Oral anticoagulants for primary prevention, treatment and secondary prevention of venous thromboembolic disease, and for prevention of stroke in atrial fibrillation: systematic review, network meta-analysis and cost-effectiveness analysis

Atrial fibrillation, stroke and myocardial infarction

Atrial fibrillation (AF) is the most common cardiac arrhythmia. 1 The prevalence of AF roughly doubles with each decade of age, rising to almost 9% at age 80–90 years. 2 AF substantially increases (by up to five times) the risk of thromboembolic stroke (annual incidence 114 per 100,000) because of blood pooling in the left atrium and systemic embolisation to the brain. More than 20% of the 130,000 annual strokes in England and Wales are attributed to AF. Approximately one-third of stroke patients die in the first 10 days, one-third recover in 1 month and one-third have disabilities necessitating rehabilitation, making stroke the leading cause of adult disability. Patients with thromboembolic stroke from AF have higher mortality and morbidity and longer hospital stay than patients with other stroke subtypes. Warfarin is an effective oral anticoagulant for the prevention of stroke in patients with AF. 3 Although the incidence and mortality of stroke continue to fall in the UK, the underutilisation of anticoagulation in patients with AF at high risk of stroke is a major gap in clinical care. 4 In patients with AF, antiplatelet and anticoagulant therapies are generally considered from the perspective of mitigation of stroke risk. However, the presence of AF is also associated with an approximately twofold higher risk of future acute myocardial infarction (MI),5 whose annual incidence in England (130 and 55.9 per 100,000 for men and women, respectively)6 is higher than that of stroke.

Venous thromboembolic disease

The annual incidence of venous thromboembolic disease from a study7 conducted in Europe is 183 per 100,000. It encompasses clot formation in deep veins of the legs or pelvis [deep-vein thrombosis (DVT); annual incidence 124 per 100,000] and the displacement of clots to the pulmonary arteries [pulmonary embolism (PE); annual incidence 60 per 100,000]. Important risk factors for venous thromboembolism (VTE) include major surgery, particularly lower limb orthopaedic surgery and surgery for cancer, as well as hospitalisation in acutely ill general medical patients (approximate incidence 15%). VTE costs the UK NHS £640M and is responsible for approximately 30,000 deaths each year in hospitals in England. DVT is an important cause of long-term morbidity, being a risk factor for chronic leg ulceration. PE may also lead to long-term morbidity due to pulmonary hypertension. There is an approximately 30% risk of recurrence of VTE within 8 years.

The risk of VTE during hospitalisation for surgical or medical treatment can be reduced by low-molecular-weight heparin (LMWH), fondaparinux (Arixtra^®, GlaxoSmithKline, London, UK) or unfractionated heparin. 8 Warfarin is the most frequently prescribed anticoagulant for the initial treatment and for the long-term secondary prevention of VTE in those who are deemed to be at high risk of recurrence.

Study designs

In all reviews we included Phase II or Phase III RCTs using either a superiority or a non-inferiority design.

Participants

In all reviews we included adults (≥ 18 years) who were eligible for oral anticoagulation or (antithrombotic) treatment. Trials in participants who were eligible for only parenteral (injected) anticoagulation were excluded. Unless otherwise specified, anticoagulation services may have been delivered in hospital-, primary care- and pharmacy-based clinics or through home monitoring and telephone support. The review was not limited to NHS anticoagulation services.

Specific criteria for inclusion in the four reviews were as follows.

1. Stroke prevention in AF Adults with non-valvular AF.

2. Primary prevention of VTE Adults admitted to hospital who were considered to be at high risk of VTE, including those with a medical condition (e.g. cancer, major trauma, stroke) or undergoing a surgical procedure (e.g. total knee or hip arthroplasty, hip fracture surgery) that carries a high risk of VTE.

3. Acute treatment of VTE Adults who have received a new or recurrent objectively confirmed diagnosis of acute symptomatic VTE.

4. Secondary prevention of VTE Adults who have completed a minimum of 3 months of anticoagulant treatment for objectively confirmed first VTE without recurrence (secondary prevention).

Interventions and comparators

Five NOACs were the focus of all reviews: dabigatran, apixaban, edoxaban, betrixaban, rivaroxaban. NOACs not considered were eribaxaban (the current stage of development was unclear), ximelagatran (withdrawn because of liver toxicity), darexaban (YM150) and AZD0837 (both discontinued), LY517717 and letaxaban (TAK442) (no available information on any further clinical development for both); and otamixaban (parenteral administration).

As the reviews were conducted to inform NMAs, we determined the comparator interventions to ensure that they would provide information on the relative effectiveness of the interventions of interest. We constructed preliminary networks of available treatment comparisons from trials included in previously published NMAs (irrespective of the outcome data available from them). Comparators were chosen based on the possibility of informing indirect evidence on the relative effectiveness of oral anticoagulants, and on the ‘distance’ of these comparators from our interventions of interest in the network, which relates to the likely increase in precision in the estimates of relative effectiveness of the oral anticoagulants.

Specific comparators in the four reviews were as follows:

1. Stroke prevention in AF Therapeutic doses of warfarin or other VKA [with optimal international normalised ratio (INR) range 2–4]; aspirin; clopidogrel (Plavix^®, Sanofi, USA).

2. Primary prevention of VTE Standard dose LMWH; therapeutic doses of warfarin or other VKA (with optimal INR range 2–4); placebo.

3. Acute treatment of VTE Therapeutic doses of warfarin or other VKA (with optimal INR range 2–4).

4. Secondary prevention of VTE Therapeutic doses of warfarin or other VKA (with optimal INR range 2–4); placebo; no treatment.

Studies evaluating fixed-dose administration of warfarin were excluded. Studies evaluating warfarin with suboptimal target INR compared with UK guidelines were excluded from the main analyses but combined with studies evaluating warfarin with standard target INR in sensitivity analyses. Unfractionated heparin and fondaparinux were excluded from the primary prevention of VTE review, as they would be distant from the NOACs in the network and hence contribute very little information. Non-standard doses of LMWH that were excluded from this review included enoxaparin (Lovenox^®, Clexane^®, Sanofi-Aventis, France) at 20 mg twice daily (bd), ardeparin (Normiflo, Wyeth-Ayerst, USA) at 25 anti-Xa units/kg bd or 35 anti-Xa units/kg bd and nadroparin (Fraxiparine^®, Sanofi-Synthelabo, France) 3800 IU anti-Xa once daily (od).

Prevention of stroke in atrial fibrillation

We sought data on the following outcomes:

• stroke or systemic embolism (SE)*

• all stroke

• ischaemic stroke (major ischaemic stroke or minor ischaemic stroke)*

• fatal stroke

• non-fatal stroke

• haemorrhagic stroke (major haemorrhagic stroke or minor haemorrhagic stroke)

• any bleeding

• minor bleeding

• major bleeding*

• clinically relevant non-major (CRNM) bleeding

• clinically relevant bleeding (CRB)* (defined as CRNM bleeding or major bleeding)

• intracranial bleeding*

• extracranial major bleeding

• extracranial minor bleeding

• fatal bleeding

• bleeding from surgical site

• thrombocytopenia

• MI*

• transient ischaemic attack (TIA)

• arterial event

• quality-of-life outcomes

• hospital admission

• death (cardiovascular)

• all-cause mortality.*

The outcomes addressed in NMAs are marked with an asterisk in the list above. These were chosen based on three considerations: (1) their clinical importance; (2) the consistency of reporting across studies included in the network; and (3) the number of data that were available for inclusion in NMAs.

Venous thromboembolism

For all VTE reviews we sought data on the following outcomes.

Search strategy

Scoping searches that were conducted during protocol development identified some previously published NMAs of oral anticoagulants. We rescreened the studies included in these NMAs against our eligibility criteria, and developed searches to identify any additional studies published beyond the search dates of the most recent NMAs in each population. 8,16–18

We used two separate search strategies, one for the review of stroke prevention in AF (for which the search was run on 12 March 2014 and updated on 15 September 2014, and covered the period 2010 to September 2014) and one for the three reviews in VTE (for all three of which the search was run on 19 March 2014, updated on 15 September 2014, and covered the period 2008 to September 2014). In each search strategy we combined terms for either AF or VTE with terms for the interventions and comparators of interest and added a filter to focus the search on RCTs. We searched MEDLINE and PREMEDLINE In-Process & Other Non-Indexed Citations, EMBASE and The Cochrane Library. The stroke prevention in AF review search was run on the 12 March 2014 and updated on 15 September 2014, and covered the period 2010 to September 2014. The search for the three reviews in VTE was run on the 19 March 2014, updated on the 15 September 2014 and covered the period 2008 to September 2014. We applied no restrictions on language. The principal search strategy is included in Appendix 1. We removed duplicate records, identified by title, authors, journal citation and date published.

We sought information on studies in progress, unpublished research or research reported in the grey literature from www.clinicaltrials.gov (to September 2012). We screened reference lists of retrieved studies and relevant review articles. We also searched the NHS Economic Evaluation Database (NHS EED) and NICE Technology Appraisals.

Assessing relevance and inclusion

Two reviewers independently screened the results of the searches by title and abstract. We resolved disagreements through consensus or referral to a third reviewer where necessary. We obtained full texts of all potentially relevant reports and two reviewers assessed these independently against the eligibility criteria, with disagreements resolved by a third reviewer. We collated multiple reports of the same study and mapped them to unique studies.

Choice of interventions

To perform NMAs we had to allocate each intervention group in each trial to a category, with each intervention category forming a ‘node’ in the network. We kept different doses or frequencies of administration (i.e. od or bd) of oral anticoagulants in separate nodes. We assigned different VKAs to one node (named ‘Warfarin’), but separated intended INR range 2–3 from intended INR range 3–4 and from other ranges. For LMWH interventions in the review of primary prevention of VTE, we separated preoperative (pre-op) LMWH from postoperative (post-op) LMWH. The intervention categories (or network nodes) are labelled throughout the report using drug, frequency and dose, or INR range, as appropriate.

Choice of time points

When outcome data were presented for multiple time points, we took the longest period of follow-up, except for bleeding events in the review of primary prevention of VTE, which we assessed at the end of the treatment period.

Choice of outcomes

When outcome data were not presented directly, we computed or substituted them, using data for other outcomes, making assumptions that we considered to be reasonable. When we could not extract data for the outcome ‘stroke or SE’ in the review of stroke prevention in AF, we used ‘all stroke’. When CRB was not reported but both major bleeding and CRNM bleeding events were, we used the total number of events across these two categories. If symptomatic PE was not reported in any of the three VTE reviews, we used symptomatic non-fatal PE if available, or the sum of fatal PE and non-fatal PE. Additionally, in the review of primary prevention of VTE, when symptomatic VTE was not reported, we added across symptomatic DVT and symptomatic PE, if available.

Investigation of heterogeneity

We had planned to use subgroup and meta-regression24 analyses to examine the extent to which patient- and study-level characteristics explain between-study heterogeneity. We prespecified the important characteristics to be age, gender, ethnicity/race, body mass index (BMI) or weight, renal status or creatinine clearance, blood pressure, diabetes mellitus, hypertension, previous thrombotic event, liver disease, chronic heart failure, cancer, pregnancy, intervention dose, average TTR in the warfarin group, and summary assessment of risk of bias for each outcome. Additional factors for AF trials were CHADS₂, CHADS₂ VASC, HAS-BLED, history of previous stroke or TIA and previous MI. Additional factors for primary prevention of VTE were general surgery compared with orthopaedic surgery, elective emergency surgery compared with non-elective emergency surgery, and medical trials compared with surgical trials. An additional factor for acute treatment or secondary prevention of VTE was the nature of the index event (whether PE or DVT). When available, inferences about subgroup effects would be based on within-trial subgroup analyses (e.g. comparing relative intervention effects in older and younger participants). Investigation of between-study variation using these characteristics could not be studied in most cases because of the lack of multiple trials of the same pairwise comparison, although we conducted some sensitivity analyses for the review of stroke prevention in patients with AF. Specifically, we performed several meta-regressions using the average TTR in the warfarin group as a covariate.

Investigation of inconsistency

The validity of a NMA depends on the assumption that there is no effect modification of the pairwise intervention effects or, that the prevalence of effect modifiers is similar in the different studies. This key assumption has been referred to variously as exchangeability,22 transitivity,25 similarity26 and consistency. 27,28 For a clinical and epidemiological judgement of the plausibility of this assumption we examined whether or not the trials were similar in ways that might modify treatment effect, based on the prespecified list of potential effect modifiers (see Investigation of heterogeneity).

‘Evidence inconsistency’ can be considered an additional layer of heterogeneity that occurs in networks of evidence when there is a discrepancy between the direct and indirect estimates of relative intervention effects. Therefore, inconsistency is a property of ‘closed loops’ of evidence, in which both direct and indirect evidence are available for each comparison. We visually inspected the network diagrams to identify potential for inconsistency (closed loops), and used model fit and selection statistics to informally assess whether or not it was evident. Where there was potential for inconsistency, we compared the residual deviance from the consistency model (providing NMA evidence) with the residual deviance from an ‘inconsistency model’, without consistency constraints (in which only direct evidence is analysed for each comparison). When both direct and indirect evidence were available, and the direct evidence had a standard error that differed (beyond the second decimal place) from the NMA estimate, we used results from these two analyses to back-compute the indirect estimates, on the basis that the NMA estimates (from the consistency model) would be equivalent to a weighted average of the direct estimate (from the inconsistency model) and the indirect estimate. In the results tables we present all three of these estimates. The extent of the disagreement between the direct and indirect estimates can be used as a local measure of inconsistency for that comparison. Note that for the vast majority of comparisons there was either only direct evidence or only indirect evidence, so that the NMA estimates correspond to one of these.

Atrial fibrillation: patient population

We considered patients with non-valvular AF who were eligible for anticoagulation. We made no distinction between paroxysmal, persistent and permanent AF. The RCTs identified in the systematic review did not distinguish between AF type, but patients with paroxysmal AF are less likely to be included in RCTs than those with other AF types; therefore, our results are most applicable to patients with persistent and permanent AF. We consider a cohort of patients receiving first-line anticoagulation at the age of 70 years, based on the mean age observed in the RCTs identified in the systematic review [mean age 70 years, standard deviation (SD) 8 years], and consider costs and benefits over a lifetime. We assume a 60 : 40 split in favour of males, similar to that observed in the RCTs.

Atrial fibrillation: interventions

The first-line treatments for AF included in the CEA, alongside their standard or licensed doses, are listed in Table 4. We consider only licensed treatments and doses in our analysis. Although a few small RCTs have compared betrixaban with warfarin in AF, there was not enough evidence to include it in the economic model. Standard care for patients with AF, before the introduction of NOACs, was warfarin. 68

Treatment switching may occur as a result of treatment failure, indicated by ischaemic stroke or serious AEs, such as intracranial haemorrhage (ICH). For patients on warfarin first-line treatment, the only second-line intervention available was assumed to be no treatment. For patients on a NOAC first-line treatment, second-line treatment may be either warfarin or no treatment. No treatment is the only third-line treatment. These rules are illustrated in Figure 1, where the events that may lead to treatment switching are indicated.

FIGURE 1.

Treatment strategies and switching/discontinuation rules. The events that may lead to treatment switching are indicated next to the arrows between treatments.

Venous thromboembolism: patient populations

For primary prevention, we estimated cost-effectiveness in two distinct subpopulations: patients undergoing elective THR or TKR. We considered including other populations (e.g. patients who were hospitalised for medical treatment) but there was not enough evidence identified in the literature review to inform a model.

After a confirmed VTE event, patients receive acute treatment. The population of patients in the acute treatment model includes those for whom a non-fatal symptomatic VTE event (DVT or PE) followed a THR or TKR, as well as patients with a symptomatic VTE from other causes. Patients who completed at least 3 months of anticoagulant treatment for symptomatic VTE without recurrence are included in the secondary prevention model.

We assumed an average age of subjects entering the primary prevention model of 68.7 years (SD 11.4 years) and the split between males and females of 40 : 60, based on estimates from the National Joint Registry. 69 The assumed age is in line with the median of the mean age of patients enrolled in the primary prevention RCTs (median 64.6 years). The starting age in the acute and secondary prevention populations was 57.35 years, the median (across RCTs) of the mean age of patients enrolled in the acute treatment and secondary prevention RCTs. We assumed that the index VTE event on entry to the acute treatment and secondary prevention models was split between DVT and based on the proportion of non-fatal PE and DVT in the acute treatment population.

Venous thromboembolism: interventions

For each indication we compared first-line treatments for which we have sufficient evidence to estimate model parameters. There are seven comparators evaluated in the secondary prevention model (Table 5), four in the acute treatment model (Table 6) and four in each of the two primary prevention subpopulations (Table 7). Before the introduction of NOACs, standard practice70 for primary prevention was LMWH, and for acute treatment was LMWH and warfarin for at least 5 days, then continue with warfarin only. In secondary prevention, NICE guidance71 recommends that clinicians, after discussion with patients, consider extending warfarin therapy beyond 3 months if the risk of VTE recurrence is high and there is no additional risk of major bleeding. However, NICE also acknowledged the need for further research to establish the cost-effectiveness of long-term anticoagulation after unprovoked VTE. In clinical practice, patients may be offered long-term anticoagulation after a second VTE event. Owing to this uncertainty about best practice, we compared all anticoagulants to a ‘no pharmacotherapy’ secondary prevention strategy in the base-case model. In a sensitivity analysis, we assumed that patients in this reference group would receive warfarin after a second VTE event. We assumed that all treatment will be stopped for subjects who have an ICH and that no other treatment switching occurs. This assumption differs from the AF population for whom treatment can be stopped or switched for other reasons (see Atrial fibrillation: interventions, above).

Outcomes of atrial fibrillation and venous thromboembolism models

We present results on total costs and quality-adjusted life-years (QALYs), both discounted at 3.5%. We present a probabilistic analysis, for which model parameters are given probability distributions to reflect uncertainty in their values. 72 We summarised the results with the expected costs, expected QALYs and expected net monetary benefit (NMB) for a range of willingness to pay per additional QALY gained (where expected values are an average over the joint distribution of the model parameters). NICE has a stated willingness-to-pay threshold of £20,000–30,000 per QALY. 73

Uncertainty in the model input parameters is captured using simulation [Monte Carlo simulation for parameters with assumed distributions, and Markov chain Monte Carlo (MCMC) simulation for parameters estimated from the NMA]. We represent decision uncertainty using the cost-effectiveness plane, cost-effectiveness acceptability curves (CEACs), and cost-effectiveness acceptability frontiers (CEAFs). The cost-effectiveness plane plots incremental effects (QALYs) against incremental costs for each simulation sample. The CEAC plots the proportion of the simulation samples where each strategy had the highest net benefit (i.e. was most cost-effective) against willingness-to-pay-per-QALY threshold. These proportions are estimates of the probability that the treatment is the most cost-effective. If this probability is close to one for a particular treatment, this suggests very little uncertainty as to the most cost-effective treatment, whereas if it is low the choice of most cost-effective treatment is uncertain. This allows decision-makers to identify interventions that are unlikely to be cost-effective at any plausible threshold and to judge how sensitive treatment choice is to the amount that the NHS is willing to pay for a QALY. The CEAC is not robust when there is a treatment with a high degree of uncertainty in net benefit, giving high probabilities of being both most cost-effective and least cost-effective. For this reason the CEAF has been proposed. 74 This plots, for each willingness-to-pay threshold, the probability of being most cost-effective only for the treatment with the highest expected net benefit at that willingness-to-pay threshold.

We use value of information (VOI) methods to explore how sensitive the optimal treatment is to uncertainty in the model inputs, and guide research recommendations. We estimate the expected value of perfect information (EVPI) and the expected value of partial perfect information (EVPPI). EVPI and EVPPI measure the expected improvement to our decision-making (in monetary units) if we were to eliminate uncertainty in all (EVPI) or some (EVPPI) of the model input parameters. We present EVPI per person per year and also per population over 10 years discounted at 3.5%, for given annual incidence for each of our populations. EVPPI for subsets of parameters are computed using the Sheffield Accelerated Value of Information (SAVI) version 2.0.9 (University of Sheffield, Sheffield, UK) web application. 75,76 This method gives only approximate results, which can be interpreted as indicative of the relative sensitivity of the decision to different groups of parameters.

Venous thromboembolism model structure: secondary prevention

A Markov model with half-cycle correction78 was used to evaluate the cost-effectiveness of prophylaxis in patients who have experienced a previous non-fatal VTE event (Figure 4). The model has a cycle length of 1 year and includes eight health states (Table 8). Subjects enter the model in post PE or post DVT. Subjects in the ‘post DVT’ (or ‘post PE’) state can have an additional non-fatal DVT (or PE) event with a transient utility decrement and cost, but remain in the same health state. Subjects in the ‘post DVT’ state who experience a non-fatal PE and subjects in ‘post PE’ who experience a non-fatal DVT transition to ‘post PE DVT’. Subjects in the ‘post DVT’ and ‘post PE DVT’ states can develop post-thrombotic syndrome (PTS) and transition to either ‘mild/moderate PTS’ or ‘severe PTS’. Subjects who have had a PE may experience chronic thromboembolic pulmonary hypertension (CTPH). Subjects can transition to CTPH from ‘post PE’ and ‘post PE DVT’.

FIGURE 4.

Venous thromboembolism secondary prevention Markov model. Nodes represent the health states; lines between nodes represent possible transitions; all health states can transition to death. ICH, other clinically relevant bleeds, DVT and PE are acute events, which may lead to a change in chronic health state (e.g. post ICH).

All subjects who are receiving treatment can transition to the ICH health state. After entering this state, we assumed that anticoagulation therapy will be stopped and subjects will remain there until death, as this is considered to be the state with the lowest quality of life.

Venous thromboembolism model structure: acute treatment

The acute treatment of symptomatic VTE was modelled using a decision tree covering the first 6 months of therapy, in line with current guidelines for the duration of acute treatment (Figure 5). There is a probability that patients will experience recurrent symptomatic VTE during the acute treatment period and, regardless of VTE recurrence, all patients are at risk of other CRB or ICH. Longer-term costs and outcomes following acute treatment were estimated using the secondary prevention Markov model for patients who are alive at the end of acute treatment.

FIGURE 5.

Venous thromboembolism acute treatment decision tree. At the end of each branch in the decision tree, patients progress to the secondary prevention model. ICH branches enter in ‘post ICH’ state; treated symptomatic DVT (with bleed or no AE) will enter the post DVT state; treated symptomatic PE (with bleed or no AE) will enter the post PE state; recurrent symptomatic VTE post DVT will enter the post DVT or post DVT PE state, depending on the previous event; and recurrent symptomatic VTE post PE will enter the post PE or post DVT PE state, depending on the previous event.

Venous thromboembolism model structure: primary prevention

The primary prevention model consists of a decision tree covering the first 180 days of prophylactic anticoagulation (Figure 6). After this initial period, the long-term cost and outcomes of patients who do not have a symptomatic VTE are tracked using a two-state Markov model (Figure 7). This Markov model has two health states: no VTE/asymptomatic VTE and dead. The Markov model has a lifetime time horizon and yearly cycles. The longer-term costs and outcomes of patients who have a symptomatic VTE are tracked in the acute treatment model (see Figure 5) and the secondary prevention model (see Figure 4).

FIGURE 6.

Primary prevention decision tree. At the end of the decision tree, subjects will have experienced a symptomatic VTE or not; if they have then they will enter the acute treatment model. Those who did not experience a symptomatic VTE will enter the two-stage Markov model.

FIGURE 7.

Primary prevention Markov model.

Patients enter the primary prevention model after having elective surgery (TKR or THR). They then experience either a symptomatic VTE event or no VTE/asymptomatic VTE. Patients who experience a symptomatic event have a fatal PE, non-fatal PE or DVT, and are treated. Regardless of VTE incidence, all patients are at risk of another CRB during the initial 90-day period of anticoagulation. Because treatment duration is short for primary prevention, the risk of ICH is very low. and there is no evidence of a relative effect of NOACs compared with LMWH in this patient population. Therefore, we have not incorporated ICH in the primary prevention model.

Cost of pharmacotherapy

Average drug costs were based on the British National Formulary (BNF) March 2015 update,67 using the most economical pack size (Tables 9–11). Edoxaban does not currently have a list price in the UK. For the base case we assume that the 6-monthly cost is equivalent to dabigatran. We tested this assumption in a sensitivity analysis. As all of the NOACs are taken orally, it was assumed that there are no administration or monitoring costs, following the costing report in AF of Ali et al. 79 Average drug and monitoring cost of warfarin comes from a costing report by NICE68 and is cited in the study by Kansal et al. 42 The cost of LMWH was an average over all of the LMWHs included in the meta-analyses and listed in the BNF.

The unit costs of drugs are assumed to be fixed and known, so that point estimates – rather than distributions – are entered into the models. However, the administration and monitoring cost of warfarin is uncertain, and in the absence of other information we assumed a uniform distribution ranging from 50% to 150% of the estimated cost from the NICE costing report. 68 We performed a sensitivity analysis for the assumed cost of warfarin monitoring.

Cost of acute venous thromboembolism, atrial fibrillation and anticoagulant-related events

All costs of acute and chronic care were inflated to 2013–14 values using the Office for National Statistics (ONS) Consumer Price Inflation Index for Medical Services (DKC3). 80 Acute management costs for SE, MI, TIA, DVT, PE and CRB come from the 2012–13 NHS reference costs. 81 The reference costs for MI account for only direct hospitalisation; we assumed that total costs would be double this amount to account for follow-up costs. 82 The cost of a sudden fatal PE is assumed to be zero and the patients who have a non-fatal PE are assumed to accrue the full cost of a PE. Acute management costs for ischaemic stroke and ICH come from a study of patients with AF on a UK stroke registry. 83 Normal distributions are assumed for the mean acute costs, with SDs defined by the standard errors of the source data (Table 12).

Cost of chronic care for venous thromboembolism, atrial fibrillation and anticoagulant-related events

Long-term management costs of stroke (ischaemic stroke or ICH) also come from the UK stroke registry83 (Table 13). This registry83 stratified the severity of ischaemic strokes by disability (non-disabling, moderately disabling, totally disabling) and we averaged their annual costs and SDs, weighted by the number of events. As in the study by Kansal et al. ,42 we assumed the same cost for ICH as for ischaemic stroke. Normal distributions are assumed, with SDs defined by the standard errors of the source data. For states with a history of multiple events, we assumed that the additional post-event management costs were the maximum of the management costs for the constituent events. We divided sampled costs by four to obtain 3-monthly cycle costs.

Costs for mild to moderate and severe PTS have previously been estimated in a NICE technology appraisal,84 which looked at the clinical effectiveness and cost-effectiveness of dabigatran for the prevention of VTE after a TKR or TKR in adults. This study84 converted and inflated costs from a US economic burden study85 of long-term complications of primary prevention of DVT after a THR. This study84 estimated the cost of mild to moderate PTS to be £541 for the first year and £220 for subsequent years, and severe PTS to be £2461 for the first year and £602 for subsequent years. Inflating to 2013–14 values resulted in a cost of £689 for the first year and £280 for subsequent years for mild/moderate PTS, and £3136 for the first year and £767 for subsequent years of severe PTS. NICE guidance for the management of VTE71 estimated a 4-weekly cost of CTPH to be £2173, equivalent to an annual cost of £33,028 in 2013–14 prices.

Utilities

The AF and VTE models used utility weights combined with survival to estimate QALYs. Utility weights are anchored at 1 (best health) and 0 (as bad as death), such that a year spent in an intermediate health state with a utility weight of 0.5 would be considered equivalent to 6 months in the best health state with a utility value of 1. The models have a number of acute health events that affect patients for a short period, followed by a partial or full recovery and a number of chronic health states from which patients do not recover. Several of these health events and health states are shared between the AF and VTE models.

Utilities were identified from a previous NICE technology appraisal submission on rivaroxaban33 and from a rapid literature review to identify quality-of-life studies in VTE. The rivaroxaban technology appraisal submission33 conducted a systematic literature search for evidence on EQ-5D (European Quality of Life-5 Dimensions) utility index in health states related to AF. For VTE events (DVT and PE), Locadia et al. 86 estimated health utilities, using time trade-off methods, from a cohort of 53 patients who had experienced a VTE event.

Utilities of venous thromboembolism, atrial fibrillation and anticoagulant-related acute health events

The acute health event disutilities for AF for other CRB, SE and TIA are reported in Table 14. The remaining acute health event disutilities for AF (acute ICH and acute MI; see Table 14) are obtained by subtracting ‘Stable AF’ from the utility of the event. For example, the disutility for MI would be 0.683 – 0.779 = –0.096.

These disutilities are capped above at 0. Further acute event utility values extracted from the literature for VTE primary prevention, acute treatment and secondary prevention models are reported in Table 15. When uncertainty estimates were reported, we assumed that mean utilities would be normally distributed, as indicated by the central limit theorem. When uncertainty estimates and sample sizes were not available (acute ischaemic stroke, TIA, SE), we assumed mean utilities to follow a uniform distribution ranging from 50% to 150% of the reported mean. The duration of the decrements for DVT and PE was assumed to be 6 months71 and 3 months for ICH, before moving to the post-ICH health state. 51 Duration of decrements was generally not reported for the AF disutilities, so they were assumed to last one cycle.

In both the AF and VTE model, to account for quality of life decreasing with age, all utility decrements were multiplied by the ratio of the utility for a given age range relative to a reference age (65–75 years), based on general population utilities estimated in Kind et al. 92 (Table 16). Utilities were also adjusted by gender in this way for the VTE models, whereas for the AF models all utilities were weighted averages across gender.

Utilities of venous thromboembolism, atrial fibrillation and anticoagulant-related chronic health states

In the AF model, for which patients can have more than one chronic health condition, utilities for chronic health states are assumed to be multiplicative. For example, the utility of a patient who has experienced both an ischaemic stroke and a MI will be the product of the two utility scores (see Table 14): 0.690 × 0.718 = 0.495.

Utilities are multiplied by 0.25 to get a QALY for 3-month cycle.

For the VTE-related chronic health states, we used estimates from the study by Lenert and Soetikno,89 who elicited preferences in 30 volunteers and 30 medicine physicians with mild/moderate PTS and severe PTS, and the study of Meads et al. ,93 who used the Cambridge Pulmonary Hypertension Outcome Review utility index94 to estimate a utility value for CTPH from 308 patients (see Table 14). Further chronic VTE-related chronic health-state utilities extracted from the literature are reported in Table 17.

Stroke or systemic embolism

Sixteen studies reported the number of stroke or SE events, and the other seven trials reported the number of stroke events, so that the resulting network was based on data from all 23 trials, comparing a total of 26 interventions (Figure 9). There were 3217 stroke or SE events. Twenty studies were included in the main analysis, with the remaining three included only in sensitivity analyses. The thicker lines joining interventions, which mainly correspond with comparisons between licensed doses of NOACs and warfarin (INR 2–3) represent the larger (mainly Phase III) trials. Similarly, the larger green circles represent the interventions to which the largest number of patients were randomised. Importantly, there were no direct comparisons between different NOACs, although there were numerous comparisons between different doses of the same NOAC in mainly Phase II trials, and some such comparisons in larger trials. Therefore, comparisons between the effects of different NOACs need to be inferred from the network (indirect evidence).

FIGURE 9.

Network plot for stroke or SE (stroke prevention in AF). a, Excluded interventions that were included in sensitivity analyses.

Table 24 shows risk-of-bias judgements for studies reporting stroke or SE. The studies were at mixed risks of bias: there were concerns about lack of blinding of participants for most trials, and about lack of allocation concealment and blinding of outcome assessment in some.

Table 25, which shows comparisons of licensed doses with warfarin (INR 2–3), suggests that both low- and high-dose antiplatelet drugs increase the risk of stroke or SE compared with warfarin (INR 2–3). Among NOACs, there was some evidence that apixaban [5 mg bd (bd)], dabigatran (150 mg bd), edoxaban (60 mg od) and rivaroxaban (20 mg od) reduce the risk of stroke or SE compared with warfarin (INR 2–3). Most other comparisons were imprecisely estimated. Comparisons among licensed doses of NOACs were almost all based on indirect evidence (Table 26). Among the comparisons that were not classified as imprecisely estimated, there was some evidence that edoxaban (60 mg od) and rivaroxaban (20 mg od) increase the risk of stroke or SE compared with dabigatran (150 mg bd).

Results from a supplementary analysis taking into account the differences in duration of follow-up within and between trials, and the differences in the definition of event used across trials (e.g. total number of events vs. first events only), are presented in Tables 27 and 28. They are very similar to those for ORs.

As a post hoc sensitivity analysis, we fitted a fixed-effects meta-regression model using the mean TTR for warfarin patients (see Table 19) as a covariate and the mean log-odds ratio (log-OR) from each pairwise comparison (with warfarin as the reference category) as the response variable. There was little evidence of effect modification due to mean TTR (estimated coefficient 0.0021 with 95% CI −0.07 to 0.08 per 1% increase). The model fit indices were very similar with and without the covariate.

Ischaemic stroke

Fourteen studies reported on 2228 ischaemic stroke events, leading to a connected network comparing a total of 15 interventions (Figure 10). Twelve studies were included in the main analysis, with the remaining two included only in sensitivity analyses. The studies were at mixed risks of bias (Table 29). There were concerns about lack of blinding of participants for most trials, and about lack of allocation concealment and blinding of outcome assessment in one trial (AF-ASA-VKA-CHINA,135 only included in sensitivity analyses due to implementation of warfarin within non-standard INR range).

FIGURE 10.

Network plot for ischaemic stroke (stroke prevention in AF). a, Excluded interventions that were included in sensitivity analyses.

Table 30, which shows comparisons of all interventions with warfarin (INR 2–3), suggests that both low- and high-dose antiplatelets increase the risk of ischaemic stroke compared with warfarin (INR 2–3). Among NOACs, there was some evidence that dabigatran (150 mg bd) reduces the risk of ischaemic stroke compared with warfarin, whereas edoxaban (30 mg od) increases that risk. There was little evidence that the risk of ischaemic stroke differed between licensed doses of NOACs (Table 31).

In a sensitivity analysis to take into account the differences in duration of follow-up, NMA results were as presented in Tables 32 and 33, and show very similar results.

Myocardial infarction

A total of 15 studies reported 1334 MI events, leading to a network of 16 interventions (Figure 11). Thirteen studies were included in the main analysis, with the other two included only in sensitivity analyses. The studies were at mixed risks of bias (Table 34). There were concerns about lack of blinding of participants for most trials, and about lack of allocation concealment and blinding of outcome assessment in some.

FIGURE 11.

Network plot for MI (stroke prevention in AF). a, Excluded interventions that were included in sensitivity analyses.

Table 35 shows weak evidence that dabigatran (110 mg bd), dabigatran (150 mg bd) and edoxaban (30 mg od) increase the risk of MI compared with warfarin (INR 2–3), and weak evidence that rivaroxaban (20 mg od) decreases risk of MI compared with warfarin (INR 2–3). None of the interventions was superior or inferior to warfarin (INR 2–3). The pairwise comparisons of licensed NOACs, presented in Table 36, show weak evidence that dabigatran (150 mg bd) increases the risk of MI compared with apixaban (5 mg bd), and evidence that rivaroxaban (20 mg od) reduces the risk of MI compared with dabigatran (150 mg bd). Results were similar in a sensitivity analysis, taking into account the differences in duration of follow-up within and between trials, and the differences in the definition of event used across trials (e.g. total number of events vs. first events only) (Tables 37 and 38).

Major bleeding

Eighteen studies reported 4314 major bleeding events, leading to a network of 24 interventions (Figure 12). Seventeen studies were included in the main analysis, with the remaining study included only in sensitivity analyses. These studies were at mixed risks of bias (Table 39). There were concerns about lack of blinding of participants for most trials, and about lack of allocation concealment and blinding of outcome assessment in some.

FIGURE 12.

Network plot for major bleeding (stroke prevention in AF). a, Excluded interventions that were included in sensitivity analyses.

There was weak evidence that antiplatelet therapy (< 150 mg od) reduced major bleeding compared with warfarin (INR 2–3). There was evidence that apixaban (5 mg bd), dabigatran (110 mg bd), edoxaban (30 mg od) and edoxaban (60 mg od) reduced major bleeding risk compared with warfarin (INR 2–3) (Table 40). Comparisons among licensed doses of NOACs, presented in Table 41, suggest that dabigatran (150 mg bd) increases risk of major bleeding compared with apixaban (5 mg bd), whereas rivaroxaban (20 mg od) increases risk of major bleeding compared with apixaban (5 mg bd) and edoxaban (60 mg od).

In a sensitivity analysis to take into account the differences in duration of follow-up, NMA results were as presented in Tables 42 and 43, and show very similar results. Another sensitivity analysis involved fitting a fixed-effects meta-regression model using the mean TTR for warfarin patients (see Table 19) as a covariate and the mean log-OR from each pairwise comparison (with warfarin as the reference category) as the response variable. We found no evidence of an effect modification according to mean TTR (estimated coefficient 0.04 with 95% CI −0.03 to 0.12 per 1% increase). The model fit indices yielded almost identical values for the models with and without the covariate.

Clinically relevant bleeding

Twelve studies reported 9556 CRB events, leading to a network of 23 interventions (Figure 13). Eleven studies were included in the main analysis, with the remaining study included only in sensitivity analyses. These studies were at mixed risks of bias (Table 44), the concerns being due to lack of blinding of participants for most trials.

FIGURE 13.

Network plot for CRB (stroke prevention in AF). a, Excluded interventions that were included in sensitivity analyses.

Results presented in Table 45 suggest that antiplatelet therapy (< 150 mg od) reduces CRB compared with warfarin (INR 2–3). Note that the licensed dose for of antiplatelet therapy for AF is ≥ 150 mg od: no studies provided data for that dose for CRB. Among NOACs, there was evidence that apixaban (5 mg bd), edoxaban (30 mg od) and edoxaban (60 mg od) reduce CRB compared with warfarin (INR 2–3). However, edoxaban (30 mg bd) and edoxaban (60 mg bd) increased CRB compared with warfarin (INR 2–3). Among licensed NOACs (Table 46), there was evidence that edoxaban (60 mg od) and rivaroxaban (20 mg od) increase CRB compared with apixaban (5 mg bd) and that rivaroxaban (20 mg od) increased CRB compared with edoxaban (60 mg od).

Supplementary NMAs of HRs rather than ORs show very similar results (Tables 47 and 48).

Intracranial bleeding

Eight studies reported a total of 757 intracranial bleeds, leading to a network of 10 interventions (Figure 14). Seven trials were included in the primary analysis, with the remaining study included only in sensitivity analyses. These studies were at mixed risks of bias (Table 49), the concerns being due to lack of blinding of participants and, in one study, lack of blinding of outcome assessment.

FIGURE 14.

Network plot for intracranial bleeding (stroke prevention in AF). a, Excluded interventions that were included in sensitivity analyses.

There was strong evidence that apixaban (5 mg bd), dabigatran (110 mg bd), dabigatran (150 mg bd), edoxaban (30 mg od), edoxaban (60 mg od) and rivaroxaban (20 mg od) reduced risk of intracranial bleeding compared with warfarin (INR 2–3) (Table 50). For each of these NOAC doses except rivaroxaban (20 mg od), the estimated reduction in risk was > 50%. There was weak evidence that risk of intracranial bleeding was increased for rivaroxaban (20 mg od) compared with apixaban (5 mg bd), dabigatran (150 mg bd) and edoxaban (60 mg od) (Table 51). Analysing HRs rather than ORs led to similar results (Tables 52 and 53).

All-cause mortality

Eighteen studies reported 6479 all-cause mortality events, leading to a network of fifteen interventions (Figure 15). Fifteen studies were included in the primary analysis, with the remaining three studies included in sensitivity analyses. These studies were at mixed risks of bias (Table 54). There were concerns about lack of blinding of participants for most trials, and about lack of allocation concealment and blinding of outcome assessment in some studies.

FIGURE 15.

Network plot for all-cause mortality (stroke prevention in AF). a, Excluded interventions that were included in sensitivity analyses.

Table 55 suggests that all NOAC doses with comparisons that were not imprecisely estimated [apixaban (5 mg bd), dabigatran (110 mg bd), dabigatran (150 mg bd), edoxaban (30 mg od), edoxaban (60 mg od) and rivaroxaban (20 mg od)] were associated with a reduced risk of all-cause mortality compared with warfarin (INR 2–3). There was little evidence that the risk of all-cause mortality differed between licensed doses of NOACs (Table 56). Analysing HRs rather than ORs produced similar results (Tables 57 and 58).