Type of Economic Analysis
We performed both a cost-effectiveness analysis and a cost-utility analysis where the ICER gained was derived from estimated life-years and QALYs gained and direct healthcare costs. Costs are expressed in Canadian dollars, year 2018 values. The cost-utility analysis was performed in accordance with established guidelines for economic evaluations  and adopting a Canadian provincial healthcare system perspective .
The patient population from the AZA-AML-001 clinical trial  served as the basis for the economic analysis. Specifically, these were patients aged ≥ 65 years with a new diagnosis of de novo AML with > 30% blasts in the bone marrow who were not considered candidates for allogeneic stem cell transplantation.
Treatment with azacitidine is compared with CCR, namely the three accepted therapies for AML in transplant-ineligible patients: BSC, low-dose cytarabine (LDAC) and standard-dose chemotherapy (SDC). In the AZA-AML-001 trial, patients were first assigned by their treating physician to receive the CCR that best suited them. Patients were then randomized in a 1:1 ratio to receive either azacitidine or the CCR to which they had been assigned. Patients remained stratified according to their pre-randomization arm to ensure that similar patient groups were compared.
Model Structure and Outcomes
We developed a Markov model to evaluate the health outcomes and costs of azacitidine compared with each CCR option (Fig. 1). The model was developed using Microsoft Excel and underwent verification and validation consistent with recommendations by Vemer et al.  and the International Society for Pharmacoeconomics and Outcomes Research Task Force . Hypothetical patients entered the model in the AML state (> 30% blasts in the bone marrow) at age 69 years, thus matching the AZA-AML-001 population and transited either to complete remission (CR) or death (absorbing state). Patients in CR could in turn transit back to the AML state or progress to the absorbing state (Fig. 1). Response evaluations in the AZA-AML-001 trial were scored as per the International Working Group in AML , with CR defined as ≤ 5% blasts in the bone marrow with absolute neutrophil counts ≥ 1 × 109/L and platelets ≥ 100 × 109/L.
The primary endpoints were OS, QALYs and direct medical costs and are expressed as ICERs. OS and QALYs were selected as outcomes since they are most clinically relevant in patients with AML.
In the model, patients were assumed to receive treatment for AML until the end of the treatment cycle, after which they could either remain on treatment or have the treatment put on hold (Fig. 1).
Efficacy, Safety and Adverse Events
The base-case model inputs and probability distribution are shown in Table 1. Probabilities of outcomes in each treatment arm were derived from the AZA-AML-001 clinical trial . The AZA-AML-001 study enrolled 488 patients with de novo AML; these patients were then followed for a median of 24.4 months .
The transition probabilities to the absorbing state of each cycle were extracted from the Kaplan–Meier curves of the AZA-AML-001 study using the WebPlot Digitizer web app v3.10  [see Electronic Supplementary Material (ESM) 1]. The transition probabilities from the time of randomisation to the CR state and from the CR to the AML state for both the azacitidine and the CCR arms were derived from the AZA-AML-001 randomized study comparing the two arms  (see ESM 2 for details). The measurement of the transition probabilities was made relative to the randomisation and not from the time of diagnosis since the delay between diagnosis and randomisation was not available to the authors.
For the model, we assumed that patients with AML treated with azacitidine transitioned to the CR state between cycles 4 and 6. This reflected what has been documented in the literature [13, 14]. To reflect clinical experience, in the CCR arm, we assumed that only patients who received induction chemotherapy could achieve CR; only 2–3% of patients treated with LDAC achieved CR, and none of the patients managed with BSC achieved CR . In the model, the transition to a CR state was assumed to occur between cycles 2 and 6. This assumption was made based on the observation that patients assigned to receive induction chemotherapy most commonly need to receive sequential cycles of chemotherapy before achieving remission and that the treatment cycles are frequently delayed by 1–3 months (corresponding to the duration of 1–3 cycles). This implies that patients assigned to the induction chemotherapy arm had a chance of achieving CR only after 2 cycles until cycle 6.
Baseline utility values were obtained from results presented in the AZA-AML-001 study . The subsequent changes in utility scores were obtained after converting the EORTC QLQ-C30 scores collected alongside the AZA-AML-001 clinical trial at cycles 3, 5, 7 and 9, thus ensuring optimal internal validity. Moreover, trial participants included adult Canadian patients with AML with > 30% blasts. These health scores were converted to utility scores using the mapping algorithm described in McKenzie and van der Pol . We conducted a linear regression analysis to derive the impact on utilities of both treatment arms and whether the patient achieved complete response (see ESM 3). This allowed estimation of the utility values for the whole time horizon of the model based on the proportion of patients for each treatment that remained with complete response. Adverse events were not specifically modelled since they were explicitly included in the QOL questionnaire.
Direct medical costs were adjusted to year 2018 values using the Bank of Canada inflation calculator . Healthcare costs were derived from the Levy et al.  study, in which the health costs were derived from a Canadian panel of four clinical experts who estimated the costs associated with health resource utilization borne to the Ministry of Health and Long-Term Care in Ontario, Canada. These costs included medication, human health resource, routine physician follow-up, hospitalization and laboratory investigations. The cost of 1 cycle of azacitidine in year 2018 values was estimated at $Can8701.25, with a unit price of azacitidine of $Can662 (average dose of 1.34 vials; Celgene Ltd). For the CCR comparators, the following assumptions were made: BSC is exclusively received as an outpatient, LDAC is administered 50% of the time by an outpatient clinic nurse and induction chemotherapy is received exclusively as an inpatient .
Uncertainty in model parameters was characterized by probability distribution using simulations, and 5000 iterations were used to include parameter uncertainty in the results. In addition, to examine the effects of individual parameter uncertainty, one-way sensitivity analysis varying each parameter included in the model was performed, with parameters varied by 20%, giving a range of values that we felt was reasonable. Results of the one-way sensitivity analysis are presented in a tornado diagram.
The cycle duration was 35 days, corresponding to the average treatment cycle duration in the AZA-AML-001 clinical trial. A time horizon of 25 months was selected for our base case as the Kaplan–Meier survival curves for azacitidine compared with CCR converged at this point, with no difference in survival beyond this time horizon. Thus, expanding the time horizon is unnecessary as it likely would have little impact and introduce greater uncertainty because of the need to extrapolate beyond the known clinical data . However, we conducted a further analysis to assess the impact of adopting a lifetime or 10-year horizon using survival data beyond 25 months from the same study. For this analysis, an exponential model was fitted to the data between 24 months and 28 months and then used to extrapolate for the estimation of lifetime outcomes. The exponential model provided the best fit for the available data. The analysis found both no difference in survival between the two regimens post 24 months and a consistent probability of death across the time period (see ESM 4).
The model assumed that patients received the assigned treatment until the end of the treatment cycle, after which they could either remain on treatment or have the treatment put on hold (Table 2).
Costs and benefits are discounted at 1.5% per annum, with sensitivity analysis adopting rates of 0 and 3% .