The target population of the study was patients with advanced or metastatic esophageal squamous cell carcinoma who previously failed to receive first-line chemotherapy. The patients were assigned to receive either camrelizumab or chemotherapy (docetaxel or irinotecan). A partitioned survival model was established to reflect the disease progression. The model included three states: progression-free disease (PFD), progressive disease (PD), and death. The three states are mutually exclusive. All patients were assumed to enter the model in the PFD state, and that they could maintain their designated health state or develop into another health state in each cycle (Figure 1). The relative 5-year survival rate is 8% or less for patients diagnosed with metastatic disease; thus, the time horizon of the model was set to 10 years (ASCO, 2020; Cancer Information Service, 2020). The model period was set to 1 month to facilitate model operation and parameter calculation. The main results of the model output were total cost, incremental cost-effectiveness ratio (ICER), and quality-adjusted life years (QALYs). ICER refers to the additional cost required for each additional QALY. Cost and utility were discounted at a rate of 5% (Liu, 2020). All costs were converted to USD, with an average RMB exchange rate of $1 to 6.8974 Yuan for the full year of 2020 (National Bureau of statistics of China, 2020). In addition, 1–3 times the national per capita GDP in 2020 ($10,503.52) was used as the willingness-to-pay (WTP) threshold (World Health Organization, 2011; Liu, 2020; National Bureau of statistics of China, 2020). The TreeAge Pro 2020 software package was used to build the model and conduct statistical analysis.

The clinical efficacy and safety data of the second-line treatment for advanced or metastatic esophageal squamous cell carcinoma were mainly from the ESCORT clinical trial (Huang et al., 2020). The ESCORT study is a randomized, open, multi-center phase III clinical study. It is the world;s first and largest clinical study of camrelizumab in Chinese patients with advanced/metastatic esophageal squamous cell carcinoma who previously failed first-line chemotherapy. A total of 228 patients received camrelizumab monotherapy, and 220 patients received chemotherapy (docetaxel or irinotecan). The eligible patients had histological or cytological diagnosis of esophageal squamous cell carcinoma; advanced, recurrent, or distant metastatic disease; and showed progression on previous first-line chemotherapy. The main exclusion criteria included CNS metastases and a history of anti-PD-1 or anti-PD-L1 therapy (Huang et al., 2020). GetData Graph Digitizer 2.26 (http://www.getdata-graph-digitizer.com) was used to obtain points on the survival rates of the Kaplan–Meier curves. Following Hoyle et al.(Hoyle and Henley, 2011), R software was used to reconstruct the individual data, which were then fitted by exponential, gamma, gompertz, Weibull, loglogistic and lognormal distribution. The best fitting distribution was selected by visual inspection, the lowest value of the Akaike information criterion (AIC), and Bayesian information criterion (Zhang et al., 2020a) (Supplementary Table S1). The log-logistic distribution function was used to simulate the PFS and OS curves of the two schemes. We performed internal and external model validations (Goldstein et al., 2015). The internal validation showed that the PFS and OS curves closely approximated those presented in the clinical trials (Supplementary Figures S1, S2). In the external validation, we compared the survival curves used in this study with those for the same treatment in other published studies. PFS and OS curves were compared with those from ORIENT-2 study (Xu et al., 2020), as shows in Supplementary Figures S3, S4. The survival function of log-logistic distribution at time t was S(t) = 1/(1+λt^γ) to calculate the scale parameter λ and the shape parameter γ (Ishak et al., 2013; Diaby et al., 2014). The Weibull distribution as the next best fitting function was also used to project lifetime survival curves in our model. The key clinical inputs are given in Table 1.

The study only considered direct medical costs, including drug acquisition, follow-up, best supportive treatment, and severe adverse event (SAE) management costs. In accordance with the ESCORT clinical research and guidelines (CSCO, 2020; Huang et al., 2020), 200 mg carrelizumab was administered intravenously on the first day of every 2 weeks, 75 mg/m² docetaxel was provided on the first day of every 3 weeks, and 180 mg/m² of irinotecan was administered intravenously on the first day every 2 weeks. Treatment was continued until disease progression or unacceptable toxicity. The cost effectiveness of the two scenarios was discussed to avoid the effect of the course of drugs on the results. In the first scenario, camrelizumab was assumed to be used for six (IQR 4–13) courses, docetaxel for three (2–3) courses, and irinotecan for four (2–5) courses in accordance with the results of ESCORT (Huang et al., 2020). A shorter time horizon (3, 5 and 7 years) was also considered in this scenario. In the second scenario, both groups continued treatment until the disease progressed. The proportion of patients receiving specific chemotherapy regimens was not defined in the clinical trials. The model assumed that the patients had equal opportunities to receive docetaxel and irinotecan. The average body surface area of the patients in the model was 1.72 m² (1.5–1.9 m²) (Zeng et al., 2013). After the failure of second-line treatment, the best third-line treatment was not clear, and the specific scheme was not shown in the ESCORT study. Therefore, the best support treatment was regarded as the treatment after progression.

The cost of camrelizumab was derived from the negotiated price of China’s national medical insurance (Nation Healthcare Security Administration, 2020). The cost of docetaxel and irinotecan was the median of the bidding price of drugs in different provinces (YaoZH, 2020). Only the SAE of grade ≥3 was considered (Guy et al., 2019; Zhang et al., 2020b). The incidence rate of anemia in the camrelizumab group was 3%, while the incidence rates of anemia, decreased neutrophil count, and vomiting in the chemotherapy group were 5.0, 15.0, and 5%, respectively (Huang et al., 2020). Other costs are shown in Table 1.

The utility value represents the health-related quality of life for each health state. The ESCORT trail did not involve health utility. Thus, the utility in the model was obtained from other public literature (Tan et al., 2018; Zhang et al., 2020b; National Institute for Health and Care Excellence, 2021), utility values for the PFD and PD health states were taken from EQ-5D data from a global, randomised, placebo-controlled, double-blind, phase 3 study, which recruited adults with advanced gastric cancer or gastro–oesophageal junction adenocarcinoma. The utility of PFD in the two groups was assumed to be consistent, but SAE (grade ≥3) could affect the utility. After disease progression, the utility of all patients in PD state was 0.581 (Zhang et al., 2020b; National Institute for Health and Care Excellence, 2021). All utility values are shown in Table 1.

One-way sensitivity analysis was performed to determine the influence of different parameters on ICER when changing within a certain range. The current price of camrelizumab fluctuated by 50% downward as the value range. The variation range of other parameters was the 95% confidence interval or the base value of the parameter ±25%. The discount rate was 0–8% (Liu, 2020). The results were presented in the form of tornado diagram. The horizontal axis of the cyclone graph represented the influence range of each parameter on ICER, and the vertical axis represented the parameter name. The degree of influence of the factors that have an influence on the evaluation result decreased from top to bottom.

In the probability sensitivity analysis, the parameters were set as random variables with specific distribution, and 10,000 Monte Carlo simulation was used to run the model to evaluate the overall robustness of the research results. The utility and the transition probability parameter were assumed to conform to the β distribution, and the cost parameter was assumed to conform to the γ distribution (Briggs et al., 2012). The results were represented by scatter plots and cost-acceptance curves.

Compared with the chemotherapy group, camrelizumab group showed an incremental cost of $1,439.64. The incremental effectiveness was 0.36 QALY and the ICER was $3,999.00/QALY in the first scenario. When both groups continued treatment until the disease progressed, the incremental cost of the camrelizumab group was $2,319.44 and the ICER was $6,442.89/QALY. In both scenarios, the ICER was far less than the WTP threshold of one time the GDP ($10503.52/QALY), that is, the camrelizumab group had an absolute cost-effective advantage. The results of basic-case analysis are shown in Table 2. The results of scenario analysis on a shorter time horizon are shown in Supplementary Table S2.

The results of one-way sensitivity analysis are shown in Figure 2. The main factors with a great effect on ICER were the cost of docetaxel, irinotecan, best supportive treatment, and camrelizumab. The ICER value of the model did not exceed the threshold of one time per capita GDP with the change in all uncertainty parameters, basically consistent with the conclusion of basic-case analysis.

The results of probabilistic sensitivity analysis are shown in Supplementary Figure S3. All the scatter points in the scatter plot were below the threshold, Supplementary Figure S3 indicating that camrelizumab was more cost-effective than chemotherapy in all cases. The cost-effectiveness acceptance curve is showed in Figure 3. When the WTP threshold was $3,151.06, the probability of cost-effectiveness advantage of camrelizumab was 14.4%. When the WTP threshold was $6,302.11, the probability of cost-effectiveness advantage was 92.5%. With the increase in the threshold, the possibility of camrelizumab to cost-effective increased; when the WTP threshold was $10,503.52, the probability of cost-effectiveness advantage was 100%.