Basic probabilities of patient’s progression between NYHA classes for the first cycle, second cycle, and onwards were the same for both arms (however, they were adjusted for the BAT arm, using the approach described below) and were obtained from the CARE-HF trial [20]. The probability of hospitalization and excess mortality (on top of normal mortality) by NYHA class was obtained from the literature [19] (Table 1). No excess risk of death was assumed for NYHA class I. Mortality in the overall population was obtained from German sex-specific life tables.

The effectiveness of BAT was modeled via application of relative risks for all-cause mortality and hospitalizations compared with OMT. These relative risks were assumed to be independent of NYHA class. The relative risk for mortality was obtained from the application of anonymized individual patient-level data from the BAT Randomized Controlled Trial [17] to the Meta-analysis Global Group in Chronic Heart Failure (MAGGIC) risk prediction model [21]. The MAGGIC model provides a comprehensive opportunity to develop a prognostic model in patients with HF, with reduced and preserved LVEF. This model uses readily available risk factors based on 39,372 patients with HF from 30 cohort studies, six of which were randomized controlled clinical trials (24,041 patients) and 24 were registries (15,331 patients). A final model included 13 highly significant independent predictors of mortality in the following order of predictive strength: age, lower ejection fraction, NYHA class, serum creatinine, diabetes, no prescription of a beta-blocker, lower systolic blood pressure, lower body mass, time since diagnosis, current smoker, chronic obstructive pulmonary disease, male sex, and no prescription of an angiotensin-converting enzyme inhibitor or angiotensin-receptor blocker. Although these independent predictors of mortality in HF have been previously identified, the MAGGIC model and risk score are the most comprehensive and generalizable items available in the literature. Analysis of 1-year risk of mortality was performed in 82 patients (who had available LVEF and NYHA class information at 6-month follow-up), 52 patients were in the device arm, and 30 patients in the control arm. For each patient, an individual total integer risk score was calculated. From this score, a predicted probability of death was estimated. Three variables were not available in the BAT Randomized Controlled Trial (smoking status, the presence of chronic obstructive pulmonary disease, and first diagnosis of HF in the previous 18 months) and they were obtained from the original cohort in the MAGGIC meta-analysis [21]. For these variables, a list of values was developed following appropriate distribution and values from the list were randomly assigned to each patient in the cohort.

The relative risk of mortality at 1 year was 0.61. The relative risk for hospitalization (relative risk = 0.40) was obtained from a study by Zile et al. [18] (16% were hospitalized in the BAT arm and 40% in control arm). Relative risks were assumed to be constant and were applied over a lifetime horizon.

All cost inputs were based on German data (Table 2). The cost of the BAT implant procedure (G-DRG 901D, German Diagnosis Related Groups) and the battery replacement (G-DRG F17B) procedures were derived from the German DRG in 2013. To estimate the cost of the procedure (except for the cost of the device), the length of hospital stay was assumed to be 2 days, and the cost of the implant was subtracted from the overall DRG. The remaining value was summed up with the cost of the full Barostim neo™ system or the cost of the battery for a replacement procedure.

The cost of short-term complications was assumed as the difference between 2 and 5 days of hospital stay, where it was assumed that complications will lead to a prolonged hospital stay. The cost of pre- and post-implant physicians’ visits were derived from the out-patient statutory health insurance reimbursement catalog (Einheitlicher Bewertungsmaβstab, EBM) using codes 13,542 and 07212. We assumed that one pre-implant and two post-implant visits will be required. We also assumed that long-term complications (pain near stimulator) do not lead to consumption of healthcare resources and do not have any cost.

The cost of routine management of HF (including costs of visits to physician, medication, and rehabilitation, but excluding costs of hospitalization owing to worsening of HF) according to different NYHA classes was obtained from the literature [22]. The costs of hospitalization with intensive care units, coronary care units, coronary artery bypass grafting, percutaneous transluminal coronary angioplasty, transplantation, and no interventional procedures performed were obtained from the German DRGs. The distribution of reasons for hospitalizations was obtained from the literature [20]. All cost data are presented in 2013 Euros. Inflation adjustment was performed using the German consumer price index [23].

Utility scores were assigned for each NYHA class irrespective of received treatment and hospitalization status. Utility values and corresponding distribution parameters were obtained from the CARE-HF trial, in which they were measured using the EQ-5D [19, 20]. The lower and upper bounds of the range for the one-way deterministic sensitivity analysis were obtained by reducing and increasing the baseline values of the utilities by 20%. Utility weights were combined with life-years to estimate QALYs.

The target population included patients who were not eligible for CRT according to established guidelines. The patients had a mean age of 63 years, 83.2% were male, and had chronic heart failure in NYHA class III.

The incremental cost-effectiveness ratio (ICER) was calculated by comparing the difference in average total costs with the difference in average quality-adjusted life years (QALYs) between two patient cohorts. The intervention was considered cost-effective if the ICER was below €35,000 per QALY. All costs and outcomes beyond the first year were discounted 3.0% annually based on the recommendations of the German National Institute for Quality and Efficiency in Health Care (IQWiG) [24]. The model was constructed using Microsoft Excel 2013 (Microsoft Corp., Redmond, WA, USA).

In the base-case analysis results were obtained from probabilistic sampling of the model with 5000 simulations. Unlike in probabilistic sensitivity analysis, only patient’s gender and age were sampled, while keeping all other inputs (costs, transition probabilities, utilities) fixed. Mean values and 95% credible intervals were reported for costs, QALY and LYG. A normal distribution was assigned to patient age, whereas patient gender was assumed to follow a beta distribution.

One-way sensitivity analysis was performed to assess the effect of varying model parameters while holding other variables fixed at base-case values. Cost drivers (variables with major effect on ICER) were identified and results were presented using a Tornado diagram. In addition, a probabilistic sensitivity analysis was performed using Monte Carlo simulations. Five thousand simulations were performed. Beta distribution was used for the probabilities and utilities. Gamma distribution was used for cost data. Log-normal distribution was used for the relative risks. Dirichlet distribution was used for the probabilities of transition between NYHA classes.

The performance of the model was tested against data from within-trial, all-cause mortality outcomes in the medical treatment group of the CARE-HF trial [25]. To perform validation, the baseline patients’ characteristics in the model were adjusted to mean (standard deviation) age (66 ± 6.6 years) and the proportion of male individuals (73%) of patients in the CARE-HF trial, and 5000 Monte Carlo simulations were performed.