Dacomitinib in first-line treatment of advanced EGFR-mutated non-small-cell lung cancer: a cost–effectiveness analysis

A partitioned survival model model was constructed using clinical data from the ARCHER 1050 randomized study [15]. The quality-adjusted life-years (QALYs), the costs, and ICER of two different treatment strategies (dacomitinib vs gefitinib) were estimated in two hypothetical cohorts of patients with newly diagnosed advanced NSCLC and one EGFR mutation (exon 19 or Leu858Arg). We modeled the health states of patients with similar criteria to those enrolled in the ARCHER 1050 study: patients in IIIB/IV stage, with new or recurrent diagnosis, and histological or cytopathological confirmation. The presence of at least one documented EGFR mutation (exon 19 deletion or the Leu858Arg mutation, with or without the Thr790Met mutation) was required. The model was developed from the perspective of the Spanish National Health System. The threshold for determining the cost–effectiveness of a strategy was €24,000/QALY [20]. All the costs were estimated in euros (€) 2019, and a discount rate of 3% was used for costs and effects throughout the model. The partitioned survival model was developed in Microsoft Excel 2011 (Microsoft Corp., WA, USA) using a 15-year time horizon, which was selected because it was sufficient to collect all the costs and benefits generated in the model. The results were presented in terms of costs (€), QALYs gained and ICER.

The model included three mutually exclusive health states: stable disease, progressive disease and death. As shown in Figure 1, all patients were initially on stable disease and received one of the two treatment strategies (dacomitinib or gefitinib). On each 28-day simulation cycle, the model redistributes the hypothetical cohort of patients among the three health states according to the transition probabilities. On progressive disease, the patients received a second-line regimen, and after the occurrence of progressive disease, they could remain in this state or die. Progressive disease was simulated until all the patients died. A half-cycle correction was applied.

The model cycle length was 28 days (4 weeks), consistent with the labeled dose frequency of the two treatments. Patients in the dacomitinib group were treated with oral dacomitinib 45 mg once daily in 28-day cycles. Of the patients in this group, 66% experienced a dose reduction, 38% received the lowest dose of 30 mg/day and 28% received the lowest dose of 15 mg/day, according to the ARCHER 1050 trial [15]. The patients in the gefitinib group received gefitinib 250 mg orally once daily in 28-day cycles.

The clinical effectiveness data of PFS and OS were obtained from the ARCHER 1050 trial [15,17], via the techniques outlined in Guyot et al. [21]. WebPlotDigitizer was used to recreate Kaplan–Meier graphs to project outcomes to the end of the 15-year time horizon using Flexurv, an R package for the fully parametric modeling of survival data [22,23] (R version 3.3.3). The following parametric distributions were considered to determine the most appropriate parametric survival curve as recommended by the NICE Decision Support Unit [24]: gamma, log-logistic, Weibull, lognormal, Gompertz, exponential, generalized F and generalized gamma. The distributions were selected based on statistical tests (Akaike Information Criterion [AIC] and Bayesian Information Criterion [BIC]), visual inspection of fit to Kaplan–Meier plots, goodness-of-fit statistics and clinical plausibility. (Figure 2 & Supplementary Table 2).

For PFS and OS, a log-logistic distribution was selected. This distribution has the best visual and statistical fit; the hazards consistent with the observed hazards in the ARCHER 1050 trial [15], do not yield implausible projections with the survival curves for the two arms crossing (Supplementary Material & Figures 1–4). For each TKI, the probability that patients remain in stable disease at each time is determined by the values of the PFS curve at that time. In addition, the probability of patients to achieve the death state is determined as 1 minus the OS curve at that time. From this, the probability of patients in progressive disease follows, as the three states together should always add up to 100%.

Table 1 outlines the calculated costs. Direct medical costs include treatment costs, disease management costs, end-of-life care costs, adverse events costs and second-line treatment costs. The cost of gefitinib and dacomitinib were calculated according to the officially notified listed prices ([drug price – 7.5% official discount in Spain] x Value added tax [VAT]) [25,26]. The cost of dacomitinib can be considered part of the problem because up to the date of publication of this article, an official Spanish price for this drug has not been published, and it has been obtained from NICE guidance [27].

Disease management costs were estimated according to an expert panel’s advice. Disease management cost per patient and cycle was calculated by multiplying the cost of healthcare resources employed by the unit cost of each resource consumed over a 15-year time horizon. The unit costs were obtained from an official database published in Spain [32].

End-of-life care costs were applied to each patient entering the death state and were obtained from an article published in Spain [28].

The costs of side effects management from the perspective of the Spanish National Health System (Table 1) were obtained from published articles [29,30]. Adverse effect (grade 3/4 events) frequencies associated with dacomitinib and gefitinib treatments and reported in at least 3% of patients were obtained from the ARCHER 1050 study [15].

The second-line therapy regimens were obtained from Supplementary Table 1 ARCHER 1050 study [15] as is shown in Table 1. In dacomitinib second-line arm, 23.8% of the patients were treated with pemetrexed, 13.7% with carboplatin, 13.2% with cisplatin and 7.9% with osimertinib. In gefitinib second-line arm, 25.9% of the patients were treated with pemetrexed, 13.8% with carboplatin, 17.9% with cisplatin and 12.9% with osimertinib. The patients were assumed to have a body height of 170 cm and a weight of 70 kg, resulting in a body surface area of 1.73 m².

All cost inputs from prior years were inflated to 2019 Spanish values using the Consumer Price Index. The model costs are presented in Euros (€) 2019 (Table 1).

The ARCHER 1050 study has not reported health state utilities. Thus, utility inputs and disutility values for the base case were estimated from the recent data published in the literature [30,31]. In order to estimate QALYs, utility and disutility values were applied considering the different health states (stable disease and progressive disease) and are summarized in Table 1. A health utility of zero was applied to the health state of death.

The disutility values associated with grade 3/4 adverse events while the patients remained in stable disease were adopted from a recently published international study that evaluated the disutilities and complications for advanced NSCLC in different countries like the UK, France, Australia and Republic of China by employing a time trade-off technique [33]. To calculate the disutility values associated with grade 3/4 in stable disease, the disutility parameters of each adverse event extracted from Nafees et al. were multiplied by the relative frequency of the corresponding event obtained from the ARCHER 1050 trial to calculate a weighted average disutility value for each event profile as is shown in Supplementary Table 1. The disutility values calculated for each grade 3/4 adverse event were subtracted from the utility values while the patients remained in stable disease.

Deterministic sensitivity analysis (DSA) was performed to explore the impact of the essential variables on the ICER estimated value. Thus, a single parameter in the model (drug costs, utilities or discounts) was varied to test the effect on the ICER result. The utilities values were varied in a range of ±20%. The drug acquisition costs were modified in three different ranges (±15, ±20 and ±25%). The end-of-life care costs were varied in two different ranges (±10 and ±20%), and the discounts values in the DSA were modified in a percentage of 0 and 6%. The transition probability values of PFS and OS were varied in a range of ±5. The results of the DSA were presented in a tornado diagram.

A probabilistic sensitivity analysis (PSA) was conducted to assess the influence of parameter uncertainties using 10,000 Monte Carlo simulations. Different parameters (side effects management costs, disease management costs, second-line treatment costs, acquisition costs, end-of-life care costs, utilities and transitions probabilities) of the model were varied to determine the robustness of the model. The results of the PSA was employed to obtain the cost–effectiveness acceptability curves, showing the probability of each alternative being cost-effective across a range of possible values of willingness-to-pay (WTP) for an additional QALY [34].

The different types of probability distributions were applied to variate the model parameters according to the characteristics of each variable [35]. Gamma distributions were employed for costs, beta for utilities and Dirichlet distributions for transitions probabilities. The number of postprogression systemic treatments per patient were assumed to follow a triangular distribution.