Implantable cardiac monitors to detect atrial fibrillation after cryptogenic stroke: a systematic review and economic evaluation

Population: cryptogenic stroke or transient ischaemic attack

Stroke is the third most common cause of premature death in the UK2 and a major cause of preventable disability. 3 Improvements in care have greatly improved mortality and morbidity over the last two decades, but there are still around 30,0002 stroke-related deaths each year in England, and around one-quarter of patients leave hospital with moderate to severe disability. 4

Strokes and transient ischaemic attacks (TIAs) are caused by the interruption of the blood supply to part of the brain, either due to the narrowing or blockage of a blood vessel by a blood clot (ischaemic stroke), or due to a bleed from a blood vessel in the brain (haemorrhagic stroke). The main difference between a stroke and a TIA is that the symptoms caused by damage to brain tissue from a TIA resolve within 24 hours, whereas in an untreated stroke, the symptoms last for longer. Common symptoms of stroke include numbness, weakness or paralysis, slurred speech, blurred vision, confusion and severe headache. 5

The causes of stroke are manifold and include the build-up of plaque in the artery supplying the ischaemic region of the brain (atherosclerosis), occlusion of small arteries deep in the brain (lacunar) and a clot (embolism), which often originates in the heart and travels to a blood vessel in the brain (e.g. as a result of AF). Other less common causes of stroke are tumours in the heart, heart abnormalities, recent myocardial infarction (MI), migraine, malignancy and drug misuse. 6 However, up to one-third of first-time strokes are cryptogenic, meaning no known cause can be identified, which is most common in younger patients. 6 The Evidence Assessment Group’s (EAG’s) clinical experts reported that patients in the UK who have had a stroke will generally undergo a series of tests to identify a cause before the event is classed as cryptogenic, although some definitions include insufficient testing or identification of more than one cause. 6,7 Diagnostic tests to identify the cause of stroke generally include blood tests, inpatient electrocardiography, echocardiography and Doppler ultrasonography of the carotid arteries.

In 1994, the Oxfordshire community stroke project reported a stroke recurrence rate of ≈30% by 5 years and that people are at highest risk of a subsequent stroke in the first year, when mortality rates are also at their highest. 8 However, a systematic review9 from 2011 suggests that there is a temporal reduction in the 5-year risk of stroke recurrence from 32% to 16.2% across its included studies and that risk of stroke recurrence increases over time, with higher rates at 10 years than at 30 days post stroke. It should also be noted that the review authors reported substantial heterogeneity in the included studies and recurrence rates varied depending on the definition of stroke applied. 9 Establishing the cause of a stroke is paramount to decrease the risk of recurrence by selecting appropriate preventative care. 10

Current pathway of care

The EAG’s clinical experts reported that there is no standard guideline on the diagnostic tests required in the UK to further investigate patients who have had a CS or TIA for underlying AF and there is no consensus on the duration or mode of monitoring for AF. The National Institute for Health and Care Excellence (NICE)’s guideline on stroke and TIA in those aged > 16 years (NICE guideline 128)5 was updated in May 2019 and provides no specific recommendations on the diagnosis of AF in people who have had an acute stroke.

The NICE guideline on AF19 recommends that people with asymptomatic suspected paroxysmal AF undetected by standard electrocardiography recording have a 24-hour ambulatory electrocardiographic monitor, although this recommendation is not specific for patients with CS or TIA. The European Society of Cardiology Guidelines20 for the management of AF recommend that patients with ischaemic stroke (IS) or TIA are investigated for AF using a short-term electrocardiography recording and then continuous electrocardiographic monitoring for a minimum of 72 hours.

The EAG’s clinical experts reported that patients with a CS or TIA diagnosis will typically have short-term electrocardiography as an inpatient to detect cardiac arrhythmias, such as AF, as part of the standard suite of diagnostic tests to identify the cause of stroke or TIA. Patients with no AF during inpatient monitoring will often receive outpatient external ambulatory electrocardiographic monitoring for 24–48 hours (e.g. using a Holter monitor). Clinical experts reported that, in some areas, this may be extended to 2 weeks or even 30 days of monitoring depending on local practices and patient–clinician preferences. Clinical experts reported that ICMs are not routinely used in UK clinical practice for AF detection after CS or TIA and that they are likely to be used in the NHS only after patients have received an initial period of at least 24 hours’ external ambulatory monitoring.

Patients with AF detected after stroke or TIA can be treated to reduce the risk of a further stroke. NICE recommendations19 for stroke prevention therapy include rate or rhythm control and anticoagulation based on bleeding risk and CHA₂DS₂-VASc score {Congestive heart failure (or left ventricular systolic dysfunction), Hypertension [blood pressure consistently > 140/90 mmHg (or treated hypertension on medication)], Age ≥ 75 years (doubled), Diabetes mellitus, prior Stroke or transient ischaemic attack or thromboembolism (doubled), Vascular disease (e.g. peripheral artery disease, myocardial infarction, aortic plaque), Age 65–74 years, Sex category (e.g. female)}. CHA₂DS₂-VASc is measure of stroke risk in patients with AF based on age; sex; and history of congestive heart failure, stroke or TIA, vascular disease and diabetes. 21 Patients with prior stroke or TIA have a minimum CHA₂DS₂-VASc score of 2 and automatically qualify for anticoagulation according to current NICE guidance,19 regardless of the presence of other stroke risk factors. The NICE pathway for preventing stroke in people with AF22 recommends anticoagulation with apixaban (Eliquis^®, Bristol-Myers Squibb Company, New York, NY, USA), dabigatran etexilate (Pradaxa^®, Boehringer Ingelheim, Ingelheim am Rhein, Germany), edoxaban (Lixiana^®, Daiichi Sankyo Company Ltd, Tokyo, Japan), rivaroxaban (Xarelto^®, Bayer AG, Leverkusen, Germany) or a vitamin K antagonist, and the NICE guideline for AF management19 recommends review at least annually, and recommends against aspirin monotherapy. If anticoagulation is contraindicated because of bleeding risk, NICE recommends rate or rhythm control measures, annual review to assess stroke and bleeding risk, and consideration for left atrial appendage occlusion. 19 Clinical experts reported that patients who have had a CS and are diagnosed with AF during follow-up with an ICM are most likely to have paroxysmal AF, for which the management would usually be anticoagulation. Clinical experts also reported that patients identified in advance as being unsuitable for anticoagulation, for example because of their risk of bleeding, may not receive an ICM. However, clinical experts also reported that some patients diagnosed with AF may receive a left atrial appendage occlusion device as an alternative to OAC therapy.

Description of the technologies under assessment

Implantable cardiac monitors, also known as insertable cardiac monitors or implantable loop recorders, are small devices inserted beneath the skin of the chest. The devices allow extended monitoring and automatic recording of heart rhythm. The devices are inserted under local anaesthetic via a small incision and capture continuous electrocardiograms (ECGs) to detect various arrhythmias, including AF. ICMs are currently used in the NHS primarily as a method of monitoring patients experiencing syncope (fainting) to detect and treat underlying arrhythmias. The devices offer the possibility of continuous rhythm monitoring of people who have had a CS or TIA to increase the detection of intermittent or paroxysmal AF to help guide appropriate treatment for secondary stroke prevention.

The devices are usually inserted by cardiologists, cardiac physiologists and nursing staff in a sterile environment such as a catheterisation laboratory (hereafter referred to as cath lab), but clinical experts report that there is variation across devices and with the ICM experience of the service in which the patient is being treated. Devices can be explanted once an arrhythmia has been detected or at the end of the battery life, but can also be left in situ. Adverse events (AEs) are rare, but can include infection or reaction at the insertion site, bleeding, excessive fibrotic tissue growth, extrusion, hematomas or cysts, keloid formation, and erosion or migration of the device.

Once implanted, the devices automatically capture continuous ECGs, and record and transmit detected arrhythmia episodes for clinical review. Recording of episodes can also be activated manually by the patient if symptoms occur using optional external handheld patient devices or smartphone applications (hereafter referred to as ‘apps’), depending on the ICM. Detection parameters, data storage, method of data transmission and notification settings vary by device (Table 1), but all have capabilities to recognise a range of arrhythmias and alert clinicians when an episode is detected. Data are transmitted via internet or cellular networks and encrypted for online storage. Clinical experts reported that programming of ICMs in relation to use of inbuilt automatic programmes varies depending on the patient characteristics and clinician preference. The clinical experts reported that, often, the ICMs’ standard setting for arrhythmia detection in CS patients is used to start with and this is then adjusted as necessary. The clinical experts also reported that the patient activator device is generally of little benefit if used in CS patients, as they are generally asymptomatic in terms of AF and other cardiac arrhythmias.

Characteristics of the three ICMs included in the NICE scope1 – BioMonitor 2-AF23 Confirm Rx24 and Reveal LINQ25 – are summarised in Table 1. The EAG has also included information about the Reveal XT device, which is an earlier Medtronic model, because it was the device used in the only randomised controlled trial (RCT) identified in the clinical evidence search. Earlier Biotronik and Abbott ICM models are also available, but have not been included because no relevant evidence in the CS or TIA population was submitted by the companies, and the capabilities of these earlier models were not considered relevant to the decision problem. However, it should be noted that some data on the Confirm DM202 in a non-CS population is discussed in Chapter 3, in the absence of data on the Confirm Rx in a CS or non-CS population.

Comparators and the reference standard

The diagnostic accuracy and clinical outcomes of ICMs are considered for patients who have had a CS or TIA in whom no AF has been detected following a minimum of 24 hours of external electrocardiographic cardiac monitoring. The clinical outcomes for ICMs (after a minimum of 24 hours of external electrocardiographic monitoring) will be compared with no further monitoring (also after a minimum period of 24 hours of external electrocardiographic monitoring). The diagnostic test accuracy (DTA) of the ICMs will be compared with 24-hour external ambulatory electrocardiographic monitoring or other commonly used electrocardiographic monitoring regimens, such as 7-day Holter monitoring, which is the reference standard. External electrocardiographic monitoring is most commonly conducted with a Holter monitor, a portable battery-operated device that records continuous ECGs, usually for 24–48 hours, via electrodes that attach to the skin.

Study selection and data extraction

The titles and abstracts of all identified studies from the electronic database searches were independently assessed for inclusion by two reviewers to identify the potentially relevant full-text articles to be retrieved. Full-text copies of the selected studies agreed for inclusion after title and abstract screening were obtained and all full-text articles were again assessed independently by two reviewers for inclusion using the eligibility criteria outlined in Eligibility criteria. Any disagreements were resolved by discussion; it was not necessary to consult with the third reviewer.

Data for the comparative studies were extracted independently by two reviewers using a standardised data extraction form. Data for five of the single-arm and observational studies were extracted independently by two reviewers to pilot the data extraction form. After agreeing the final data extraction form, one reviewer completed the data extraction for the remaining studies and the second reviewer validated 25% of the included studies. Information extracted included details of the study’s design and methodology, intervention and comparator tests, reference standard, baseline characteristics of participants, and outcome measures, including clinical outcome efficacy and any AEs (see Appendix 3). If there was incomplete information, attempts were made to contact authors with a request for further details. Discrepancies in the data extraction were resolved by discussion, and a third reviewer was available if necessary, although they were not required.

Quality assessment

The quality of included comparative studies has been independently assessed by two reviewers and any differences were resolved by consensus with a third reviewer who was consulted if necessary. The included RCT was assessed according to recommendations by the CRD14 and the Cochrane Handbook for Systematic Reviews of Interventions,18 and recorded using the Cochrane Risk of Bias 2.0 tool. 35 The observational studies were not quality assessed as the majority of them were single-arm studies and there is no standardised quality assessment tool suitable for assessing single-arm clinical effectiveness studies. It should also be noted that their results are reported only narratively or in tables (no evidence synthesis was conducted using them). There were no diagnostic accuracy studies in CS patients included; therefore, quality assessment with the quality assessment of diagnostic accuracy studies–2 (QUADAS-2) tool36 was not required.

Methods of analysis and evidence synthesis

Details of results on clinical effectiveness and quality assessment for each included study are presented in structured tables and as a narrative summary. There were insufficient clinically and methodologically homogenous data available to enable data to be pooled and meta-analysed. Clinical and methodological heterogeneity were investigated and are discussed narratively.

For test accuracy data, positive predictive values (PPVs), negative predictive values (NPVs), sensitivity values and specificity values, with 95% confidence intervals (CIs) are presented for each study, when available.

The CRYSTAL-AF trial details

The CRYSTAL-AF37 trial was an open-label, parallel-group RCT sponsored by the company, Medtronic. There were various conflicts of interest relating to the authors of the different publications of the trial, including employment, grants and personal fees from Medtronic. The EAG also noted that the CRYSTAL-AF trial formed the basis of the clinical data in the company submission from Medtronic for this DAR, despite CRYSTAL-AF being a trial of the Reveal XT, a predecessor model of the Reveal LINQ, the model under review in this DAR. The differences between the two models are discussed in Chapter 1 and data provided by the company on the DTA of the two devices (albeit not from an exclusively CS population) is discussed in Chapter 3, Medtronic.

In the CRYSTAL-AF trial, patients were randomised 1 : 1 to receive the Reveal XT ICM or conventional follow-up care. Details of the follow-up received by both groups is reported in Table 2. Randomisation was stratified in the trial groups according to the type of index event (stroke or TIA) and the presence or absence of a patent foramen ovale (PFO). The EAG’s clinical experts reported that the rationale for stratification by PFO is likely to be because its presence is associated with CS. There is no known difference in the incidence of AF in patients with TIA compared with patients with stroke as their index event, although clinical experts considered it reasonable for it to also be applied as a stratification factor.

Patients were enrolled to the CRYSTAL-AF trial between June 2009 and April 2012 from 55 centres in 14 countries across Europe, Canada and the USA. The study closure was planned to be at 12 months after the last patient was randomised, with the primary study follow-ups scheduled at 6 and 12 months. The study inclusion criteria were as follows:

• A recent episode of cryptogenic symptomatic TIA or a recent episode of cryptogenic IS; recent was defined in a protocol amendment as from 60 to 90 days prior to enrolment. TIAs were required to have a visible lesion on a magnetic resonance imaging (MRI) or computerised tomography (CT) scan that fitted the symptoms of the TIA, and associated speech problems, or weakness of arm or leg, or hemianopia.

• The patient or their legally authorised representative had to be willing to sign a patient consent form.

• The patient had to be aged ≥ 40 years.

The definition of a CS in the CRYSTAL-AF trial was that no possible cause could be determined despite extensive work-up according to the standard protocol of the participating study centre. Before randomisation, the following clinical tests were required to establish the diagnosis of CS:

• A MRI or CT scan.

• 12-lead electrocardiography for AF detection.

• 24-hour electrocardiographic monitoring for AF detection and premature atrial contraction analysis (e.g. Holter monitoring).

• Transoesophageal echocardiography (TOE).

• Computed tomography angiography (CTA) or magnetic resonance angiography (MRA) of the head and neck to rule out other causes of stroke pathologies. A later protocol amendment allowed ultrasonography of cervical arteries and transcranial Doppler ultrasonography of intracranial vessels, in place of MRA or CTA of the head and neck in patients aged > 55 years.

The EAG’s clinical experts reported that the tests required in the CRYSTAL-AF trial to define CS were broadly consistent with the tests expected to be conducted in England. The clinical experts also reported that there are standard blood tests that would be required as part of the diagnostic work-up, and that all patients should receive transthoracic echocardiography prior to TOE; a small minority of patients may not receive TOE because of its invasive nature, but they may still be classified as having had a CS and go on to have an ICM.

The actual pre-enrolment screening for AF in the CRYSTAL-AF trial consisted of Holter monitoring, with a median duration of 23 hours [interquartile range (IQR) 21–24 hours] in 71.2% of patients {n = 314, mean 31.0 ± 66.7 hours [assumed to be the standard deviation (SD), although this was not specified in the paper]}, and inpatient telemetry monitoring, with a median duration of 68 hours (IQR 40–96 hours) in 29.7% of patients [n = 131, mean 74.6 ± 51.4 hours (assumed to be the SD, although this was not specified in the paper)]. The EAG considers it important to highlight that, in the DAR protocol, it was specified that patients were required to have a minimum of 24 hours of outpatient external electrocardiographic monitoring to be diagnosed as having had a CS. The EAG notes that 29.7% of patients in the CRYSTAL-AF trial did not receive outpatient electrocardiographic monitoring and that even the patients who did receive the outpatient Holter monitoring did not necessarily receive it for a full 24 hours (median 23 hours).

The main exclusion criteria for the CRYSTAL-AF trial were a history of AF or atrial flutter, an indication or contraindication for permanent OAC therapy at enrolment, or an indication for a pacemaker or implantable cardioverter defibrillator (full exclusion criteria are presented in Sinha et al. 48). The EAG’s clinical experts reported that these exclusion criteria are as expected for a clinical trial and in keeping with what would be expected in clinical practice in England and Wales, with the exception of a recent history of MI; if left ventricular function remained good, then MI would not necessarily be a reason for not implanting an ICM device in CS patients in clinical practice in England and Wales.

In total, 447 patients were enrolled to the CRYSTAL-AF trial, although only 441 underwent randomisation, with 221 randomised to the ICM trial arm and 220 to the conventional follow-up arm. Only 208 randomised participants (94.1%) in the ICM arm received the ICM device; 5.4% of these had withdrawn from the trial by the 6-month follow-up assessment. Reasons for withdrawals are presented in Table 2; with the exception of cross-over, there were similar numbers of withdrawals between the two trial arms. In relation to cross-over, 2.7% of participants in the conventional follow-up arm received an ICM, whereas 5.4% of participants in the ICM arm received conventional follow-up. In addition, there was an issue relating to delayed implantation of the ICM device in 11.5% of participants, which may have affected the AF-detection results of Reveal XT in the CRYSTAL-AF trial. The possible impact of the withdrawals and delayed ICM implantation on the results is discussed further in The CRYSTAL-AF trial: quality assessment.

The standard scheduled follow-up for patients in both of the arms of the CRYSTAL-AF trial was follow-up visits at 1, 6 and 12 months, and every 6 months thereafter until trial closure, with unscheduled visits in the event of symptom occurrence or after the transmission of ICM data, if advised by the investigator. If patients reported AF, then source documentation was acquired for adjudication, when possible. As reported in Table 2, the number of patients who reached 36 months’ follow-up was low in both trial arms, although the numbers were balanced across the two arms (24 patients in each trial arm).

The primary efficacy outcome in the CRYSTAL-AF trial was the time to first detection of AF (lasting > 30 seconds) at 6 months’ follow-up and the secondary outcome was AF detection at 12 months’ follow-up. The rate of AF detection was estimated with the use of the Kaplan–Meier (KM) method and compared between groups on an intention-to-treat (ITT) basis with the use of a log-rank test. Participants were censored in the primary analysis at the time of death, trial exit or completion of 6 months of follow-up. Pre-planned subgroup analyses were age, sex, race or ethnic group, type of index event, presence or absence of PFO, and CHADS₂ score. As only the type of index event was relevant to the NICE final scope,1 the results for the other subgroups are not discussed in detail in this report; however, they are summarised in Diagnostic yield: atrial fibrillation detection rate.

The baseline characteristics of the randomised participants in the CRYSTAL-AF trial are presented in Table 3. The EAG notes that, although there were no significant differences between the trial arms at baseline (p < 0.05), there were some small baseline differences, for example in the distribution of participants with PFO and history of prior stroke. These differences were small and unlikely to be a result of any systematic issues with randomisation.

In terms of applicability of the patients in the CRYSTAL-AF trial to the equivalent patients in the UK who may be eligible for an ICM for AF detection following a CS, the EAG’s clinical experts reported that, as expected in a clinical trial, the patients in the CRYSTAL-AF trial were slightly younger than those likely to be eligible for an ICM after CS in the UK. In addition, clinical experts reported that if the CRYSTAL-AF criteria for cryptogenic TIA are used, then, possibly, a higher proportion of people who had a TIA would be expected to be eligible for an ICM in clinical practice, and estimated the proportion of people who had a TIA to be closer to 20% of the total ICM-eligible CS population. In addition, all patients would be expected to be on an antiplatelet agent. If patients are contraindicated to antiplatelets, they are likely to also be unsuitable for OAC (the treatment likely to be provided if AF is detected).

Device sensitivity and specificity

There were no data on the sensitivity or specificity of the Reveal XT reported in the identified CRYSTAL-AF trial publications. Information from the CRYSTAL-AF trial and advice from the EAG’s clinical experts indicate that alerts generated by an ICM will need to be reviewed by a clinician to confirm AF before the initiation of anticoagulation treatment, and so there are essentially no false positives with the ICM.

One study (Choe et al. )14 conducted simulations using the the CRYSTAL-AF trial data to establish the relative sensitivity of the Reveal XT compared with various simulated external monitoring strategies, including one-off 24-hour Holter monitoring and 30 days’ continuous Holter monitoring, assuming that the Reveal XT had a sensitivity of 100%. This study, along with its results, is discussed further alongside the observational studies, as it was not a RCT.

Diagnostic yield: atrial fibrillation detection rate

The AF detection rate at 6 months was the primary outcome of the CRYSTAL-AF trial. The definition of AF in the CRYSTAL-AF trial was an episode of irregular heart rhythm, without detectable P-waves, lasting > 30 seconds. However, AF episodes are detected by an ICM using an automatic algorithm that is based on R-wave interval variability detected within 2-minute analysis windows. 49,50 It is therefore possible that some AF episodes of between 30 seconds’ and 2 minutes’ duration may have been missed in the ICM arm because of the 2-minute analysis window of the ICM. 49,51 As a result, there was a potential discrepancy in the duration of episodes of AF between the ICM and conventional follow-up arms in the CRYSTAL-AF trial that potentially bias the results in favour of conventional follow-up. In addition, as discussed in The CRYSTAL-AF trial: quality assessment, the open-label nature of the CRYSTAL-AF trial may have resulted in bias in the conventional follow-up arm as the outcome assessor was aware of the intervention assignment and was able to influence the electrocardiography or other assessment of AF. The ICM arm was unlikely to be affected by bias relating to the outcome assessor as all episodes of AF that qualified for analysis were adjudicated by an independent committee. These factors should therefore be taken into consideration when interpreting the results for AF detection, along with the risk-of-bias assessment findings. However, it is unclear what the resulting direction of the potential biases would be on the results. For the 6-month and 12-month results, it is most probable that the bias would favour AF detection with conventional follow-up, although beyond 12 months it is much less certain what direction the bias would be because of the large number of people censored in the analyses.

The results for AF detection demonstrated a trend in favour of the ICM across all time points (Table 5). At 6 months, 8.6% of patients were diagnosed with AF in the ICM arm compared with only 1.4% of patients in the conventional follow-up arm. The number of patients with AF diagnosed had risen to 19.0% in the ICM arm at 36 months, compared to only 2.3% in the conventional follow-up arm; this is despite small numbers of patients followed up at 36 months. The estimated AF detection rates are therefore higher in the 36-month KM analysis because of the non-informative censoring of patients lost to follow-up (the AF detection rate was estimated as 30% in the ICM arm and 3% in the conventional follow-up arm).

Only one patient was diagnosed with AF beyond 12 months’ follow-up in the conventional follow-up arm, whereas in the ICM arm, a further 13 patients were diagnosed with AF (nine patients between 12 and 24 months and four patients between 24 and 36 months; see Table 5). These results would suggest that long-term monitoring with an ICM, such as the Reveal XT, is beneficial in detecting more cases of AF; thus, enabling the treatment of AF to help reduce the risk of a further stroke or TIA.

Atrial fibrillation detection with the ICM compared with conventional follow-up was reported to be consistent across all the prespecified subgroups in the CRYSTAL-AF trial (age, sex, race or ethnic group, index event, presence or absence of PFO and CHADS₂ score), with no significant interactions. In addition, it was reported that the subgroup analysis results at 12 months were consistent with those at 6 months. The EAG notes that the subgroup results by index event (i.e. stroke or TIA) suggest a higher incidence of AF in the ICM arm of the TIA subgroup than in the stroke subgroup, although it is also noted that the number of patients in the TIA subgroup was very small (21 patients in the ICM arm). The trend favouring ICM over conventional follow-up seen in the primary study results was consistent in both the TIA and stroke subgroups.

Diagnostic yield: detection of other cardiac pathologies

There were no results reported for the detection of other cardiac pathologies in the CRYSTAL-AF trial.

Atrial fibrillation

Anticoagulant use

Time to initiation of anticoagulants

There were no results reported for the time to initiation of anticoagulants in the CRYSTAL-AF trial.

Ease of use of devices for clinicians

There were no results reported for the ease of use of devices for clinicians in the CRYSTAL-AF trial.

Mortality

There were no results reported for mortality in the CRYSTAL-AF trial.

Further strokes or transient ischaemic attacks

Outcome data on recurrent stroke or TIAs during the CRYSTAL-AF trial follow-up were presented for the composite of recurrent stroke or TIA and demonstrated a non-significant trend in favour of fewer recurrent events in the ICM arm than in the conventional follow-up arm (p > 0.05) (Table 7). It should also be noted that, in the ICM arm, there were fewer recurrent strokes or TIAs than the number of patients with AF detected at each of the time points, whereas, in the conventional follow-up arm, there were higher numbers of recurrent stroke and TIA events than the number of patients diagnosed with AF at each time point. However, outcome data were not reported by intervention and diagnosis of AF, and so it is unclear whether the recurrent stroke or TIA events occurred in patients diagnosed with AF or in the undiagnosed subgroup.

Other thromboembolisms

There were no results reported for other non-stroke- or TIA-related thromboembolisms in the CRYSTAL-AF trial.

Heart failure

There were no results reported for the diagnosis of heart failure in the CRYSTAL-AF trial.

Adverse events

Health-related quality of life

The EuroQol 5-Dimensions (EQ-5D) tool was used to collect HRQoL data during the CRYSTAL-AF trial. However, these results are confidential and therefore cannot be reproduced by the EAG.

Acceptability of the devices to patients

There were no results reported, beyond the HRQoL data mentioned in the previous section, for the acceptability of the devices to participants in the CRYSTAL-AF trial.

Device sensitivity and specificity

None of the observational studies provided comparative DTA between a group of patients who were monitored with an ICM and a group that received standard monitoring. However, two studies14,15 used AF detection data for a group of patients who had had a CS who were monitored for AF with an ICM to estimate the sensitivity of intermittent monitoring strategies if the ICM is assumed to have a sensitivity of 100%. Choe et al. 14 used data from 168 patients who received the Reveal XT in the CRYSTAL-AF trial (those with adequate follow-up from the 221 randomised to the ICM group), and Ziegler et al. 15 used data from a large registry of patients with a Reveal LINQ device15 (n = 1247). Choe et al. 14 used a 30-second episode threshold and Ziegler et al. 15 used a 2-minute threshold, but both studies used the same technique of modelling episodes of AF detected by the ICM; repeated iterations (10,000) were run to estimate the number of patients whose AF would not have been detected had alternative intermittent monitoring strategies been used.

Based on the assumption that the ICMs had 100% sensitivity for AF after a CS, Table 10 shows the estimated sensitivity of other monitoring strategies from the model simulations. Ziegler et al. 15 found sensitivities of between 2.9%, from a single 24-hour Holter monitor, and 29.9%, from quarterly 7-day Holter monitoring, and results were similar in Choe et al. 14 based on the CRYSTAL-AF trial cohort. As a result, even the best-performing intermittent monitoring strategy detected less than one-third of the AF detected by the ICM.

Two other studies reported false-positive rates as the proportion of episodes detected by ICM algorithm that were not subsequently verified as AF by a clinician. Li et al. 57 reported a 79.7% false-positive rate from the Reveal LINQ and Israel et al. 74 reported that > 90% of detected episodes were not confirmed by manual review (Reveal XT and BioMonitor). In their response to queries about individual studies identified by the EAG, Medtronic emphasised that false positive rates vary considerably depending on the model of device, sensitivity configuration and episode detection threshold.

Time to atrial fibrillation diagnosis

Eighteen observational studies reported time from device insertion to AF detection: five with the Reveal LINQ, seven with Reveal LINQ or XT, five with the Reveal XT and one with Reveal XT or BioMonitor (Table 13). Overall, average follow-up ranged from 7 to 20 months and median time to first AF detection was highly variable, ranging from 21 to 217 days. Where reported, IQRs also indicate a high degree of variability within studies.

Anticoagulant use

In seven studies of Reveal LINQ and/or XT, uptake of OACs in patients detected with AF was consistently high (see Table 13). Most of the studies had small populations, but the evidence suggests that uptake of anticoagulation is in the region of 90% to 100% once AF is detected. Christensen et al. 65 (SURPRISE) reported the overall uptake of OACs regardless of whether AF was detected and did not report whether or not the 19 patients starting OACs included all 14 patients with AF.

Incidences of device failure and removal

Three studies of Reveal LINQ and/or XT52,65,72 reported the number of device removals during follow-up. Ritter et al. 72 (Reveal XT) offered removal to patients once AF was detected, but, for 18 out of 60 removals (30%), it was not clear if the removal was for this reason or other reasons such as tolerability or battery life. Christensen et al. 65 (Reveal XT) reported that the device was prematurely explanted in 5 of 87 (5.7%) patients (three owing to skin reactions and two owing to discomfort) and that the median time to removal was 45 days; a further three patients (3.4%) chose to have the device removed after > 1 year of monitoring without AF being detected. Asaithambi et al. 52 reported that, of the 234 patients implanted with Reveal LINQ, 5.1% of the ICMs were removed from patients who died or required palliative care, 2.6% were removed electively, 1.3% of patients were lost to follow-up and 0.9% of ICMs migrated or fell out.

Further strokes or transient ischaemic attacks

Six studies reporting recurrent stroke indicate that a minority of patients have recurrence in the first year after device implantation, although the data were variable (0–14.6%). The data for recurrent stroke or TIA in patients with AF suggest higher rates than in those without AF detected, but it is unclear how many of these strokes and TIAs in the AF patients occurred prior to the detection of AF. Table 14 summarises the recurrent strokes or TIAs reported in those observational studies.

Adverse events

In addition to the device removal data summarised in Incidences of device failure and removal, some of which related to tolerability, five studies reported AEs. Three studies of Reveal XT,69,71,72 one of Reveal LINQ and XT,62 and one of Reveal XT and BioMonitor74 reported that no complications of the procedure or insertion site were noted during follow-up.

Abbott

The information provided in the company submission by Abbott regarding the Confirm Rx was that the only relevant study was the Diabetes Cardiovascular Risk Evaluation: Targets and Essential Data for Commitment of Treatment (DETECT) AF study (Nölker et al. ). 75 The EAG notes that the patient group in the DETECT AF study was not restricted to CS patients; therefore, the EAG did not consider this study to meet the review inclusion criteria. In addition, the ICM device used in the DETECT AF study was the Confirm ICM, Model DM2102, whereas supporting documents for the company submission included a user guide for the Confirm Rx Model DM3500. The company clarified that Confirm DM2102 was an older and larger model of the Confirm Rx, which is the model specified in the NICE final scope. 1 The EAG is unsure how the firmware in the models differs, but, in the absence of any suitable clinical data for the Confirm Rx DM3500, the data for the Confirm DM2102 are summarised below.

In addition, the company reported that Healey et al. 76 may provide some useful clinical data for the assessment of the Confirm Rx. The EAG notes that this was an observational cohort study that used the Confirm DM2100, the predecessor to the DM2102. The study population in Healey et al. 76 comprised patients at risk of AF who were aged ≥ 65 years and attending outpatient cardiology and neurology clinics. Healey et al. 76 do not specifically report whether or not the study population included any CS or cryptogenic TIA patients, although 48.0% had a history of stroke, TIA or systemic embolism. The EAG considers the data from the DETECT AF study75 to be more appropriate given that they are based on a more recent model of the Confirm ICM, and so the results from Healey et al. 76 are not discussed further.

The DETECT AF study75 was a prospective observational study conducted to assess the diagnostic accuracy of the Confirm ICM in detecting AF compared with Holter monitoring. The intervention comprised 4 days of simultaneous monitoring for AF using the Confirm ICM and a Holter monitor, and was required to take place at least 2 weeks after ICM implantation. A total of 90 patients were enrolled from 12 centres in Germany and the Netherlands between September 2012 and December 2013, although only 79 patients were deemed eligible for inclusion in the analyses. Reasons for exclusion from the analysis included clock synchronisation issues due to batteries running low in the Holter or patient ICM external symptom activator (for a total of five patients) and insufficient duration of analysable Holter recordings (four patients). Patients were required to have been diagnosed with paroxysmal AF or to have a clinical suspicion of paroxysmal AF. In total, eight of the enrolled patients had a history of prior stroke or TIA, although it is unclear whether any of these were CS or cryptogenic TIA patients.

The ICM monitored for AF episodes lasting at least 2 minutes and the Holter monitor data were analysed by a blinded, independent core laboratory. Patient and episode sensitivity, specificity, PPV and NPV were calculated.

At least one AF episode was detected in 16 of the 79 patients analysed, and all 16 patients had episodes of AF recorded by both the ICM and the Holter monitor. There were no incidences in which the Holter monitor detected additional episodes of AF compared with the ICM, but nine patients had at least one 2-minute AF detection by the ICM, without any corresponding AF episode detected on the Holter recording. However, most of these false positives were due to irregular sinus rhythms and not noise (44/58 episodes; number of patients was not reported and the clinical consequence of detecting these was not reported). In a per-patient analysis, the sensitivity was 100% (95% CI 79.4% to 100%), PPV was 64.0% (95% CI 42.5% to 82.0%), specificity was 85.7% (95% CI 74.6% to 93.3%) and NPV was 100% (95% CI 93.4% to 100%) for the Confirm ICM using Holter monitoring (minimum of 45 hours’ analysable data) as the reference standard. The results of the per-patient analysis, therefore, suggest that the Confirm DM2102 ICM can detect AF with a high sensitivity and a reasonably high specificity.

The DETECT AF study also reported no AEs during the follow-up time for any of the 90 enrolled patients who received the ICM device.

Biotronik

Biotronik provided 12 publications in support of their company submission with clinical data that they deemed to be of relevance to the assessment of the BioMonitor 2-AF in patients who had a CS. However, (confidential information has been removed) were deemed to meet the inclusion criteria for a discussion of non-CS or mixed population data. The key characteristics of the five included studies (two publications and six personal communications)77–84 are summarised in Table 15 and their results are discussed in the paragraphs below. The EAG notes that only (confidential information has been removed) and that the primary indication for the ICM is CS (confidential information has been removed) of the study participants for each of the included studies. As discussed in Quantity and quality of the available evidence, AF detection in ICM devices is dependent on the patient population, incidence rate of AF, duration of monitoring and type of AF, and so these results may not be a true reflection of the ICM performance in CS patients. 34 It is also unclear in (confidential information has been removed) what proportion of the study participants received the BioMonitor 2-AF model of the BioMonitor 2 as specified in the NICE final scope and it should be noted that (confidential information has been removed).

(Confidential information has been removed.)

(Confidential information has been removed.)

Clinician time taken to insert the BioMonitor 2 was reported in five studies. 77–84 Reinsch et al. 80 reported that all devices were successfully implanted in a cath lab with a median time from skin cut to last suture of 8 minutes (IQR 7–10 minutes) and Ooi et al. 78 reported that all implantations were successfully performed in cath labs on the first attempt, with a median time from incision to last suture of 9 minutes (IQR 5–14 minutes). (Confidential information has been removed.)

(Confidential information has been removed.)

Four studies77,78,80,83 reported data on AEs, with one study77 also reporting subgroup data for CS patients. Ooi et al. 78 reported that there was one pocket infection observed and successfully treated with oral antibiotics. Reinsch et al. 80 reported that no devices had migrated by the 3-month follow-up, but two patients experienced AEs: one patient (3%) who was immunosuppressed developed a device-related pocket infection requiring implantable loop recorder explantation and oral antibiotic treatment; the second patient (3%) complained of slight discomfort in the area of the flexible ICM antenna. (Confidential information has been removed.)

(Confidential information has been removed.)

One study80 also provided data on additional patient-related outcomes of interest to the NICE final scope. Reinsch et al. 80 reported that at least one therapeutic intervention was performed in 23% of patients following the recording of arrhythmias during follow-up, and this included initiation of OAC in one patient (3%). Reinsch et al. 80 also reported results from patient satisfaction surveys at 1 day and 3 months. The results were generally good, with only 7% reporting moderate to severe pain and 20% reporting mild pain within 24 hours post intervention at the implantation site. Sustained paraesthesia was moderate in 7% and mild in 17% of patients and moderate impairment in daily life was reported by one (3%) patient. The cosmetic result was mostly reported to be very satisfying (63%) or satisfying (30%).

In summary, the studies of the BioMonitor 2 suggest that it is clinically effective in detecting AF and that it is associated with low levels of AEs and reasonably good levels of patient satisfaction. However, it should be noted that these results are not exclusively for the BioMonitor 2-AF or for a CS population; therefore, they should be interpreted with caution.

Medtronic

The documents supplied by Medtronic were reviewed for data and all relevant studies relating to the CS population have been included and discussed in the CRYSTAL-AF and observational studies sections; however, owing to time constraints and the large volume of citations of potential relevance for data in a non-CS or mixed population provided by the company, the EAG took the pragmatic decision to limit the inclusion of non-CS studies to those studies directly referred to and reporting clinical outcome data that were reported in the request for information document received on 28 July 2017. Five studies34,50,85–87 in non-CS or mixed populations were identified from the company submission, of which four34,50,86,87 reported data on the DTA of the Reveal ICMs and two85,87 provided AE data for the Reveal LINQ. These five studies are discussed briefly in the following paragraphs, along with their relevant results. As discussed earlier, it is important to remember when interpreting these results that the performance (e.g. PPV and NPV) of AF detection in ICM devices is dependent on the patient population, incidence rate of AF, duration of monitoring and type of AF, and so these results may not be a true reflection of the ICM performance in CS patients. 34

The Reveal XT Performance Trial (XPECT) (Hindricks et al. )50 was a single-arm prospective observational study of 247 patients conducted to assess the performance of the Reveal XT in detecting AF (of at least 2 minutes). Patients were enrolled between September 2007 and July 2008 from 24 medical centres, mainly in Europe and Canada. Eligible patients were those who fulfilled at least one of the following criteria:

• scheduled for pulmonary vein ablation or surgical rhythm control intervention

• had documented frequent AF or frequent symptoms attributable to AF

• had undergone pulmonary vein ablation in the previous 6 months and still had symptoms attributable to AF.

The study protocol required the enrolled patients to be implanted with the Reveal XT; 4 to 6 weeks after the ICM implantation, they were to receive 46 hours of Holter monitoring (with a minimum of 18 hours’ Holter recording required for inclusion in the analyses). A total of 206 patients had analysable Holter recordings, of which 76 (37%) had at least one episode of AF, although only 73 (96.1% of the Holter-detected AF patients) of these patients were also identified as having AF by the ICM. The XPECT study results demonstrate that the dedicated AF detection algorithm in the Reveal XT identified the presence or absence of AF with an accuracy of 98.5%, compared with the Holter monitor (Table 16). In addition, statistical analyses demonstrated that the AF burden measured with the ICM was well correlated with the reference value derived from the Holter monitor (Pearson coefficient = 0.97).

Pürerfellner et al. 86 used the XPECT trial data set and applied a change to the ICM AF detection algorithm so that it also incorporated data on P-waves when classifying patients with AF (this algorithm change was applied in the Reveal LINQ). The revised data set was compared with the original Holter monitor data and the results are presented in Table 16.

The Reveal LINQ usability study (Sanders et al. 87) was a non-randomised, single-arm prospective multicentre observational study to assess the diagnostic accuracy of the new Reveal LINQ ICM for AF detection using 24 hours of Holter monitoring as the reference standard. The Holter monitoring was scheduled to occur at the 1-month follow-up visit, 1 month following ICM insertion, and AF had to be of ≥ 2 minutes’ duration. The patients enrolled in the Reveal LINQ usability study comprised 30 patients with any indication for an ICM and 121 patients with a documented history of AF (including patients awaiting AF ablation). The reference standard and patient population are, therefore, different in this study to those in the XPECT study. The results of the Reveal LINQ usability study are summarised in Table 16. A total of 138 patients had Holter monitor recordings that were suitable for inclusion in the analyses and the ICM correctly identified 37 of the 38 patients with Holter-detected AF (diagnostic sensitivity of 97.4%). The results of the new AF detection algorithm in the Reveal LINQ ICM demonstrate an improvement in terms of AF detection compared with the Reveal XT.

The Pürerfellner et al. 34 study was similar to Pürerfellner et al. 86 study in that it was applying a further P-wave-related algorithm enhancement for the Reveal ICMs AF detecting capability to existing data sets to see what impact it had on the diagnostic accuracy of the ICMs. The Pürerfellner et al. 34 study used both the XPECT and Reveal LINQ usability study data sets. The first 56 patients in the XPECT study with suitable data were used as the development data set for testing the algorithm enhancement and then data from 176 patients were used as the validation data set. In addition, the algorithm enhancement [adaptive P-sense (TruRhythm)] was applied to the Reveal LINQ. The per-patient results were reported in the paper only for the LINQ usability study data set (see Table 16), although the EAG notes that no explanation was provided for the discrepancy in the number of patients with AF diagnosed on Holter monitor in the TruRhythm analysis reported in Pürerfellner et al. 34 compared with in the Sanders et al. 87 publication for the LINQ usability study (37 vs. 38 patients, respectively). Nonetheless, assuming the results of the Pürerfellner et al. 34 study are accurate, they suggest that the new adaptive P-sense (TruRhythm) enhancement results in an improvement in sensitivity, specificity and accuracy in the Reveal LINQ in detecting AF. Medtronic reported in their company submission that the Reveal LINQ with TruRhythm detection was rolled out in 2017.

The results of the DTA studies in the non-CS population suggest that the enhancements over time to the AF diagnosis algorithm in the Reveal XT and Reveal LINQ ICMs have improved the DTA of the ICMs (sensitivity and specificity; see Table 16). However, it should be noted that these data are not in the CS population and the data in the XPECT and Reveal LINQ usability studies used to make some of these comparisons are heterogeneous owing to differences in the way in which the reference standard was applied (Holter monitoring 48 hours vs. 24 hours for the XPECT and Reveal LINQ usability studies, respectively) and differences in the patient populations (e.g. reasons for ICM insertion). Nonetheless, these data suggest that the Reveal LINQ is likely to be as effective at, if not better than, detecting AF as the Reveal XT (as the Reveal LINQ has fewer false positives and fewer false negatives); therefore, the AF detection rate from the CRYSTAL-AF trial is potentially a conservative estimate for the Reveal LINQ, given that it was the Reveal XT that was used in the CRYSTAL-AF trial.

Mittal et al. 85 reported AE data for two observational studies of the Reveal LINQ; one was the LINQ usability study87 already discussed and the second was the Reveal LINQ registry. The registry is a post-market surveillance study of patients with a Reveal LINQ ICM for any indication (the proportion of people with a CS indication is not reported) and the AE data discussed in Mittal et al. 85 for this study were limited to 122 patients from seven centres who were enrolled pre device insertion. The combined cohort of 273 patients from the two studies had an infection rate of 1.5% (n = 4), an AE rate of 4.0% (n = 11) and a SAE rate of 1.1% (n = 3). The company highlighted that the definition of an AE varies across studies and the EAG notes that the analysis in Mittal et al. 85 does not take into account the differences between the two study populations; thus, the analysis is subject to clinical heterogeneity. The results do, however, suggest that the Reveal LINQ is associated with a low rate of AEs and SAEs.

Quantity and quality of evidence

• The clinical evidence searches sought to identify RCTs and comparative observational studies that compared any of the three devices [Confirm Rx (Abbott), BioMonitor 2-AF (Biotronik) and Reveal LINQ (Medtronic)] with at least 24 hours of Holter monitoring to detect AF in people with CS. Electronic database searches were run on 13 September 2018 and results were assessed together with reference lists of systematic reviews and evidence submitted by the companies.

• A single RCT37 assessing an earlier Medtronic Reveal model (XT rather than LINQ) met the original eligibility criteria (the CRYSTAL-AF trial), so the criteria were widened to find evidence for the BioMonitor 2-AF, Confirm Rx and Reveal LINQ. First, non-comparative observational studies were sought within the correct CS population, and then evidence was considered from studies of mixed populations submitted by each company. Only the CRYSTAL-AF trial37 falls within the eligibility criteria outlined in the original published protocol for this diagnostics assessment, so the additional evidence should be interpreted with caution.

• The CRYSTAL-AF trial37 (n = 441 participants) represents the most robust clinical evidence to inform the decision problem, despite assessing the Reveal XT. The study was open-label and compared the ICM with conventional follow-up in a population of people who had a CS or TIA and no history of AF after extensive diagnostic work-up. The study was conducted in North America and Europe and the population was considered generally applicable to patients who would be eligible for an ICM in UK clinical practice.

• Twenty-six single-arm observational studies were identified after widening the eligibility criteria to include non-comparative studies. The studies were conducted in North America and western Europe (one in the UK) and all assessed the Reveal XT and Reveal LINQ in populations of people who had had a CS or TIA; none provided evidence to assess the efficacy of the BioMonitor 2-AF (other than a mixed device study that did not provide separate results) or Confirm Rx for patients who had had a CS. The observational studies represent a wide sample of patients who have received an ICM in practice (n = 3414) and provide evidence for the Reveal LINQ and for additional outcomes that were not available from the CRYSTAL-AF trial.

• In total, one study65 for Confirm Rx (of an older model, the Confirm DM2102), five studies of the BioMonitor 2-AF77–84 (all BioMonitor 2 but only one80 of which we can be certain was of the ‘-AF’ model) and five studies34,50,85–87 of the Reveal LINQ [three studies34,85,87 of the Reveal LINQ and three50,86 studies of the Reveal XT (one study34 included both devices)] in mixed populations were included based on the recommendations of the companies. All of these mixed population studies are either single-arm observational studies or they provide DTA data for the ICM using Holter monitoring as the reference standard.

• All the observational studies were single-arm and therefore were deemed to be at a high risk of bias. Three conducted within-patient comparisons of ICM versus other monitoring strategies. 14,15,72 Key sources of heterogeneity between the observational studies include patient demographics (mean or median age 52–72 years), rigour of stroke assessment, stroke risk score (CHA₂DS₂VASC score of 3–5), definition and adjudication of AF and length of follow-up; all are likely to affect AF detection and other clinical outcomes.

• Eight ongoing studies40–46 of potential relevance were identified, although only five (three RCTs40,42,47 and two observational studies41,45) reported details of their current status and the ICM being studied and none relates to the BioMonitor 2-AF. The three ongoing RCTs all involve the Reveal LINQ, with only one RCT solely in a CS population: a Canadian randomised trial comparing the clinical effectiveness and cost-effectiveness of the Reveal LINQ ICM with external loop recording in 300 CS patients, which is estimated to complete in December 2019 (PERDIEM). 47 There was only one ongoing study identified relating to the Confirm RX: the SMART registry,45 a post-approval study planning to recruit at least 2000 patients with Confirm RX across multiple indications, but with a planned subgroup analysis for CS; completion is expected in December 2020.

Methods

A systematic review of the literature was undertaken in September 2018 to identify published economic evaluations of ICMs to detect AF in people who have had a CS. The sources identified in those searches were also used to identify resource use and cost data that could be utilised in the economic model. In addition, one further systematic review was conducted, in September 2018, aiming to identify studies providing utility (generic, preference-based) data on the HRQoL of people with AF and stroke, that could be used for the estimation of quality-adjusted life-years (QALYs) in the economic model.

The following databases were searched for relevant studies:

• MEDLINE and Epub Ahead of Print, MEDLINE In-Process & Other Non-Indexed Citations, Daily and Versions (via Ovid)

• EMBASE (via Ovid)

• EconLit (via Ovid)

• NHS Economic Evaluation Database (via the CRD)

• CDSR (via The Cochrane Library)

• CENTRAL (via The Cochrane Library)

• DARE (via the CRD)

• HTA database (via the CRD).

Further to the database searches, experts in the field were contacted with a request for details of relevant published and unpublished studies of which they may have knowledge; reference lists of key identified studies were also reviewed for any potentially relevant studies.

The search strategy for existing economic evaluations and studies reporting resource use or cost data combined terms capturing the interventions of interest (ICM, i.e. Reveal LINQ, BioMonitor 2-AF and Confirm Rx) and the target population (patients who have had a CS) with economic or health-care resource use terms, applied to all electronic databases. The search strategy for HRQoL data was not restricted by intervention, and combined terms capturing the target population with HRQoL terms.

The search for resource use and cost data was limited to the UK NHS setting, as the aim of this search was to identify data directly relevant to the NHS context that could inform the economic model; however, no country restrictions were applied to searches for existing economic evaluations.

Owing to the high volume of hits in the searches for HRQoL evidence, searches were restricted by date, starting from 1997; the year 1997 was selected as this was the year the utility index for the EQ-5D was published. Studies were then restricted to those collecting data in the Organisation for Economic Co-operation and Development countries, as HRQoL data collected in low income countries were unlikely to be generalisable to the UK.

Initially, the EAG considered studies reporting utility data elicited using a generic, preference-based measure [EQ-5D, Health Utilities Index, Short Form questionnaire-36 items (SF-36), Short Form questionnaire-6 Dimensions (SF-6D)] or self-reported validated, choice-based technique for valuation (i.e. time trade-off or standard gamble). However, given the availability of relevant EQ-5D data in this population (made apparent to the EAG during the first sift) and NICE’s preference for EQ-5D data, the EAG decided to restrict studies to primary sources of EQ-5D data.

Limits were applied to all searches to remove animal studies, letters, editorials, comments or case studies. Only conference abstracts published in the previous 2 years were considered for inclusion; it was assumed that any high-quality studies reported in abstract form before that date would have been published in a peer-reviewed journal. Full details of the search strategies are presented in Tables 38–46 in Appendix 5.

The titles and abstracts of papers identified through the searches were independently assessed for inclusion by two health economists using predefined eligibility criteria. Owing to the high volume of studies retrieved by the HRQoL search, one health economist reviewed all identified citations and a second health economist reviewed 20% of citations to confirm that the same studies were included for second pass.

The inclusion and exclusion criteria for each of the three systematic reviews described above are outlined in Box 1. The methodological quality of the full economic evaluations identified in the review was assessed using the Drummond checklist. 88

Economic evaluations

The SLR identified a total of 41 papers after de-duplication and, based on title and abstract, a total of nine papers (including one unpublished report supplied by Biotronik) were identified as potentially relevant and were obtained for full-text review based on the criteria listed in Box 1. Of the nine papers identified for full-text review, five papers were included for data extraction (see Appendix 7, Table 50). 89–93 The results of the searches are summarised in Figure 2. Reasons for exclusion of the ordered papers are provided in Appendix 6 (see Tables 47 and 48).

FIGURE 2.

The PRISMA flow diagram of the economic evaluation SLR. NHS EED, NHS Economic Evaluation Database.

All economic evaluations meeting the inclusion criteria in the SLR were based on a Markov model structure with model cycles ranging from 1 to 3 months. 89–93 Two studies assessed the cost-effectiveness of Medtronic’s Reveal XT device with standard of care monitoring (SoC). 89,90 One study was based on Biotronik’s BioMonitor 2-AF device compared with SoC. 91 (Confidential information has been removed.) Two studies did not indicate which model or brand of ICM was being assessed in the economic evaluations. 92,93

Of the five studies, only one was based on the UK (NHS) payer perspective; therefore, it will be the focus of a more in-depth analysis of model structure and parameter estimation. 90 In addition to the search, the manufacturer of BioMonitor 2 made an unpublished report and economic model available to the EAG, which assessed the cost-effectiveness of BioMonitor 2-AF in patients who had had a CS. (Confidential information has been removed.)

The study by Diamantopoulos et al. 90 is a cost–utility analysis assessing the use of an ICM (Reveal XT) to detect AF in patients who have had a stroke or TIA that is considered cryptogenic after an initial 24-hour period of non-invasive external Holter monitoring. The comparator in the study was no further monitoring after the initial 24-hour period. The perspective of the analysis was the UK NHS, and the time horizon of the model was lifetime.

The model was developed using a Markov structure with three main health states for AF status: AF free, AF detected and AF undetected. Figure 3 presents the model schematic.

FIGURE 3.

Model schematic of the CRYSTAL-AF trial cost-effectiveness analysis. CRNMB, clinically relevant non-major bleeding; DOAC, direct oral anticoagulant; ECH, extracranial haemorrhage; HS, haemorrhagic stroke; IS, ischaemic stroke. Reproduced from Diamantopoulos et al. 90 International Journal of Stroke (volume 11, issue 3), pp. 302–12, copyright © 2016 by SAGE Publications. Reprinted by Permission of SAGE Publications, Ltd.

Reproduced from Diamantopoulos et al. 90 International Journal of Stroke (volume 11, issue 3), pp. 302–12, copyright © 2016 by SAGE Publications. Reprinted by Permission of SAGE Publications, Ltd.

Patients start in the AF-free state, from which they can move to the AF-undetected or AF-detected states at any given model cycle. From the AF-undetected state, patients can either remain or move to the AF-detected state, and a patient remains in the AF-detected state unless she/he experiences a subsequent cerebrovascular event or bleeding event as follows.

The consequences of these subsequent events were modelled in two separate categories for either temporary or permanent effects. Events with temporary consequences were non-fatal extracranial haemorrhage (ECH) or intracranial haemorrhage (ICH), or a clinically relevant non-major bleed. Events with permanent consequences were non-fatal IS, non-fatal haemorrhagic stroke (HS), or fatal ECH, ICH, IS or HS events. Deaths of any cause could also occur from any heath state in the model.

Following a temporary event, patients return to their previous AF status health state and can continue to move between these health states as described previously. For patients who moved to a post-stroke health state following a permanent event, patients were assumed to remain there and face no further risk of stroke or bleeding events, with the only possible remaining transition being to the death state.

Treatment in the AF-free and AF-undetected states was assumed to be aspirin. In the AF-detected state, treatment was assumed to change to a directly acting oral anticoagulant (DOAC) until a bleeding event (HS, other ICH or ECH) occurs, at which point patients were assumed to revert to aspirin.

The risk of subsequent IS was determined by AF status, virtual CHADS₂ score, age and treatment received. Evidence was synthesised from six studies, which included systematic reviews, RCTs and registry data. The severity of IS was considered to measure the expected impact on quality of life and resource use. The distribution of severity (mild, moderate, severe and fatal) was taken from two published cost-effectiveness analyses,94,95 comparing anticoagulant treatments for stroke prevention in patients with AF. The distribution of severity was assumed to be independent of treatment, so the average across all treatments was used.

Bleeding consequences were also included in the model and the risks were assumed to be treatment and age related. Data for these risks were derived from five studies96–100 including a systematic review plus various trials comparing anticoagulants for patients with AF. The same cost-effectiveness analyses used to inform the distribution of IS severity were used to inform the distribution of type and severity of bleeding events, which were also assumed to be independent of treatment.

Age-dependent mortality was applied in the model and based on interim UK life tables. 101 It was adjusted, when applicable, to exclude deaths caused by cerebrovascular events, as these were modelled separately. Following a non-fatal stroke, the mortality risk was increased depending on the severity of stroke and the treatment received for it.

Health-related quality-of-life data for patients experiencing stroke events were collected in the Oxford Vascular Study (OX-VASC). 102 Disutilities associated with bleeding events were also included and informed by two published models. 94,95,103 Utilities were adjusted to account for age and sex using previously published methods.

The price year of the model was 2013. Costs for the insertion of the ICM (£1836) were included in the economic analysis, as were per-cycle costs to account for follow-up visits and monitoring, as well as drug treatments. The resource use required was determined by an unpublished post hoc analysis of the CRYSTAL-AF trial data. The lifetime of the ICM was assumed to be 3 years, at which point the device was removed. The cost of removal (£491) was also included. These costs were sourced from NHS Reference Costs 2012 to 2013. 104 Costs associated with events such as stroke were included, as well as estimated long-term costs associated with living in a post-stroke health state.

Results of the deterministic base-case analysis showed that an ICM was £2587 more expensive than SoC and provided a benefit of 0.151 QALYs, resulting in an incremental cost-effectiveness ratio (ICER) of £17,175 per QALY gained. A probabilistic sensitivity analysis (PSA) was performed, which reduced the incremental cost to £2574 and increased the QALY gain to 0.161, thereby reducing the ICER.

The EAG considers that the results produced by the model are potentially unreliable as there is significant uncertainty around the estimation of the clinical parameters in the model, particularly around the estimation of treatment effects by indirect comparison, AF incidence and detection rates used in Medtronic’s analysis. The authors conducted an indirect comparison to estimate HRs for IS, bleeding events, ICHs, ECHs and mortality that are conditional on treatment received. The EAG attempted to validate the HRs used in the model but could not verify the source data used by the authors: details of how the indirect comparison was conducted, as well as how the publications informing the analysis were identified, were not sufficiently described. Furthermore, the EAG considers that estimation of some of the HRs could be flawed; for example, the authors estimate a HR to adjust mortality in the model, but the source data used is based on standardised mortality ratios.

Finally, patients who are not detected as having AF are assumed to be given aspirin as their treatment option. However, the EAG’s clinical experts stated that patients would be given clopidogrel (75 mg) as their antiplatelet treatment.

However, the EAG considers that the initial health states of the Diamantopoulos et al. 90 model to determine AF status is useful to inform a short-term model, in which the time horizon is linked to the battery life of an ICM device. From the short-term model, patients with AF (whether detected or undetected) will then feed into a long-term (lifetime) model, assessing the costs and benefits of anticoagulation therapy.

Models assessing the long-term impact of anticoagulation therapy for patients with atrial fibrillation

In addition to the SLR, the EAG were notified by NICE of an ongoing DAR for lead-I ECG devices for detecting AF using single-time point testing in primary care [Diagnostic Assessment Programme 39 (DAP39)]. 105 The population considered in DAP39105 is adults presenting to primary care with signs and symptoms of AF who have an irregular pulse. Lead-I ECG devices are handheld instruments that can be used in primary care to detect AF at a single time point in people who present with relevant signs and symptoms (i.e. palpitations, dizziness, shortness of breath and tiredness). 105 If a lead-I device detects AF, the patient initiates anticoagulation and rate control therapy (unless contraindicated) and a 12-lead ECG is conducted to provide more diagnostic information and inform treatment. The EAG assumed anticoagulation therapy would be with apixaban, which is a simplifying assumption.

The comparator in this study was no further immediate testing after manual pulse palpation (MPP), with patients referred for a 12-lead ECG if the general practitioner (GP) was suspicious of AF after MPP (standard care pathway). In the standard care pathway, no AF treatment is initiated if the GP is suspicious of AF until after the results from the 12-lead ECG are available, confirming diagnosis.

Although the technology and population under assessment are not relevant to the decision problem of the current report, the EAG was interested in the approach taken to estimate long-term costs and benefits of anticoagulation therapy once patients have been identified as having AF using lead-I devices. For DAP39,105 once patients had a diagnosis of AF confirmed, they entered a post-diagnostic Markov model with either no history of cardiovascular events (CVEs), one CVE or two CVEs. In each cycle, patients can remain in their current health state, have a CVE and move to a worse health state, or die. Patients with two CVEs can remain in their current health state only until death.

The model parameters used to estimate the transition probabilities for the post-diagnostic Markov model were derived mainly from a cost-effectiveness study by Sterne et al. ,106 which assessed the long-term costs and benefits of anticoagulation therapy for prevention of stroke in patients with AF (hereafter referred to as the DOAC model). The EAG reviewed the publication for the DOAC model and deemed it relevant for the current decision problem, and contacted the authors to obtain a copy of the model for assessment. After reviewing the DOAC model and discussing it with the model developer, the EAG was made aware of an adapted version of the DOAC model, which was used for another publication,107 that would be appropriate to review and potentially use for development of the ICM model.

The adapted DOAC model was developed to assess the cost-effectiveness of screening strategies for AF. 107 The model structure employed by the authors was a hybrid model, with a short-term decision tree that used sensitivity and specificity estimates of different screening strategies to detect AF (confirmed by a 12-lead ECG) and initiate anticoagulation therapy and a long-term adapted version of the DOAC Markov model. In the analysis, it was assumed that 75% of patients not contraindicated to anticoagulation therapy and who are prescribed anticoagulants use DOACs, with the remaining 25% prescribed warfarin. Patients who are diagnosed with AF, but who are contraindicated to anticoagulation therapy, not prescribed or choose not to take OACs, would receive aspirin.

The results of the screening decision-tree model feed directly into the adapted DOAC model (Figure 4). The discrete-time Markov multistate model implemented a cycle length of 3 months and employed a lifetime horizon with a cut-off point at 100 years. Patients who are prescribed an OAC enter the model either on first-line apixaban or warfarin [international normalised ratio (INR) range 2–3], with the remainder on aspirin. The authors assumed the use of apixaban, as it was determined to be the most cost-effective DOAC in the anticoagulation therapy cost-effectiveness analysis, but state that the results are similar when considering other available DOACs.

FIGURE 4.

Prevention of stroke in AF model structure. MB, major bleed; S, stroke. Reproduced from Sterne et al. 106 Contains information licensed under the Non-Commercial Government Licence v2.0.

Reproduced from Sterne et al. 106 Contains information licensed under the Non-Commercial Government Licence v2.0

Depending on the occurrence of IS or SAEs (such as ICH), treatment switching can occur (Figure 5). For patients on first-line apixaban, second-line treatment may be either warfarin or no treatment. No treatment is the only third-line treatment available. For those who fail on warfarin, no further treatment would be given. 106

FIGURE 5.

Treatment strategies and switching/discontinuation rules for the prevention of stroke in the AF model. MB, major bleed; SE, systemic embolism. Reproduced from Sterne et al. 106 Contains information licensed under the Non-Commercial Government Licence v2.0.

Reproduced from Sterne et al. 106 Contains information licensed under the Non-Commercial Government Licence v2.0

The same model structure is used for each treatment option (see Figure 4), but is adjusted for the different costs, utilities and transition probabilities relevant to treatment. Patients start the model in the AF well health state (no event). From any health state in the Markov model, patients can have an IS, a MI, a clinically relevant (extracranial) bleed (CRB), an ICH, a systemic embolism, a TIA or die. The authors of the DOAC model assumed that systemic embolism and TIA have only short-term impacts on future risks, costs and utilities, but IS, ICH, CRB and MI have long-term impacts that will change future risks, costs and utilities. For example, a patient who experiences a MI and ICH will have different risks, costs and utilities compared with a patient who experiences only a MI or ICH. In addition, the model does not distinguish between minor and major ISs because of limited published evidence on the relative rates of these events from the RCTs.

As with all Markov models, patient history through the model is not recorded; therefore, future health state transitions depend only on the current health state the patient occupies. An assumption is made in the model that transition probabilities do not change with time but that, as the cohort ages, mortality risk increases in line with general population life tables.

The authors of the screening model adapted the long-term DOAC model by including HRs for events (stroke, systemic embolism, TIA) affected by AF type (paroxysmal relative to permanent or persistent). Furthermore, the DOAC model depends on age, sex, previous history of IS or TIA and previous history of MI.

Treatment effects implemented in the model were based on a competing risks network meta-analysis (NMA) to jointly estimate log HRs of each treatment relative to warfarin for the different possible health states in the model.

The costs included in the analysis comprised pharmacotherapy costs and costs of acute and chronic AF and anticoagulant-related events. Sources of cost data included the British National Formulary (BNF) for drug costs (March 2015 update),108 NHS reference costs109 and other published sources. When necessary, cost data were inflated to 2015 prices using the Office for National Statistics (ONS) Consumer Price Inflation Index for Medical Services (DKC3). 110

Quality-adjusted life-years were estimated by applying health-state utility values to the proportion of patients occupying each health state per model cycle. Utilities were identified from a previous NICE technology appraisal submission on rivaroxaban (TA256),111 which included a systematic literature search for evidence on EQ-5D utility index scores in health states related to AF. For acute health states (such as CRB, systemic embolism, TIA, ICH, acute IS and acute MI), disutilities were applied for one model cycle. For patients who have multiple chronic health conditions, utilities for the health states were assumed to be multiplicative. All utilities were adjusted for age.

Total costs and QALYs estimated for each first-line anticoagulation therapy were generated as well as incremental results compared with warfarin. The authors did not calculate ICERs, but instead calculated the incremental net benefit (INB) of each DOAC compared with warfarin, when a QALY is valued at either £20,000 or £30,000. Compared with warfarin, all DOACs had a positive INB, with apixaban (5 mg twice daily) estimated to have the highest expected INB (£7533), followed by dabigatran (150 mg twice daily; £6365), rivaroxaban (20 mg once daily; £5279) and edoxaban (60 mg once daily; £5212). The 95% CI around the INB for apixaban was positive, suggesting that apixaban is cost-effective compared with warfarin.

The EAG considers that the adapted DOAC model is suitable to inform the long-term costs and benefits of anticoagulation treatment versus antiplatelet treatment in the CS population, who have suspected AF; therefore, the adopted DOAC model will be incorporated into the model structure assessing ICMs in this population. More detail on the integration of the DOAC model into the EAG’s Microsoft Excel^® (Microsoft Corporation, Redmond, WA, USA) model is given later in the subsection that describes the development of the economic model (see Model structure and Chapter 5, Probabilistic sensitivity analysis).

Health-related quality-of-life evidence

The systematic literature search identified a total of 7641 papers after deduplication. Based on a review of titles and abstracts, 112 papers were identified as potentially relevant and were obtained for full-text review based on the criteria listed in Box 1. An additional two papers were identified from the reference lists of identified papers. Of the 114 papers identified for full-text review, 25 papers were included for data extraction (see Appendix 7, Table 51). Reasons for exclusion of the 89 papers are provided in Appendix 6, Table 49. The results of the process to identify HRQoL evidence is summarised in Figure 6.

FIGURE 6.

The PRISMA flow diagram of the HRQoL SLR. NHS EED, NHS Economic Evaluation Database.

Data from patients with cerebral infarction, IS, haemorrhagic stroke (intracranial, intracerebral or subarachnoid) or TIA were collected by 21 studies,102,103,112–130 and data on patients specifically with AF (with or without stroke) were collected in six studies. 112,113,121,131–133 Four of the included studies also provided data for patients with MI103,113,121,131 and two assessed the impact of additional bleeding events. 121,133 The studies differed in how stroke was defined, with some having much broader definitions than others, which hindered comparisons by type of stroke.

All studies reported EuroQol-5 Dimensions, three-level version, (EQ-5D-3L) data and two118,119 also collected EuroQol-5 Dimensions, five-level version, (EQ-5D-5L) data. EQ-5D responses were converted into utilities using UK population tariffs developed by Dolan134 in nine studies. 102,103,117,122–124,127,128,131 Two of those studies were undertaken in the UK. 102,132 The remaining studies were undertaken in Canada, Finland, Germany, Korea, Norway, Poland, Spain, Sweden, the USA, the Netherlands or in multiple countries.

The EAG considers that the most relevant utilities for the model are those from OX-VASC,102 which were also utilised in the CRYSTAL-AF trial economic evaluation90 and Berg et al. 131

The OX-VASC study consisted of stroke and TIA patients whose quality of life was assessed using the EQ-5D-3L questionnaire at regular follow-ups of 1, 6, 12, 24 and 60 months after their stroke or TIA. The baseline population consisted of 748 patients with stroke and 440 TIA patients. EQ-5D-3L responses were available for 759 patients at 1 month, 723 patients at 12 months and 479 at 60 months. EQ-5D-3L responses were converted into utilities using UK population tariffs. 102 The mean age of the population was 75 years and 44% were female. Utilities were estimated for different events including, TIA, all stroke, IS, ICH and subarachnoid haemorrhage, as well as different severities of stroke.

The study by Berg et al. 131 assessed HRQoL for patients with AF, based on data from the Euro heart survey. 135 The mean age of the population was 66 years and 41.9% were female. HRQoL was measured at baseline and at 1-year follow-up using the EQ-5D-3L questionnaire. At baseline, 5050 EQ-5D-3L responses were recorded, with 3045 responses recorded at 1-year follow-up. EQ-5D-3L responses were then converted into utilities using UK population tariffs. Baseline utility for patients with AF were estimated, as well as utility decrements for AEs during follow-up, including MI, stroke, congestive heart failure and other major AEs.

The utility values for events that have been included in the economic model are given in the subsection that describes the development of the economic model (see Utility values).

Population

The population considered in the model are patients who have had a CS, including TIAs, for whom there is a suspicion of paroxysmal AF, and who have received at least 24 hours of outpatient external ambulatory electrocardiography monitoring that has not detected AF. The diagnostic data included in the model are based on the results of the SLR that identified the CRYSTAL-AF RCT37 assessing the Reveal XT ICM compared with SoC in the patient population of interest. The mean age (61 years) and sex split (≈65% male) of patients in the model is based on data from the CRYSTAL-AF trial. The EAG’s clinical experts considered that, in general, the population in the CRYSTAL-AF trial is reflective of UK patients; some inconsistencies were noted but were not deemed significant. See Chapter 3 for further detail.

Intervention and comparator

As per the NICE final scope, the interventions included in the model are as follows:

• BioMonitor 2-AF

• Confirm Rx

• Reveal LINQ.

The comparator for the analysis listed in the NICE final scope was no further monitoring after at least 24 hours of outpatient external ambulatory electrocardiography monitoring that has not detected AF. Data for the comparator arm are taken from the CRYSTAL-AF trial, in which patients in the control arm underwent assessment at scheduled visits (every 3 months) and unscheduled visits, if patients were experiencing symptoms of AF. 90 To match the monitoring period of the ICM devices (3 years), the SoC period was also 3 years. Tests for the control arm included ECGs and Holter monitoring (for 24 hours, 48 hours or 7 days). Table 17 presents the tests performed per person per year in the control arm of the CRYSTAL-AF trial.

Model structure

The EAG developed a two-stage economic model to assess the cost-effectiveness of using ICMs to detect AF in patients who have had a CS. The comparator in the analysis was 24 hours of external ambulatory electrocardiography monitoring.

The development of the model was informed by published models identified in the SLR. The first stage of the model (short-term model) outlines the initial patient-flow over a 3-year period (i.e. the battery life of the Reveal XT). The second stage of the model (long-term DOAC model) estimates the lifetime risks, costs and benefits for patients on either long-term anticoagulation or antiplatelet therapy. Figure 7 presents the schematic for the short-term model. The long-term DOAC model is presented in Figure 4. Each stage of the model is described in more detail below.

FIGURE 7.

Short-term patient flow model.

All patients enter the model as CS patients who have received at least 24 hours of outpatient external ambulatory electrocardiography monitoring that has not detected AF. The initial cohort is a mixture of patients with and without pre-existing paroxysmal AF who are given antiplatelet therapy for stroke prevention. Over the time horizon of the short-term model (3 years), patients who have an episode of AF may have that AF detected and thus will move to the AF-detected health state, where they enter the anticoagulation arm of the long-term DOAC model. However, episodes of AF may not be detected; thus, patients will then move to the AF-undetected health state where they enter the antiplatelet arm of the long-term DOAC model. Patients who do not have an episode of AF over the 3-year time horizon of the model remain in the CS patient health state (see Figure 7) and are not considered in the long-term model, as treatment, costs and benefits would be the same with or without an ICM. Therefore, the incremental analysis would equal to zero. However, in the short-term, ICM patients incur the costs of having an ICM and the associated costs of follow-up appointments and SoC patients incur the costs of monitoring and associated follow-up appointments.

The proportion of patients who are identified as having AF in each of the 12 3-month cycles is informed by data from the CRYSTAL-AF trial. It is assumed that patients in the SoC arm are detected either during follow-up appointments or owing to developing symptoms of AF. Explicit transitions from the AF-undetected to the AF-detected states are not modelled, as the data from the CRYSTAL-AF trial present cumulative detection rates. However, as a simplifying assumption, patients who receive SoC and have undetected AF by the end of the short-term model (3 years) will remain undetected and on antiplatelet treatment for the remainder of the modelled time horizon.

Sensitivity data for the Reveal LINQ device, in a broader AF population, indicate a sensitivity of 100%, enabling the calculation of the AF-undetected health state occupancy for the SoC arm of the model (see the evidence from Medtronic on ICMs in non-CS populations given in Chapter 3, Medtronic, for more detail). 34 Furthermore, based on the sensitivity and on data on AF detection rates for the ICM arm, patients from the initial cohort who do not have AF are excluded from the long-term analysis of outcomes as the proportion is assumed to be the same in each arm of the model (i.e. no false positives), and thus incremental costs and QALYs are zero. The assumption of no false positives in the model is based on information from the CRYSTAL-AF trial and advice from the EAG’s clinical experts that state alerts generated by an ICM will need to be reviewed by a clinician to confirm AF and initiate anticoagulation treatment. It should be noted that all patients in the ICM cohort incur the cost of the device, implantation, removal of device and follow-up. All SoC patients incur the cost of monitoring. Patients who do not have AF detected are assumed to incur the cost of follow-up appointments with a consultant cardiologist at 1, 3, 6 and 12 months as per the advice of the EAG’s clinical experts.

The second-stage long-term model uses the adapted version of the DOAC model [previously described in the subsection that reports the results of the SLR (see Economic evaluations)]. The DOAC model is a probabilistic model that outputs total costs and QALYs for DOAC treatments (apixaban, rivaroxaban, edoxaban and dabigatran etexilate), warfarin and antiplatelet treatment. The cycle length of the DOAC model is 3 months. The EAG adapted the model code to allow the output to be given as per-cycle costs and QALYs over a lifetime time horizon. This enabled the application of costs and QALYs specific to each cycle (i.e. time dependent per cycle costs and QALYs) in the ICM Excel model from the point when patients have AF detected and start anticoagulation treatment or are in the AF-undetected state and continue with antiplatelet treatment.

Data inputs were updated to reflect the CRYSTAL-AF trial population, for example the starting age was set at 62 years and the ratio of males-to-females was set at 65 : 35 to weight the general mortality death rates. The model was adapted to include all DOACs plus warfarin. Costs in the DOAC model were updated or inflated to 2018 prices, where appropriate, to reflect current values. Based on the HRQoL SLR, utility inputs were also updated. Life tables were also updated to the most recent year available (2015–17). 101

Mean costs and benefits per cycle (based on 10,000 samples run in the long-term model) related to AF patients treated with either anticoagulation or antiplatelet medication and are estimated for each individual DOAC treatment in the model. Figure 4 outlines the model schematic for the adapted DOAC model. The same structure is used for each treatment included in the model (i.e. DOACs for patients with detected AF and antiplatelet treatments for patients with undetected AF) and is adjusted for treatment-specific transition probabilities, costs and utilities. It should be noted that, for the current model, the adapted DOAC model population was prespecified for previous history of IS and paroxysmal AF (e.g. the risks of events in the model were adjusted to reflect a secondary stroke population with paroxysmal AF).

The mean, time-dependent per-cycle costs and benefits of anticoagulation treatment are then applied to the proportion of patients in each cycle of the AF-detected health state and the mean, time-dependent per-cycle costs and benefits of antiplatelet therapy are applied to the proportion of patients in each cycle of the AF-undetected health state. The economic assessment is taken from the perspective of the NHS and Personal Social Services and both costs and benefits are discounted at 3.5% per annum.

Diagnostic efficacy of implantable cardiac monitors

The clinical effectiveness SLR identified diagnostic yield data only for the Reveal XT device from the CRYSTAL-AF RCT. AF detection rates for the comparator arm of the model are also derived from the CRYSTAL-AF trial. In the CRYSTAL-AF trial, an episode of AF was defined as irregular heart rhythm lasting > 30 seconds. 37 Table 18 presents the cumulative AF detection rate per model cycle (3 months) for both Reveal XT and SoC implemented in the short-term Excel model. These data are based on a KM analysis and include non-informative censoring of patients lost to follow-up. See Chapter 3, Diagnostic yield: atrial fibrillation detection rate, for further details. As no data were identified for BioMonitor 2-AF and Confirm Rx, the EAG sought advice from clinical experts as to whether or not there would be any differences in the detection rates between the devices. The EAG’s clinical experts acknowledged that the main source of efficacy for ICMs is the CRYSTAL-AF trial, but that there would not be any substantial differences in detection rates for the devices. As a result, the EAG has assumed equal efficacy for all devices.

The BioMonitor 2-AF has the longest battery life of all the devices, at 4 years. The battery life of the Reveal LINQ is 3 years and Confirm Rx has the shortest battery life, at 2 years. The EAG’s clinical experts advised that it is improbable that a device will be replaced once the battery has expired. This means that those implanted with the Reveal LINQ or BioMonitor 2-AF who have AF detected between 24 and 36 months would not have AF detected with the Confirm Rx. Thus, the EAG adjusted the detection rates of the Confirm Rx to reflect the number of AF cases that would be missed because of the relatively shorter battery life of the device.

The battery life of BioMonitor 2-AF is 4 years; however, data for AF detection are available for only 3 years. Therefore, in the absence of additional data, the EAG has capped the BioMonitor 2-AF detection rate to 3 years. It is difficult to predict what impact this assumption has for the cost-effectiveness of the BioMonitor 2-AF, as the additional year of monitoring with the device could mean that there is potential (if limited) for additional cases of AF to be picked up compared with the SoC arm.

Diagnostic accuracy data for the Reveal LINQ device indicates a sensitivity and specificity of 100% and 98.1%, respectively. 34 However, the sensitivity and specificity values for the Reveal LINQ are based on an update to the Reveal XT algorithm that incorporates P-waves and is applied to the data set of the XPECT trial, which assessed the performance of the Reveal XT device in a population with known AF. 34 Although the sensitivity estimated in a population with known AF may not be a reliable measure for a population with paroxysmal AF, the EAG’s clinical experts advised that the ICM will probably pick up all cases of paroxysmal AF. As a result, an assumption has been made in the model that the detection rate of the device estimates the true prevalence of AF in the CS population. Please see the evidence from Medtronic on ICMs in non-CS populations given in Chapter 3, Medtronic, for further detail.

Therefore, the detection rate in each cycle of the ICM arm provides an estimate of the proportion of patients in the cohort who have AF at any given cycle. The proportion of AF-undetected per cycle in the SoC arm can then be calculated as the difference between the detection rate of the ICM and the SoC arm per cycle.

It should be noted that, based on the CRYSTAL-AF trial data, the proportion of patients at the end of the 3-year follow-up period who have AF is estimated to be 30%. Theoretically, the subset of CS patients who have AF is known at the start of the model; therefore, all patients could enter the model in the AF-undetected state and over time this would reduce as patients are detected in each arm. However, the EAG chose to start all patients in the model without AF status known and to use the per-cycle incidence of AF, based on the detection rate and sensitivity of the ICM, to calculate the number of patients with undetected AF (calculated as per-cycle incidence of AF minus the per-cycle AF-detection rate). If the overall AF prevalence is used, then the calculation of AF-undetected patients in the ICM arm infers that there is a large proportion of AF patients whom the ICM devices miss, which contradicts the 100% sensitivity. In fact, because of the nature of paroxysmal AF, it may not be true that all patients have AF at the start of the model, particularly those that are detected by the ICM late in the model time horizon, as they may have developed AF as they age in the model. However, using either method to define the starting population of the model, based on the detection rates of the CRYSTAL-AF trial has no impact on the results.

Long-term clinical outcomes

Long-term outcomes for patients with AF (whether detected or undetected) are modelled using the adapted DOAC model. 106,107 Outcomes assessed in the model include IS, MI, TIA, systemic embolism, CRB, ICH and death (all causes). The long-term model is structured so that, as well as experiencing single events, patients can have multiple events (up to a maximum of three).

As discussed earlier in Model structure, each treatment considered in the model (OACs and antiplatelets) has the same long-term model structure, but adjusted for treatment-specific risks, costs and benefits. The authors of the model estimated treatment effects by performing a competing-risks NMA, based on the clinical effectiveness SLR conducted for the study, to estimate HRs for the different events considered. 106

The clinical effectiveness SLR conducted by Sterne et al. 106 identified 23 completed RCTs for inclusion in the review. Seventeen types of events were included in the NMA to account for correlation and competing risks. However, as mentioned previously, the events of interest for the economic model are IS, MI, TIA, systemic embolism, CRB, ICH and death (all causes). Three types of outcome data were incorporated into the model to estimate the HRs: number of first events, number of individuals experiencing at least one event of a given type and total number of events.

Baseline treatment in the adapted DOAC model is warfarin (INR 2–3); therefore, the authors developed a competing-risks model for warfarin separately to estimate the baseline hazard for the outcomes of interest in the model. Further detail on the methodology and estimates used in the long-term model can be found in the publication by Sterne et al. 106

The treatment effects for antiplatelet therapy used in the model have been estimated in the competing-risks NMA using outcomes for aspirin treatment. The EAG consulted with clinical experts to confirm that patients would be given aspirin after stroke and if, in lieu of any diagnosis of AF, they would remain on lifetime treatment with aspirin. The clinical experts advised that, in current clinical practice, treatment for CS patients is, in fact, with clopidogrel (75 mg, once daily) and would be the long-term treatment if patients are not diagnosed with AF.

Consequently, the EAG performed targeted searches to identify evidence on the relative efficacy of clopidogrel and aspirin in AF patients at risk of ischaemic events, as this population reflects the cohort that occupies the AF-undetected health state and would therefore be receiving antiplatelet treatment. The EAG found that much of the literature assesses clopidogrel in combination with aspirin. 136,137 The EAG identified a systematic review and NMA by Cameron et al. 138 comparing antithrombotic agents for the prevention of stroke and major bleeding in patients with AF. The review included aspirin and aspirin plus clopidogrel (dual antiplatelet therapy). Compared with standard dose vitamin K antagonist (e.g. warfarin), aspirin and dual antiplatelet therapy produced similar odds ratios (ORs) for an increase in risk of all-cause stroke or systemic embolism (aspirin OR 1.87, 95% CI 1.26 to 2.8; dual antiplatelet therapy OR 1.93, 95% CI 1.42 to 2.64). Similar results were seen for major bleeding (aspirin OR 1.05, 95% CI 0.60 to 1.87; dual antiplatelet therapy OR 1.1, 95% CI 0.83 to 1.47).

Although the review did not assess clopidogrel on its own, the NMA demonstrated a non-significant increase in risk with dual antiplatelet therapy. Thus, in the economic analysis, the EAG has used this evidence to form the assumption that, in the long-term model, clopidogrel is as effective as aspirin for patients with undetected AF; therefore, the effectiveness estimates used for antiplatelet therapy in the adapted DOAC model remain unchanged. The EAG acknowledges that this is a simplifying, conservative assumption, and a limitation of the analysis.

Mortality

Mortality risk implemented in the long-term DOAC model was estimated using the competing-risks NMA, as described previously for the age and sex split obtained from the CRYSTAL-AF trial. The mortality risk is then adjusted for general population mortality, using life tables for England and Wales101 for each age group beyond the baseline age and weighted by the proportion of males and females in the CRYSTAL-AF trial.

Anticoagulation treatment

For patients diagnosed with AF, anticoagulation treatment with either a DOAC or warfarin would be prescribed. However, analysis of prescribing trends, based on data from the openprescribing.net database139 published by the University of Oxford, has shown that prescriptions of warfarin have been declining, with prescriptions of DOACs overtaking warfarin in April 2018. It should be noted that the prescribing data are not broken down by indication, but an assumption can be made that DOACs are becoming the preferred treatment for patients requiring anticoagulation. Therefore, for the base-case analysis, the EAG assumed that all patients with detected AF will start treatment on a DOAC (i.e. apixaban, dabigatran etexilate, rivaroxaban or edoxaban).

The results from the DOAC model indicate that DOACs are more cost-effective than warfarin and that apixaban is the most cost-effective DOAC treatment (see the subsection that reports the results of the SLR: Models assessing the long-term impact of anticoagulation therapy for patients with atrial fibrillation). However, prescribing data show that apixaban accounts for only 48% of all DOAC prescriptions, with the remainder distributed between rivaroxaban (44%), dabigatran etexilate (6%) and edoxaban (3%). 139 In the base-case analysis, these proportions are used in the short-term Excel model. The proportion of patients on each of the treatments is based on the proportion of prescriptions of each drug between September 2017 and September 2018 from the openprescribing.net database. 139 In the short-term Excel model, the proportion of patients on each DOAC treatment is used to weight long-term costs and benefits. The EAG performed a scenario analysis including a proportion of patients on warfarin, as there may be clinicians who would prescribe this treatment to newly diagnosed AF patients. See Chapter 5 for further details.

Treatment switching probabilities

In the long-term DOAC model, depending on the occurrence of IS or SAEs (such as ICH), treatment switching can occur (see Figure 5). For patients on first-line DOAC treatment, second-line treatment may be either warfarin or no treatment. For patients who fail on warfarin, no further treatment is given. 106 The probability of a patient switching treatment after experiencing an event was based on clinical expert opinion obtained by the authors of the DOAC model.

Device costs

The device costs of the Reveal LINQ, BioMonitor 2-AF and Confirm Rx ICMs used in the model are £1800, £1030 and £1600, respectively. These costs were supplied by the manufacturers of each device. The manufacturer of the Reveal LINQ also provides an optional triage service, FOCUSON. The company provided two cost options for FOCUSON: the first option is £187 per patient per year and the second option is £374 per patient per device. Both options are explored in scenario analyses, presented in Chapter 5.

The manufacturers of each of the devices have indicated that no additional training is required to perform the insertion procedure, as it is expected that staff will already have the necessary skills, competencies and experience in performing sterile device insertion procedures. However, the manufacturer of the Reveal LINQ device stated that training for staff is included in the cost of the Reveal LINQ ICM system. The manufacturer of BioMonitor 2-AF also stated that training is offered by the company but did not indicate whether the cost of this is covered by the device cost. Therefore, the EAG has assumed no additional costs of training for the base-case analysis.

Furthermore, the costs of reviewing alerts generated by the ICM have not been included in the base-case analysis as no data were available regarding the average number of alerts generated per day/month that require review. However, based on discussions during the scoping phase of the topic, clinical experts advised that reviewing alerts is relatively quick and would form part of the clinician’s normal workload, although it is dependent on volume.

Implantation and device removal costs

Table 21 presents the ICM implantation costs per patient implemented in the short-term Excel model. The costs of implantation for the base-case analysis are based on resource use assumptions provided by the EAG’s clinical experts for this report. The clinical experts also provided alternative resource use assumptions that are explored in a scenario analysis. The company for the Reveal LINQ device also provided a costing study comparing the costs of the Reveal XT implanted in a cath lab setting versus the Reveal LINQ implanted in a sterile procedure room setting. 141 The resource use assumptions for this study were costed by the EAG and used in a scenario analysis. The cost for removal of an ICM device implemented in the Excel model is £238, taken from the NHS reference costs schedule 2017–18 (EY13Z – removal of electrocardiography loop recorder, outpatient setting, treatment function code 320). 109

The most common AEs associated with implantation of an ICM, based on data from the CRYSTAL-AF trial, include infection (1.4%), pain (1.4%) and irritation or inflammation at the insertion site (1.9%). 37 However, there was no further detail on how severe the AEs were in the CRYSTAL-AF trial; as a result, the EAG has not included AE costs in the base-case analysis. However, it is anticipated that, given the relatively small proportions of patients experiencing each AE, it is not expected to have a substantial impact on the cost-effectiveness of the ICM devices.

Comparator arm costs

Costs of the comparator are based on the standard of care arm from the CRYSTAL-AF trial. 90 As presented in Table 17, the standard of care arm comprises various monitoring tests that are performed at 3-monthly intervals for the duration of the trial (3 years). The unit cost of monitoring is estimated to be £141, based on the NHS reference costs schedule 2017–18 [Healthcare Resource Group (HRG) code EY51Z – ECG monitoring or stress testing (outpatient procedures, service code 320)]. 109 Table 22 presents the weighted cost of monitoring per cycle based on number of tests per patient recorded in the CRYSTAL-AF trial (see Table 17).

In UK clinical practice, it is probable that monitoring tests will be performed only if a patient presents with symptoms. Therefore, the EAG explored a conservative scenario in which the cost of SoC is zero. See Chapter 5 for more detail.

Follow-up costs

In the CRYSTAL-AF trial, follow-up visits were scheduled at 1, 6 and 12 months and then every 6 months until trial closure, for both arms of the trial. However, the EAG’s clinical experts advised that patients with an ICM are likely to have a follow-up visit 1 month post surgery only, and then after that will be remotely monitored, unless patients request a face-to-face appointment. The clinical experts’ advice aligns with information provided in the company submissions. As a result, because of the nature of virtual continuous follow-up with the ICM device, there is a reduction in the need for physical follow-up visits. However, once AF is detected, patients will need to be seen by a clinician to start anticoagulation treatment.

Therefore, the EAG assumed, for the base case, that all patients with an ICM will have one face-to-face follow-up appointment after 1 month and then a subsequent follow-up appointment when AF has been detected. For the SoC arm, follow-up is at 1, 3, 6 and 12 months, as per advice from the EAG’s clinical experts, and the costs of these follow-up appointments are applied to all patients who do not have detected AF. However, after 12 months, any newly AF-detected patients in the SoC arm will have the cost of a subsequent follow-up appointment applied to account for being identified. Table 23 presents the unit cost of follow-up appointments implemented in the short-term Excel model.

Pharmacotherapy costs

As mentioned previously, the DOACs considered in the model are apixaban, dabigatran etexilate, edoxaban and rivaroxaban. Based on clinical expert opinion, antiplatelet treatment in the model is clopidogrel. Warfarin (INR 2–3) was considered only in a scenario analysis. Drug costs used in the DOAC model are presented in Table 24. The costs of DOACs and clopidogrel used in the DOAC model were updated using prices obtained from the BNF September 2018–March 2019 edition. 142 The original cost of warfarin used in the DOAC model (which including monitoring costs) was uplifted to 2018 prices for the current analysis. 106 All drugs considered in the model are taken orally; therefore, it has been assumed that there are no administration or monitoring costs.

Acute and chronic care costs of atrial fibrillation and anticoagulant-related events

In the long-term adapted DOAC model, acute management costs for IS, ICH, systemic embolism, TIA, MI, deep-vein thrombosis (DVT), pulmonary embolism (PE) and CRB are considered. 106 The acute costs of IS and ICH in the DOAC model are derived from a UK-based population study, which estimated the acute and long-term costs of stroke in AF patients. 143 For the current analysis, costs were uplifted to 2017 prices using the ONS Consumer Price Inflation Index for Medical Services (DKC3). 2 All other event costs were derived from NHS reference costs and updated using the latest schedule (2017–18). 109 Acute event costs are presented in Table 25.

To ensure consistency, cost assumptions from the original model have been maintained. The authors of the original model assumed that the cost of MI obtained from NHS reference costs accounts for only direct hospitalisation, and, therefore, doubled the total costs to account for follow-up costs. Furthermore, the cost of sudden fatal PE is assumed to be zero, and patients who have a non-fatal PE are assumed to accrue the full cost of PE.

The costs of chronic IS and ICH management in the DOAC model are also derived from the study by Luengo-Fernandez et al. 143 The study estimated the annual cost of stroke, stratified by severity, in the post-acute phase (3 months post index event). The mean cost was calculated by weighting the cost of stroke by severity by the number of events, excluding deaths within 90 days, uplifted to 2018 prices (Table 26) for the current analysis. As per the original model, it is assumed that the cost for ICH is the same as stroke.

Addition of optional FocusOn triage costs

For the Reveal LINQ device only, the company provides a triage service, which can be provided in two ways. Option 1 provides the service at a cost of £187 per patient per year, whereas option 2 provides the same service but at a one-off fee of £374 per patient per device. Each option was considered as a separate scenario.

Addition of optional BioMonitor 2-AF devices

The BioMonitor 2-AF has the option to include a remote assistant device and CardioMessenger transmitter, at a cost of £230 and £400, respectively. These costs were included as part of the intervention cost and considered as separate scenarios.

Different time horizons (1 year, 2 year)

This scenario assumed that the ICM devices detect for a period of 1 year and 2 years only. This means that any detections that were identified in the CRYSTAL-AF trial beyond these time points were assumed to be missed by the devices, thereby reducing the benefits of the ICMs in comparison with SoC.

Constant detection rate

As an alternative to using the detection data directly from the CRYSTAL-AF trial, this scenario uses the 36-month detection proportion to calculate a constant monthly detection rate using the following formula:

where r_m is the monthly rate and p₃₆ is the proportion who are detected at 36 months. The monthly proportions, p_m, are then calculated as:

where t is the time in months.

Using each directly acting oral anticoagulant separately to determine the long-term outcomes following atrial fibrillation detection

Instead of taking the weighted long-term DOAC outcomes based on the usage data, this applied the outcomes for each DOAC alone as separate scenarios.

Inclusion of warfarin as a treatment option for patients diagnosed with atrial fibrillation

Currently, warfarin is still in use for the treatment of AF, although, based on clinical expert opinion, the current primary treatments for newly diagnosed AF patients are DOACs. However, given that data suggest that ≈50% of anticoagulation usage comprises warfarin, the EAG conducted a scenario to test the impact on the ICER of this usage. 139 This scenario applied the same approach to weight the costs and QALYs for DOAC treatment from the DOAC model, but also included warfarin as an option in this weighting. Therefore, this applied 50% of the warfarin outcomes and reduced the weighted DOAC outcomes used in the base case by 50%.

No removal of devices

The base-case analysis assumes that all devices are removed at the end of their battery life. This scenario assumes that the devices will not need to be removed at all, as clinical expert advice suggests that they are safe to remain in place indefinitely.

Implanter and implanter assistant assumptions

Two separate scenarios were conducted, which assumed that the implantation was performed by a cardiac physiologist (band 7) and assisted by a cardiac physiologist (band 5), respectively.

Implantation assumptions based on Kanters et al.141

This scenario assumes that a cardiac physiologist (band 7) performs the implantation, assisted by a nurse (band 5). The assumed time required for the cardiac physiologist (band 7) is 25.6 minutes, and for the nurse (band 5) is 43.1 minutes, based on data from Kanters et al. 141

No monitoring for standard of care

This scenario removes all monitoring costs from the SoC group and assumes that no incidences of AF are detected, that is assuming a greater benefit for the ICM groups but also an increased total cost relative to the SoC group.

A discounted ICER for each scenario analysis is given in Table 30.

One-way and two-way sensitivity analyses

The EAG conducted a number of sensitivity analyses around the cost inputs that were based on estimates (e.g. NHS reference costs), the outcomes applied from the long-term DOAC model, that is total costs and QALYs per cycle obtained from the long-term DOAC model, and the discount rate applied.

The most recent publication of NHS reference costs (2017–18) no longer gives an IQR for the costs associated with each HRG. Given the lack of data to inform the variation around the mean estimate, the EAG assumed a standard error of 20% of the mean value for each parameter. For DOAC outcomes (costs and QALYs), two-way sensitivity analyses around the 2.5th and 97.5th percentiles of the 10,000 samples for each cycle were used as the lower and upper limits, respectively, and the discount rate was lowered to 1.5% (as per the NICE Guide to the Methods of Technology Appraisal 2013145), as well as being increased to 6%. The summary of the inputs along with the results is given in Table 31.

Clinical

The clinical evidence systematic review sought to identify RCTs and comparative observational studies that compared any of the three devices (Confirm Rx, BioMonitor 2-AF and Reveal LINQ) with at least 24 hours of Holter (external electrocardiography monitoring to detect AF in people who have had a CS). Only a single RCT assessing an earlier Medtronic Reveal model (XT rather than LINQ) met the original review eligibility criteria (the CRYSTAL-AF trial),37 and so the criteria were widened in an attempt to find evidence for the BioMonitor 2-AF, Confirm Rx and Reveal LINQ. First, non-comparative observational studies were sought within the correct CS population, and then evidence was considered from studies of mixed populations submitted by each company. Only the CRYSTAL-AF trial37 falls within the eligibility criteria outlined in the original published protocol for this DAR, and so the additional evidence should be interpreted with caution. Therefore, the CRYSTAL-AF trial37 (n = 441) represents the most robust clinical evidence available to inform the decision problem, albeit assessing the Reveal XT.

The CRYSTAL-AF trial was an open-label study that compared the Reveal XT ICM with conventional follow-up in a population that had had CS and no history of AF after extensive diagnostic work-up. The study was conducted in North America and Europe and the population was considered by the EAG’s clinical experts to be generally applicable to patients who would be eligible for an ICM in UK clinical practice.

Twenty-six single-arm observational studies were identified after widening the eligibility criteria to include non-comparative studies. The studies were conducted in North America and western Europe (one in the UK) and all assessed the Reveal XT and Reveal LINQ in populations that had had a CS; none provided evidence to assess the efficacy of the BioMonitor 2-AF (other than a mixed-device study that did not provide results by individual device) or the Confirm Rx for patients who have had a CS. The observational studies represent a wide sample of patients who have received an ICM in practice (n = 3414) and provide evidence for the Reveal LINQ and for additional outcomes specified in the NICE final scope that were not available from the CRYSTAL-AF trial.

All of the observational studies were single-arm and therefore were at a high risk of bias, although three conducted within-patient comparisons of ICM versus other monitoring strategies. 14,15,72 Key sources of heterogeneity between the observational studies include patient demographics (mean or median age of 52–72 years), rigour of stroke assessment, stroke risk score (CHA₂DS₂VASC score of 3–5), definition and adjudication of AF, and length of follow-up; published data and the EAG’s clinical experts suggest that these are all likely to affect AF detection and other clinical outcomes. 34

Mixed population studies recommended by the companies as having potential data for inclusion in the review have also been included, as no data were identified for the Confirm Rx or BioMonitor 2-AF in the CS population and only limited outcome data were identified in the CS population for the Reveal LINQ. A full systematic literature search was not conducted to validate their inclusion owing to time constraints and concerns regarding the applicability of their results to the CS population. 34 The data presented from these studies may therefore be subject to study selection bias as well as clinical heterogeneity due to the variation in the patient populations of each of the studies. In total, there was one study of the Confirm DM2102 (an older model of the Confirm Rx), five studies of the BioMonitor 2-AF (all used BioMonitor 2 but only one specified it as the ‘-AF’ model) and five studies of the Reveal LINQ or XT [three studies used the Reveal LINQ and three studies used the Reveal XT (note that one study included both devices)]. All of these mixed population studies are either single-arm observational studies or they provide DTA data for the ICM using Holter monitoring as the reference standard.

The CRYSTAL-AF trial was designed to measure diagnostic yield rather than accuracy, and none of the observational studies provided comparative DTA between an ICM and standard monitoring. Two studies modelled patient AF detection data from the CRYSTAL-AF trial (Choe et al. 14) and a large patient registry (Ziegler et al. 15) with repeated iterations (10,000) to estimate the number of patients whose AF would not have been detected should an intermittent monitoring strategy have been used (based on the assumption that an ICM has 100% sensitivity). The studies found that the best-performing intermittent monitoring strategy detected less than one-third of the AF detected by the ICM (ranging from around 3% for a single 24-hour Holter monitor to 30% with a quarterly 7-day Holter monitor). Studies reporting false positive rates as the proportion of episodes detected by an ICM algorithm that were not subsequently verified by a clinician were highly dependent on the ICM model and device programme settings.

The results of the mixed population DTA studies suggested that the enhancements over time to the AF diagnosis algorithm in the Reveal XT and Reveal LINQ ICMs has improved the DTA (sensitivity and specificity) of the ICMs. However, it should be noted that these data are not exclusively in the CS population and the data in the XPECT50 and Reveal LINQ Usability87 studies used to make some of these comparisons are heterogeneous owing to differences in the way in which the reference standard was applied (Holter monitoring for 48 hours vs. 24 hours, respectively) and differences in the patient populations (e.g. reasons for ICM insertion). Nonetheless, these data suggest that the Reveal LINQ is likely to be as effective as, if not better than, the Reveal XT at detecting AF (as the Reveal LINQ has fewer false positives and false negatives). Therefore, the AF detection rate from the CRYSTAL-AF trial is potentially a conservative estimate for the Reveal LINQ, given that it was the Reveal XT that was used in the CRYSTAL-AF trial.

A naive comparison of the sensitivity and specificity data from non-CS or mixed populations in the studies flagged of relevance by the respective companies of the Confirm DM2102 (an older model of the Confirm Rx) and Reveal LINQ suggests that they both have 100% sensitivity for AF detection, whereas specificity varies (85.7% and 99.0%, respectively); the BioMonitor 2 (confidential information has been removed). However, it should be noted that the studies are subject to clinical heterogeneity in terms of the patient populations, interventions and study designs, as well as the reference standards. The device-related performance of ICMs is known to be dependent on the patient population and the incidence rate of AF, as well as the reference standard; therefore, this naive comparison should be interpreted with caution as these data are not necessarily reflective of the respective ICMs performance in CS patients. In addition, they do not necessarily reflect the performance of the current device model firmware; for example, the Confirm Rx data are based on an earlier model.

Atrial fibrillation detection rate was the primary outcome in the CRYSTAL-AF trial (at 6 months) and in all 26 observational studies. In the CRYSTAL-AF trial, AF detection was higher with the Reveal XT than with conventional follow-up at all time points. At the primary 6-month analysis, AF had been detected in 19 (8.6%) patients with an ICM and in 3 (1.4%) patients in the conventional follow-up group. By 36 months, the number of patients detected were 42 (19%) with an ICM and 5 (2.3%) receiving conventional follow-up, demonstrating the continued and increasing benefit of ICM monitoring. AF detection rates reported at the primary follow-up (6–24 months) across the 26 observational studies were highly variable, ranging from 6.7%38 (Reveal LINQ and XT at 12 months) to 40.9%68 (Reveal XT, unknown follow-up). These data demonstrate that, even within a CS population, AF detection rates are highly variable, and it is impossible to make any meaningful comparison between the observational studies and the CRYSTAL-AF trial. Observational studies reporting AF detection at different lengths of follow-up indicate that a minority of patients are diagnosed within the first month (mostly in the region of 10% of those detected by 1 year), around 70–80% by 6 months, and a small number beyond 1 year of monitoring. 15,52,58,72,74 In comparison, the 36-month data from the ICM arm of the CRYSTAL-AF trial show higher proportions of AF diagnosed at 1 month (19.0%) and beyond 12 months (31.0%), and a lower proportion at 6 months (45.2%), than with the observational studies. When described, all or most AF detected was asymptomatic and so would probably not have been picked up without continuous ICM monitoring.

Median time to AF detection was longer for patients with the Reveal XT in the CRYSTAL-AF trial than for those receiving conventional follow-up at 6, 12 and 36 months (p-value not reported). Nevertheless, the benefit of the ICM increased with length of follow-up because very few patients in the conventional follow-up arm were diagnosed, whereas detection increased steadily in the group with an ICM (36 months: HR 8.8, 95% CI 3.5 to 22.2; p < 0.001). The observational studies showed highly variable median time to first AF detection, ranging from 21 to 217 days (average follow-up of between 7 and 20 months); nevertheless, the results are still broadly consistent with the results from the CRYSTAL-AF trial, in which median time to AF diagnosis was 41 (IQR 4–84) days at 6 months’ follow-up, 84 (IQR 18–265) days at 12 months’ follow-up and 8.4 months (IQR not reported) at 36 months’ follow-up.

Three of the observational studies, primarily of the Reveal LINQ, suggest that the proportion of patients for whom the ICM detected other non-AF cardiac arrhythmias is in the region of 10%; these arrhythmias consisted mainly of bigeminy, pause and bradycardia. No information was presented about whether or not and how the detected arrhythmias were treated to prevent related complications, and data on detection of other arrhythmias were not available from the CRYSTAL-AF trial. The value of this additional potential benefit of the ICMs is, therefore, unclear.

In the CRYSTAL-AF trial, > 90% of patients diagnosed with AF in the ICM arm started an OAC. Data were available for the conventional follow-up group irrespective of AF diagnosis, indicating that 8.3% were on an anticoagulant by 36 months (24 patients, whereas five had been diagnosed with AF by that time point). These data, along with the data from the observational studies, suggest that most patients with ICMs diagnosed with AF go on to receive long-term OACs. Time to anticoagulation and AEs related to anticoagulant use were not reported in any of the identified evidence. Subsequent stroke or TIA rates in the CRYSTAL-AF trial were reported to be 5.0% in the ICM arm versus 8.2% in the conventional follow-up arm at 6 months, 6.8% versus 8.6% at 12 months and 9.0% versus 10.9% at 36 months (p > 0.05). None of these data suggests statistically significant stroke prevention benefits of the Reveal XT versus conventional monitoring, although there is a trend to fewer events in the ICM arm. It was unclear how many recurrent strokes occurred in patients diagnosed with AF or on OACs, and no studies reported other thromboembolisms or related morbidities.

In the CRYSTAL-AF trial, the overall rate of serious AEs was similar between groups (≈25–30%), but more patients in the ICM group than in conventional follow-up group had non-serious AEs (18.6% vs. 4.1%, respectively). The CRYSTAL-AF trial reported that five devices (2.4%) were removed because of infection or pocket erosion, which was in line with the premature removal rates seen in the observational studies (0.9–5.7%). At 12 months’ follow-up, 3.4% of ICMs had been removed in the CRYSTAL-AF trial; this contrasts with results reported in the Ritter et al. 72 (Reveal XT) study, in which removal after AF detection was offered in addition to removal for other reasons; 30% of patients had their ICM device removed during the study (median follow-up time in the study for all patients was 13 months). In the absence of further data, it is unclear why the removal rate was so high in Ritter et al. 72 However, device-related AEs, such as pain and infection, were consistently low in the CRYSTAL-AF trial, the single-arm observational studies and mixed population studies, suggesting that ICMs are generally well tolerated.

The EQ-5D data collected throughout the CRYSTAL-AF trial were (confidential information has been removed). The CRYSTAL-AF trial did not collect any other ease of use or patient acceptability data, and information from the observational studies was anecdotal. However, company submissions and the EAG’s clinical experts reported that the newer models of the ICMs (e.g. Reveal LINQ and Confirm Rx) were easier to insert and were suitable for insertion by trained nurses and cardiac physiologists.

Eight ongoing studies of potential relevance were identified, although only five (three RCTs and two observational studies) reported details of their status and the ICM being studied. None of the ongoing studies include the BioMonitor 2-AF. The three ongoing RCTs all include the Reveal LINQ but only one RCT is in a discrete CS population; this is a Canadian trial comparing the clinical effectiveness and cost-effectiveness of the Reveal LINQ ICM with external loop recording in 300 CS patients, which is estimated to complete in December 2019 (PERDIEM). 47 There was only one ongoing study identified relating to the Confirm Rx: the SMART registry,45 a post-approval study planning to recruit at least 2000 patients with Confirm Rx across multiple indications, but with a planned subgroup analysis for CS; completion is expected in December 2020. These studies may help to provide further clinical data for these two ICMs, although they will not address the lack of comparative data between the ICMs and do not provide any comparative data for the Confirm Rx or BioMonitor 2-AF against either Holter monitoring or other ICMs.

Economic

As mentioned previously, only one RCT (CRYSTAL-AF) was identified in the clinical effectiveness SLR, which assessed the impact of using an ICM compared with SoC, in a CS population in which there was a suspicion of paroxysmal AF. The CRYSTAL-AF trial reported data on AF detection rates for SoC and the Reveal XT device, which is an earlier model of the Reveal LINQ device. No data were obtained for BioMonitor 2-AF or Confirm Rx. As a result, a strong assumption was made in the economic analysis, based on clinical expert opinion, that the effectiveness of different ICMs is similar; thus, the detection rates obtained from the CRYSTAL-AF trial were used for all the ICM devices under assessment.

The results from the de novo economic model were ICERs, also known as cost per QALY gained. The results of the pairwise analysis, that is each ICM device compared with SoC, demonstrate that ICMs could be considered cost-effective at a £20,000–30,000 threshold compared with SoC. When each device is compared incrementally, the BioMonitor 2-AF dominates the Reveal LINQ and Confirm Rx. However, the results for the BioMonitor 2-AF and Confirm Rx should be viewed with caution, as no data were available for any version of these devices in the CS population and, as a result, there is substantial uncertainty in the results.

The EAG conducted various scenario and sensitivity analyses and found that the scenario that caused the most substantial change in the ICER for all three devices was the inclusion of warfarin. From the one-way sensitivity analysis, the key driver of the cost-effectiveness results relates to outcomes (i.e. total costs and QALYs) obtained from the long-term DOAC model, which for Reveal LINQ and Confirm Rx exceeded the £30,000 cost-effectiveness threshold.

The EAG conducted a SLR to identify any published economic evaluations of ICM devices for the detection of AF in a CS population that could be used to inform the current analysis. One study was identified that assessed the cost-effectiveness of the Reveal XT ICM (a predecessor of the Reveal LINQ) compared with SoC in a CS population from the UK perspective.

The model was developed using a Markov structure with three main health states for AF status: AF free, AF detected and AF undetected. Patients start in the AF-free state, from which they can move to AF-undetected or AF-detected states at any given model cycle. From the AF-undetected state, patients can either remain or move to the AF-detected state, and patients remain in the AF-detected state unless the patient experiences a subsequent cerebrovascular event or bleeding event. Detection rates of AF were based on data from the CRYSTAL-AF trial. 37

Results of the deterministic base-case analysis showed that the ICM was £2587 more expensive than SoC and provided a benefit of 0.151 QALYs, resulting in an ICER of £17,175 per QALY gained. This ICER is lower than the EAG’s results for the Reveal LINQ, which estimated that the ICM was £1687 more expensive than SoC and provided a benefit of 0.07 QALYs, resulting in an ICER of £24,875. The EAG’s short-term model was informed by the model structure used by Diamantopoulos et al. ,90 as it includes the health states of AF detected and AF undetected, with data informing the proportions in each health state per model cycle based on the results from the CRYSTAL-AF trial. 90 However, the approach to modelling long-term outcomes for patients with AF who are either detected and on anticoagulation treatment or undetected and on antiplatelet treatment, is based on a published DOAC cost-effectiveness model. 106

The EAG’s model produces incremental costs that are lower than those of Diamantopoulos et al. 90 and this can be attributed to a lower baseline hazard of IS used in the long-term DOAC model and, therefore, lower health-state costs. Furthermore, there were differences between the two models in the way in which monitoring costs were estimated. The EAG used data on the monitoring tests performed per person per year in the control arm of the CRYSTAL-AF trial, obtained from Diamantopoulos et al. ,90 to estimate costs for SoC in the current analysis. Minor differences in SoC costs between the two models are attributed to a change in the NHS reference cost used in the analysis (£137 in 2016, increased to £141 in 2018). 90,109 In addition, the EAG used a different methodology of calculating the per-cycle cost of SoC, by calculating the cost per year of the monitoring tests and dividing the costs by the number of model cycles per year. In the Diamantopoulos et al. 90 model, the per-cycle probability of each test was estimated and used to weight the unit cost per cycle.

In addition, the incremental QALY gained for the EAG model is lower than that of the Diamantopoulos et al. 90 model. The EAG considers that the difference in QALYs can also be attributed to a lower baseline hazard of IS used in the long-term DOAC model.

It should be noted that, in the model by Diamantopoulos et al. ,90 the entire cohort (no AF, AF detected and AF undetected) is modelled for clinical outcomes. However, the EAG considered that, clinically, outcomes for the no-AF cohort would be the same in each arm of the model (ICM and SoC), and so essentially cancel out, hence a focus on the overall incremental costs and QALYs between the two models.

Clinical expert opinion suggests that an additional benefit of ICMs devices is the ability to detect non-AF arrhythmias, potentially preventing other events. However, data on incidental findings from ICMs were found only in single-arm observational studies, as previously mentioned, and are of poor quality. As a result, it is unclear how detection of other non-AF arrhythmias differs between standard care and ICMs and, furthermore, how a patient’s treatment pathway changes. Therefore, understanding the differences in costs and benefits for incidental findings for ICMs is problematic. However, the EAG considers that, if without an ICM, some of these arrhythmias remain undetected, then the impact on the cost-effectiveness estimates would be favourable towards ICMs, but the extent of the impact is difficult to determine.

Strengths and limitations

Economic

One of the main strengths of the economic analysis is that outcome data were available from a RCT on the effectiveness of an ICM compared with SoC in the CS population. Therefore, reliable estimates of AF detection rates for an ICM (Reveal XT in this case) and SoC were used to estimate the long-term outcomes, costs and benefits of anticoagulation therapy versus antiplatelet therapy, using a previously developed and established economic model.

Even though the strength of the economic analysis is data being available for an ICM device in the correct target population, data were not available for each of the devices in the NICE final scope of this DAR (i.e. Reveal LINQ, BioMonitor 2-AF and Confirm Rx). For the Reveal LINQ, this is less of a limitation, as the data used in the analysis are based on an earlier model: the Reveal XT. However, for the BioMonitor 2-AF and Confirm Rx, the assumption of clinical equivalence with the Reveal XT is a strong assumption. The manufacturer of the Reveal devices advised that, with each iteration of the device, improvements are made to the algorithm to improve sensitivity and specificity, such that the Reveal LINQ has been estimated to have 100% sensitivity. Furthermore, the EAG’s clinical experts advised that the detection rates for each of the devices will be at least as good as the rates seen in the CRYSTAL-AF trial. With this in mind, caution should be applied when interpreting the cost-effectiveness results for the BioMonitor 2-AF and Confirm Rx, as the strong assumption of equivalence with the Reveal LINQ is not based on evidence of the performance of any version of these devices in the CS population, resulting in substantial uncertainty around the ICER.