3 Evidence

The diagnostics advisory committee considered evidence from several sources on implantable cardiac monitors (BioMonitor 2‑AF, Confirm Rx and Reveal LINQ) to assess for suspected paroxysmal atrial fibrillation in people who have had a cryptogenic stroke. Full details of all the evidence are in the committee papers.

Clinical effectiveness

3.1 The external assessment group (EAG) did a systematic review to identify evidence on the clinical effectiveness and diagnostic accuracy of implantable cardiac monitors to detect suspected atrial fibrillation after cryptogenic stroke. The devices reviewed were:

  • BioMonitor 2‑AF

  • Confirm Rx

  • Reveal LINQ.

3.2 Studies were included if they assessed the devices in people who had had a cryptogenic stroke or cryptogenic transient ischaemic attack (TIA), and paroxysmal atrial fibrillation was suspected. Because of the small number of studies identified, the requirement for at least 24 hours of outpatient external ambulatory electrocardiogram (ECG) monitoring without atrial fibrillation detection before the devices were implanted (as per current practice) was not applied. Also, data from earlier versions of the devices were considered.

3.3 Because only 1 study (the CRYSTAL‑AF randomised controlled trial) met the EAG's initial eligibility criteria, the EAG relaxed the study inclusion criteria to consider single-arm observational studies. The EAG did not change the population inclusion criteria because it considered that data from non-cryptogenic stroke populations would not represent the device's performance in people with cryptogenic stroke or TIA. This is because non-cryptogenic stroke populations have different incidence rates of atrial fibrillation. However, the EAG provided a summary of studies highlighted by the device manufacturers that were not done in a cryptogenic stroke population. The EAG further highlighted that the patient population, duration of monitoring and the type of atrial fibrillation would affect estimates of device performance.

Comparative studies

3.4 One study (reported in 6 publications) compared the effectiveness of using 1 of the devices with conventional follow up: the CRYSTAL‑AF study. This was an open-label, parallel group randomised controlled trial that used Reveal XT (an earlier version of Reveal LINQ). The XT model is larger. The EAG said that evidence from diagnostic accuracy studies (see sections 3.48 and 3.49) suggests that Reveal LINQ has better specificity and sensitivity than the XT, is easier to implant and causes fewer complications.

3.5 People aged 40 or older who had a recent episode of cryptogenic symptomatic TIA or recent episode of cryptogenic ischaemic stroke had Reveal XT (n=221) or conventional follow-up care (n=220). People with TIA were only enrolled if they had a visible lesion on MRI or CT that fitted the symptoms of the TIA, and at least 1 of the following symptoms: speech problems, limb weakness or hemianopsia. Follow up in the control group was ECG monitoring at the discretion of the site investigator. The study was done in 55 centres across 14 countries in Europe (none in the UK), Canada and the US. The EAG said that there were similar numbers of withdrawals between the 2 arms (except for crossovers, see section 3.12). Data were collected for up to 36 months of follow up, but relatively few people reached this point (24 in each arm). Mean duration of follow up was 20.3 months for Reveal XT and 19.2 months for conventional follow-up care.

3.6 The EAG noted that, although there were no significant differences in baseline characteristics between study arms, there were differences in the numbers of people with patent foramen ovale and history of prior stroke. However, these were small and unlikely to be because of systematic issues with randomisation. The clinical experts said that the population was slightly younger than people expected to be eligible for an implantable cardiac monitor in the UK. Also, a higher proportion of TIA (rather than stroke) would be expected in clinical practice (closer to 20%, instead of about 9% seen in each of the study arms). All patients would be expected to be taking an antiplatelet agent (about 96% in each arm were using an antiplatelet agent at baseline).

3.7 The EAG's clinical experts said that the tests used in the trial to define a stroke as cryptogenic were broadly the same as what would be done in the NHS. Pre-enrolment screening for atrial fibrillation was Holter monitoring for 71.2% of people (median duration of 23 hours, interquartile range 21 hours to 24 hours), and the remaining people had inpatient telemetry monitoring only. The EAG said that this meant almost 30% of people did not have any outpatient ECG monitoring (as specified in the scope) and that not all patients who did have outpatient ECG monitoring had it for at least 24 hours.

3.8 This trial was sponsored by Medtronic, manufacturers of the device used in the study. The EAG said that the authors of publications for this study reported employment, grants and personal fees from this company. The EAG considered that this was the most robust clinical evidence for Reveal LINQ, even though it relates to an earlier version of the device.

Non-comparative studies

3.9 Comparative data were not identified for BioMonitor 2‑AF, Confirm Rx or the current Reveal LINQ version. Therefore, the EAG reviewed single-arm observational studies in cryptogenic stroke (including TIA) populations to identify available data on these devices. Biotronik submitted a technical validation report comparing the accuracy of BioMonitor 2‑AF with Reveal LINQ during consultation (see section 3.21).

3.10 Twenty six observational studies (reported in 60 publications) were found. All but 1 study assessed either Reveal LINQ or Reveal XT. In 1 study (Israel et al. 2017), 13% of people used the BioMonitor (an earlier version of BioMonitor 2‑AF), but results were not reported by device. The EAG said that these studies therefore do not provide any data for BioMonitor 2‑AF or Confirm Rx, but that they did supplement data from the CRYSTAL‑AF study.

3.11 Sample sizes in the studies ranged from 14 to 1,247. Only 1 study (Cotter et al. 2013) was done in the UK. Most studies (17) were prospective single-arm observational studies. There were 5 retrospective studies (Asaithambi et al. 2018, Chalfoun et al. 2016, Heckle et al. 2018, Li et al. 2018 and Salahuddin et al. 2015). One did not report a clear methodology (Cotter et al. 2013). Ritter et al. (2013) did a within-patient comparison of Reveal XT and 7-day Holter ECG monitoring. Choe et al. (2015) used the CRYSTAL‑AF dataset to predict how many cases of atrial fibrillation detected by Reveal XT would have been detected by shorter length intermittent ECG monitoring strategies using simulations. Ziegler et al. (2017) presented data from a registry of people who had Reveal LINQ and used simulations to predict how many people with atrial fibrillation detected by the device would have been identified by shorter (non-continuous) ECG monitoring.

Quality assessment of studies

Randomised controlled trials

3.12 The CRYSTAL‑AF study was assessed using the Cochrane risk of bias 2.0 tool. The full quality assessment is in the diagnostics assessment report starting from page 28. There was some concern about risk of bias because the trial was open-label and not all people had the randomised intervention required by the study protocol (5.4% of people assigned to Reveal XT got conventional follow up; 2.7% of people assigned conventional follow up got Reveal XT). Also, device implantation was delayed for 11.5% of people who had Reveal XT (median length of delay was 6 days, interquartile range 1 day to 32 days). The EAG noted that results were analysed by intention-to-treat population, which included patients who did not have Reveal XT, received it late, or crossed over to conventional follow up. This means the estimated benefit of having the device may be underestimated. Delays in implanting Reveal XT were mostly short and unlikely to affect outcomes. The EAG said that the lack of blinding was unlikely to affect relative atrial fibrillation detection rates between groups. It noted that only a small number of people were followed up after 12 months, so the 24‑month and 36‑month results are likely to be less reliable than results from 6 months and 12 months, but the direction of this bias is unclear.

Observational studies

3.13 The EAG said that it was not able to formally quality assess the 26 additional studies identified that were not randomised controlled trials. However, it considered them all to be at high risk of bias because of their single-arm designs. Because of heterogeneity between the studies, the EAG did not consider it appropriate to pool results from these studies. This included the model of device used, detection settings, patient characteristics, rigour of stroke assessment, severity of index stroke, definition and adjudication of atrial fibrillation, and length of follow up.

Evidence on ability to detect atrial fibrillation

Diagnostic yield (atrial fibrillation detection rate)

3.14 Atrial fibrillation detection rate at 6 months was the CRYSTAL‑AF study's primary outcome (episodes had to last more than 30 seconds). At 6 months, 19 people were diagnosed with atrial fibrillation in the Reveal XT arm and 3 people in the conventional follow-up arm. More atrial fibrillation was detected with Reveal XT at all time points (see table 1). Reveal XT increased atrial fibrillation detection across all pre-specified subgroups (age, sex, race or ethnic group, index event, presence or absence of patent foramen ovale, and CHADS2 score), with no significant interactions. Most people who had atrial fibrillation detected by Reveal XT were asymptomatic (34 out of the 42 detected by 36 months). Estimated detection rates are higher in the 36‑month Kaplan–Meier analysis because of the non-informative censoring (that is, people who dropped out for reasons unrelated to the study) of patients lost to follow up (atrial fibrillation detection rate estimated as 30% with Reveal XT and 3% with conventional follow up).

Table 1 Atrial fibrillation detection in CRYSTAL-AF

Months

Reveal XT: cumulative number of patients with AF detected

(n [% ITT])

Conventional follow up: cumulative number of patients with AF detected

(n [% ITT])

1

8 (3.6)

1 (0.5)

6

19 (8.6)

3 (1.4)

12

29 (13.1)

4 (1.8)

24

38 (17.2)

5 (2.3)

36

42 (19.0)

5 (2.3)

Abbreviations: AF, atrial fibrillation; ITT, intention to treat.

3.15 All 26 observational studies reported atrial fibrillation detection rate. Detection rates varied widely, ranging from 6.7% to 40.9% (length of monitoring varied between studies). Several studies reported atrial fibrillation detection rates over multiple time points. The EAG said that the studies generally show that a minority of patients are diagnosed in the first month (about 10% of those detected by 1 year). Around 70% to 80% (of the total number of people with atrial fibrillation detected in a study) are diagnosed by 6 months, and a small number after a year of monitoring. All or most of the detected atrial fibrillation in the observational studies (when stated) were asymptomatic, as in CRYSTAL‑AF.

3.16 Two observational studies estimated how many atrial fibrillation episodes would have been detected by intermittent ECG monitoring. These used datasets generated by Reveal XT (in CRYSTAL‑AF; Choe et al. 2015) or Reveal LINQ (from a large registry of patients with the device [n=1,247]; Ziegler et al. 2017). The studies assumed Reveal devices had 100% sensitivity. The studies estimated that even the best intermittent ECG monitoring strategies would detect less than a third of atrial fibrillation detected by Reveal devices.

Diagnostic yield: other (non-atrial fibrillation) cardiac pathologies

3.17 CRYSTAL‑AF did not report any results for the detection of other cardiac pathologies.

3.18 Five non-comparative observational studies reported incidental detection of other arrhythmias. The EAG said that the proportion of patients detected with other arrhythmias is about 10% of the total number of people in a study. This mainly consists of bigeminy, pause and bradycardia. Two studies reported the breakdown of arrhythmias and gave rates of 1% (atrial flutter, cardiac arrest, sick sinus node, bigeminy, ventricular tachycardia) to 7% to 8% (atrioventricular block and ventricular extra systole). Full details are on page 51 of the diagnostics assessment report. The studies did not say if the other detected arrhythmias were treated, or if outcomes were improved because these arrhythmias were identified. Also, because these were non-comparative studies, the extent of any increase in detection compared with conventional follow up could not be determined.

Diagnostic accuracy

3.19 No data relevant to this outcome were reported in CRYSTAL‑AF.

3.20 Two non-comparative observational studies reported the proportion of episodes detected by the devices that were not verified as atrial fibrillation by a clinician. Li et al. (2018) reported 79.7% for Reveal LINQ and Israel et al. (2017) reported that over 90% of detected episodes were not confirmed by review (Reveal XT and BioMonitor). The EAG noted that Medtronic had said that the number of false positive alerts varies depending on the device model used, and the configuration for detection (including episode duration) that is programmed by the operator. Data on device accuracy (for all devices) in non-cryptogenic stroke populations from studies identified by manufacturers are presented later (see section 3.41).

3.21 During the first consultation on this guidance, the manufacturer of the BioMonitor 2‑AF submitted an unpublished technical validation report comparing the accuracy of the Reveal LINQ and BioMonitor devices. This was done by replaying ECG data recorded by a Holter monitor in a previous trial into the sensing electrodes of the Reveal LINQ and BioMonitor devices. The report stated that because the atrial fibrillation detection algorithm of the BioMonitor 2‑AF and BIOMONITOR III are the same, the results are applicable to both devices. People enrolled in the original trial had documented atrial fibrillation episodes or symptoms attributable to atrial fibrillation, were scheduled for catheter ablation or had had it, but were still experiencing atrial fibrillation-related symptoms. Of the participants, 70% had a history of paroxysmal atrial fibrillation. The rest had a history of persistent atrial fibrillation. However, the EAG highlighted that the report also stated that people with long-standing persistent or permanent atrial fibrillation were excluded. The EAG pointed out that this was contradictory and meant that the characteristics of the population in the study were not clear. At consultation on the draft guidance, BioMonitor's manufacturer commented that all ECG data fed into the devices as part of the study were for less than 48 hours. It said that therefore they represented paroxysmal atrial fibrillation episodes only. In the study, atrial fibrillation episodes detected, or not, by clinician assessment of the Holter monitor ECG trace was used to classify true and false positives and false negative atrial fibrillation episodes detected by the Reveal LINQ and BioMonitor. Atrial fibrillation episode sensitivity for BioMonitor and Reveal LINQ were 78.0% and 79.0% respectively. Patient-averaged positive predictive values were 98.7% for BioMonitor and 99.7% for Reveal LINQ.

Evidence on clinical outcomes

Time to diagnosis of atrial fibrillation

3.22 The EAG said that atrial fibrillation was detected in only 5 people in the conventional follow-up arm of CRYSTAL‑AF (and none after 24 months; see table 1). This means it is difficult to make any conclusions about the effect of using Reveal XT from the median time to atrial fibrillation detection data. Atrial fibrillation was detected in more people with longer follow up, and therefore the median time to detection also increased. There was a greater increase in the median time to atrial fibrillation detection with Reveal XT compared with conventional follow up across all time points. The EAG said that the low detection rate of atrial fibrillation in the conventional follow-up arm was the likely cause of this difference.

3.23 There were 18 observational studies that reported time from device insertion to atrial fibrillation detection. Average follow up ranged from 7 months to 20 months, and median time to first atrial fibrillation detection had a wide range, from 21 days to 217 days. When reported, interquartile ranges also showed high variability within studies.

Atrial fibrillation-related hospitalisation

3.24 No data were reported in CRYSTAL‑AF or the observational studies.

Incidence of outpatient monitoring

3.25 No data were reported in CRYSTAL‑AF or the observational studies.

Uptake of anticoagulants

3.26 Most people diagnosed with atrial fibrillation using Reveal XT started having an oral anticoagulant (more than 90%) in the CRYSTAL‑AF study. The reasons people did not start on anticoagulants after being diagnosed with atrial fibrillation were not clear. The EAG noted that some people who were not diagnosed with atrial fibrillation in the trial were also started on anticoagulants. Reasons for this were not provided.

3.27 Seven observational studies (Asaithambi et al. 2018, Carrazco et al. 2018, Christensen et al. 2014, Etgen et al. 2013, Li et al. 2018, Merce et al. 2013 and Seow et al. 2018) reported that uptake of anticoagulants for people with atrial fibrillation detected by Reveal XT or LINQ was high: between 83.3% and 100%.

Time to start of anticoagulants

3.28 No data were reported in CRYSTAL‑AF or the observational studies.

Incidences of device failure and removal

3.29 No incidence of Reveal XT failure was reported in CRYSTAL‑AF. The device had to be removed early because of infection or pocket erosion from 5 out of 208 (2.4%) people by 36 months.

3.30 Three non-comparative observational studies reported how many devices were removed during follow up. In Christensen et al. (2014), Reveal XT was removed prematurely in 5.7% people because of skin reactions and discomfort. A further 3.4% of people chose to have the device removed after more than 1 year without atrial fibrillation being detected. In Asaithambi et al. (2018), 2.6% of people chose to have Reveal LINQ removed, and for 0.9% of people the devices migrated or fell out. In Ritter et al. (2013), study participants were offered removal of Reveal XT once atrial fibrillation was detected. But they did not report how many of the 30% of removals were because of this, or for other reasons such as discomfort.

Ease of use of devices for clinicians

3.31 No data were reported in CRYSTAL‑AF or the observational studies.

Mortality

3.32 No data were reported in CRYSTAL‑AF or the observational studies.

Further strokes or TIAs

3.33 In CRYSTAL‑AF, a non-significant trend of fewer recurrent events (stroke or TIA) in the Reveal XT arm was reported (see table 2). The study was not powered for this outcome. It is not clear if the recurrent stroke or TIA events occurred in people who were diagnosed with atrial fibrillation or not.

Table 2 Cumulative incidence of further strokes or TIAs in CRYSTAL-AF

Month

Reveal XT (n=221): people having another stroke or TIA
(n [%])

Conventional follow up (n=220): people having another stroke or TIA
(n [%])

Hazard ratio
(95% CI)

6

11 (5.0)

18 (8.2)

Not reported

12

15 (6.8)

19 (8.6)

0.63 (0.22 to 1.80)

36

20 (9.1)

24 (10.9)

0.77 (0.30 to 1.97)

Abbreviations: CI, confidence interval; TIA, transient ischaemic attack.

3.34 Of the studies, 6 non-comparative observational studies reported variable incidences of secondary stroke or TIA in people with an implantable cardiac monitor: from 0% to 14.6%.

Other thromboembolisms

3.35 No data were reported in CRYSTAL‑AF or the observational studies.

Device-related adverse events

3.36 The EAG said that the incidence of device-related adverse effects (such as pain and infection) was relatively low for people who had Reveal XT implanted in CRYSTAL‑AF. However, adverse events did lead to the device being removed in 2.4% of patients. The proportion of people with serious adverse events was slightly higher for Reveal XT (30.8%) than conventional follow up (27.9%). More people had non-serious adverse events in the Reveal XT arm (18.6%) than in the conventional follow-up arm (4.1%). No details of these events were reported, and the EAG said that it was unclear why there was a difference between the study arms. Reveal XT is larger than Reveal LINQ.

3.37 For 5 non-comparative observational studies, there were no complications from the procedure or insertion site reported at follow up (length of follow up was not specified). These were Merce et al. (2013), Reinke et al. (2018) and Ritter et al. (2013) for Reveal XT; Poli et al. (2016) for Reveal LINQ and XT; and Israel et al. (2017) for Reveal XT and BioMonitor.

Anticoagulant-related adverse events

3.38 No data were reported in CRYSTAL‑AF or the observational studies.

Evidence on patient-reported outcomes

Health-related quality of life

3.39 Health-related quality of life data were collected in CRYSTAL‑AF using the EuroQol 5-Dimensions (EQ-5D) tool. Unpublished data were provided by the company as academic in confidence so they cannot be reported here.

Acceptability of the devices to patients

3.40 No data were reported in CRYSTAL‑AF or the observational studies.

Evidence from non-cryptogenic stroke populations

3.41 The EAG provided a narrative summary of studies in non-cryptogenic stroke populations identified by manufacturers of the devices. These studies were not done in populations who had exclusively had a cryptogenic stroke or TIA although some were in a 'mixed population' (less than 50% of the study population had a cryptogenic stroke or TIA and subgroup analysis was not provided). All studies were either single-arm observational studies or assessed the diagnostic accuracy of the devices compared with Holter monitoring. The EAG highlighted that the performance of the devices depends on the patient population, atrial fibrillation incidence rate, and the type of atrial fibrillation. Therefore, the results from these studies do not necessarily represent the devices' performance in people with cryptogenic stroke.

3.42 The EAG did not do a full systematic literature search to validate the inclusion of the studies. This was because of time constraints and concerns about the applicability of results to the cryptogenic stroke population. The EAG said that the data may have study selection bias as well as clinical heterogeneity caused by the variation in the patient populations of each of the studies.

Abbott Medical

3.43 The company said the Detect AF study (Nölker et al. 2016) was potentially relevant for assessing Confirm Rx. The EAG noted that the device used in Detect AF was the Confirm model DM2102. This is an older and larger model of Confirm Rx. The EAG was unsure how the software in this earlier version compared with the current Confirm Rx.

3.44 Detect AF was a prospective observational study. It assessed the diagnostic accuracy of the Confirm system in detecting atrial fibrillation compared with Holter monitoring (reference standard) with simultaneous use of the devices. In per-patient analysis, sensitivity of the Confirm system was 100%, positive predictive value was 64.0%, specificity was 85.7% and negative predictive value was 100%. Most of the episodes of atrial fibrillation detected by the Confirm system but not confirmed by the Holter monitor were because of irregular sinus rhythms. No adverse events associated with the device were reported.

Biotronik SE & Co

3.45 The EAG discussed 5 single-arm prospective observational studies (in 8 publications) provided by Biotronik. Three of these studies (which included data on diagnostic accuracy) were unpublished and were provided as academic or commercial in confidence so details cannot be reported here.

3.46 Reinsch et al. (2018) reported that BioMonitor 2 was successfully implanted in a catheterisation laboratory with a median time from first cut to final suture of 8 minutes (interquartile range 7 minutes to 10 minutes). Ooi et al. (2017) reported that all insertions of the device were made on first attempt in a catheterisation laboratory with a median time of 9 minutes (interquartile range 5 minutes to 14 minutes). Ooi et al. reported that 1 pocket infection occurred when using the device. Reinsch et al. reported that no devices implanted in the study migrated, and 1 person needed the device removing because of device-related pocket infection. Another patient complained of slight discomfort.

3.47 Reinsch et al. reported results from patient satisfaction surveys. Of the respondents, 7% reported moderate to severe pain and 20% reported mild pain within 24 hours of device insertion. One person reported a moderate impairment in daily life. Of the respondents, 63% said that the cosmetic result was 'very satisfying' and 30% said 'satisfying'.

Medtronic Limited

3.48 The EAG discussed 5 studies highlighted by the company. Two compared the diagnostic accuracy of Reveal devices (per-patient analysis) with Holter monitoring for detecting atrial fibrillation (Hindricks et al. 2010 and Sanders et al. 2016). In Hindricks et al. (2010), Reveal XT was used (the XPECT trial). Another study (Puerefellner et al. 2014) used data from this trial and recalculated accuracy estimates when changes were made to the atrial fibrillation detection algorithm. This incorporated data on P waves when classifying patients, and this algorithm change was applied in Reveal LINQ. Sanders et al. (2016) used Reveal LINQ. A subsequent study (Puerefellner et al. 2018) was published using this dataset (and the XPECT data) to calculate the accuracy of a modified algorithm for detecting atrial fibrillation (using the TruRhythm algorithm that has now been incorporated in the device). Data on the diagnostic accuracy reported in these studies are shown in table 3.

Table 3 Diagnostic accuracy estimates for Reveal XT and Reveal LINQ

Measure

XPECT study

Hindricks et al. 2010

(Reveal XT)

XPECT dataset

Puerefellner et al. 2014

(Reveal XT with P-sense enhancement)

LINQ usability study

Sanders et al. 2016

(Reveal LINQ)

LINQ usability dataset

Puerefellner et al. 2018

(Reveal LINQ with adaptive P-sense; TruRhythm)

Sensitivity (%)

96.1

96.1

97.4

100

Specificity (%)

85.4

90.0

97.0

99.0

Positive predictive value (%)

79.3

84.9

92.5

97.4

Negative predictive value (%)

97.4

97.5

99.0

100

Accuracy (%)

89.3

92.2

97.1

99.3

Positive predictive value, negative predictive value and accuracy for the XPECT dataset calculated by the EAG using data in Puerefellner et al. (2014).

3.49 The EAG said that the studies showed improved detection of atrial fibrillation by Reveal LINQ compared with Reveal XT. Changes made to the algorithm also improved detection. But the results should be interpreted with caution because these studies were not done in people who had had a cryptogenic stroke. However, the EAG said that these data suggest that Reveal LINQ is likely to be as effective as Reveal XT, if not better, at detecting atrial fibrillation. Therefore, the clinical data from CRYSTAL‑AF (which uses the Reveal XT) could be a conservative estimate of the clinical effectiveness of the device.

3.50 Mittal et al. (2015) reported adverse event data from 2 observational studies that used Reveal LINQ. An infection occurred in 1.5% of people, an adverse event in 4.0% and a serious adverse event in 1.1%.

Ongoing studies

3.51 The EAG identified 8 potentially relevant ongoing studies from searches of trial registries and electronic databases, in addition to company submissions. There are 3 ongoing randomised controlled trials assessing Reveal LINQ. Of these, 1 is in people with cryptogenic stroke. This is a Canadian randomised trial comparing the clinical and cost effectiveness of Reveal LINQ with external loop recording in 300 people who have had cryptogenic stroke. It was estimated to complete in December 2019 (PERDIEM; NCT02428140). One ongoing study identified is assessing Confirm Rx: the SMART registry (NCT03505801). This is a post-approval study planned for at least 2,000 patients with Confirm Rx across multiple indications, with a planned subgroup analysis for cryptogenic stroke. Completion was expected during 2019. At consultation on the draft guidance, a stakeholder submitted a recent conference abstract (Yokokawa et al. 2019). The abstract gave only limited methodological detail. In this study, people were randomised to have either Confirm Rx or Reveal LINQ implanted (n=80; 52 had cryptogenic stroke but no subgroup analysis was provided). The abstract reported that 28 of 51 atrial fibrillation events (55%) were detected accurately by Reveal LINQ and 131 of 301 atrial fibrillation events (44%) were accurately detected by Confirm Rx (p=0.13).

Cost effectiveness

Systematic review of cost-effectiveness evidence

3.52 The EAG did a systematic review to identify any published economic evaluations of implantable cardiac monitors to detect atrial fibrillation in people with cryptogenic stroke. There were 5 studies that met the EAG inclusion criteria. Of these, 2 assessed the cost effectiveness of Reveal XT compared with standard care monitoring (DeAngelis et al. 2016 and Diamantopoulos et al. 2016). Another study assessed BioMonitor 2‑AF (Maervoet et al. 2017; further details provided as unpublished report and model by the device manufacturer as commercial in confidence), and 2 studies did not indicate which implantable cardiac monitor was being assessed (Quiroz et al. 2017 and Thijs et al. 2018). Only 1 study (Diamantopoulos et al. 2016) was based on an NHS payer perspective and was discussed in the diagnostics assessment report.

Diamantopoulos et al. (2016)

3.53 This study was a cost–utility analysis. It compared use of Reveal XT in people who have had a cryptogenic stroke or TIA with conventional follow up, as assessed in the CRYSTAL‑AF study. A Markov model structure was used with 3 main health states for atrial fibrillation status: free, detected and undetected. The deterministic base case produced an incremental cost-effectiveness ratio (ICER) of £17,175 per quality-adjusted life year (QALY) gained for Reveal XT compared with standard care (£2,587 higher costs, 0.151 additional QALYs). The probabilistic ICER was lower.

3.54 The EAG considered that results from this model were potentially unreliable because there was uncertainty about how parameters in the model had been estimated. The estimation of treatment effects by indirect comparison, atrial fibrillation incidence and detection rates used in the analysis were particularly unclear. The study authors used indirect comparisons to estimate hazard ratios for the benefit of anticoagulants on the occurrence of ischaemic stroke, bleeding events, intracranial haemorrhages, extracranial haemorrhages and mortality. The EAG tried to verify these figures but was unable to because there were insufficient details in the publication about how the indirect comparisons were done and how publications that informed the analysis were identified. The EAG also considered that estimation of some of the hazard ratios could be flawed. For example, the authors estimated a hazard ratio to adjust mortality in the model, but the source data used are based on standardised mortality ratios. Furthermore, people without atrial fibrillation detected were assumed to be offered aspirin, but the EAG's clinical experts said that clopidogrel would be used as an antiplatelet treatment.

Modelling approach

3.55 The EAG developed a de novo economic model to assess the cost effectiveness of using implantable cardiac monitors (BioMonitor 2‑AF, Confirm Rx or Reveal LINQ) to assess for suspected paroxysmal atrial fibrillation in people who have had a cryptogenic stroke (including TIA).

Model structure

3.56 The EAG developed a 2-stage economic model. The first stage (an Excel model developed by the EAG) modelled people having either monitoring for suspected paroxysmal atrial fibrillation after a cryptogenic stroke (including TIA) with the implantable cardiac monitors or conventional follow up. Everyone starts the model having antiplatelet therapy (clopidogrel) for stroke prevention. At every 3‑month cycle in the model, a proportion of people have atrial fibrillation. For people with an implantable cardiac monitor, all cases of atrial fibrillation are detected, and treatment is switched to anticoagulants (atrial fibrillation detected). For people with conventional follow up, a proportion of people with atrial fibrillation are detected (and switch to anticoagulants) but most are not (atrial fibrillation undetected) and remain on antiplatelet therapy.

3.57 For the subsequent long-term anticoagulation model, the EAG adapted a published economic model to model the long-term effect of people with detected atrial fibrillation (anticoagulant treatment) or undetected atrial fibrillation (remain on antiplatelet therapy with clopidogrel). This is the 'adapted direct oral anticoagulant (DOAC) model' (Sterne et al. 2017 and Welton et al. 2017). People enter the model after having atrial fibrillation in an 'atrial fibrillation well' state. After this, clinical events can occur. These are TIA, ischaemic stroke, intracranial haemorrhage, myocardial infarction, clinically relevant (extracranial) bleed or systemic embolism (multiple events can happen to one person over the course of the model). The risks of these events happening in the model were based on a population with a history of ischaemic stroke and paroxysmal atrial fibrillation. The model structure is the same for people with detected and undetected atrial fibrillation. However, the probability of the events happening depends on the treatment used (anticoagulants or antiplatelet therapy).

Model population

3.58 The population in the model was people who had had a cryptogenic stroke (including TIA), when there was suspected paroxysmal atrial fibrillation. These people had had at least 24 hours of outpatient external ambulatory ECG monitoring that had not detected atrial fibrillation. Characteristics were based on the population in the CRYSTAL‑AF study, with a mean age of 61 years and about 65% people assumed to be men.

Comparator

3.59 In the model, the EAG used data from the control arm of CRYSTAL‑AF for the comparator. People in the study were assessed at scheduled visits (every 3 months) and unscheduled visits if they were having symptoms of atrial fibrillation. Tests included ECGs and Holter monitoring (for 24 hours, 48 hours or 7 days).

Model inputs

3.60 Diagnostic yield data from CRYSTAL‑AF were used for the number of people with atrial fibrillation detected by an implantable cardiac monitor or by conventional follow up. No equivalent data were identified for BioMonitor 2‑AF or Confirm Rx (or the current Reveal LINQ version). Therefore, the EAG assumed equal efficacy for all devices. A published model (Sterne et al. 2017 and Welton et al. 2017; the 'adapted DOAC' model) was used to model longer-term clinical outcomes for people with atrial fibrillation that is detected (treatment with an anticoagulant) or not detected (treatment with an antiplatelet drug). Outcomes included were ischaemic stroke, myocardial infarction, TIA, systemic embolism, clinically relevant extracranial bleed, intracranial haemorrhage and all-cause mortality.

Costs

3.61 All costs in the model were valued in 2018, in UK pounds sterling. Device costs are shown in table 4.

Table 4 Cost of the implantable cardiac monitors

Device

Unit cost (£ excluding VAT)

BioMonitor 2‑AF

1,030

Confirm Rx

1,600

Reveal LINQ

1,800

3.62 Medtronic also offers an optional triage service for use with Reveal LINQ (FOCUSON) that was included in scenario analyses. There were 2 cost options included: £187 per patient per year or £374 per patient per device. The EAG did not include the cost of reviewing alerts generated by the devices in the base case.

Implantation and device removal costs

3.63 In the base case, the EAG estimated the cost of implanting the devices as £24.17. This was based on advice from the clinical experts about the staff involved (cardiologist and nurse) and time taken for the procedure (10 minutes). The cost of removing the devices was assumed to be £238, based on NHS reference costs schedule 2017/18 (EY13Z – removal of electrocardiography loop recorder, outpatient setting, treatment function code 320). Costs associated with adverse events from implanting the devices were not included in the EAG's analysis.

Comparator arm and follow-up costs

3.64 The EAG based costs for the comparator on the conventional follow-up arm of CRYSTAL‑AF. Costs per cycle in the model were calculated based on the proportion of people having testing every 3 months or no testing in the study. The unit cost of monitoring was £141, based on the NHS reference costs schedule 2017/18 (HRG code EY51Z – ECG monitoring or stress testing [outpatient procedures, service code 320]). The EAG assumed that people with an implantable cardiac monitor will have 1 face-to-face follow up a month after the procedure and then will be remotely monitored. For people in the conventional follow-up arm who do not have atrial fibrillation detected, follow-up appointments are assumed to happen after 1, 3, 6 and 12 months, based on clinical expert advice. If atrial fibrillation is detected, a follow-up appointment is assumed to discuss treatment. The cost of an initial follow up (£163.36) and subsequent follow up (£128.05) were taken from NHS reference costs.

Treatment and clinical event costs

3.65 The costs of DOACs and clopidogrel were taken from the BNF September 2018 to March 2019 edition. Costs of acute and chronic health events were taken from NHS reference costs or Luengo-Fernandez et al. (2013).

Health-related quality of life and QALY decrements

3.66 The EAG did a systematic review to identify relevant utility values to update the adapted DOAC model. There were 2 papers (Berg et al. 2010 and Luengo-Fernandez et al. 2013) with relevant utility values for ischaemic stroke, intracranial haemorrhage, myocardial infarction and TIA events. These were included in the model and were used to update the adapted DOAC model. The utility value used for people with atrial fibrillation in a 'well' health state (that is, when no clinical events such as stroke have occurred) was 0.78 (Berg et al. 2010). The duration of disutility for an acute event was assumed to be 3 months (1 model cycle).

Base-case assumptions

3.67 The following assumptions (in addition to those described in previous sections) were applied in the base-case analysis:

  • The prevalence of atrial fibrillation in this population was equal to the detection rate in CRYSTAL‑AF.

  • Reveal LINQ was as good as Reveal XT (the device used in CRYSTAL‑AF) for detecting atrial fibrillation.

  • BioMonitor 2‑AF and Confirm Rx were equivalent to Reveal XT or Reveal LINQ for detecting atrial fibrillation.

  • The detection of atrial fibrillation was capped at 3 years for BioMonitor 2‑AF even though the manufacturer said the battery life is expected to be 4 years. This was because atrial fibrillation detection data were only available for 3 years of follow up.

  • Atrial fibrillation detection was capped at 2 years for Confirm Rx because this is the expected battery life of the device, and the clinical experts advised that devices are unlikely to be replaced once a battery expires.

  • After 3 years, detection rates of atrial fibrillation are the same in both the implantable cardiac monitors and conventional follow-up arms.

  • Once atrial fibrillation was detected, all patients accepted anticoagulation.

  • DOACs were the only anticoagulation therapies offered (use of warfarin was investigated in a scenario analysis).

Base-case results

3.68 During the first consultation on this guidance, errors were identified in the economic model. NICE commissioned a review of the model by the NICE Decision Support Unit (DSU) who validated the coding and corrected a further minor error. The updated cost-effectiveness results produced by the DSU were provided for the third committee meeting (see table 5 for deterministic results). Probabilistic results (shown in section 3.72) and deterministic results were similar.

Table 5 Base-case deterministic pairwise cost-effectiveness analysis (compared with conventional follow up) – from DSU report (November 2019 committee meeting)

Type of monitoring

Total costs (£)

Total QALYs

Incremental costs (£)

Incremental QALYs

ICER (£)

Conventional follow up

7,600

1.74

Reveal LINQ

9,092

1.88

1,492

0.14

10,342

BioMonitor 2‑AF

8,322

1.88

722

0.14

5,006

Confirm Rx

8,866

1.84

1,267

0.10

12,879

Abbreviations: ICER, incremental cost-effectiveness ratio; QALY, quality-adjusted life year.

3.69 The ICERs in table 5 were produced by separate comparisons of each of the 3 implantable cardiac monitors with conventional follow up. The lower number of QALYs generated by Confirm Rx is because the battery is assumed to last 2 years, rather than 3 years. The EAG noted that if BioMonitor 2‑AF battery life was 4 years, rather than 3 years, as assumed in the model, the device might detect more cases of atrial fibrillation than are captured in the analyses.

3.70 The fully incremental analysis is shown in table 6. The EAG advised that the BioMonitor 2‑AF and Confirm Rx results should be viewed with caution because they are based on a strong assumption of equivalence with Reveal LINQ. The difference in costs between BioMonitor 2‑AF and Reveal LINQ is because of the difference in costs of the devices alone.

Table 6 Base-case deterministic incremental cost-effectiveness analysis – from DSU report (November 2019 committee meeting)

Type of monitoring

Total costs (£)

Total QALYs

Incremental costs (£)

Incremental QALYs

ICER (£)

Conventional follow up

7,600

1.74

Reveal LINQ

9,092

1.88

226

0.00

Dominated

BioMonitor 2‑AF

8,322

1.88

722

0.14

5,006

Confirm Rx

8,866

1.84

544

-0.05

Dominated

Abbreviations: ICER, incremental cost-effectiveness ratio; QALY, quality-adjusted life year. Dominated means that using the device costs more but produced fewer, or the same number of, QALYs than the comparator.

Scenario analyses

3.71 The EAG did some scenario analysis to assess the effect of some of the assumptions made in the model. Selected results are shown in table 7.

Table 7 Selected scenario analyses – from DSU updated model (November 2019 committee meeting)

Scenario

Reveal LINQ ICER

(£ compared with conventional follow up)

BioMonitor 2‑AF ICER

(£ compared with conventional follow up)

Confirm Rx ICER

(£ compared with conventional follow up)

Base case

10,342

5,006

12,879

Addition of FOCUSON triage service provided by Medtronic for Reveal LINQ

Option 1: £187 per patient per year

14,100

NA

NA

Addition of FOCUSON triage service provided by Medtronic for Reveal LINQ

Option 2: one-off fee of £374 per patient per device

12,934

NA

NA

No monitoring in conventional follow-up arm (monitoring costs and cases of AF detected in the conventional follow-up arm removed)

11,617

6,823

14,304

Abbreviations: AF, atrial fibrillation; DOAC, direct oral anticoagulant; ICER, incremental cost-effectiveness ratio; NA, not applicable.

Probabilistic sensitivity analysis

3.72 The DSU provided updated probabilistic sensitivity analysis for the third committee meeting. The ICERs in table 8 were produced by separate comparisons of each of the 3 implantable cardiac monitors with conventional follow up.

Table 8 Probabilistic pairwise cost-effectiveness analysis (compared with conventional follow up)

Type of monitoring

Total costs (£)

Total QALYs

Incremental costs (£)

Incremental QALYs

ICER (£)

Conventional follow up

7,600

1.74

Reveal LINQ

9,093

1.88

1,493

0.14

10,350

BioMonitor 2‑AF

8,323

1.88

723

0.14

5,014

Confirm Rx

8,867

1.84

1,268

0.10

12,888

Abbreviations: ICER, incremental cost-effectiveness ratio; QALY, quality-adjusted life year.

3.73 From the cost-effectiveness acceptability curves (each device was compared independently with conventional follow up), at a maximum acceptable ICER of £20,000 per QALY, all 3 devices had an almost 100% probability of being cost effective.

  • National Institute for Health and Care Excellence (NICE)