Cost–effectiveness of an insertable cardiac monitor to detect atrial fibrillation in patients with cryptogenic stroke

Movement of patients through the model is shown in Figure 1. In each cycle, the cohort has a probability of having AF detected, or developing it without detection. If AF is detected, patients receive a nonvitamin-K oral anticoagulants (NOAC), as recommended by current clinical guidelines [17]. If AF is undetected, patients continue to receive only aspirin.

Secondary stroke and bleeding risks are based on the cohort’s cerebrovascular risk and the efficacy and safety of treatments received. Patients can experience up to two recurrent strokes within the model; first and second recurrent strokes each have an associated level of permanent disability as captured in distinct ‘post-stroke’ health states. AF detection and treatment change can still occur in these states, but the only transition patients can make following a second recurrent stroke is to death.

Patients in the delayed ICM arm started by receiving a 7-day external Holter monitor. After 3 months (upon the beginning of the next model cycle), an ICM is implanted, if no AF has been detected and patients have not experienced a hemorrhagic stroke. If a patient has experienced a hemorrhagic stroke they are assumed to become ineligible for NOACs, so diagnosing AF does not provide the benefit of a change in treatment; thus these patients do not progress to an ICM and transition to SoC instead.

Event probabilities used in the model are shown in Table 1. Values were derived from published studies of baseline risks and calculated via application of drug treatment effects on those risks. Further details of model input parameters are given in Supplementary Figures 1–4 & Supplementary Tables 1–13.

The definition of AF considered in this analysis is the same as used in the CRYSTAL-AF trial (episodes >30 s) [11]. Data from the first 36 months of CRYSTAL-AF were used to model AF detection with an ICM (REVEAL^® XT, Medtronic), a 7-day Holter and SoC. The relationship between ICM and SoC strategies was defined by the hazard ratio for AF detection observed in the trial: 8.8 (95% CI: 3.5–22.2) [11]. Detection probabilities vary between cycles; for the first cycle, they are 8% for ICM, 1% for the Holter in the delayed ICM arm and 0.9% for SoC. Since an ICM may not accurately identify all episodes of AF, we applied a 96.1% sensitivity rate to the device to estimate the true incidence of AF [31]. For patients in the delayed ICM arm who transition to ICM, the rate of detection following transition to ICM is assumed to catch up with the rate of detection in the immediate ICM arm within one cycle.

The risk of ischemic stroke in the model depends on AF status, CHADS₂ score, treatment and age. Gage et al. [18] reported the annual risk of ischemic stroke among patients with diagnosed AF receiving aspirin divided by CHADS₂ score. These data were used as baseline risks, with the effect of AF status, treatment and age applied using a range of sources [19,20,22,29,32]. The risks of a second recurrent stroke (either ischemic or hemorrhagic) and a first recurrent stroke are independent. This is a conservative assumption, made based on clinical expert opinion in the absence of empirical data.

A distribution of the severity for a secondary ischemic stroke was defined based on data from recent cost–effectiveness analyses [23,28] and was independent of treatment. Based on expert opinion, it was assumed that a second recurrent ischemic stroke will be at least as severe as the first recurrent ischemic stroke. The likely severity of a second recurrent stroke was modeled by redistributing the probable severity of a first recurrent stroke.

Bleeding risks were taken from several studies and were treatment- and age-related [24,30,32–34]. The breakdown of different major bleeding events, in terms of type and severity, was taken from two cost–effectiveness analyses [23,28] and was assumed to apply independently of treatment.

In the model, patients experiencing clinically relevant nonmajor bleeds are assumed to remain on treatment without discontinuation. Patients experiencing hemorrhagic stroke while on NOAC, discontinued and switched to aspirin. Hemorrhagic stroke patients on aspirin did not discontinue treatment following major bleeding events but were precluded from ever changing to NOAC, even if they were subsequently diagnosed with AF. The severity of a second recurrent hemorrhagic stroke was modeled using the same method as for ischemic stroke outlined in the previous section.

Among model patients on NOACs, a nonfatal extracranial hemorrhage or other intracranial hemorrhage would prompt 25 and 56%, respectively, to permanently switch to aspirin [35,36]. The remainder of the patients stopped NOAC temporarily, resuming after 6 weeks. The same percentages were precluded from subsequently receiving NOAC therapy if the bleeding event occurred while on aspirin.

Nonfatal extracranial hemorrhage, other intracranial hemorrhage and clinically relevant nonmajor bleeds carry temporary consequences lasting only the cycle in which the bleed occurred. The consequences of ischemic stroke and hemorrhagic strokes were assumed to be permanent, with the impact on costs, HRQoL and long-term mortality dependent on the severity of the stroke (Table 2).

Age-dependent mortality risks in the model were based on US life tables [41] and were adjusted, where applicable, to exclude deaths due to cerebrovascular events which were explicitly modeled [41,42]. Following a nonfatal stroke, an individual’s mortality risk increases depending on the stroke’s severity and their treatment.

Calculating QALYs requires estimating patients’ utility, or HRQoL, over time on a scale between 1 (perfect health) and 0 (equivalent to death). We used evidence from the OXVASC study and Gage et al., to obtain HRQoL values for stroke events and from other published models for disutility values, or utility decrements, associated with bleeding (Table 2) [23,28,38,39]. These were applied to baseline utility estimates taken from the CRYSTAL-AF trial [11]. All utilities were adjusted to account for the age and sex of the population, using published formulae [43]. Further details are available in Supplementary Tables 14 & 15.

The cost–effectiveness of ICM compared with delayed ICM and SoC was estimated with the incremental cost–effectiveness ratio (ICER), which synthesizes QALYs and healthcare costs accrued over the cohort's lifetime. We use the US payers’ willingness-to-pay thresholds of $50,000, 100,000 and 150,000 per added QALY [45,46]. The base–case analysis was based on CRYSTAL-AF patient characteristics [11] and patients diagnosed with AF were switched to a NOAC. We also performed subgroup analyses by CHADS₂ score [11]. A deterministic one-way sensitivity analysis was performed on all model parameters: baseline characteristics, diagnostic accuracy, treatment efficacy and safety, utility and cost values. A scenario was tested in which the battery life of an ICM is increased to 4.5 years, by extrapolating AF detection rates via two methods: assuming a linear increase in AF detection at the end of the trial and assuming 50% of the linear rate. Probabilities of treatment discontinuation for reasons other than bleeding were also considered. A range of different Holter lengths were considered as part of the delayed ICM strategy, as was lower Holter compliance and dropout of patients being lost to follow-up before progressing to an ICM (Table 4). Use of warfarin instead of NOACs was considered, with additional results presented in the Data Supplement. A probabilistic sensitivity analysis was also performed (1,000 samples) and appropriate distributions were fitted to all model variables (Supplementary Table 18).

The deterministic analysis comparing the immediate ICM versus SoC strategies over the patient’s lifetime showed an incremental benefit of 0.198 QALYs and 0.226 life years, a reduction of 60 strokes per 1,000 patients and an incremental cost of US$6,211 (Table 3). These benefits were driven by the 3 years of ICM monitoring. The immediate ICM strategy was associated with higher initial costs for the technology and its use and higher bleed-related costs compared with SoC. It was also associated with lower stroke event costs and lower long-term health related costs. Based on these deterministic results, the comparison of the immediate ICM and SoC strategies showed an estimated ICER of $31,345 per QALY and US$27,438 per life year gained, making the ICM a cost-effective intervention in the US setting.

Compared with delayed ICM, the immediate ICM strategy was associated with 0.001 additional QALYs and US$135 lower costs on average over a patient’s lifetime. This means that an immediate ICM strategy had comparable effectiveness yet was cost-saving compared with delayed ICM. In the delayed ICM strategy, 7-day monitor costs totaled US$233,300 per 1,000 patients but captured only 3% of patients ultimately diagnosed with AF (the remaining 97% were detected via ICM).

A detailed breakdown of the results is available in Supplementary Tables 19–27.

Relative to the base–case analysis, the ICER increased when we considered patients with CHADS₂ scores of 2 and 3 (US$53,492 and 38,415 vs 31,345, respectively). When we considered a higher risk population (CHADS₂ scores 4–6), the ICER decreased to US$19,514 per QALY gained.

We used deterministic sensitivity analysis to test the importance of several input parameters (Figure 2). The most important driver of cost–effectiveness was the efficacy of NOACs versus aspirin, followed by total utility estimates. Increasing the hazard ratio for ischemic strokes on NOACs versus aspirin to its upper confidence interval limit (0.49), increased the ICER to $60,767 (compared with US$20,299 at the lower bound of 0.22). An increase in utilities of 20% led to an ICER of $26,120, compared with US$39,181 for a 20% decrease. Varying the cost of ICM device and insertion cost yielded a range of ICERS from $28,114 to 36,191. The rate of NOAC discontinuation for reasons other than bleeding was varied from 0 to 3.2%, leading to modest changes in the ICER (US$30,249, 32,458). Varying the hazard ratio for ischemic stroke rate with and without AF, also led to modest changes in the ICER ($30,444, 32,584).

We tested a number of relevant scenarios. One with a significant impact was the ICM battery life. Increasing the ICM battery life from 3 to 4.5 years and assuming a linear extrapolation of AF detection after the 3-year trial data, yielded an ICER of US$26,717. If the linear increase is reduced by 50%, the ICER increases to US$29,304, though is still improved from the base case. Another important parameter was the rate of uptake of NOACs following AF detection. Reducing this from 100 to 90% increased the ICER to US$35,508. One scenario tested the situation in which patients received warfarin instead of NOACs. This increased the ICER to US$32,183 because although the acquisition cost of warfarin is lower, it is counterbalanced by additional INR monitoring costs and higher bleeding risks.

The length of external Holter monitoring used as part of the delayed ICM strategy was also varied. For one-time monitors ranging from 24 h to 30 days, the immediate ICM strategy remains cost-saving, with a range of US$75–221 average cost savings per patient. For all quarterly monitoring scenarios tested, immediate ICM extendedly dominates delayed ICM, meaning that immediate ICM is more effective than delayed ICM and has a lower ICER versus SoC.

Table 3 shows results from the probabilistic sensitivity analysis. The probabilities that the ICM strategy would be cost-effective under willingness-to-pay thresholds of US$50,000, 100,000 and 150,000 per QALY were 85.3, 98.1 and 99.0%, respectively. SoC was more effective and less costly than immediate ICM in only 3% of samples. Delayed ICM was never found to be more cost-effective than immediate ICM. A cost–effectiveness acceptability curve is shown in Figure 3.

Model input data were derived from a clinical trial which, although providing the best available evidence for a cryptogenic stroke population, had patient selection criteria and treatment pathways that may not reflect current US clinical practice. However, the ICER for immediate ICM was consistently below US$50,000 per QALY gained even after adjusting for variations expected within US practice, such as stroke risks and treatment costs. If AF detection rates are higher in real-world use than in CRYSTAL-AF, as reported by Ziegler et al. then the cost–effectiveness of ICMs is likely to be even better [57].

The delayed ICM strategy represents a scenario in which clinicians attempt to detect AF without immediately resorting to internal monitor placement and thus is likely to be highly relevant. The detection efficacy of a delayed approach is unknown. In the model, it is conservatively assumed that cumulative detection under delayed ICM catches up with the immediate ICM arm in the same cycle that ICM is inserted. No data are available to inform alternative assumptions, but it is likely that this represents the highest possible detection efficacy that could be expected in a delayed ICM strategy. Further, it is assumed that 100% of eligible patients receive an ICM, whereas in reality a portion of patients might opt out or be lost to follow-up. If fewer patients receive an ICM after short-term monitoring, the delayed ICM strategy becomes less expensive, but also less effective. The costs saved in this scenario are not worth the health benefits forgone and immediate ICM remains the more cost-effective strategy.

CHA₂DS₂-VASc has been introduced as an updated alternative to CHADS₂. Although this score was not directly captured in the CRYSTAL-AF trial, a future model could assess the impact of mapping and modeling the new risk score.

This analysis included a number of subgroups, but future analyses could further explore subgroups defined according to different clinical criteria that might affect their likelihood of having AF or their stroke risk, for example, age and atrial size. There has been discussion about the length of AF that must be detected to consider it clinically meaningful [58]. The CRYSTAL-AF study used a 30-s threshold, although the ICM was automatically limited to detect only AF episodes >2 min, while the SoC arm included episodes of 30 s or more, which is a conservative assumption. Different definitions could be tested within further research. A subgroup analysis of CRYSTAL-AF found 95% of patients in the ICM arm who were diagnosed with AF experienced an episode >6 min [59]. This suggests that results would be robust to assuming a higher AF threshold [59].

In US clinical practice, decisions can vary about whether and when to discontinue NOACs after patients experience different types of major bleeds, including hemorrhagic stroke. In our model, we took a conservative approach in which hemorrhagic stroke leads to immediate and permanent discontinuation of NOACs with no exceptions. A future analysis could investigate the effect of allowing a proportion of patients to resume NOACs after a hemorrhagic stroke or take up a NOAC at the point of AF diagnosis despite a history of hemorrhagic stroke on ICM’s cost–effectiveness.

Finally, use of a payer perspective means that our model does not capture the societal costs of ischemic strokes in a relatively young population, many of whom might be employed. Although the ICER was highest in the younger subgroup of patients with CHADS₂ score of 2, the societal costs may be greater. Future evaluations taking a societal costing perspective might show a reduced ICER in this subgroup and overall.