Rivaroxaban for preventing adverse outcomes after acute management of acute coronary syndrome

The company submitted a de novo Markov cohort model comparing rivaroxaban 2.5 mg twice daily with standard care (clopidogrel plus aspirin or aspirin alone) in adults who had a recent acute coronary syndrome with elevated cardiac biomarkers and who had not had a previous stroke or TIA. The model used a time horizon of 40 years that was divided into 2 periods: an observation period, which was intended to replicate the duration of the trial data, and an extrapolation period. The extrapolation period started after 96 weeks and had a cycle length of 6 months. In the observation period the initial 2 cycles had a cycle length of 4 weeks and 8 weeks respectively and the remaining cycles used a cycle length of 12 weeks. The company based the analysis from an NHS and personal social services perspective, and costs and benefits were discounted at an annual rate of 3.5%. Half‑cycle correction was performed on the Markov trace.

The company's model consisted of a number of health states corresponding to whether or not the hypothetical patient had another acute coronary syndrome event. The acute coronary syndrome events considered in the model were: MI, ischaemic stroke, haemorrhagic stroke or intracranial haemorrhage (HS/ICH); a bleeding event measured on the TIMI scale; and revascularisation. These acute coronary syndrome events fell into 2 broad categories: those with longer term implications for the relative risks of developing further conditions, utility and costs and those deemed to be transient events where the impacts were limited to 1 model cycle. Patients could die at any time in the model and there were multiple causes of death simulated in the model. Patients could die from an MI, ischaemic stroke or HS/ICH or other CV death, which included deaths related to bleeding. Patients could also die from non‑CV causes at any time point in the model.

The long‑term acute coronary syndrome events included the MI, ischaemic stroke and HS/ICH conditions. The long‑term events had 2 subsequent tunnel states to allow for the patient's health‑related quality of life to improve over time, and for the cost of treatment and the relative risk of having a subsequent event to fall over time. Patients could have up to 3 acute coronary syndrome events; the specific types of event were recorded when patients had 2 or fewer events. When 3 events happened, it was assumed that there was 1 event of each type (that is an MI, an ischaemic stroke and a HS/ICH).

The health states corresponding to bleeding and revascularisations were assumed to be transient health states and when a patient entered these states a one‑off cost and utility decrement was applied. These transient health states were applied to only the patients in the observation period of the model, implicitly assuming that the bleeding and revascularisation rates for the 2 interventions were comparable after rivaroxaban treatment was discontinued for all patients at the end of the second year.

The population modelled was the subgroup of patients who had acute coronary syndrome with elevated cardiac biomarkers and had not experienced a prior stroke or TIA in the ATLAS‑ACS 2‑TIMI 51 trial. The data were not pooled from both rivaroxaban trial arms; the model was based on the subgroup data from patients who had 2.5 mg rivaroxaban twice daily only. Data from both trial strata were used to inform the model. In accordance with ATLAS‑ACS 2‑TIMI 51, it was assumed that in the base case 93% of patients had clopidogrel plus aspirin and 7% of patients had aspirin alone. A scenario analysis was presented considering only those patients who had clopidogrel and aspirin.

In the base case the transition probabilities for future acute coronary syndrome‑related events in the observation period (2 years) were determined by fitting a Weibull distribution to the trial data. This was undertaken for both the rivaroxaban 2.5 mg twice‑daily arm and for the placebo arm. The company stated that because patient numbers diminished over time, particularly visible towards the end of ATLAS‑ACS 2‑TIMI 51, it was difficult to estimate transition probabilities directly from the data for the later cycles within the observation period. The company commented that by fitting a Weibull distribution to the ATLAS‑ACS 2‑TIMI 51 trial data, it was able to remove the data fluctuations caused by a decline in the numbers of observations over the trial.

The company's model assumed that patients could discontinue treatment in the observation period after they had an acute coronary syndrome event. The probability of discontinuation following an event was obtained from ATLAS‑ACS 2‑TIMI 51. This was calculated by using the total trial population, not the licensed population.

The UK marking authorisation for rivaroxaban states that 'extension of treatment beyond 12 months should be done on an individual patient basis as experience up to 24 months is limited'. To reflect this, the company adjusted the proportion of patients who continued on rivaroxaban treatment in the second year (that is from 48 to 60 weeks on wards). The proportion of patients selected was to allow an overall continuation rate of 19% after 12 months and that this would decline to 0% at the end of the second year. No treatment effect or cost was applied to those patients who discontinued rivaroxaban treatment.

In the model patients had clopidogrel 75 mg once a day, aspirin 75 mg once a day and rivaroxaban 2.5 mg twice daily as appropriate. Because rivaroxaban entered the treatment pathway after stabilisation of an acute coronary syndrome event, any further differences in costs between the intervention and the comparator were a result of acute coronary syndrome events and discontinuations related to an acute coronary syndrome event.

Costs of acute coronary syndrome events were determined by the NHS reference costs (2012 to 2013) of treating the event and the cost of follow up for the patient. An assumption was made that if a patient had multiple acute events in the long term, then the cost of hospitalisation and the follow up of two events were applied. This was the case irrespective of the time between the events. It was assumed that, on average, patients experienced 5, 14 and 28 days rehabilitation after MI, ischaemic stroke and HS/ICH respectively. These rehabilitation costs occurred in the first 3 months after the event. Transient event costs were also included in the model.

The utility values associated with long‑term health states were obtained from the literature, primarily from NICE's technology appraisal guidance on ticagrelor for the treatment of acute coronary syndromes. A study by Ara and Brazier was used to calculate the improvement in health‑related quality of life that patients would experience in the stroke health states. The study was used to obtain the utility values of patients with stroke in the UK at baseline and at 12 months after the stroke occurred. Based on the utility values from these 2 time points, a 33% improvement in health‑related quality of life over 12 months was calculated for patients who had strokes. To calculate the utility values for patients who had a stroke 6 months after a previous stroke, the average of the stroke first 6 months and the stroke (post 12 months) health states was taken. Utility values were assumed to be the same for both rivaroxaban and standard care following any event. For multiple event states, the utility values of two events that had occurred were multiplied together. The company stated that this allowed for worsening health‑related quality of life following multiple events to be taken into account. The utility value for the event‑free health state was assumed to remain constant over time. The resulting utility estimates were 0.842 for no event, 0.779 for non‑fatal MI, 0.821 for after MI, 0.703 for non‑fatal stroke, and 0.703 for after stroke.

For the licensed population, the company reported a deterministic incremental cost‑effectiveness ratio (ICER) of £6,203 per quality‑adjusted life year (QALY) gained (incremental costs £764, incremental QALYs 0.12) for a life time horizon of 40 years for rivaroxaban compared with clopidogrel plus aspirin or aspirin alone.

The company conducted a series of 1‑way deterministic sensitivity analyses. Changes to the cost parameters, discount rates, utility values and risk estimates for MI impacted on the base case ICER, but no factor increased the ICER to over £10,000 per QALY gained. The company's probabilistic analysis showed that rivaroxaban had a 99.9% probability of being cost effective compared with clopidogrel plus aspirin or with aspirin alone if the maximum acceptable amount for an additional QALY was £20,000.

The ERG stated that the company undertook a comprehensive systematic review of rivaroxaban for the prevention of adverse outcomes in patients after the acute management of acute coronary syndrome. The ERG considered ATLAS‑ACS 2 TIMI 51 to be a well‑designed, multicentre RCT of reasonable quality.

The ERG questioned the generalisability of the population enrolled in ATLAS‑ACS 2‑TIMI 51 to the population seen in clinical practice in England. The ERG noted that of all patients randomised in ATLAS‑ACS 2‑TIMI 51, 74% were men and the mean age was 61.8 years. The ERG commented that patients with acute coronary syndrome in England are usually older, with a mean age of 65 years and 72 years for patients with STEMI and NSTEMI respectively. It highlighted that the EMA's assessment report noted that patients in the trial were considered to be at low risk. Patients in the trial had little comorbidity, lower than usual use of percutaneous coronary intervention and included a relatively small proportion of people aged over 75 years or who had impaired renal impairment with creatinine clearance less than 50 ml/min. As a result, the findings from ATLAS‑ACS 2‑TIMI 51 may not be applicable to an older population or to those with a greater incidence of renal impairment and a higher baseline bleeding risk.

The ERG commented that mean treatment duration with rivaroxaban in ATLAS‑ACS 2‑TIMI 51 was 13.1 months. As a result, the evidence on efficacy and safety of rivaroxaban 2.5 mg twice daily beyond this time is limited. The ERG noted that this is reflected in the summary of product characteristics, which recommends that extension of treatment beyond 12 months should be done on an individual patient basis because experience up to 24 months is limited.

The ERG stated that the test 'elevated cardiac biomarker' is less sensitive than if a patient exhibits a rise or fall in their cardiac biomarkers (preferably troponins), because many patients have persistently elevated biomarkers outside the context of acute coronary syndrome. In current practice, the diagnosis of NSTEMI requires evidence of myocardial ischaemia with a rise or fall in the blood level of a cardiac biomarker (troponin). In addition, the sensitivity of biomarker assays has increased since the trial was conducted. If more sensitive assays had been available during ATLAS‑ACS 2‑TIMI 51, more patients might have been diagnosed with NSTEMI rather than unstable angina and therefore included in the licensed population.

The ERG noted that there were numerical inconsistencies between the 2 dose groups (2.5 mg twice daily and 5 mg twice daily) for the components of the composite efficacy end points in the licensed population (see table 2). When the components of the primary efficacy end points were analysed individually, rivaroxaban 2.5 mg twice daily significantly reduced the risk of death from CV causes compared with placebo, but did not reduce the risk of MI or stroke. In contrast, rivaroxaban 5 mg twice daily significantly reduced the risk of MI, but did not reduce the risk of CV death or stroke. The ERG noted that the numerical inconsistencies between the 2 dose groups had been extensively discussed in a US FDA briefing document (albeit in the whole trial population of ATLAS‑ACS 2‑TIMI 51, rather than the licensed population). This briefing document states that 'the proposition that a lower dose of an antithrombotic drug is significantly more effective than a higher dose lacks biological plausibility'. The ERG also noted that the EMA's assessment report concluded that these findings may partly have been due to chance. The ERG therefore considered the hazard ratios from the combined dose to be more plausible than those of the individual doses.

The ERG considered the validity of the results from the ATLAS‑ACS 2‑TIMI 51 study to be questionable as a result of the high discontinuation rates from the trial. The ERG noted that 15.5% of the total randomised population (n=15,526) withdrew from the trial (2.5 mg rivaroxaban twice daily 15%; 5 mg rivaroxaban twice daily 16.3%; placebo 15.1%). The ERG highlighted that the rates of premature withdrawal in the ATLAS‑ACS 2‑TIMI 51 trial were higher than other similar randomised trials in patients with acute coronary syndrome: APPRAISE‑2 (apixaban, 1.8% [131 out of 7392]); TRACER (vorapaxar, 5.9% [761 out of 12,944]); PLATO (ticagrelor, 3.0% [562 out of 18,624]) and TRITON (prasugrel, 5.9% [804 out of 13,619]).

The ERG commented that because data were missing for people who withdrew from the trial (proportion of patients with unknown vital statistics 3.2% [ITT analysis, 495 out of 15,526]) there was a potential risk that this may have led to informative censoring. That is, patients who drop out (and therefore are censored) are more or less likely to experience the primary outcome of interest compared with those remaining in the study, and this happens in a non‑random manner. This may be compounded if the reason for, or frequency of, discontinuation differs between treatment groups. The ERG highlighted that no detailed discussion was provided in the EMA's assessment report regarding this issue. The ERG considered that the efficacy analyses were at risk of bias because prognoses may differ in those patients who withdrew from the trial. The ERG highlighted that the likely bias introduced by informative censoring in the clinical outcomes and cost‑effectiveness analyses was unknown.

The ERG stated that the structure of the company's model led to the potential for systematic errors to occur, because the time between multiple events is not tracked. This causes the potential for systematic errors in 3 ways. First, patients who had 2 events in 1 time cycle were not distinguished from those patients who had multiple events in separate time cycles. Second, for patients who had multiple events in separate time cycles any improvement over time that they may have experienced was ignored. Finally, for those patients who transition into the multiple event states from the single event states, the first event was not tracked. The ERG commented that there were 2 solutions to this problem. First, a more complicated state transition cohort model could be developed so that cost and utilities for each multiple event state can be varied by the preceding health state and by the time between the events. Second, a patient‑level simulation approach could be taken.

The ERG stated that the population modelled by the company was the patient subgroup who had elevated cardiac biomarkers and had not experienced a prior stroke or TIA. Therefore, all issues with the generalisability of the results of ATLAS‑ACS 2‑TIMI 51 (see sections 3.3 and 3.28) and informative censoring (see sections 3.10 and 3.33) also apply to the economic model results.

The ERG stated that the approach used by the company to calculate the transition probabilities for the transient health states was inappropriate because the cost and QALYs of the events that occurred in the second year were not appropriately discounted. Also, there was no clear adjustment for the number of additional patients who were assumed to discontinue rivaroxaban in year 2 or for those patients who were assumed to discontinue clopidogrel or rivaroxaban treatment after an acute coronary syndrome event.

The ERG commented that it was not clear in the submission how the patients who continued rivaroxaban treatment after 1 year were selected from the rest of the patient population. The ERG stated that it was unknown whether the base‑case parameters for the change in efficacy and costs represent patient discontinuation in the second year of treatment.

The ERG had concerns with the methodology used by the company to calculate the utility values for patients who had a stroke. The ERG stated that it was unclear why the values from Ara and Brazier were appropriate to calculate the improvement in health‑related quality of life of patients who experienced a stroke, but not considered appropriate to be used as the utility values in the economic model.

The ERG had concerns about how the improvement in utility values over time was modelled in the multiple event states. If a hypothetical patient transitioned into the multiple event states from a single event state, their utility in the multiple event state could be understated, because improvement in utility after the first event had been ignored. The ERG stated that that this problem was again related to the inability of the model to distinguish when events had occurred. The ERG noted that this was not the only assumption that the company could have made to calculate the utility value in the multiple event states. It could have assumed that the lowest utility value of the 2 events applied to the patients or, if the model could track the chronicity of events, it could have assumed that the utility of the most recent event applied.

The ERG expressed concerns about the appropriateness of the methods used to model the costs of rivaroxaban, clopidogrel and aspirin and the efficacy data (shape and scale parameters of the Weibull curve) in the probabilistic sensitivity analysis. The ERG recalculated the probabilistic sensitivity analysis and found that the results were generally more favourable to rivaroxaban, producing more incremental QALYs at a lower incremental cost.

The ERG commented that there were a number of key parameters that could not be adjusted within the model, which may have changed the ICER to a greater extent. These included: amendments to the hazard ratio for fatal bleeds; using pooled efficacy data rather than the 2.5 mg rivaroxaban twice‑daily dose alone; and adjusting for the possibility of informative bias.

The ERG conducted an exploratory probabilistic sensitivity analysis in which published levels of uncertainty around the utility value estimates and the reference costs, rather than an arbitrary range, were taken into account. The resulting probabilistic ICER was £6,150 per QALY gained for rivaroxaban plus clopidogrel plus aspirin or rivaroxaban plus aspirin, compared with clopidogrel plus aspirin or aspirin alone. The ERG noted that its probabilistic ICER was similar to the company's deterministic ICER.

The ERG addressed its concerns about informative censoring (see section 3.33) by conducting a 'crude' exploratory sensitivity analysis, to explore the effects on the ICER of increasing the number of patients who experienced a fatal bleeding event with rivaroxaban (assuming that the event occurred immediately on taking rivaroxaban). The ERG considered a range of additional fatal bleeding events ranging from no additional fatal bleeding events (company's base case) to 20 additional bleeding events. Because there were 21 fatal bleeding events in the combined rivaroxaban treatment arms of the total population in ATLAS‑ACS 2‑TIMI 51, the ERG considered that 20 additional fatal bleeding events was an unfavourable scenario for the rivaroxaban 2.5 mg twice‑daily dose. The result of the ERG's exploratory analysis showed that even if rivaroxaban 2.5 mg twice a day caused an additional 20 fatal bleeding events compared with the event rate observed in the trial, the ICER was not estimated to be greater than £10,000 per QALY gained.

The ERG also undertook a series of exploratory scenario analyses. When the following changes to the model were implemented, the ERG's preferred deterministic base‑case ICER was £5,622 per QALY gained (compared with the company's deterministic base‑case ICER of £6,203 per QALY gained):

• The transition probabilities were estimated from the trial data rather than using the Weibull curves.

• The treatment duration of rivaroxaban was limited to 1 year.

• Age‑adjusted utility values for the whole population from Ara and Brazier's formula were used to adjust the no‑event health state.

• Only 1 cost was applied to the multiple event states. Where there were 2 different costs added together in the company's base case, the maximum of the 2 costs was applied.

• No improvement over time in the stroke utility was modelled.

• The relative risk of having a subsequent event, given that an event had already occurred, was amended.

• The life‑years gained matrix and the costs were adjusted for the 12‑week cycle length in the observation period.

Full details of all the evidence are in the committee papers.