of apixaban with warfarin, rivaroxaban and dabigatran.

The ERG considered that the inclusion and exclusion criteria, follow-up and statistical analysis of ARISTOTLE were acceptable and that the baseline characteristics of the randomised populations were well balanced between trial arms. The ERG commented that, based on advice given by clinicians on the time in therapeutic range expected in a UK population, the mean time in therapeutic range in ARISTOTLE (62.2%) was acceptable. It also considered the INR monitoring in ARISTOTLE to be consistent with that which would occur routinely in the UK. The ERG additionally considered that the distribution of CHADS₂ scores in ARISTOTLE was comparable to the UK population. However, the ERG highlighted that no data on transient ischaemic attack or health-related quality of life were collected in ARISTOTLE or AVERROES, and that the effectiveness of apixaban in reducing transient ischaemic attacks and improving health-related quality of life was therefore unclear.

The ERG noted that the results of the manufacturer's base case were generated deterministically rather than probabilistically. Therefore, the ERG used the manufacturer's probabilistic sensitivity analysis to estimate the manufacturer's probabilistic base case. The ICER for apixaban compared with warfarin in the probabilistic base case was £16,852 per QALY gained. The ERG considered that the manufacturer had presented a robust and predominantly conservative (direction of bias more likely to be against rather than towards apixaban) economic evaluation of apixaban compared with warfarin, dabigatran 110 mg, dabigatran blend and rivaroxaban in the warfarin suitable population. However, the ERG commented on the plausibility of some of the assumptions and inputs used in the manufacturer's model. The ERG considered whether certain outcomes would be expected to be dependent on the treatment a person received. It noted that severity of stroke event and bleed type was assumed to be dependent on the treatment received. The ERG considered that this may not be clinically appropriate and that there may be limitations to the data that informed these assumptions. The ERG also noted that the within-trial rate of other-cause mortality was different for patients treated with warfarin than apixaban, dabigatran or rivaroxaban. Although patients treated with warfarin may be at a higher risk of event-specific death, the ERG did not expect that they would be at a different risk of other-cause mortality.

The ERG noted that patients who had a stroke (haemorrhagic or ischaemic), systemic embolism or myocardial infarction were assumed to be at risk of fewer types of subsequent clinical events than patients in other health states. The ERG accepted the risk limitation applied to patients who experienced a stroke but that patients with systemic embolism or myocardial infarction would remain at risk of further events (in particular ischaemic stroke). The ERG considered that some people who stop therapy with apixaban, dabigatran or rivaroxaban may be eligible for treatment with warfarin or a different oral anticoagulant rather than aspirin which was the second-line treatment in the manufacturer's model. The ERG additionally commented that the risk profile of people on second-line therapy was not adjusted for characteristics such as age or CHADS₂ score in the manufacturer's model, but it accepted that adjusting for characteristics in second-line treatment may be beyond the reasonable scope of a Markov model.

The ERG commented that utilities were not age adjusted in the manufacturer's model, meaning that a person's quality of life would be affected by events experienced but not by increasing age. The ERG considered that the assumption of equivalent disutility between the apixaban, rivaroxaban and dabigatran may not be robust but that any resultant bias was likely to be against apixaban.

The ERG noted that the acute cost of systemic embolism in the manufacturer's model (£4,077.98) was approximately double the acute costs used in the submissions for NICE's technology appraisal guidance on dabigatran etexilate (dabigatran £2,772 [fatal and non-fatal acute costs]) and NICE's technology appraisal guidance on rivaroxaban (rivaroxaban £1,658). These submissions had used NHS reference costs.

For its revised base case the ERG changed some of the assumptions used in the manufacturer's model. The ERG assumed that other-cause mortality, stroke severity and bleed type were independent of the type of anticoagulant treatment received. The ERG adjusted utility for increasing age by -0.00029 per year. The ERG assumed that people who had myocardial infarction or systemic embolism were at risk of recurrent stroke, and used the same acute costs for systemic embolism as in NICE's technology appraisal guidance on dabigatran etexilate, as this was the most conservative cost used in the submissions for NICE's technology appraisal guidance on rivaroxaban and NICE's technology appraisal guidance on dabigatran etexilate. The ERG also assumed the time horizon was 26 years.

The ERG noted that after these amendments, dabigatran 110 mg continued to be strictly dominated by the dabigatran blend and rivaroxaban and the dabigatran blend remained extendedly dominated by apixaban. The ICER for apixaban compared with warfarin for each individual amendment was relatively consistent with the manufacturer's base-case ICER. The ERG noted that assuming stroke severity was independent of treatment increased the ICER of apixaban compared with warfarin to £12,277 per QALY gained, whereas assuming bleed type was independent of treatment decreased this ICER to £9,771 per QALY gained. When all of the amendments were combined to form the ERG's revised base case, this resulted in an ICER for apixaban compared with warfarin of £12,757 per QALY gained. This represented an incremental cost of £1,823 compared with warfarin for an additional 0.14 QALY.

The ERG carried out 3 further exploratory analyses that were not included in its revised base case. These were:

• Age adjustment of event risks for people on second-line therapy, using the same risk adjustment factors as for people receiving first-line therapy. Dabigatran blend and rivaroxaban continued to be extendedly dominated by apixaban. The ICER for apixaban compared with warfarin fell slightly from the manufacturer's base case of £11,008 to £10,779 per QALY gained.

• Removal of treatment-related disutility. Dabigatran blend and rivaroxaban continued to be extendedly dominated by apixaban but the ICER for apixaban compared with warfarin increased from £11,008 to £14,530 per QALY gained.

• Changes to the treatment sequence to allow second-line treatment with warfarin, apixaban, rivaroxaban or dabigatran. The results of these analyses were highly variable, with ICERs for apixaban varying between £287 per QALY gained (compared with warfarin when dabigatran 110 mg was the second-line treatment) and £60,366 per QALY gained (compared with dabigatran blend when rivaroxaban was the second-line treatment). However, the ERG commented that the results of this analysis should be interpreted with caution because the main driver of the ICERs was discontinuation rates associated with first-line therapy and, consequently, treatments with higher discontinuation rates such as dabigatran appeared more effective than in the manufacturer's base case.

Full details of all the evidence are in the manufacturer's submission and the ERG report.