Cost-Effectiveness of Nivolumab Plus Ipilimumab versus Chemotherapy for Previously Untreated Metastatic NSCLC Using Mixture-Cure Survival Analysis Based on CheckMate 227 5-Year Data

The population modelled in this study represents the CheckMate 227 (NCT02477826) Part 1 intention-to-treat (ITT) population in a US healthcare setting. This consisted of adults with stage IV or recurrent NSCLC who had received no previous systemic anticancer therapy for metastatic disease. Patients with squamous or non-squamous disease were eligible, but excluded if suitable for targeted therapy due to presence of sensitising EGFR mutations or ALK translocations. Patients were required to have an Eastern Cooperative Oncology (ECOG) performance status of 0 or 1, and were excluded in the presence of autoimmune disease, or untreated or symptomatic central nervous system metastases. The Part 1 population from the CheckMate 227 trial [14] modelled here comprised patients randomised to receive NIVO (3 mg/kg/2 weeks) + IPI (1 mg/kg/6 weeks) up to 2 years, or chemo (platinum doublet, every 3 weeks for 4 cycles). Patients were included regardless of tumour PD-L1 expression, as the benefits of NIVO+IPI were observed in both populations within CheckMate 227. Patient characteristics from CheckMate 227 were assumed to be representative of patients with previously untreated metastatic NSCLC eligible for systemic therapy in the US. This population allows results to be directly compared with that of Berling et al. [24].

A mixture-cure economic model with health states defined by progression-free (LTR), progression-free (non-LTR), post-progression, and death was implemented in Microsoft Excel^® (Office 365). At model initiation, all patients were in the progression-free states, with a fraction (\(\theta\)) in the LTR state and the complement in the non-LTR state. Patients in the LTR state transitioned to the death state at a pre-specified rate identical to the general population. Patients in the non-LTR state transitioned to the post-progression and death states at a rate compliant with the at-risk fraction of parametric mixture-cure models of PFS and OS. A schematic of the model structure is provided in Fig. 1.

A 3-week cycle length was used, corresponding to the treatment cycle length for chemo. Half-cycle correction was employed for state-based cost and quality-adjusted life-year (QALY) aggregation, but was not used for first-line drug acquisition and administration costs, as these were incurred at the start of each model cycle. A lifetime time horizon (up to age 100 years) was applied. Adverse events related to first-line therapy, treatment-related adverse events (TRAEs), were assumed to occur over the first four model cycles. A summary of the core assumptions of the model is shown in Supplementary S1 (see the electronic supplementary material).

OS and PFS per Blinded Independent Central Review (BICR) data from the 5-year database lock of CheckMate 227 were used to fit parametric mixture-cure survival models. In these models, all patients were considered at a baseline hazard of general population mortality informed by contemporary general population life tables stratified by country of study centre and patient sex and age. In scenario analyses, hypothesised mortality ratios were applied to these hazards in order to represent a residual excess hazard of mortality amongst the LTR. In addition to this hazard, an unidentified at-risk (non-LTR) fraction was assumed to experience an additional parametrically defined hazard. Maximum likelihood parameters defining this non-LTR hazard function and the proportion within the LTR fraction were sought via optimisation within the flexsurvcure package [30] within the R statistical environment [31].

A non-mixture-cure survival analysis was considered, in which the cure fraction is approached according to an exponential asymptotic model scaling a standard cumulative distribution function [32]. However, this study derives extrapolative validity from the assumption of similarity of survival time distributions in the fraction without LTR with a standard parametric distribution. This contrasts with the assumption of continuity of a hazard function representing an improper statistical distribution, as used for extrapolation of non-mixture survival models, and represents a class of models more similar to those suggested as the defaults in model selection guidelines [33].

In all survival approaches, the parametric analysis was conducted in reference to Decision Support Unit (DSU) guidance, Technical Support Document 14 [33], and implementation of mixture-cure and relative survival modelling was considerate of the recommendations of Technical Support Document 21 [34]. As the objective of this study was to employ a mixture-cure modelling strategy to explain heterogeneity in survival times, it was known that the resultant models would not display marginal proportional hazards (PHs) or the accelerated failure time (AFT) property, as all models would asymptotically converge to the baseline (general population) hazard. However, it was considered that such scaling rules may be appropriate for the non-LTR fraction, and three model structures were explored:

Several models were considered for the representation of survival among the non-LTR fraction, i.e. the exponential, Weibull, Gamma, Gompertz, lognormal, log-logistic, and generalised Gamma models. Model fit was assessed for each model using plots of the predictions versus observations (hazard rates, cumulative hazard rates, survival rates), Akaike information criterion (AIC), and Bayesian information criterion (BIC) statistics. The face validity of survival predictions was determined by comparison of results to the existing cost-effectiveness model and a comparison of estimates to an existing conditional survival curve derived from published literature (supplemental materials of [24]).

Choice of the outcome (OS, PFS) defining the LTR fraction imposes a restriction on the LTR fraction of the other outcome, as the model structure precludes a long-term post-progression state. This decision was made in order to allow for subsequent modelling of indirect comparators for which patient-level data were unavailable. As such, to ensure good calibration to observed data, after selection of the LTR-determining outcome, the other outcome was fitted using the defining model LTR fraction. In the base case, OS was chosen as the LTR-determining outcome as fitting to PFS resulted in consistent under-prediction of OS on the NIVO+IPI arm from approximately 2 years.

As mixture-cure survival models trend asymptotically towards the baseline hazard, the absolute difference in hazards between NIVO+IPI and chemo was constrained to demonstrate a decreasing treatment effect in both PFS and OS. As this rate of loss of treatment effect was driven by trends in the CheckMate 227 data, and as previously used assumptions such as those in National Institute for Health and Care Excellence (NICE) technical assessment 724 [35] required an implausible loss of treatment effect within the observed period of CheckMate 227, this was considered the most plausible expression of a loss of treatment effect practicable within the model.

Values are detailed in Supplementary S2 (see the electronic supplementary material). The utility in the LTR state was assumed to be equivalent to the general population according to US population norms [36]. Utility values applied in the non-LTR states of the model were estimated using an analysis of EQ-5D-3L data from CheckMate 227, using US preference weights for the progression-free and progressed-disease states [24]. Mixed model regression methods were used to estimate both progression-based and time-to-death (TTD)-based utility values. These models were consistent with those used by Berling et al. [24] to isolate the impact of the survival modelling strategy on the cost-effectiveness results. A scenario analysis setting LTR utility equal to PFS utility was run; as general population utility was only 2% greater than PFS utility, this had a minor impact on the incremental cost-effectiveness ratio (ICER). Disutility due to TRAEs was applied uniformly over the first 12 weeks, i.e. assuming 25% of incidence during each 3-week cycle up to the expected end of chemo. An end-of-life disutility based upon observations from CheckMate 227 was applied.

This study adopted a US third-party payer perspective to align with other previously published non-mixture-cure survival analyses [24, 25]. The model considered direct healthcare costs including drug-acquisition and drug-acquisition-cost modifiers (Supplementary S3 and S4; see the electronic supplementary material), drug administration costs (Supplementary S5), drug monitoring costs (Supplementary S6), subsequent systemic therapy costs (Supplementary S7), drug cost of resolving TRAEs (Supplementary S8), and disease management costs (progression-free and progressive disease health states and end-of-life care; Supplementary S9 and S10). All unit costs were based on US-specific costs (2021 dollars, inflated where required using the US medical care consumer price index) [37]. Resource use estimates and costing year were matched with an existing cost-effectiveness analysis of NIVO+IPI based on CheckMate 227 to allow comparison of results [24]. Annual discount rates of 3.0% have been applied to all costs and health benefits, as recommended by the US Institute for Clinical and Economic Review [27]. Baseline hazards were calculated from US general population life tables [28]. A willingness-to-pay threshold of US$150,000/QALY was assumed [29].

Acquisition, administration, and monitoring costs of first-line treatment were estimated based upon duration of study treatment recorded in CheckMate 227. These costs have been aggregated in the progression-free non-LTR state. For NIVO+IPI, a per-cycle cost was calculated and was assumed to apply to all patients remaining on first-line treatment until a maximum of week 108, the maximum time on treatment observed in CheckMate 227.

For chemo, an initial per-cycle rate of cost aggregation was calculated, weighted based upon the treatment mix among the ITT population, and applied to all patients on first-line treatment for the first four model cycles (i.e. until week 12). After this point, all patients remaining on therapy for the duration of study records of CheckMate 227 were assumed to be receiving maintenance treatment in the form of the chemo agent pemetrexed. Costs were subsequently aggregated according to this schedule until the end of CheckMate 227 follow-up (i.e. until week 246). Wastage was assumed for each dose of each treatment based upon mean patient weight and body surface area in CheckMate 227; these values are consistent with other cost-effectiveness analyses using alternative survival modelling techniques, but it is noted that the mean patient weight may be lower than that expected in a US cohort, and usage of the method of moments would be expected to improve accuracy of wastage calculations [38].

In addition to first-line treatment-related costs, costs due to TRAEs were aggregated over the first 12 weeks of the model. State-specific disease management costs were accrued in the progression-free and progressed-disease states, and a terminal care cost applied upon transition from any state to the death state. Upon transition to the progressed-disease state, patients were assumed to have a chance of commencing a subsequent therapy line, informed by the distribution of subsequent systemic therapies used by patients in CheckMate 227, shown in Supplementary S7. The weighted mean cost of an average treatment course with these therapies, reduced to the proportion observed commencing any subsequent systemic therapy, was applied at the time of progression to all patients. A majority of chemo patients receiving subsequent therapy received nivolumab or pembrolizumab, whilst re-challenge with immunotherapy was less common for NIVO+IPI. As a result, chemo has a higher per-patient subsequent treatment cost, but the efficacy of subsequent-line immunotherapy is reflected in the OS outcomes of CheckMate 227.

Due to missed or reduced doses, there was a lower mean dose intensity for each component of treatment in CheckMate 227 than that predicted by the complete duration of study therapy. To compensate for this reduction in effective resource use, an overall mean relative dose intensity, the ratio of observed exposure to expected exposure for each treatment component in the clinical trial, was applied to the acquisition costs of each therapy (Supplementary S4). A no-wastage assumption was made on missed or reduced doses, and sensitivity to this was tested in a scenario analysis.

The incidence rate of TRAEs was informed by safety analysis of the CheckMate 227 Part 1 safety population, with adverse events of neutropenia, anaemia, neutrophil count decrease, thrombocytopenia, and febrile neutropenia incurring costs and loss of health-related quality of life within the model. The majority of events occurred over the first 12 months of the trial, but within this economic evaluation were assumed to accrue uniformly over the first 12 weeks for both arms. The pre-maintenance treatment period for chemo and the assumed compression of TRAEs for NIVO+IPI were assumed in this 12-week window, and would provide no discounting benefit to the latter therapy. This was considered a conservative assumption.

Deterministic sensitivity analyses were conducted to assess the impact of individual parameters on results, detailed in Supplementary S11 (see the electronic supplementary material). In addition, a number of scenario analyses were conducted to assess the impact of modelling decisions, such as the survival distribution, baseline hazard of mortality, and utility model. Finally, probabilistic sensitivity analysis (PSA) was conducted, varying the majority of uncertain parameters simultaneously.

Due to the linking PFS and OS LTR fraction and the expected failure of the multivariate normal assumption for parameters of the survival models, a non-parametric bootstrap was used to generate 1000 linked parameter estimates for the PFS and OS models, which were resampled in PSA.

The OS outcome was chosen to determine the LTR fraction due to the maturity of the data at the 5-year data cut-off. Whilst fully independent models fitted to the NIVO+IPI and chemo arms had the highest statistical goodness-of-fit per AIC and BIC, the scaled non-LTR and shared non-LTR models did not differ by more than 10 units of these measures and demonstrated similar visual goodness-of-fit (Supplementary S12; see the electronic supplementary material). Under consideration of parsimony, and the desirability of fitting only a single “treatment effect” parameter (i.e. the difference in LTR fraction between treatments) with consideration of future incorporation of comparators through indirect treatment comparison, the shared non-LTR model was preferred.

The model chosen was log-logistic, due to its low AIC and BIC, and due to the observation that predictions from models fitted to data from the 3-year data cut-off were best for “long-tailed” (i.e. log-logistic, lognormal) models. It has been previously noted that despite poor initial prognosis, many real-world distributions of survival time after diagnosis with advanced cancer tend to a decreasing hazard function [39]. In this publication, based upon the US Surveillance, Epidemiology, and End Results (SEER) database, patients with advanced NSCLC who survived beyond 5 years in the 60- to 69-years age group were at a hazard of mortality of approximately 0.02/month, decreasing to approximately 0.01/month at year 10. This is very consistent with the presently modelled hazard of mortality in the chemo arm at year 10 of 0.0098/month (Supplementary S13). NIVO+IPI was estimated to have a marginal hazard of mortality at this time approximately half that of chemo and double that of the baseline hazard. As the PFS model shared the LTR fractions estimated from the OS model, a log-logistic model was also used for PFS. Predictions from this model are shown in Fig. 2. The fitted NIVO+IPI PFS model demonstrated over-estimation from year 3.

The LTR fraction varied according to the choice of model for the non-LTR, with short-tailed distributions generating higher LTR fractions (e.g. exponential 22.1% NIVO+IPI, 9.9% chemo; Weibull 22.9% NIVO+IPI, 10.7% chemo) than long-tailed distributions (e.g. log-logistic 15.4% NIVO+IPI, 2% chemo; lognormal 14.6% NIVO+IPI, 1.9% chemo) (Supplementary S14). The Weibull distribution was used in scenario analysis to characterise sensitivity of the model to this choice. An additional scenario using treatment-independent estimates of the non-LTR distribution parameters was used to characterise sensitivity of the model to the shared non-LTR distribution assumption.

The base-case results for NIVO+IPI and chemo are reported in Table 1. Patients gained 4.04 total life-years (LYs) with NIVO+IPI, versus 2.35 with chemo. NIVO+IPI analysis resulted in an increase in total LYs of 1.69 (4.04 NIVO+IPI; 2.35 chemo), and increased QALYs of 1.42 (3.18 NIVO+IPI; 1.76 chemo), with incremental and payer total costs versus chemo of US$125,321 (US$233,335 NIVO+IPI; US$108,014 chemo), for an estimate cost-utility ratio of US$88,219/QALY (Table 1). Benefits to LYs and QALYs were mostly accrued in the LTR state (Table 2), as a smaller proportion of patients entered the non-LTR progression-free and progressed-disease states. It was structurally imposed by the shared non-LTR survival models that for the proportion at risk, the time in the progression-free without LTR state and progressed-disease states was identical between arms. Costs were higher in the NIVO+IPI arm, primarily due to increased first-line drug acquisition costs (Table 1). The incremental cost per LY gained was US$74,053.

Deterministic sensitivity analysis identified that the ICER was most sensitive to the discount rate applied to QALYs, with lower discounting improving cost-effectiveness, reflecting the long-term survival predicted for NIVO+IPI (Fig. 3). The costs of NIVO and IPI were also highly influential on the cost-effectiveness results, as expected, due to their high impact on the total costs of the NIVO+IPI arm. The model was also sensitive to the utility value assumed for those with LTR; structurally, a higher proportion of patients in this state was the only way in which the survival time benefits due to NIVO+IPI could be expressed. Notably, the model was insensitive to the utility value assumed for the progression-free without LTR and progressed-disease states.

Probabilistic sensitivity analysis was conducted with 1000 iterations of the input parameters. Mean incremental costs were US$125,789 and mean QALY gain was 1.422, resulting in an incremental cost per QALY gained of US$88,454, consistent with the deterministic base case. In 72% and 98% of iterations, respectively, NIVO+IPI was cost-effective versus chemo at a willingness-to-pay threshold of US$100,000 and US$150,000 (Fig. 4).

Most scenario analyses explored sensitivity to survival modelling decisions (Table 3). Despite the higher LTR fraction for the NIVO+IPI arm, use of Weibull survival models resulted in slightly higher ICERs due to the proportionally higher increase in LTR for the chemo arm. Health state utility for the pre-progression (uncured) state in the model is based upon data collected in PFS; use of the PFS endpoint to determine the LTR fraction improved calibration to observed PFS and increased the ICER by US$15,000; however, between years 2 and 6, the difference between the predicted OS rates for NIVO+IPI and chemo were consistently lower than that between their respective Kaplan-Meier estimators in this scenario (Supplementary S15; see the electronic supplementary material). Scenarios employing a higher rate of mortality in the LTR fraction decreased cost-effectiveness. For these scenarios, models were re-fitted to account for the baseline hazard forming a higher proportion of the total and so ensure good fit to the observed OS data; however, calibration to observed PFS became worse, as higher estimated LTR fractions over-estimated PFS. Two scenarios were run. One had a baseline hazard twice that of the general population and thus a baseline hazard at year 10 that approximated the total NIVO+IPI hazard in the base case. A second scenario investigated a baseline hazard five times that of the general population, and thus was similar to the total hazard of mortality among (historical) survivors in the SEER advanced NSCLC cohort at year 10. A theoretical scenario investigating maximum possible treatment costs by removal of the dose-intensity modifier (ratio of true drug exposure relative to expected dose) increased acquisition and administration costs to US$186,625 and US$48,751 for NIVO+IPI and chemo, respectively, commensurate with other analyses and generating an ICER of US$93,135.

The model was not sensitive to differences in sources of health state utility values, or replacement of the health state utility evaluation with a TTD utility model. Removal of the treatment-cost modifier, reflecting the ratio of observed to time-on-treatment expected use of first-line therapies, impacted the total costs of NIVO+IPI more than chemo and resulted in a higher ICER.

The fraction of progressors assumed to incur subsequent treatment costs is in the base case estimated by the fraction of the ITT population to have received a subsequent systemic therapy, per arm in CheckMate 227. A scenario analysis where the fraction is derived with a denominator of the number of patients observed to survive at least 1 day after progression results in higher post-progression costs and decreases the ICER.