4 Committee discussion
Technologies that improve paroxysmal atrial fibrillation detection after cryptogenic stroke or TIA could have substantial benefits for people
4.1 The patient expert told the committee how, when someone has a stroke or transient ischaemic attack (TIA) with no identifiable cause, they can live in fear of having another stroke. This is because they know that the cause of the stroke is not being treated. This can make them anxious and want to visit the GP often for reassurance. Paroxysmal atrial fibrillation is often a cause of cryptogenic stroke. But it's often not detected because it's not present when someone has their initial assessment. If atrial fibrillation is detected, the clinical experts highlighted the importance of offering anticoagulants, rather than antiplatelet therapy, to reduce the risk of a further stroke or TIA. The patient expert explained that people who have had a cryptogenic stroke tend to be younger than people who have had a stroke with a known cause. Therefore, they're more likely to be working and have dependants, such as elderly parents or children. They pointed out the benefits of preventing further strokes, including reducing post-stroke dementia and the psychological impact of sudden illness. The clinical experts said that current practice is to monitor for suspected atrial fibrillation for up to about 14 days at most using Holter monitors if implantable cardiac monitors are not available. A patient expert said that at the moment, monitoring often misses atrial fibrillation in people who have had a stroke, who could benefit from treatment. The committee concluded that identifying the cause of a cryptogenic stroke is important to reduce risk of a further stroke or TIA. Technologies that can identify paroxysmal atrial fibrillation missed by current post-stroke follow-up testing could have substantial benefits for people who have had a cryptogenic stroke.
Implantable cardiac monitors can reassure people who have had a cryptogenic stroke or TIA, and their carers
4.2 The patient expert said that, if atrial fibrillation is suspected after a stroke or TIA, people can often be anxious that new symptoms may be related to the condition, and that they should report them to their doctor. A continuous electrocardiogram (ECG) monitor can reassure people that if they have symptoms, the monitor will detect any atrial fibrillation that caused them, and can be used to confirm or rule out the condition. Because the devices can remotely monitor people, they may need fewer follow-up appointments after a cryptogenic stroke. This could particularly benefit people living in remote areas far from a hospital. The patient expert said that after a stroke people are fatigued for a long time. Travelling to follow-up appointments can be tiring, costly and time-consuming. People may also need to go with a carer to help them and describe symptoms. The committee concluded that implantable cardiac monitors could have quality of life benefits beyond preventing another stroke.
4.3 The only study identified by the external assessment group (EAG) that compared the effectiveness of using an implantable cardiac monitor with conventional follow up after a cryptogenic stroke was the CRYSTAL‑AF study. The clinical experts said that it's important that non-invasive ECG monitoring is done before an implantable cardiac monitor is considered. They also said that the length of monitoring in the NHS can vary. Holter monitors are typically used for 24 hours to 7 days. Not everyone in CRYSTAL‑AF had outpatient ECG monitoring before having an implantable cardiac monitor fitted. Those who did were monitored for a median of 23 hours. The committee also considered that participants in CRYSTAL‑AF were younger than would be expected for people who have had a stroke (mean age about 61.5 years). However, the clinical experts explained that people with cryptogenic stroke are usually younger than the overall stroke population. The committee concluded that the population in the CRYSTAL‑AF study broadly represented people with cryptogenic stroke who would have an implantable cardiac monitor fitted in the NHS.
People in the control arm of CRYSTAL-AF may have been tested more for atrial fibrillation than is usual in the NHS
4.4 Some people in the control arm of the CRYSTAL‑AF study, who did not have a cardiac monitor implanted, were tested for atrial fibrillation every 3 months using ECG, including Holter monitoring. The clinical experts said that in current practice, the amount of testing for atrial fibrillation varies if an implantable cardiac monitor is not used, but it is likely to be less than in CRYSTAL‑AF. They also said that people may only be tested again for atrial fibrillation if they have another stroke. The committee concluded that testing for atrial fibrillation in the control arm of CRYSTAL‑AF may be more than is done in the NHS, which may underestimate the increased yield of people with atrial fibrillation reported for the intervention arm.
Reveal XT increases atrial fibrillation detection, but the effect on further stroke or TIA reduction is uncertain
4.5 In the CRYSTAL‑AF study, Reveal XT detected more people with atrial fibrillation than conventional follow up (see table 1). There were also fewer strokes or TIAs in the Reveal XT arm of the study (see table 2). However, because of the length of follow up and sample size, the true effect of the device on reducing stroke or TIA incidence is uncertain; the 95% confidence interval for the hazard ratio at 12 months was 0.22 to 1.80. The committee concluded that there was good evidence that Reveal XT detected more people with atrial fibrillation than conventional follow up, and that this was likely to be seen in clinical practice. However, the extent of a subsequent reduction in stroke or TIA occurrence is uncertain.
CRYSTAL-AF data can be used to assess how well Reveal LINQ detects atrial fibrillation in people who have had a cryptogenic stroke, but not BioMonitor 2‑AF or Confirm Rx
4.6 The CRYSTAL‑AF study used Reveal XT, a predecessor model of Reveal LINQ. Changes have been made to the atrial fibrillation detection algorithm that is now used in Reveal LINQ. There was some evidence that suggested this had improved its ability to detect atrial fibrillation. The clinical experts said that the atrial fibrillation detection algorithms in other manufacturers' devices may use the same features of an ECG to detect potential atrial fibrillation. But how these features are used to determine if atrial fibrillation is likely to be present, or to classify an arrythmia as atrial fibrillation or another type of arrythmia, is likely to differ between devices. At consultation, the manufacturer of the BioMonitor submitted an unpublished technical validation report that compared the ability of the BioMonitor 2‑AF and Reveal LINQ to detect atrial fibrillation from a Holter monitor recording (see section 3.21). The EAG commented that this was not a clinical comparison of the devices, which might perform differently when implanted. The study was also not done in a cryptogenic stroke population, where the device may perform differently because of different patient characteristics. The committee noted that the study had a small population size and had not been published and so was not peer reviewed. Clinical experts commented that electrode positioning is different for Holter monitors and implantable cardiac monitors. So the ECG output from a Holter monitor is not the same as the signal that an implantable cardiac monitor receives. The results could therefore be considered to be artificial and not reflect clinical reality. The committee considered that this study did not show that the Reveal LINQ and BioMonitor devices were comparable in detecting atrial fibrillation in a cryptogenic stroke population. The committee concluded that it is feasible that data from Reveal XT are likely to apply to the updated version from the same manufacturer, Reveal LINQ. But there is too much uncertainty over whether the data can be used to show the performance of the BioMonitor 2‑AF or Confirm Rx to detect atrial fibrillation in people who have had a cryptogenic stroke. Therefore, the committee did not accept that evidence from the CRYSTAL‑AF study could be applied to these devices.
It is not appropriate to use data from CRYSTAL-AF to model the performance of BioMonitor 2-AF or Confirm Rx
4.7 The EAG used diagnostic yield data (a measure of how many people with atrial fibrillation were diagnosed) from the CRYSTAL‑AF study in the economic model for all 3 devices. The committee considered data from this study to be appropriate to assess how well Reveal LINQ detected atrial fibrillation in people who have had a cryptogenic stroke. But it did not think it was appropriate to use it for BioMonitor 2‑AF or Confirm Rx (see section 4.6). In the absence of clinical or comparative data for these devices in people who have had a cryptogenic stroke, the committee concluded that it was not appropriate to consider the cost-effectiveness estimates for BioMonitor 2‑AF or Confirm Rx.
Not including adverse events caused by implanting Reveal LINQ is unlikely to have a large impact on cost-effectiveness estimates
4.9 The EAG's economic model did not include the effect of any adverse events caused by implanting the devices. The EAG explained that the proportions of people who had non-serious adverse events was reported in CRYSTAL‑AF, but there were no details on what these events were. Therefore, the EAG could not include any costs or disutilities caused by these events in the model. The clinical and patient experts said that there are some minor issues caused by the devices, such as irritation and pain when they are fitted or removed, or if someone unintentionally exposes the device through the skin, but that these are not common and do not have severe consequences. The committee concluded that not including any adverse events caused by implanting the devices was unlikely to have had a large effect on cost-effectiveness estimates.
There may be uncaptured benefit in detecting non-atrial fibrillation arrythmia, but the impact of this on patient outcomes is uncertain
4.10 The model did not include detection of non-atrial fibrillation arrythmias. There was not much evidence on the number of these arrhythmias detected by the devices, and what evidence there was came from non-comparative observational studies. The clinical experts explained that most asymptomatic non-atrial fibrillation arrythmia detected by the implantable cardiac monitors would not lead to any change in care. The committee concluded that using the devices may increase detection of non-atrial fibrillation arrythmias, but that the extent of this increase, and the clinical significance of these arrythmias and consequent impact on patient outcomes, is highly uncertain.
The base case may overestimate how much monitoring for atrial fibrillation is done in current practice, which lowers the ICER
4.11 The EAG used the control arm of the CRYSTAL‑AF study to model current practice ('conventional follow up'). The committee had earlier concluded that this may include more monitoring for suspected atrial fibrillation than would be done in the NHS (see section 4.4). This may make the increased diagnostic yield of atrial fibrillation for Reveal LINQ in the model a conservative estimate. But it may also mean that the model overestimates the cost of monitoring for atrial fibrillation in current practice. The EAG did a scenario analysis in which no further monitoring for atrial fibrillation was done in current practice (that is, no people with atrial fibrillation were detected or cost of monitoring included). This increased the incremental cost-effectiveness ratio (ICER) for Reveal LINQ by about £1,300 per quality-adjusted life year (QALY) gained. The committee concluded that the EAG's scenario may be too extreme, in that some monitoring is likely to be done in the NHS for people with no implantable cardiac monitor fitted. However, the amount of assessment for atrial fibrillation in current practice is likely to have been overestimated in the base-case model, which lowered the base-case ICER by up to about £1,300 per QALY gained.
The number of false positive alerts from Reveal LINQ is uncertain, and including this in the economic model increases the base-case ICER
4.12 The base-case model did not include the cost of interpreting alerts produced by Reveal LINQ. The EAG explained that this was because of a lack of data on the number of alerts produced by the device. The clinical experts said that the device would produce false positive alerts. Anecdotal evidence differed on the impact of false positive alerts on workload. One clinical expert said that alerts from the devices can generate several hours of work per day for electrophysiologists to review, although this was based largely on alerts from people with syncope. However, another clinical expert said that it takes minimal time to review alerts generated for possible atrial fibrillation (less than 10 seconds), and that the increase in workload for technicians would be minimal. The clinical experts highlighted that the number of alerts can vary widely between people and noted that cardiac physiologists need to triage the alerts. The EAG did 2 scenario analyses that included the costs of an optional triage service for alerts offered by Medtronic for Reveal LINQ. This increased the ICER by about £2,600 to £3,800 per QALY gained, depending on the cost option used. The clinical experts said that the costs used (£187 per patient per year or £374 per patient) are likely to be a realistic estimate and could be considered a reasonable proxy for the costs of triaging alerts in the NHS. The committee concluded that there is uncertainty about the likely number of false alerts that Reveal LINQ generates in people who have had a cryptogenic stroke if used in routine clinical practice, and the impact on services. Including costs for reviewing alerts in the economic model would increase the ICER for Reveal LINQ, although it is uncertain by how much.
4.13 At the first committee meeting, the committee was concerned by the difference between the deterministic and probabilistic base-case results provided by the EAG. There was a large difference in the incremental costs and QALYs generated for the devices (compared with conventional follow up) between these analyses. For the second committee meeting, the EAG provided updated analyses in which an error in the model code used to run the probabilistic sensitivity analysis was corrected. The updated base-case probabilistic sensitivity analysis results were now very similar to the deterministic results. At the first committee meeting, the committee was concerned that the model results may not be realistic (lacking face validity) because of the small number of total QALYs generated in the model. At the second committee meeting, the EAG explained that this was because QALYs in the model were only generated by people who had episodes of atrial fibrillation, which was 30% of the total population. No QALYs were considered for the remaining 70% because there would be no difference in the number of QALYs generated between people with implantable cardiac monitors fitted and conventional follow up (and therefore no impact on ICERs). The EAG further explained that the cohort modelled had a starting age of 62, had all had a stroke or TIA, and all had atrial fibrillation. Therefore, they did not consider that the number of QALYs generated was unrealistic. During the first consultation on this guidance, an error was identified in the model. On reviewing the model, the EAG identified another error. NICE commissioned a review of the model by NICE's Decision Support Unit (DSU). The DSU checked the model and corrected another small error. Updated model results were presented at the third committee meeting. The committee concluded that, considering the DSU's review of the model, the corrections made to the model and the explanations provided, the revised model was suitable for decision making.
The different parameters used in the EAG's and Diamantopoulos et al. models are unlikely to affect decision making
4.14 The updated cost-effectiveness estimate for the Reveal LINQ provided for the third meeting was now lower than the results of a previous economic model that also used data from CRYSTAL‑AF (Diamantopoulos et al. 2016; see sections 3.53 and 3.54). The EAG's updated base-case ICER was £10,342 compared with £17,175 per QALY gained for Diamantopoulos et al. and the incremental QALYs were similar (0.14 and 0.15). The EAG explained that the difference between the results of the models was driven by differences in how the impact of anticoagulant treatment was modelled. Because of differences in model structure, the outcomes included, and the mechanisms used to estimate outcomes, the EAG considered a direct comparison of the parameters used in each model to be difficult and potentially not very informative. The DSU's amended version of the EAG's model estimated that the number of strokes that would be avoided by using an implantable cardiac monitor was 52 per 1,000 people with cryptogenic stroke, compared with 40 per 1,000 people estimated by the Diamantopoulos et al. model. The committee recalled that the size of any reduction in further stroke or TIA caused by using the devices was uncertain (see section 4.5). The committee concluded that there was uncertainty about which was the most appropriate approach to modelling the impact of anticoagulant treatment. The acute and post-stroke utilities were lower in the Diamantopoulos et al. model, which would also have contributed to the difference in incremental QALYs between models. At consultation, several stakeholders commented that the underlying model used (reported in Sterne et al.) was developed for a primary stroke population. The EAG explained that they had adjusted parameters in the model (for example, risk of further stroke, TIA, systemic embolism, intracranial haemorrhage and bleeds) for a secondary stroke population. Stakeholders also highlighted that the impact of a secondary, rather than primary, stroke may have been underestimated in the model. For example, the costs of ongoing treatment and impact on someone's health-related quality of life. The committee concluded that there is uncertainty about the most appropriate parameters to use to model the longer-term effects of anticoagulant and antiplatelet treatment in this population, but that the different parameters used in the EAG's and Diamantopoulos et al. models are unlikely to affect decision making.
4.15 The committee only considered cost-effectiveness estimates for Reveal LINQ (see section 4.7). The probabilistic ICER for Reveal LINQ in the EAG's model was almost identical to the deterministic value. The deterministic base case for Reveal LINQ compared with conventional follow up was £10,342 per QALY gained. If assessment for atrial fibrillation in the conventional follow-up arm is removed from the EAG's base-case model, the ICER increases by about £1,300 per QALY gained. However, the assumption that no longer-term monitoring for atrial fibrillation is done in standard monitoring is unlikely (see section 4.11). In addition, costs of reviewing alerts produced by Reveal LINQ were not included in the base-case model. If the cost of a triage service was included, the EAG's base-case ICER increased by about £2,600 to £3,800 per QALY gained (see section 4.12). The committee concluded that there was uncertainty about the most plausible ICER for Reveal LINQ. Including its preferences in the EAG's model would increase the base-case ICER, but this was unlikely to increase to over £20,000 per QALY gained. The committee concluded that the most plausible ICER for Reveal LINQ is likely to be less than £20,000 per QALY gained.
4.16 The committee agreed that Reveal LINQ was likely to be clinically effective because it identifies more people who have atrial fibrillation after a cryptogenic stroke or TIA than current practice. The committee recalled its conclusion that technologies that improve the detection of paroxysmal atrial fibrillation after cryptogenic stroke or TIA could have substantial benefits for patients. In addition, there is an unmet need for longer-term monitoring for atrial fibrillation after a cryptogenic stroke or TIA (see sections 4.1 and 4.2). The committee considered that the most plausible ICER for Reveal LINQ is likely to be less than £20,000 per QALY gained (see section 4.15). Therefore, the committee concluded that Reveal LINQ is likely to be a cost-effective use of NHS resources.
4.17 The inclusion criteria for CRYSTAL‑AF included a requirement for a 12‑lead ECG and 24‑hour ECG monitoring for atrial fibrillation detection to establish the diagnosis of cryptogenic stroke (before use of an implantable cardiac monitor). The amount of atrial fibrillation detected in this study population was used for the cost-effectiveness estimates for Reveal LINQ in this assessment. During consultation, stakeholders highlighted that longer duration non-invasive monitors (that is, monitors that are not implanted) are increasingly available and questioned if this would impact the cost effectiveness of Reveal LINQ. The committee recalled that clinical experts had emphasised that Reveal LINQ would only be used after all available non-invasive monitoring had been done. Therefore, these non-invasive monitors were not comparators to implantable cardiac devices. However, longer duration non-invasive monitoring is likely to detect some cases of atrial fibrillation that shorter duration non-invasive monitoring would miss, and therefore there may be a lower yield of people with atrial fibrillation subsequently detected by implantable cardiac monitors. The EAG commented that longer duration non-invasive monitoring of up to a month was unlikely to have a large impact on the cost effectiveness of Reveal LINQ. They based their comment on exploratory model analysis that assumed that anyone with atrial fibrillation in the first month of CRYSTAL‑AF would not have had an implantable cardiac monitor (reducing the diagnostic yield for Reveal LINQ in the model). Clinical experts highlighted that it is important that non-invasive ECG monitoring is done first before Reveal LINQ is considered, and that the type and duration of non-invasive monitoring will vary by local availability across the NHS. The committee concluded that it's important that Reveal LINQ is only used if paroxysmal atrial fibrillation is still suspected after non-invasive ECG monitoring has been done.
Further evidence is needed to show the effectiveness of BioMonitor 2-AF and Confirm Rx to detect atrial fibrillation in people with cryptogenic stroke
4.19 The committee considered that there was no evidence plausibly showing that BioMonitor 2-AF and Confirm Rx (or previous versions) were as effective as Reveal devices at detecting atrial fibrillation in people with cryptogenic stroke. And the committee noted that it was difficult to get good comparative data on this. Only diagnostic yield data from a Reveal device were available to model cost effectiveness. A randomised controlled trial comparing Reveal LINQ with Confirm Rx was highlighted during consultation (Yokokawa et al. 2019; see section 3.51). Most of the people in the trial had had a cryptogenic stroke. But the study was only available as a conference abstract. Details of the methodology were limited, and it was not clear why the number of events detected in the Reveal LINQ and Confirm Rx arms were so different. Abbott Medical UK, which makes Confirm Rx, said it was not involved in the study and could not give any more information. The committee considered that it did not have enough information to be able to use the study to assess if Reveal LINQ and Confirm Rx had similar effectiveness. However, it concluded that the study did show that it was feasible to do a trial comparing the effectiveness of different implantable cardiac monitors to detect atrial fibrillation in a cryptogenic stroke population.