Model overview
The analysis included cost-effectiveness and cost-utility analysis of crizotinib compared with pemetrexed-based platinum-doublet chemotherapy for patients with previously untreated, ROS1+ advanced NSCLC. The outcomes included total costs, life-years (LYs), quality-adjusted life-years (QALYs), and incremental cost-effectiveness ratios (ICERs) per LY gained and per QALY gained. LYs measure the average expected survival time with each strategy, and the difference between strategies represent incremental LYs gained with crizotinib. QALYs incorporate effects on both quality (morbidity) and quantity (mortality) of life [10]. To calculate QALYs, LYs are multiplied by a quality of life weighting that measure preference for a given health state, known as health utility [10]. The ICER is a ratio of the difference in expected costs divided by difference in expected outcomes (either LY or QALYs), representing the additional cost associated with each additional (quality-adjusted) year of life gained with the crizotinib strategy A compared to usual care strategy B [10].
$$ ICER=\frac{Cost_A-{Cost}_B\ }{Effects_A-{Effects}_B}=\frac{\Delta C}{\Delta E} $$
The model was developed from the perspective of the publicly-funded Canadian health care system over a 10 year time horizon using a one-week model cycle length. Future costs and outcomes were discounted at a rate of 1.5% per year as per Canadian guidelines [11].
The cost-effectiveness model was developed in Microsoft Excel® (version 16.0) using a decision tree for ROS1 testing and a five-health state Markov model for the disease, which included health states for PFS, two post-progression survival states for further lines of therapy, palliation and death (Fig. 1).
Molecular testing was conducted using immunohistochemistry (IHC) followed by confirmatory fluorescence in situ hybridization (FISH) of positive staining results for the ROS1 mutation, based on test characteristics and frequency of ROS1+ (estimated to be 1.64% of NSCLC) [12, 13]. Patients with positive test results enter the Markov model in the PFS state and receive first-line therapy with crizotinib or pemetrexed-based platinum-doublet chemotherapy. Clinicians have suggested that a subset (estimated to be 50%) of patients receiving platinum-doublet chemotherapy may go on to receive pemetrexed maintenance, based on patients’ fitness and willingness to continue therapy. Patients who progress after receiving initial treatment with crizotinib who are eligible for further therapy are assumed receive treatment with standard of care options, starting with platinum-doublet chemotherapy and option of pemetrexed maintenance for second-line treatment. Patients initially treated with platinum-doublet chemotherapy receive second-line treatment with either a checkpoint inhibitor or docetaxel. Clinical input has suggested there is little evidence that these therapies differ in efficacy. Upon progressing, patients who received platinum-doublet chemotherapy as second-line treatment were assumed to receive either checkpoint inhibitor or docetaxel for their third-line treatment whereas patients who received either checkpoint inhibitor or docetaxel for their second-line treatment received the alternative for third-line treatment. Patients deemed unfit for subsequent treatment or who progressed after third-line treatment receive palliation until death. Patients could progress to death from any health state in the model.
Clinical inputs
Progression-free survival
A literature review was conducted to identify relevant clinical evidence on efficacy and safety of crizotinib or pemetrexed-based platinum-doublet chemotherapy regimens for the treatment of ROS1+ advanced NSCLC (Additional file 1: Fig. 1). Given the absence of any randomized controlled trials or prospective trials for chemotherapy in patients with ROS1+ NSCLC, retrospective studies were also included. To estimate survival endpoints for the economic model, studies were excluded if they did not have Kaplan Meier (KM) time-to-event data for PFS. Seven studies (2 trials and 5 retrospective studies) were identified (Additional file 1: Table 1). Due to lack of comparative data, the small sample size of the studies, and the range of outcomes, there was no definitive source among the studies for robust efficacy data. Treatment effectiveness was estimated by pooling the time-to-event PFS data from all identified studies for crizotinib [7, 14, 15] and pemetrexed-based platinum-doublet chemotherapy [15,16,17,18,19]. PFS KM curves from each study were digitized using Engauge (version 4.1) software and individual patient data (IPD) were pooled (Additional file 1: Fig. 2). For studies with very small samples, each event and censoring time was accurately replicated from the plot. For studies with larger samples, patient-level IPD were approximated using the methods of Guyot et al [20].
After pooling PFS data, parametric curves were fitted assuming the following distributions: exponential, Weibull, log-logistic, log-normal, generalized gamma and gompertz. Statistical tests failed to reject the assumption of proportional hazards (p > 0.05), suggesting it was reasonable to model with a common treatment parameter (proportional hazards or acceleration factor). The best fitting distribution (log-logistic) was chosen based on statistical information criteria, visual inspection of the curve and clinical plausibility [21] (Fig. 2). The survival functions were used to estimate monthly transition probabilities from the initial PFS state.
Scenario analyses for the clinical inputs included alternative data sources and approaches to modelling PFS. These include survival models estimated without using a common treatment parameter (wherein parametric curves were fitted to each treatment group separately), as well as using survival data from individual studies rather than the pooled analysis. Additionally, based on a review of previous health technology assessment (HTA) reviews, where it was noted that ROS1+ and ALK+ NSCLC share similar characteristics, [22] we conducted scenario analysis using PFS data from the phase III clinical trial for previously untreated ALK+ NSCLC, PROFILE 1014, [5] for comparability (Additional file 1: Fig. 3). The IPD for the ALK+ NSCLC data were recreated using the same methods as described above, and as the proportionality was deemed to be violated (p < 0.05), parametric curves were fitted individually to each treatment arm. In all scenario analyses, parametric survival distributions were fit using the same best practice approach considering statistical criteria, visual inspection and plausibility [21].
For patients who began treatment on chemotherapy, some may receive single-agent pemetrexed maintenance. While maintenance use was not reported in two of the studies contributing to the chemotherapy arm, it was assumed that use of maintenance would have an added PFS benefit. This was considered conservative because it would further improve PFS for the chemotherapy arm beyond what was observed in the pooled retrospective studies (and therefore reduce the incremental difference in PFS between treatment arms). Based on the PARAMOUNT trial, [23] which compared patients who receive pemetrexed maintenance relative to placebo after initial treatment with platinum-doublet chemotherapy, a hazard ratio (HR) of 0.62 (95% CI: 0.49–0.79) was applied to estimate the transition probabilities for patients continuing on pemetrexed maintenance (50% of patients remaining progression-free after induction) relative to patients undergoing no maintenance (Additional file 1: Fig. 4).
Patients in each arm progressed to second-line treatment, palliation, or death based on the time-varying transition probabilities from PFS, with 70% of patients assumed to receive second-line treatment, and remaining patients assumed to discontinue further therapy and receive palliative treatment (10%) or succumb to their disease (20%).
Progressed disease
Progression following all subsequent health states were estimated assuming exponential distributions. The risk of progression during second-line platinum-doublet chemotherapy was based on the median PFS value obtained from the combined analysis of chemotherapy studies, given this represented all available evidence for ROS1+ NSCLC. In the same manner as first-line, it was assumed that patients could undergo maintenance treatment (50%) and would experience an improvement in efficacy based on a HR of 0.62 [23]. Patients who received other second-line treatments were assumed to progress based on a weighted average of median PFS for checkpoint inhibitors or docetaxel (Additional file 1: Table 2), [24,25,26,27] with 60% assumed to receive third-line treatment, and the remaining equally likely to transition to palliation (20%) or death (20%). Efficacy was assumed to be similar for docetaxel or checkpoint inhibitors in both second- and third-line treatment. After third-line treatment, patients could either progress to palliation (60%) or to death (40%). The rate at which patients died from palliation was obtained from the TAX 317 study comparing patients receiving docetaxel or best supportive care [28].
Cost inputs
Molecular test acquisition costs were used assuming upfront testing for ROS1 positivity to determine the total testing cost per case detected (Table 1). The test costs for IHC and FISH were obtained from the literature, [9] and adjusted to 2018 CAD costs using CPI health index [35].
For all regimens, an average patient body surface area (BSA) of 1.75m2 was assumed. Crizotinib costs were based on the current Ontario list price of $130.00 per 250 mg tablet, [29] given twice daily until progression or death [5,6,7, 14, 15]. Platinum-doublet chemotherapy consisted of a platinum agent (carboplatin or cisplatin) with pemetrexed, the preferred regimen of clinical experts. The cost for each treatment was obtained from Ontario’s cancer agency, Cancer Care Ontario (CCO), [30] and the dosing schedule obtained from clinical trials for a maximum of 6 cycles [5]. Carboplatin dosing was based on the target AUC according to the Calvert formula [36]. Costs for post-progression treatments with nivolumab, pembrolizumab and docetaxel were obtained from CCO, [5] and dosing and treatment schedules obtained from recent publications [23,24,25,26,27].
Administration costs, estimated from CCO costing data, included the costs of supplies, pharmacy, nursing, and administrative (clerical and management) staff for operation of the outpatient chemotherapy clinic. Monitoring costs were calculated using input from clinicians on resource utilization and the costs obtained from literature or the Physician and Laboratory Schedule of Benefits [31].
As no comparative adverse event (AE) data are available for the ROS1+ NSCLC population, comparative AE rates were obtained from the PROFILE 1014 study in ALK+ NSCLC for crizotinib and chemotherapy. Inpatient hospitalization and outpatient ambulatory care costs for diagnoses corresponding to grade III/IV AEs from the clinical trial were estimated from the Ontario Case Costing – Costing Analysis Tool [32].
The cost for palliation was obtained from a matched cohort study that estimated phase-specific net costs in Ontario, [33] and reported 12-month terminal care costs, which was used to calculate a weighted monthly cost of $3124.07.
In order to minimize bias in survival from differing number of lines of therapy available, three treatment lines were assumed in both groups followed by equal risks of death during palliation between groups. However, costs of a further line of therapy were included for some patients treated with crizotinib during the palliation state (50%), to avoid biasing the total costs in progressed states in favour of the crizotinib group.
Utility inputs
Quality of life was not measured using any utility instruments in any of the included studies for ROS1+ NSCLC. Based on the similarity between ROS1+ and ALK+ NSCLCs, it was assumed that quality of life data from the PROFILE 1014 study in ALK+ NSCLC would be similar for patients with ROS1+ NSCLC. In the PROFILE 1014 trial, patients had similar quality of life at baseline, but patients on crizotinib experienced significant improvement during treatment compared to platinum-doublet chemotherapy, with mean EQ-5D index scores on treatment of 0.81 for crizotinib and of 0.72 for platinum-doublet chemotherapy. Similar to the approach taken by the National Institute for Health and Care Excellence (NICE) evidence review, we estimated a smaller utility difference between the groups, based on the difference in disutility due to AEs (0.034), to minimize the difference between the treatment strategies, resulting in a utility value of 0.77 for platinum-doublet chemotherapy (Table 1).
Patients who progressed on the first-line treatment were assigned a utility value of 0.66, obtained from previously treated ALK+ patients receiving docetaxel [6]. It was also assumed that treatment with checkpoint inhibitors and docetaxel were associated with similar quality of life in second- or third-line treatment. Utility values for palliative care were obtained from a valuation study with 100 participants administered a standard gamble interview about health states described by oncologists and nurses [34]. Scenario analyses were also conducted with equal utility scores between arms and with PROFILE 1014 values.
Analysis
Probabilistic analysis was conducted with 5000 simulations to incorporate uncertainty in the model parameters together at once. Beta distributions were used for model parameters with values between 0 and 1 (probabilities, proportions and utilities), normal distribution for population values (e.g., BSA), and gamma distribution for costs. For the correlated uncertainty in the extrapolation parameters, we used normal distributions and the Cholesky decomposition. One-way sensitivity analyses on all input parameters of uncertainty and several scenario analyses were conducted to explore assumptions (Additional file 1: Table 3).