In order to predict the lifetime benefits and costs of different rhythm control treatment sequences, we implemented a variety of clinical input parameters in an novel economic evaluation framework that associates societal costs and health-related quality of life to the clinical pathway that is simulated by the model. The model was developed by health economists (SH, MV) and AF experts who were delegated from the Dutch Society of Cardiology (MH, DL, MR, SY). The methods and reporting of this study conform to Consolidated Health Economic Evaluation Reporting Standards (CHEERS, Electronic Supplementary Material) [13].

2.1 Patient Population

The model included Dutch patients with symptomatic AF who were referred to the hospital and were eligible for rhythm control. This means we do not include all patients with symptomatic AF, but only those for whom rhythm control is indicated. These typically include younger patients, those experiencing their first AF episode or with a short history of AF, patients with tachycardia-mediated cardiomyopathy, normal-to-moderate increased left atrial dilatation, minimal comorbidities or heart disease, those with inadequate rate control despite optimal medical therapy, and those with a preference for rhythm control over rate control [3]. The patient and intervention characteristics of our model population were based on the Dutch Heart Registration data that includes over 30,000 patients who had a CA between 2013 and 2020 [14]. The mean age was 61.4 years, 33% of patients were female, 71% had paroxysmal AF, and the most common ablation methods were conventional radiofrequency (50.6%), cryoballoon (36.8%), and phased radiofrequency (12.3%) [14].

2.2 Clinical Input Parameters2.2.1 Treatment Success and Recurrence of AF Symptoms

The relative effectiveness of AAD versus CA was based on a systematic literature review and meta-analysis of randomized controlled trials (RCTs) that compared class I/III AAD to CA (Table 1) [4]. No difference was assumed between AADs or the different CA methods [15]. The probabilities of freedom of AF symptoms after 6 months (i.e., one model cycle) were based on a meta-analysis [4] and healthcare insurance claims data (Tables 2–4 of the ESM). The pooled average patient characteristics of the included RCTs in the meta-analysis were similar to those of the patient population simulated in our model (Table 1 of the ESM).

Table 1 Relative risks of AF recurrence with AADs versus CAs derived from the meta-analysis of Cochrane Netherlands [4]

If freedom of AF symptoms is achieved, patients were exposed to the long-term risk of AF recurrence. The probability and time it takes for AF symptoms to recur following a CA were based on a sample of patients in the Dutch health insurers claims data including all patients who underwent a first CA for AF in the Netherlands in 2017–23 (n = 24,286). Based on the recorded AAD prescriptions in the health insurance claims database before the date of the CA, 85% received one or more class I and/or class III AADs (amiodaron, sotalol, propafenone, and flecainide) before the CA. The mean follow-up after the first CA was 2.6 years (standard deviation 2.0 years, range 0–7 years, median 2.1 years).

Survival methods were applied to estimate unique time to event patterns following a first, second, or third or higher CA. An event was marked as a recurrence of AF symptoms when the following activity was registered after a blanking period of 90 days in the health insurance claims database: re-ablation, cardioversion, His-bundle ablation, endoscopic MAZE-procedure, and a new prescription of amiodarone (as it is unlikely that amiodarone would be prescribed for any other reason than recurrence of AF symptoms in this patient population). A 90-day blanking period was used to exclude early post-procedural events unrelated to true AF recurrence. Patients were censored if they did not experience an event or died within the follow-up period, and Kaplan–Meier curves were fitted to the data. The Kaplan–Meier curves of each individual CA (i.e., first to fifth) showed that the Kaplan–Meier curves of the third, fourth, and fifth CA overlap, which indicates that the time to event is not significantly different after the third, fourth, or fifth CA (Fig. 1 of the ESM). Therefore, we estimated one Kaplan–Meier curve for the third, fourth, and fifth CA together (Fig. 1). The 5-year probability of freedom from recurrence of AF symptoms after the first, second, or third and higher CA was 61.0%, 51.5%, and 45.3%, respectively (Table 5 of the ESM). Table 2 shows the number of patients with recurrence of AF symptoms after the first, second, or third or higher CA.

Fig. 1

Kaplan–Meier curve of time to atrial fibrillation (AF) recurrence by first, second, and third or higher catheter ablation

Table 2 Distribution of events indicating recurrence of atrial fibrillation symptoms

To forecast long-term recurrence rates beyond the observed period, we fitted several parametric distributions to the Kaplan–Meier curves (excluding the blanking period). The best fitting distribution for the first and second ablation was the Royston-Parmar splines and for the third or more CA log-normal chosen based on the lowest Akaike information criterion and Bayesian information criterion values (Tables 6–11 and Figs. 2–4 of the ESM) and the predicted long-term survival probabilities were deemed plausible by the involved cardiologists.

The probabilities of AF recurrence after CA that are used in the model were derived from the parametric survival functions by converting the ‘survival’ probabilities to per period transition probabilities. To estimate the long-term recurrence risk of AADs, we multiply the transition probabilities derived from the parametric survival functions of CA with the relative risk (RR) of AAD versus CA derived from the meta-analysis (ESM). In other words, the meta-analysis of RCTs showed how much CA performs better than AAD and we know from health insurers claims data how well CA performs in the ‘real world’. We combine these data sources to predict how well AADs would perform in the ‘real world’ by multiplying the transition probabilities of symptoms of AF recurrence after CA with the RR of AAD versus CA.

2.2.2 Complications and Side Effects

Side effects of AADs were only included when they resulted in discontinuation of AADs. It was assumed that other side effects that did not result in discontinuation of AADs were not associated with a relevant impact on quality of life or costs and could therefore be ignored. The meta-analysis resulted in a probability of discontinuation of AADs because of side effects of 12.1% (95% confidence interval [CI] 6.7–20.6) in AAD-naïve patients and 22.5% (95% CI 16.1–30.5) in AAD-exposed patients [4]. If patients discontinued AADs, they moved to the next treatment line.

The Dutch Heart Registration reported on five ablation-related complications. Bleeding and thrombo-embolic complications were not considered as the meta-analysis showed that there were no significant differences in these risks between the treatment arms (AAD vs CA) [4]. For the other three complications, the risks were based on the occurrence rate in the most recent available year (2023) in the Dutch Heart Registration; 0.42% for cardiac tamponade, 0.40% for phrenic nerve paralysis, and 0.65% for vascular complications.

2.2.3 Mortality

The probability of death in the model was based on the age- and sex-specific mortality rates in the Dutch general population in 2019 (to prevent impact of the coronavirus disease 2019 pandemic on mortality rates) increased with the excess mortality hazard ratio of 2.01 for patients with AF based on data from the Framingham Heart Study [16]. Early mortality after CA is not considered in the model because data from the Netherlands Heart Registration showed that the 30-day mortality after CA was 0.05%, which was lower than the crude death rate for persons with a similar age [14].

2.3 Model Structure

The clinical data described above is implemented in an individual state transition model to forecast the time spent with and without symptoms under different rhythm control strategies (Fig. 2). The simulation starts with all patients in the health state ‘symptoms of AF’. The rhythm control treatment can be either AAD or CA. When the treatment is successful, patients enter the ‘symptom-free AF’ health state. After this initial treatment success, patients have a time-dependent risk of recurrence of symptoms of AF. When patients have recurrent AF symptoms, they return to the symptoms of AF health state and receive their next rhythm control treatment. Time spent in the symptom of AF health state is associated with a loss in quality of life, an increase in medical resource consumption and informal care, and a decrease in productivity relative to time spent in the symptom-free AF health state. From both the symptoms of AF and symptom-free AF health state, patients have a probability to die, which is based on background mortality of the general population adjusted for the excess mortality risk in patients with AF. Patients who are taking AADs have a probability to discontinue treatment because of side effects. Patients who underwent a CA have a probability of complications. The cycle length of the model is 6 months and the model has a lifetime horizon. Within-cycle correction was performed with the Simpson 1/3 rule [17, 18]. Different sequences with AADs and/or CA were modeled with a maximum of six treatment lines, resulting in 64 treatment sequences. If patients would still have symptoms of AF after these six treatment lines, it is assumed that AF is accepted and patients will remain in the ‘AF symptoms’ health state.

Fig. 2

Conceptual model. AF atrial fibrillation

The model was programmed in R 4.1.1 using RStudio 2024.12.1 (RStudio). The code of the disease model was adapted from the individual-based state-transition framework developed by the Decision Analysis in R for Technologies in Health workgroup [19, 20]. The code is publicly available at https://github.com/simonehuygens/AF-rhythmcontrol.git.

2.4 Costs

The healthcare costs were based on the health insurers claims data of patients with and without symptoms of AF in 2021 (Fig. 9 of the ESM). The mean additional costs of having AF symptoms were €979 per 6 months, mainly driven by cardiology-related care. The costs for 6 months of AAD use (prescriptions of flecainide, sotalol, propafenone, or amiodarone) were €60. The costs of the Diagnosis Related Group for the CA and the additional costs of the CA (6 months before and 12 months after) that were not included in the DBC costs were €14,836 in total. We estimated age- and sex-specific future medical costs using the Practical Application to Include Disease costs (PAID) version 3.0 [21]. The model also includes informal care and productivity costs. The input values and sources can be found in the Table 21 of the ESM. All costs were adjusted for inflation to 2024.

2.5 Health-Related Quality of Life

Health-related quality of life in terms of ‘utilities’ is required for the calculation of quality-adjusted life-years (QALYs). Utilities for AF symptoms were derived from a secondary data analysis using a generalized linear model with the Dutch EQ-5D-5L tariff applied to patient-level data from the AVATAR-AF trial, with AF symptoms, treatment type, and baseline utility as covariates (Tables 18–19 of the ESM) [22]. The model estimated a utility decrement of 0.068 associated with AF symptoms (standard error = 0.023, p = 0.004). The treatment variable had no direct impact on utility, meaning that the benefit is derived through the ability of the treatment to reduce symptoms, not from the treatment itself. All patients in the AVATAR-AF trial had paroxysmal AF; however, other studies suggested no differential impact of symptoms between different types of AF [23,24,25]. The utility of patients in the model without AF symptoms is based on the average utility in the general population (Table 20 of the ESM) [26]. In the absence of evidence from the literature, we assumed a disutility of 0.1 for a duration of 1 month for patients with cardiac tamponade, phrenicus paralysis, or vascular complications after CA, in line with the assumptions in other cost-effectiveness studies [11, 27].

2.6 Analyses

In accordance with Dutch economic guidelines, the analyses were performed from a societal perspective, and effects and costs were discounted with 1.5% and 3%, respectively [28]. In the Netherlands, the cost-effectiveness threshold is based on disease severity using the ‘proportional shortfall’. The applicable cost-per-QALY threshold was €20,000 per QALY (ESM).

We conducted a deterministic analysis, a one-way sensitivity analysis (ESM), and probabilistic sensitivity analyses, and calculated expected value of perfect information (ESM). For the deterministic analysis and one-way sensitivity analysis, a cohort of 50,000 simulated patients provided sufficient model stability (Fig. 10 of the ESM). For the probabilistic sensitivity analyses, we simultaneously varied all parameters across 1000 model runs with 5000 patients each. Table 21 of the ESM provides a complete overview of all parameters with their deterministic values and sampling distributions. Atrial fibrillation recurrence rates were modeled separately for the first, second, and third or higher CA and therefore they were also varied independently in the probabilistic sensitivity analyses.

We reported proportion of patients and average duration per treatment line for all treatment sequences based on the deterministic analysis. We ranked all treatment sequences with at least one CA on incremental net health benefit compared to using only AADs and reported lifetime societal costs and QALYs based on the means in the probabilistic sensitivity analyses. The incremental net health benefit can be calculated as follows: difference in QALYs − (difference in costs/cost-per-QALY threshold). If the incremental net health benefit is above zero, the treatment sequence is cost effective compared with the reference sequences and the higher the incremental net health benefit, the more cost effective the treatment sequence is. The fully incremental analysis was performed on the mean results of the probabilistic sensitivity analyses.

We performed five scenario analyses. In scenario 1, we only counted two cardioversions within 1 year as a proxy for recurrence of AF symptoms in the health insurance claims data, instead of every single cardioversion (Tables 12–16 and Figs. 5–7 of the ESM). In scenario 2, all patients were assumed to be exposed to AAD (rather than treatment naïve) by using AAD-exposed RR and treatment success proportion in line 1. Scenario 3 assumed all patients had paroxysmal AF by excluding persistent AF data. Scenario 4 assumed all patients had persistent AF, using available persistent AF parameters and approximating unavailable parameters via relative proportions of known paroxysmal versus persistent AF data (e.g., the AAD-exposed RR proportion 1.62/2.32 applied to the patients with naïve paroxysmal AF (RR = 1.76), estimated naïve persistent AF RR as 1.23). In scenario 5, AF disutility was varied from 0.01 to 0.2, in 0.01 increments (based on values from other cost-effectiveness studies, outlined in Table 17 of the ESM) and compared incremental QALYs and NHB between two treatment sequences CA-AAD-AAD-CA-CA-CA versus AAD-AAD-AAD-CA-CA-CA.

In the base case, competing risks from death were not accounted for, though death precludes future AF recurrences. A scenario analysis applied a multi-state model incorporating competing risks; 430 patients died and recurrence rates were lower than the base case (Fig. 8 of the ESM). However, transition probabilities to death showed counter-intuitive patterns (higher after the second CA, lower after later CAs), suggesting declining death rates with CA exposure. As these transitions were inappropriate for the model, we instead combined the multi-state recurrence rates with the base-case mortality rates (background mortality plus excess mortality hazard ratio).

2.7 Validation

Model validation was conducted and is reported using the ‘Assessment of the Validation Status of Health-Economic decision models’ (AdViSHE) checklist [29]. An external scientific group inspected the technical implementation of the model code using the ‘TECHnical VERification’ (TECH-VER) checklist [30]. For external validation, we compared the results of the analyses of the health insurers claims data with the information in the Dutch Heart Registration.