Overview of model
A partitioned survival model was developed in Microsoft Excel to evaluate the cost-effectiveness of atezolizumab plus nab-paclitaxel compared with nab-paclitaxel alone for first-line treatment of metastatic or unresectable locally advanced, PDL1-positive TNBC. In a partitioned survival model, the proportion of patients in each health state at each time point was derived from the OS and PFS curves to inform the benefits accrued and costs incurred over the time horizon of the model. Three health states were considered, namely progression-free (PF), progressed disease (PD) and death (Figure 1).
The model followed patients weekly (with half cycle corrections) over a time horizon of 5 years. The time horizon was considered sufficient to capture the costs and outcomes over the lifespan of the majority of patients, based on a reported 5-year survival of less than 10% observed in women with Stage IV triple negative breast cancer from the US Surveillance, Epidemiology and End Results (SEER) population-based cancer registry.8
Patients and intervention
Patients with previously untreated metastatic or unresectable locally advanced TNBC who have PDL1 expression on ≥1% of tumour-infiltrating immune cells (expressed as a percentage of tumour area) were assessed in this model. The pivotal phase III IMpassion130 trial assessed both the intention-to-treat and a priori planned PD-L1 positive subgroup of patients.5 The latter was being modelled in this analysis consistent with FDA-registered indication. The treatment dosing regimen was in line with the trial. Atezolizumab was administered intravenously at a fixed dose of 840 mg, on days 1 and 15 while nab-paclitaxel was given at a dose of 100 mg/m2 on days 1, 8, and 15 of every 28-day cycle. Nab-paclitaxel was combined with atezolizumab to potentially enhance the response to immune checkpoint inhibition. It was preferred over solvent-based paclitaxel because the glucocorticoid premedication administered alongside solvent-based paclitaxel had been hypothesized to reduce immunotherapy activity. All patients continued their respective regimens until progression or an unacceptable level of toxic effects occurred. After progression, a proportion of patients were treated with subsequent-lines regimen until palliative care or death. There was no cross-over option for patients in either treatment arm, i.e. there was no option for patients receiving placebo to cross-over to receive atezolizumab.
Cost-effectiveness analyses were conducted from the Singapore healthcare system perspective. The outcomes-of-interest were costs, life years (LYs), quality-adjusted life years (QALYs) and incremental cost-effectiveness ratio (ICER). Costs and QALYs were discounted at 3% per annum.
Clinical efficacy data
Clinical efficacy data were derived from the pivotal phase III IMpassion130 trial (first interim analysis).5 The published Kaplan Meier (KM) OS and PFS curves were digitized using a validated graphical digitizer (WebPlotDigitizer version 4.2; Ankit Rohatgi, CA, USA).9 Then, individual patient time-to-event data were reconstructed from the digitized KM curves using an algorithm developed by Guyot et al.10 Parametric survival curve fitting was subsequently performed on the data to model the outcomes beyond the trial duration. The fitting was performed separately on data from the two treatment arms to avoid proportional hazards assumption. Candidate functions for the parametric extrapolation were Weibull, exponential, Gompertz, log-normal, log-logistic and generalized gamma distributions. The Weibull and exponential distributions were found to best fit the OS and PFS data, respectively, based on the assessments of clinical plausibility by local oncologist and goodness-of-fit given by the Akaike Information Criterion (AIC) value.
Internal validation of the model was conducted by comparing the modelled clinical outcomes with empirical trial data in terms of median OS and PFS and 2-year OS. Additionally, as the updated OS results from the second interim analysis of the trial were published at the time of completion of the model,11, 12 these results were also used for internal validation of the model.
At each cycle of the model, data were processed to ensure that (1) the modelled death rate was higher than or at least equal to the general population mortality, based on Singapore life table,13 and (2) the proportion of patients who had disease progression or died, given by the PFS curve, was smaller than or at least equal to the proportion who died, given by the OS curve.
Health state utilities are numerical values that represent an individual’s preferences for different health-related outcomes, ranging from 0 (representing a state of death) to 1 (representing a state of perfect health). These values are multiplied by the time spent in the health state (i.e. number of LYs) to produce the outcome of QALYs in the model.
Utilities of 0.715 and 0.443 were used for the PF and PD states, respectively, based on a primary utility study by Lloyd et al.14 The choice of this study was consistent with prior economic analyses.15-18 Disutilities due to adverse events (AEs) were not included in the model as the reported differences in incidences of AEs between the two treatment groups were not considered significant enough to result in differences in quality of life.
Direct healthcare costs including costs of drugs, intravenous drug administration, doctor consultation visits, blood tests, scans and palliative care were considered in the analysis (Table 1). An average body surface area (BSA) of 1.6m2, estimated by local oncologist, was assumed in drug dosing calculations. The management of drug-related AEs was not expected to result in significant differences in resource consumption between the two treatment arms, based on clinical opinion. Therefore, AE-related costs were not included in the model.
Sensitivity and scenario analyses
One-way sensitivity analyses were performed to explore the impact of uncertain model parameters on the ICER. The health state utilities and cost of atezolizumab were varied arbitrarily by ± 20%. Discount rates were varied from 0% to 5% for both costs and outcomes while the time horizon was varied from 3 to 10 years.
Probabilistic sensitivity analysis was performed using 15,000 Monte Carlo simulations to further assess the robustness of the results. Uncertainty in each parameter was represented by an appropriate probability distribution that corresponded with the nature of the variable. Utility values were assumed to follow the beta-distribution, while parameter values of survival functions for OS and PFS were sampled from the multivariate normal distribution using the Cholesky decomposition matrix of the Weibull and exponential distributions, respectively. A cost-effectiveness acceptability curve was generated showing the probability of each intervention being cost-effective over a range of cost-effectiveness thresholds.
Scenario analyses were conducted to probe the effects of uncertainty in the assumptions about treatment costs and utilities. Extreme scenarios with highly favourable assumptions (utilities of 1 and substantial price reduction of atezolizumab and nab-paclitaxel) were tested to assess the lower limit of the ICER. Costs based on local treatment options (for instance, alternative use of subsequent-lines treatments and limited use of nab-paclitaxel for 6 months) were also evaluated. To explore the possibility of cure, the presence or absence of a cure fraction (a proportion of patients who were cured of disease) in the arm treated with atezolizumab plus nab-paclitaxel was tested by fitting a mixture-cure model described by Lambert et al.19