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

The primary source of evidence was ENGAGE AF‑TIMI 48, a randomised, international (46 countries, including 31 centres in the UK) double‑blind, double‑dummy, parallel‑group, non‑inferiority trial comparing edoxaban with warfarin. It included a total of 21,105 people with non‑valvular atrial fibrillation and a moderate‑to‑high risk of stroke, defined as a CHADS₂ score of 2 or more (CHADS₂ is a scoring system that measures risk factors associated with congestive heart failure, hypertension, age, diabetes and stroke). People were randomly assigned to treatment with low‑dose edoxaban (30 mg, n=7,034), high‑dose edoxaban (60 mg, n=7,035) or warfarin (n=7,036). People randomised to edoxaban who were at increased risk of bleeding because of higher drug exposure (those weighing 60 kg or less, with creatinine clearance 30 to 50 ml/min, or having concomitant treatment with potent permeability glycoprotein inhibitors) had the dose reduced, either at randomisation or during the study, to 15 mg in the low‑dose group and to 30 mg in the high‑dose group. The clinical trial results presented below focus on the higher dose treatment arm because this is the recommended dose in the marketing authorisation. This is referred to throughout as the 60 mg/30 mg treatment arm because it included people who were given 30 mg because of clinical factors. The dose in the warfarin group was adjusted to maintain an international normalised ratio [INR] of 2.0 to 3.0, and people in the trial were 'well controlled' on warfarin (median time spent in the therapeutic range [TTR] was 68.4%).

Patient characteristics were similar between the treatment groups including age, sex, ethnicity, risk factors, CHADS₂ score and renal function. The mean CHADS₂ score was 2.8 and approximately 53% of patients had a CHADS₂ score of 3 or more, indicating that the patient population was at a moderate‑to‑high risk of stroke. The median age of people in the study was 72 years and 62% were male.

The primary efficacy outcome was time to the first stroke (ischaemic or haemorrhagic) or systemic embolic event. People in the trial continued treatment and were followed up until approximately 672 primary efficacy endpoint events had been collected, which provided 87% power for confirming non‑inferiority for each edoxaban regimen. A non‑inferiority test using the modified intent‑to‑treat (mITT) population for the on‑treatment period was pre‑specified in the statistical analysis plan. To satisfy non‑inferiority, the upper boundary of the one‑sided 97.5% confidence interval for the hazard ratio of the primary efficacy endpoint comparing edoxaban with warfarin could not exceed 1.38, which was an estimate that preserved at least 50% of the benefit of warfarin over placebo. If edoxaban was shown to be statistically significantly non‑inferior to warfarin, a superiority test would be performed using the intent‑to‑treat (ITT) population and the overall study period.

For the primary efficacy outcome (prevention of stroke or systemic embolic event) in the mITT analysis set (on‑treatment and overall study period), edoxaban 60 mg/30 mg met the criteria for non‑inferiority compared with warfarin. Stroke or a systemic embolic event occurred in 182 people in the edoxaban 60 mg/30 mg arm of the trial (1.18% per year) compared with 232 people in the warfarin arm (1.50% per year, hazard ratio [HR] 0.79, 97.5% confidence interval [CI] 0.63 to 0.99, p<0.001 for non‑inferiority). In the pre‑specified superiority analysis performed in the ITT analysis set (overall study period), the rate of stroke or systemic embolic event was 1.57% per year in the edoxaban 60 mg/30 mg arm compared with 1.80% in the warfarin arm (HR compared with warfarin 0.87; 97.5% CI 0.73 to 1.04, p=0.08 for superiority). Results for the mITT overall study period were consistent with those in the ITT overall study period.

The company presented results for the analyses of the components of the primary endpoint (stroke and systemic embolism) and the subcomponents of stroke (ischaemic, haemorrhagic, fatal and disabling) for the mITT analysis set (on‑treatment period, and overall study period) and the ITT analysis overall study period. For the mITT overall study period, edoxaban was shown to be superior to warfarin for haemorrhagic stroke (p=0.001). The results were similar for the mITT population, the on‑treatment period analysis and the ITT population analysis.

The company presented analyses for the primary efficacy results using the mITT analysis set (overall study period) for subgroups according to risk of stroke (defined by CHADS₂ score) and renal function (creatinine clearance). The subgroup analysis for risk of stroke demonstrated that, compared with warfarin, the hazard ratio for the edoxaban 60 mg/30 mg dose was stable and non‑inferior across CHADS₂ scores of 2 to 6. The subgroup analysis for renal function across 3 categories of creatinine clearance (normal renal function 80 ml/min or more; mild renal impairment more than 50 to less than 80 ml/min; and moderate renal impairment 30 to 50 ml/min), suggested that renal function had a significant impact on the efficacy of edoxaban in comparison to warfarin (p=0.0042). This result was shown to be consistent across analysis sets. The hazard ratios for the primary efficacy endpoint were 0.68 (95% CI 0.54 to 0.85) and 0.86 (95% CI 0.63 to 1.17) for the subgroups of people with mild or moderate renal impairment, respectively. In contrast, the relative risk of stroke or systemic embolic event was higher with edoxaban than with warfarin in the subgroup of people with normal renal function (HR 1.31, 95% CI 0.96 to 1.79). The company noted that this analysis should be treated with caution because a variety of factors (including an unusually low event rate in the warfarin group, and potential imbalances between treatment groups because of randomisation not being performed within subgroups) could have contributed to the observed hazard ratio for stroke or systemic embolic event compared with warfarin in the subgroup of people with normal renal function.

The company also did an analysis comparing centre‑level TTR above and below 60%. The p value for interaction was 0.0361 which indicated that in centres with a TTR above 60% edoxaban had a similar effect compared with warfarin to that observed in the total study population, but there was a significant reduction in risk of stroke and systemic embolism in the subgroup with a centre‑level TTR of less than 60%. When the TTR data were examined by quartiles, however, the p value for interaction was 0.50.

Health‑related quality of life data were collected in ENGAGE AF‑TIMI 48 using the self‑administered EQ‑5D questionnaire at baseline and then every 3 months, until the end of the study. Approximately 60% of patients (11,995 patients) provided quality‑of‑life data; 164 (1.4%) patients were from the UK.

The ERG noted that the statistically significant result for non‑inferiority in ENGAGE AF‑TIMI 48 was driven largely by a reduction in haemorrhagic stroke events in patients treated with edoxaban, but there was no statistically significant difference between edoxaban and warfarin for any other listed component or subcomponent.

The ERG commented that the estimate of treatment effect (hazard ratio) for the primary outcome may not be reliable because the assumption of proportional hazards between treatment with edoxaban or warfarin for haemorrhagic stroke (one of the components of the primary outcome) appeared to be violated. The hazard trend in the warfarin group changed sharply at 6 months, in comparison with a smooth hazard trend over time in the edoxaban group.

The ERG noted that the results of the analysis for centre‑level TTR suggested that the efficacy of edoxaban in comparison to warfarin is significantly greater in the subgroup of centres achieving TTR of less than 60%, but this was not consistent across all analysis sets. There was no significant difference in the results from centres with a TTR of less than or greater than 60% when the analysis was conducted using the mITT population and the on‑treatment observation period. The ERG therefore suggested that the finding that centre‑level TTR may affect the efficacy of edoxaban in comparison to warfarin may be spurious.

The ERG stated that the health‑related quality of life data provided during the clarification process were difficult to interpret because of the low response rate and incomplete analysis. The ERG therefore suggested that it was difficult to draw any firm conclusions about any differences in patients' experiences that are attributable to the choice of treatment.

The company presented the results of the safety analyses, which included all people who had at least 1 dose of study drug for the on‑treatment period in ENGAGE AF‑TIMI 48. In the principal safety analysis for the edoxaban 60 mg/30 mg arm compared with warfarin, the company stated that edoxaban had a significantly reduced rate of major bleeding (HR 0.80, 95% CI 0.71 to 0.91; p<0.001) and of several secondary bleeding endpoints including intracranial, fatal, clinically relevant non‑major and life‑threatening bleeds (p≤0.01 for all comparisons). However, the company highlighted that major gastrointestinal bleeding occurred slightly more frequently in the edoxaban 60 mg/30 mg arm than in the warfarin arm (annualised rate of 1.51% compared with 1.23%, respectively; HR 1.23 [1.02 to 1.50]; p=0.03).

The company stated that the 5 most frequent treatment‑emergent adverse events in the edoxaban or warfarin groups were urinary tract infections, nasopharyngitis, bronchitis, dizziness and peripheral oedema. The company presented subgroup analyses for the primary safety outcome by centre‑level TTR and by risk of stroke (as defined by CHADS₂), which were consistent with the overall population.

The company did not find any head‑to‑head studies that compared edoxaban with rivaroxaban, dabigatran etexilate or apixaban so it did a network meta‑analysis to estimate the relative efficacy and safety of edoxaban for treating atrial fibrillation, that included 4 trials: ENGAGE AF‑TIMI 48, and 3 trials of other newer oral anticoagulants (apixaban 5 mg twice‑daily [ARISTOTLE]; dabigatran etexilate 150 mg twice‑daily or 110 mg twice‑daily [RE‑LY]; and rivaroxaban 20 mg once‑daily [ROCKET‑AF]). All 4 RCTs had a warfarin treatment arm. Because of significant differences in the patient characteristics and trial design between the 4 trials (for example, ARISTOTLE and RE‑LY included people with a CHADS2 score of 1 or more, whereas the CHADS2 score was 2 or more in both ENGAGE AF‑TIMI 48 and ROCKET‑AF) only data from patients with a CHADS₂ score of 2 or more from RE‑LY and ARISTOTLE were used in the network meta‑analyses.

The results of the meta‑analysis demonstrated that for the composite endpoint of stroke and systemic embolism, efficacy was similar for high‑dose edoxaban compared to other newer oral anticoagulants, but edoxaban significantly reduced major bleeding risk by 24%, 28%, and 17% compared to rivaroxaban, dabigatran etexilate 150 mg and dabigatran etexilate 110 mg, respectively. Major bleeding rates were similar between high‑dose edoxaban and apixaban.

The ERG considered that the key characteristics of the trials (study population, design, outcome measures; and effect modifiers such as age, disease severity, and duration of follow‑up) included in the network meta‑analyses were sufficiently similar to justify combining the results. The company's approach used annualised event rates and the ERG considered that this approach minimised any potential bias resulting from differences in trial duration, which ranged from 1.8 years in ARISTOTLE (apixaban) to 2.8 years in ENGAGE AF‑TIMI 48 (edoxaban).

The ERG noted that in addition to the violation of the proportional hazards assumption for some end points within ENGAGE AF‑TIMI 48, it was also violated within trials of the other 3 newer oral anticoagulants, and in the warfarin groups of the 4 trials included in the evidence network. The ERG highlighted that this meant the hazard ratios from the network meta‑analysis were not reliable and should not be used to inform the company's economic model.

The ERG considered that the assumption of proportionality that underpinned the network meta‑analysis had been shown to be violated both within and between trials (see also sections 3.10 and 3.18). The ERG highlighted that this meant the hazard ratios used to inform the company's economic model were therefore unreliable.

The ERG highlighted that the model predicted that, of a cohort of 1000 people with non‑valvular atrial fibrillation having warfarin, there will be 157 stroke events. Using the company's approach to applying the risk of acute‑stroke mortality, approximately 15 (9.6%) of these would be fatal, which is substantially less than the 16.8% reported in the study by Janes. The ERG assessed the impact of applying the acute mortality rates reported by Janes to all patients experiencing a stroke (ischaemic stroke and haemorrhagic stroke analysed separately; see sections 3.40 and 3.41). The ERG noted that mortality data from ENGAGE AF‑TIMI 48 were not used in the model and no rationale for this was given by the company. The ERG considered it more appropriate to use mortality data from ENGAGE AF‑TIMI 48. Therefore, in its exploratory analyses, the ERG extracted acute mortality data for ischaemic stroke and haemorrhagic stroke from the clinical study report (CSR50, page 132) and pooled it across the warfarin and edoxaban groups of the trial (see sections 3.40 and 3.41).

The ERG highlighted that the model overestimated overall survival for both treatment groups compared with ENGAGE AF‑TIMI 48, and that this potentially underestimated the relative effectiveness of edoxaban compared to warfarin.

The baseline quality of life for the health state of stable atrial fibrillation in the company's model (0.78) was taken from a study of patients with atrial fibrillation having warfarin in the UK (Khan, 2004). In the sensitivity analyses, health‑related quality of life data from ENGAGE AF‑TIMI 48 were used (0.836). Health‑related quality of life declined over time based on an adjustment for cohort aging (-0.00029 per year) to reflect the impact of age and thrombotic events on a patient's quality of life. Utility estimates for mild, moderate and severe stroke were derived from a published study by Gage (1996). The company assumed that patients who have a stroke, myocardial infarction or systemic embolic event experience a permanent decrement to their health‑related quality of life.

The ERG noted that the base case utility value for the stable atrial fibrillation health state had been derived from a UK study by Khan which had a modest sample size of 125 patients, with a low response rate, and was designed to assess the effectiveness of an anticoagulation education programme and self‑monitoring of patients with atrial fibrillation taking warfarin. The ERG did not consider this to be representative of a general atrial fibrillation population and it preferred the use of EQ‑5D data collected in ENGAGE AF‑TIMI 48.

The ERG highlighted that the age‑related utility decrement in the model (-0.00029 per annum) based on self‑reported health status of a US population and valued using the UK tariff may not be generalisable to a UK population. The ERG preferred to use EQ‑5D data from the Health Survey for England in its analysis, because this was a more representative population for the UK. This produced an estimated annual utility decrement of -0.00646.

The company used the BNF 68 (July 2014) to obtain drug costs in the model. All costs for the health states of ischaemic stroke, haemorrhagic stroke and systemic embolic events were based on the Oxford Vascular study (OXVASC, 2013), a large study of healthcare costs after stroke in patients with atrial fibrillation. Costs associated with myocardial infarction were based on NHS reference costs. Post‑myocardial infarction costs were based on the Electronic Drug Tariff and the BNF 68 (July 2014).

The monitoring costs for warfarin patients were adapted from the unit cost of anticoagulation monitoring used in the NHS Costing Template for dabigatran etexilate, which was also used in the apixaban technology appraisal. These costs were inflated to 2013 costs using the Personal Social Services Research Unit Hospital and Community Services Health Index (HCHS). The model assumed that 34% of patients will be seen in a secondary care setting (at a cost of £323.10) with the remaining 66% seen in primary care (£235.11) giving a weighted average annual cost of £265.03.

The ERG noted that although the cost of warfarin used in the company's model (£0.11 per day) was estimated using the list prices reported in the BNF, it is widely available to the NHS at discounted prices. The ERG did an exploratory analysis using the warfarin cost estimated from figures reported in the Department of Health's eMit database (£0.0375).

In the company's deterministic base case analysis (based on people with CHADS₂ ≥2), edoxaban, dabigatran etexilate 110 mg, apixaban and rivaroxaban were strictly dominated (less effective and more costly) by dabigatran etexilate 150 mg, which had an incremental cost‑effectiveness ratio (ICER) of £7,645 per additional QALY gained compared to warfarin (table 1).

The company presented cost‑effectiveness acceptability curves for the incremental analysis, which showed that the probability of edoxaban being cost effective was 2.9% at a maximum acceptable ICER of £20,000 per QALY gained, and 3.4% at a threshold of £30,000 per QALY gained. Warfarin had the highest probability (36.8%) of being the most cost‑effective option at a threshold of £20,000 per QALY gained. At a maximum acceptable ICER of £30,000 per QALY gained, apixaban had the highest probability of being the most cost‑effective option (32.6%). In the pairwise comparison of edoxaban compared with warfarin, the probability of edoxaban being cost effective was 47.1% at a maximum acceptable ICER of £20,000 per QALY gained and 57.1% at a maximum acceptable ICER of £30,000 per QALY gained.

The company did 14 pairwise deterministic sensitivity analyses. It highlighted that the variables that had the most impact on the deterministic base case results were patients' starting age (lower limit 52.1 years, upper limit 89.1 years), cost of treatment, monitoring costs for patients treated with edoxaban (baseline £0, upper limit £26.50), and the utility values of stable atrial fibrillation, post‑event myocardial infarction and haemorrhagic stroke.

The company presented results of subgroup analyses for people with a CHADS₂ score of 3 or more, or with a centre‑level TTR of 60% or more. In people with a CHADS₂ score of 3 or more edoxaban, dabigatran etexilate 110 mg, and rivaroxaban were strictly dominated (less effective and more costly) by apixaban and dabigatran etexilate 150 mg. For the subgroup of people with a centre‑level TTR of 60% or more, edoxaban, dabigatran etexilate 110 mg, apixaban and rivaroxaban were strictly dominated by dabigatran etexilate 150 mg, which had an ICER of £11,738 per additional QALY gained compared to warfarin.

Results from the company's base case probabilistic analysis were not explicitly included in the submission. However, they were calculated by the ERG using the company's model (table 2). Edoxaban, dabigatran etexilate 110 mg and rivaroxaban were strictly dominated by dabigatran etexilate 150 mg and apixaban extendedly dominated dabigatran etexilate 150 mg (more effective and less costly) with an ICER of £13,036 per QALY gained compared to warfarin.

The ERG highlighted that in the company's probabilistic and deterministic analyses edoxaban was dominated (less effective and more costly than at least one alternative treatment). However, in the deterministic analysis dabigatran etexilate 150 mg dominated (was less costly and more effective than) the other, newer, oral anticoagulants, whereas in the probabilistic analysis apixaban dominated dabigatran etexilate 150 mg. The ERG considered that this was because of the very small differences in QALYs between dabigatran etexilate 150 mg and apixaban in all analyses. In addition, the ERG noted that the results of the probabilistic analysis were not completely stable (repeated runs of the same analyses gave slightly different results).

The ERG considered that because the subgroup analyses for patients with a CHADS₂ of 3 or more and for centre‑level TTR of 60% or more were based on very limited data, the extent to which these results were truly representative of effects in these subgroups is unclear.

The ERG noted that the economic model appeared to be robust to the sensitivity analyses carried out by the company. The ERG carried out 17 individual exploratory scenarios, which used its preferred alternative parameter values or formulae. The ERG also combined multiple parameters to give their preferred base case, which included:

• corrected implementation of age‑related utility adjustment

• ERG‑sourced utility values for systemic embolism

• alternative utility values for myocardial infarction, transient ischaemic attack and ERG‑sourced utility values for acute and post‑stroke health states

• assumption regarding the method used to switch patient medication from dabigatran 150 mg to 110 mg at age 80

• assumption regarding treatment discontinuation after haemorrhagic stroke

• acute stroke fatality rate applied to all stroke events (16.8% for ischaemic and 31.6% for haemorrhagic stroke)

• trial data on acute stroke case fatality rates used for all ischaemic and haemorrhagic strokes

• age‑adjusted utility decrement per year amended to -0.00646 instead of -0.00029

• the daily cost of warfarin amended

• the ENGAGE trial HR applied for haemorrhagic stroke.

None of the ERG's amendments to the company's model changed the results of the full incremental analyses; edoxaban was more expensive and less effective than at least one of the alternative treatments. When all of the ERG's preferred values were used in the model the pairwise deterministic ICER for the comparison of edoxaban with warfarin was £16,008 per QALY gained, and the probabilistic ICER was £22,079 per QALY gained. When additional alternative amendments were included to reconcile the model survival outputs with the trial data, and to reflect the changing age and sex distribution over time, this changed the deterministic pairwise ICER to between approximately £15,176 and £15,807, and the probabilistic ICER to between £21,728 and £23,634 per QALY gained.

See the committee papers for full details of the evidence.