Modeling Framework
To analyze adjuvant test-treatment strategies for early breast cancer, we applied a decision-analytic computer simulation model [17] previously developed within our research center ONCOTYROL – Center for Personalized Cancer Medicine [18] (hereafter the “Oncotyrol breast cancer model”). The model validation and first application were recently published elsewhere [19, 20]. In this new model application, a hypothetical cohort of 50-year-old women diagnosed with ER and/or PR positive, HER-2/neu negative, lymph node negative breast cancer was simulated. We adopted a health care system perspective and lifetime horizon for this analysis. Outcomes of interests included survival (number of life years; LY), quality of life (number of quality-adjusted-life years; QALY), total costs (EUR) and incremental cost-effectiveness ratios (EUR/QALY). Costs and effects were discounted by 5% per year [21]. According to the ISPOR-SMDM guidelines [22], the model was implemented using a discrete event simulation approach (ARENA Version 13.90.00000, Rockwell Automation). This approach allows for individual patient pathways to be determined by multiple characteristics and test results, individual patient pathways to be recorded and time dependencies to be accounted for.
For reporting our modeling study, we followed the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) Statement [23].
Model structures
The Oncotyrol breast cancer model is divided into different modules that describe the test-treatment strategies and the respective pathways of patients, their health states and key health events (Fig. 1).
In the beginning of the simulation (Module 1), patients enter the model, patient characteristics are assigned (age, time of death from other causes) and their AO risk score (individualized breast cancer specific mortality [BCSM]) and ODX risk classification (recurrence risk score [RS]) are calculated (BCSM: L ‘low’ BCSM < 9%, I ‘intermediate’ 9% ≤ BCSM < 17% or H ‘high’ BCSM ≥ 17% [24]; RS: L ‘low’ RS < 18, I ‘intermediate’ 18 ≤ RS < 30, H ‘high’ RS ≥ 30, N ‘RS not applied’). The costs and benefits of chemotherapy are quantified for each of the twelve combinations of risk classifications (where the first letter represents AO and second letter represents ODX: L-L, L-I, L-H, L-N, I-L, I-I, I-H, I-N, H-L, H-I, H-H, H-N). Two hypothetical cohorts were simulated in which all patients within these risk groups are assumed to receive or to not receive chemotherapy. Patients that pursue chemotherapy continue to Module 2 where chemotherapy and its associated adverse events (neutropenia, fever, infections, pain, nausea and gastrointestinal complications) are modeled. After chemotherapy, these patients are considered recurrence-free and are treated with aromatase inhibitors or tamoxifen for five subsequent years (Module 3). In addition, patients who do not receive chemotherapy enter Module 3 directly. Patients who face disease recurrence continue to Module 4 where further diagnostics and treatments are considered. We assume that patients with a distant recurrence remain in this health state and in Module 4 until they die from breast cancer. Throughout the entire simulated pathway, LYs, QALYs and costs are accumulated, and analyzed in the statistical module. In addition, all patients may die due to other causes at any time point and consequently leave the model.
Model parameters
A detailed description of model parameters is provided elsewhere [19] and an overview of model parameters and sources are shown in Additional file 1: Table S1.
With respect to chemotherapeutic agents, we assumed all patients receive three cycles of FEC (5-fluorouracil, epirubicin, cyclophosphamide) followed by three cycles of DOC (docetaxel) [9]. After completion of adjuvant chemotherapy, all patients also received an aromatase inhibitor (anastozole, letrozole or exemestane) for five years. In cases in which no chemotherapy was provided, an aromatase inhibitor was started immediately.
Risk-group specific time to recurrence estimates were derived from Paulden et al. [16]. Treatment assumptions about distant recurrence were based on chart reviews by a senior gynecologist at Innsbruck Medical Hospital. The probability of death due to breast cancer in patients with distant recurrence was assumed to be identical in all patients regardless of the ER/PR status or the patient’s personal cancer history (median survival 25.8 months from time of diagnosis of recurrence [25]). Fatal toxicity of chemotherapy includes those patients who develop chemotherapy related acute myeloid leukemia (AML). All-cause mortality was applied throughout the entire simulated time horizon. Data were extrapolated using national life tables from Statistics Austria [26].
As ODX is currently not reimbursed in Austria, we relied on the manufacturer’s suggested retail price [27]. AO is available to medical experts free of charge [11]. We included direct costs for chemotherapy and related side effects (costs of chemotherapeutic agents, other supportive medications, such as pegfilgrastim and tropisetron, hospitalization, laboratory studies, and human resources), as well as costs of cancer follow-up, diagnosis and treatment of recurrent cancer [10, 25, 28] [Walter E: IPF, Vienna 2012, Report, unpublished]. Drug costs were based on pharmacy hospital prices. Utility weights were based on a recent cross-sectional observational study using the EuroQol five dimension questionnaire (EQ-5D) [29].
Model validation
Model validation is a key modeling step for judging a model’s accuracy in making accurate predictions. Following the current ISPOR-SMDM best practice recommendations, the model was validated using face validation, internal validation and cross-model validation [30]. Further details are provided in Jahn et al. [20].
Analysis
In the base-case analysis, we estimated discounted effects (LYs, QALYs) and costs of adjuvant chemotherapy in 12 different patient risk groups classified according to their AO (first letter) and ODX (second letter) risk classification (L-L, L-I, L-H, L-N, I-L, I-I, I-H, I-N, H-L, H-I, H-H, H-N). 100,000 patients were needed in the simulation in order to achieve stable results [20].
For each risk group, the simulation was run twice, the first assuming chemotherapy received by the patient and the second run assuming no chemotherapy received. The ICER was calculated by calculating the difference in discounted costs divided by the difference in discounted QALYs for these two alternatives. If one strategy is less effective but more expensive, then it is considered dominated and should not be considered. If chemotherapy is more effective but also more expensive, as compared to no chemotherapy, the ICER expresses the additional costs for one QALY gained. Chemotherapy is considered cost effective if the ratio is less than the willingness-to-pay (WTP) threshold.
As there is currently no explicit willingness-to-pay threshold for health technologies in Austria, we assumed a WTP of 50,000 EUR (alternatively 100,000 EUR) to test the robustness of our results and respective decisions in sensitivity analyses.
Parameter uncertainty was estimated using extensive deterministic one way sensitivity analyses on several parameters including age (40; 50; 70), discount rate (0; 2.5%; 5%), the cost of chemotherapy (+/− 10%), the cost of an ODX test set (+/− 10%), utilities (95% confidence intervals (CI) assuming a beta distribution), and the probability of distant recurrence (95% CI, assuming a beta distribution).
In a cross-country comparison, results were compared to the results of the Canadian modeling study by Paulden et al. [16] who applied a similar model structure. In contrast to our model, the Canadian model was designed as a probabilistic state-transition Markov [31] model for that particular health care setting which differ from Austria. For example, different chemotherapy regimens were considered (low risk patients: CMF (Cyclophosphamide, Methotrexate, 5-fluorouracil), intermediate risk patients: TC (Docetaxel, Cyclophosphamide), high risk patients: FEC-D 5-fluorouracil, Epirubicin, Cyclophosphamide, Docetaxel)). A list of parameter values for this model is provided in the Additional file 1. The modeling framework and the model structure are described elsewhere in greater detail [16].