This economic evaluation study was based on a literature review and modeling techniques, and it was deemed exempt from institutional review board approval because no real human participants were involved. The study was conducted according to the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) reporting guideline ( Supplemental Table 1 ) (12).

The target patient population was as same as that from the ASCENT trial. Included patients were adults (≥18 years of age) who had mTNBC that was relapsed or refractory to two or more previous standard chemotherapy regimens for unresectable, locally advanced, or metastatic disease. Patients with brain metastases were excluded. The median age of the patients was 56 years (8).

Included patients received either SG (10 mg per kilogram of body weight intravenously on days 1 and 8 of each 21-day cycle) or single-agent chemotherapy including eribulin (1.4 mg per square meter of body-surface area, intravenously on days 1 and 8 of a 21-day cycle, vinorelbine (25 mg per square meter intravenously on day 1 weekly), capecitabine (1000 to 1250 mg per square meter orally twice daily on days 1 to 14 of a 21-day cycle), or gemcitabine (800 to 1200 mg per square meter intravenously on days 1, 8, and 15 of a 28-day cycle). Chemotherapy was determined by the physician before randomization from one of these four single-agent treatments. Treatment would continue until disease progression or death.

A partitioned survival (PS) model was constructed using TreeAge Pro 2020 (TreeAge, Williamstown, MA) to compare costs and clinical outcomes associated with SG vs chemotherapy for treatment of mTNBC. This model is frequently used in oncology modeling. In the PS model, the proportion of patients in different health states at different time points was derived from PFS (Progression-free survival) and OS (Overall survival) curves directly (13). Three different health states were included, which were progression-free disease state (PFD), progressed disease state (PD), and death. The proportion of patients in the PFD was obtained directly from the PFS curve while the proportion of patients in the death state was derived by 1 minus the OS curve at each time point. With regard to the PD, its proportion was derived by calculating the difference between the OS and the PFS curve at each time point. The time horizon in our model was 5 years in which 99.5% of patients would be dead in both treatment arms. The cycle length was set at 4 weeks as gemcitabine used 4 weeks as a cycle. This study was conducted from the Chinese healthcare system and US healthcare payer perspective. Life-years, quality-adjusted life-years (QALYs), overall costs, and incremental cost-effectiveness ratios (ICERs) between the treatments were measured. A 5% discount rate per year was applied for both cost and effectiveness in China (14) and a 3% in the US (15). The willingness-to-pay (WTP) threshold of Chinese Renminbi ¥72447 to ¥217341 per QALY gained (3× gross domestic product (GDP) per capita) was used in China (16). And for the US, the WTP threshold was US dollars $100000 to $150 000 per QALY gained (17).

Markov model is also widely used in the economic evaluation of drugs. However, there is no consensus regarding whether the Markov model is better or not when compared with the PS model. Both the Markov and PS models are recommended to assess model structure uncertainty (18). Therefore, we also developed a Markov model to validate our results. The Markov model was built with TreeAge Pro 2020. This model contained three mutually exclusive health states, including PFD, PD, and death. All patients entered the model from the PFD. From this state, they could then either stayed in the PFD, progress to the PD, or die. Patients in the PD could either remain in the same state or transition to death. Patients were assumed to receive SG or single-agent chemotherapy when they were in the PFD and discontinue when they transitioned to the PD or death. Time-dependent transition probabilities were used in this three-state Markov model. Calculation of transition probabilities based on the fitted PFS and OS models has been described in detail by Rui et al. (19). Transition probabilities from the PFD to death were assumed to be the same as the natural death rate, and the age-specific and sex-specific death rates were obtained from the life tables for China (20) and the US (21). The Markov model used the same cycle length, time horizon, costs, and utility values as the partitioned survival model.

PFS and OS curves for patients with SG or chemotherapy were modeled based on the results of the ASCENT trial (10) according to the standard statistical analyses described by Guyot et al. (22) and Baio et al. (23). The GetData Graph Digitizer (Version 2.26; http://getdata-graph-digitizer.com/) was used to gather the data points from the PFS and OS curves. These data points were then used to fit the following survival functions including gompertz, exponential, gamma, genf, gengamma, weibull, weibullPH, loglogistic, and lognormal. The function with the best fit was determined by the Akaike information criterion, Bayesian information criterion (24), and graphical validation ( Supplemental Figures 1-4 ). We determined that the weibull and loglogistic model were the most reasonable functions for extrapolating OS and PFS in the SG arm and loglogistic model was the best for that in the chemotherapy arm ( Supplemental Table 2 ). The parameters of fitting models are provided in Table 1 . The comparison between the reconstructed Kaplan-Meier curves from the ASCENT trial and parametric fitting curves is presented in Figure 1 .

The costs were estimated from the Chinese and US perspective and only direct medical costs were considered, including costs for drugs, drug administration, follow-up, severe adverse events (AEs) management, best supportive care, and end-of-life care. The average body weight or body surface area of the Chinese and US women were used to calculate the dosage of drugs because women accounted for more than 99% percent of all participants in the ASCENT trial (10). Since SG is currently not available on the mainland Chinese market, we used the Hongkong price (DrugsHK, www.drugshk.com) to estimate the likely price as many mainlanders would go to Hongkong for drugs due to the lower prices and easy accessibility. The price of SG in the US is from the literature (11). According to the ASCENT trial, 54% of patients used eribulin, 20% used vinorelbine, 13% used capecitabine, and 12% used gemcitabine for the single-agent chemotherapy. We used these proportions to estimate the costs of chemotherapy. The price of chemotherapy in China was based on the median price from the widely used Chinese Drug Price Database (Tuling, www.315jiage.cn). Moreover, capecitabine and gemcitabine have entered the centralized drugs procurement program list in China and their mean prices were obtained from the official medical procurement network (Sunshine Medical Procurement All-in-one, http://www.smpaa.cn/). Chemotherapy costs in the US were obtained from the Centers for Medicare & Medicaid Services (35). Costs of drug administration, follow-up, severe AEs management, best supportive care, and end-of-life care were derived from previously published studies ( Table 1 ). The composition of follow-up costs included outpatient visit, laboratory evaluation, and imaging examination (29–31). According to the ASCENT trial, the computed tomography or magnetic resonance imaging examination was performed every 6 weeks for 36 weeks, then every 9 weeks thereafter, until disease progression leading to treatment discontinuation (10). We assigned follow-up costs to different cycles according to this follow-up pattern. Patients were assumed to have continuous best supportive care when diseases progressed and they also received end-of-life care before death. All costs were converted to 2020 values according to the local Consumer Price Index (44, 45).

Utility values of different health states were not reported in the ASCENT trial. We obtained the utility values from a recent similar cost-effectiveness analysis conducted by Wu et al. in which PFD, PD, and death state were assigned a value of 0.85, 0.578, and 0, respectively (43). According to Wu et al., the utility values in non-TNBC and TNBC were comparable as the quality of life was mainly affected by cancer stage but not the hormone status (43, 46). Then they estimated the utility values for PFD and PD based on values in non-TNBC (40, 47). Patients with severe AEs were assumed to have a disutility of 0.28 and all these AEs were assumed to have been incurred in the first cycle (43).

To assess the robustness of the model, one-way sensitivity analyses were performed on all parameters. The range of each parameter used in the one-way sensitivity analyses was based on either the reported 95% confidence interval (CI) in the referenced literature or a ± 25% change from the base case value. Probabilistic sensitivity analysis with a Monte Carlo simulation (10 000 iterations) was performed by simultaneously sampling the key model parameters from the prespecified distributions. All the costs were assigned with a gamma distribution and probability, proportion, and utilities were assigned with a beta distribution. A cost-effectiveness acceptability curve based on the results from 10000 iterations was created to evaluate the likelihood the SG would be considered cost-effective at different WTP thresholds.

We also performed the subgroup analyses to investigate the uncertainty of economic outcomes caused by the subgroups reported in the ASCENT trial. The ICER was calculated for each subgroup by using the reported subgroup-specific HRs for OS and PFS. The input data were assumed to be the same for all subgroups except for the HRs for OS and PFS. Details about subgroup analyses were described in Supplemental Material.

Table 2 shows the results of the base case analysis. From the Chinese perspective, the PS model predicted that SG yielded an additional 0.35 QALY at an additional cost of ¥2257842. The ICER was ¥6375856 ($924037)/QALY. From the US perspective, SG yielded the same additional QALY at an extra cost of $175393 and the ICER was $494479/QALY. For the Markov model, the ICER was ¥6407626 ($928641)/QALY in China and $507416/QALY in the US. Results from the PS model and Markov model were quite similar.

In China, when SG costs 30%, 20%, and 10% of its current price, the ICERs were ¥1817599 ($263420)/QALY, ¥1166419 ($169046)/QALY, and ¥515239 ($74672)/QALY, respectively. In the US, when SG costs 30%, 20%, and 10% of the current price, the ICERs were $135887/QALY, $84660/QALY, and $33432/QALY, respectively.

The results of the one-way sensitivity analyses are presented in Tornado diagrams, showing the effect from uncertainty in parameter input on the base case of ICER. The results were sensitive to the cost of SG, the utility of PFD, and body weight. From the Chinese perspective ( Figure 2 ), when the cost of SG was reduced to 75% of the baseline, the ICER fell to ¥4747907 ($688102)/QALY. When the cost of SG was increased to 125% of the baseline, the ICER was ¥8003805 ($1159972)/QALY. Across the broad variation in the ranges for all input parameters, the ICERs were greater than the Chinese WTP threshold of ¥217341/QALY.

Similar results were obtained from the US perspective ( Figure 3 ), When the cost of SG was reduced to 75% of the baseline, the ICER fell to $366410/QALY. When the cost of SG was increased to 125% of the baseline, the ICER was $622548/QALY. The ICERs were all greater than the US WTP threshold of ¥150000/QALY when the input parameters varied in their ranges.

At the WTP threshold of ¥217341/QALY in China, the cost-effective probability of SG treatment was 0%. When the unit price of SG was 30%, 20%, and 10% of the current price, the probabilities of cost-effectiveness for SG treatment were still 0%. When the cost was reduced to ¥10.44/mg (5.4% of the current price), the probability of cost-effectiveness of SG treatment increased to 50%. From the US perspective, with the same strategy of price reduction, the cost-effective probabilities of SG treatment were 69.3%, 98.4%, and 100% ( Figure 4 ). When the cost was reduced to $3.65/mg (32.6% of the current price), the probability of cost-effectiveness of SG treatment would increase to 50%.

ICER range of different subgroups obtained by varying the HRs for OS and PFS was presented in Supplemental Table 3 . In China, the lowest ICER among different subgroups was ¥3478680 ($504157)/QALY, which was above the WTP threshold of ¥217341/QALY. In the US, the lowest ICER among different subgroups was $366967/QALY and it was above the WTP threshold of ¥150000/QALY.