A Markov partitioned survival model with three patient health states—PFS, disease progression (PD), and death—was constructed using Treeage 2025 to evaluate the cost-effectiveness of TLE-CHM versus PLB-CHM as a first-line regimen for HER2-negative advanced G/GEJ adenocarcinoma (Figure 1). It is important to note that partitioned survival models can directly use survival curves from clinical trials (e.g., OS and PFS) to partition health states. This makes them especially suitable for scenarios with limited data or where transition probabilities are hard to obtain, such as in advanced cancer research. Health states in the model are mutually exclusive. As patients progress, they either remain in their current state or transition to another without the possibility of reverting. All patients enter the model in the PFS state. The model used a cycle length of 3 weeks, with a total simulation duration of 350 cycles, which is equivalent to approximately 20 years. This duration was chosen so that by the end of the simulation, 99% of patients in both treatment groups would have died. The transition probability from PFS to death is represented by China’s background mortality rates (Compiled by the National Bureau of Statistics of China, 2024). Outcomes measured include total costs, quality-adjusted life years (QALYs), and incremental cost-effectiveness ratios (ICERs). Where ICER represents the ratio of increased cost to increased QALY for TLE-CHM compared to PLB-CHM. Based on China’s policies regarding medical insurance reimbursement and national drug price negotiations, we followed the recommendations of Cai et al. (2022) and set China’s 2023 per capita GDP multiplied by 1.5 ($19,067/QALY) as the willingness-to-pay (WTP) threshold. Following the Guidelines for Pharmacoeconomic Evaluation in China, a willingness-to-pay (WTP) threshold of three times China’s per capita GDP in 2023 ($38,133 per QALY) was applied (Yue et al., 2021). An ICER below this preset threshold indicates that the TLE-CHM regimen is cost-effective compared to PLB-CHM and vice versa.

Patient and clinical treatment information for this study were sourced from the RATIONALE-305 trial, a randomized, double-blind, phase 3 clinical trial (Qiu et al., 2024). Conducted across 146 medical centers in Asia, Europe, and the United States, the trial enrolled patients who were ≥18 years old, HER2-negative, had locally advanced unresectable or metastatic G/GEJ adenocarcinoma, had not received prior systemic anticancer treatment, and regardless of PD-L1 expression status. Ultimately, a total of 997 patients were enrolled in the RATIONALE-305 trial, of which 748 patients (75%) were from Asia. Specifically, 499 patients were from China, 17 were from Chinese Taiwan, 131 were from South Korea, and 101 were from Japan. Patients were randomly assigned to either the TLE-CHM or PLB-CHM group. The TLE-CHM group received tislelizumab with an investigator-selected chemotherapy regimen every 3 weeks, while the PLB-CHM group received a placebo with chemotherapy every 3 weeks. The investigator-selected chemotherapy regimens were capecitabine plus oxaliplatin or 5-fluorouracil plus cisplatin. Specifically, tislelizumab 200 mg on day 1, capecitabine 1000 mg/m² twice daily on days 1–14, oxaliplatin 130 mg/m² on day 1 or 5-fluorouracil 800 mg/m² on days 1–5, and cisplatin 80 mg/m² on day 1. Tislelizumab and capecitabine were continued until PD or intolerable toxicity, with the other drugs administered for up to six cycles. The median duration of drug treatment is shown in Supplementary Table B. Since RATIONALE-305 failed to provide detailed treatment data after the patients progressed, the best supportive care was considered as an intervention after the patients’ PD. Based on the RATIONALE-305 trial (Qiu et al., 2024) and the Chinese Gastric Cancer Treatment Guidelines (2022 edition) (National Health and Safety Commission, 2022), after PD, it is assumed that 50% of patients in the TLE-CHM group and 57% in the PLB-CHM group receive docetaxel (75 mg/m² on day 1) as second-line treatment, while the remaining patients receive the best supportive care.

To obtain the probability of patients transitioning between health states, we digitized the OS and PFS survival curves for the RATIONALE-305 trial using GetData Graph Digitizer (version 1.2) software. Subsequently, we reconstructed individual patient data using R software following Guyot et al. (2012) and fitted them using a range of survival functions including exponential, gamma, gen. F, gen. Gamma, Gompertz, Weibull, log-logistic, and log-normal (Kashiwa and Matsushita, 2023). The selection of the optimal survival function was based on the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), i.e., the lower the AIC and BIC values, the better the fit (Supplementary Table C) (Ishak et al., 2013; Williams et al., 2017). Ultimately, we selected the log-logistic survival function (S(t)= (1+(λt)^γ)⁻¹; S: survival probability, t: time cycle, λ: scale parameter, and γ: shape parameter) to fit the PFS and OS survival curves for TLE-CHM and PLB-CHM (Table 1, Supplementary Figure A). Our simulation’s median PFS and OS closely match the RATIONALE-305 trial data. Specifically, the model-predicted median PFS for the TLE-CHM group is 7.7 months and 6.3 months for the PLB-CHM group, similar to the trial-reported 6.9 months and 6.2 months, respectively. For median OS, the model predicts 15.3 months for the TLE–CHM group and 13.0 months for the PLB-CHM group, near the trial’s reported 15.0 months and 12.9 months (Supplementary Table D). Additionally, external validation is crucial for assessing the generalizability of survival analysis models. By using external data sources—such as other clinical trials, long-term registries, or real-world studies—to verify model predictions, we can ensure their accuracy and clinical validity, thereby reducing decision-making risks associated with censored data or sample bias. In this study, to validate the extrapolation of the OS curve for PLB-CHM, we referenced long-term survival data from the CheckMate 649 Trial (Janjigian et al., 2024), which evaluated nivolumab plus chemotherapy in patients with advanced G/GEJ adenocarcinomas. The three-, four-, and 5-year survival rates for patients in the chemotherapy arm of CheckMate 649 were 19%, 10%, and 0.5%, respectively, which were broadly consistent with our extrapolated survival curve (23%, 12%, and 0.5%).

The study included only direct medical costs, encompassing expenses for medications, tests, routine follow-up, the best supportive care, management of serious adverse reactions (≥grade 3) with an incidence of ≥5%, and end-of-life care (Table 1). Drug costs were based on national bidding prices (YAOZH, 2024), while other costs were sourced from published literature and adjusted to 2023 values using the China Bureau of Statistics medical price index (Compiled by National Bureau of Statistics of China, 2024). The cost of managing adverse reactions was calculated by multiplying the cost of managing each adverse reaction by its probability of occurrence and summing these values across all adverse reactions, assuming all adverse reactions occur in the first treatment cycle. All costs were expressed in U.S. dollars at the average 2023 exchange rate (1 USD = 7.05 RMB). To facilitate the calculation of the patient’s medication dosage, we assumed the patient’s body surface area is 1.72 m² (Liu et al., 2022).

Utility value is an indicator for assessing patients’ social function and overall health status, including physical, psychological, and disease-related aspects. It is measured on a scale from 0 to 1, where 0 represents death and 1 represents the best possible health, with patients’ health typically falling between these two extremes. Due to the lack of quality-of-life data in the RATIONALE-305 trial, utility values for PFS and PD were derived from published Chinese literature (Table 1). The model also considered the disutility of serious adverse reactions (≥grade 3) with an incidence of ≥5%. Specifically, the disutility of each adverse reaction is calculated by multiplying the given disutility by the adverse reaction incidence rate, and then these disutility values are subtracted from the total utility value. Both costs and utilities were discounted at an annual rate of 5% (Yue et al., 2021).

To assess the robustness of the model results, a sensitivity analysis was conducted, including one-way and probabilistic sensitivity analyses. The impact of parameter variations on the model outcomes was assessed through the one-way sensitivity analysis. In the one-way sensitivity analysis, all parameters were adjusted within their reported 95% confidence intervals. For parameters without reported confidence intervals, fluctuations of ±20% from the baseline value were assumed, with the discount rate varying between 0% and 8% (Table 1). Some patients may reduce the drug dose due to intolerance during treatment. In our model, the relative intensity of the drug dose (the ratio of the actual dose intensity to the planned dose intensity) varies based on data from the RATIONALE-305 trial. The results are presented in a tornado diagram. For the probabilistic sensitivity analysis, 1,000 Monte Carlo simulations were performed based on a specific distribution, varying random and simultaneous preset parameters (Table 1). The results are depicted by a scatter plot and cost-effectiveness acceptability curve.

To assess the impact of PD-L1 tumor area positivity (TAP) scoring and race on model outcomes, we conducted an exploratory subgroup analysis targeting populations with PD-L1 TAP scores of <5% and ≥5%, as well as Asian and North American/European populations. This analysis aims to elucidate the potential influence of different scores and race backgrounds on treatment efficacy, providing more targeted insights for clinical decision-making. Due to the lack of sufficient survival data for subgroups, for subgroup survival extrapolation, we used the same survival function in all subgroups of the PLB-CHM group as in the overall population of the PLB-CHM group. Subgroup hazard ratio (HR) from the RATIONALE-305 trial were applied to calculate the ICERs for each subgroup, following the method described by Hoyle et al. (2010), and to assess the probability of cost-effectiveness acceptability. Specifically, the shape parameter in the TLE-CHM group was equivalent to that in the PLB-CHM group, and the scale parameter in the TLE-CHM group was equivalent to the scale parameter of the PLB-CHM group multiplied by the HR.

Two scenarios were analyzed for the overall population. In scenario 1, the model duration was varied to 5, 10, and 15 years to assess its impact on the results. Scenario 2 explored the economics of the two treatment options at different discount rates, using 3% and 8%. Scenario 3: Although the WTP threshold in this study was set at three times China’s per capita GDP, which is within the range recommended by the Guidelines for Pharmacoeconomic Evaluation in China (Yue et al., 2021) and commonly used in literature (Liu et al., 2025; Zhou et al., 2025; Zhu et al., 2025), Cai et al. (2022) argued that Chinese health insurance policymakers, with stronger bargaining power, tend to prefer lower thresholds. They suggested that 1.5 times the per capita GDP could serve as a reference threshold for health insurance decision-makers to minimize suboptimal decisions. Therefore, we adjusted the WTP threshold to 1.5 times the per capita GDP of China to analyze the cost-effectiveness of TLE-CHM. In scenario 3, we replace the log-logistic distribution with alternative distributions (e.g., exponential, log-normal, Weibull) to assess the robustness of model outcomes. In scenario 4, we address the dynamic uncertainty inherent in multi-state survival models by employing a Dirichlet distribution to model the probability distribution of survival states (including PFS, PD, and death) within each cycle. Robustness of the model outcomes under uncertainty is evaluated through 1,000 Monte Carlo simulations.

The study results are presented in terms of total costs, QALYs, and ICERs (Table 2). The TLE-CHM group achieved 1.53 QALYs at a cost of $23,484.39, while the PLB-CHM group achieved 1.14 QALYs at a cost of $12,123.52. The TLE-CHM group demonstrated an incremental effectiveness of 0.38 QALYs and an incremental cost of $11,361.27, resulting in an ICER of $29,608.51 per QALY gained. These findings indicate that TLE-CHM is not a cost-effective strategy for treating HER2-negative advanced G/GEJ adenocarcinoma compared to PLB-CHM, given China’s WTP threshold of $19,067 per QALY.

The results of the one-way sensitivity analysis, presented in a tornado diagram (Figure 2), indicate that the parameters with the most significant impact on the model are the PFS utility, the cost of tislelizumab, and the PD utility. However, even with variations in these parameters, the ICER consistently remains above the predetermined WTP thresholds. This stability suggests that changes in parameter values do not affect the model’s conclusions. The probabilistic sensitivity analysis results, displayed as a scatter plot (Figure 3) and cost-effectiveness acceptability curve (Figure 4), show that at a WTP threshold of $19,067 per QALY, the probability of TLE-CHM being cost-effective compared to PLB-CHM is only 0.8%.

The results of the subgroup analysis are presented in Table 2. In the subgroup with a PD-L1 TAP score ≥5%, compared to PLB-CHM, TLE-CHM increased QALYs by 0.41, with a cost of $12,425.96, resulting in an ICER of $27,581.54 per QALY gained, which is above the predefined WTP threshold, and the probability of TLE-CHM being cost-effective was 1.8%. Similarly, in the subgroup with a PD-L1 TAP score <5%, TLE-CHM only increased QALYs by 0.10, with an incremental cost of $7,030.11, leading to an ICER of $72,348.64 per QALY gained, which is substantially above the WTP threshold, with a cost-effectiveness probability of 0%. These results indicate that, compared to PLB-CHM, TLE-CHM is not cost-effective, regardless of the PD-L1 TAP score. Additionally, in the Asian population, TLE-CHM increased QALYs by 0.23 compared to PLB-CHM, with a cost of $9,181.50, resulting in an ICER of $39,818.77 per QALY gained, which is above the predefined WTP threshold. This suggests that TLE-CHM is not a cost-effective treatment option in the Asian population.

The scenario analysis results are presented in Table 3. In scenario 1, where the model duration varied to 5, 10, and 15 years, the ICER of TLE-CHM was $41,081.32/QALY, $33,248.53/QALY, and $30,863.74/QALY, respectively, compared to PLB-CHM. The ICER gradually decreases as the model duration increases. In scenario 2, adjusting the discount rate to 0.03 and 0.08 resulted in ICERs of $29,486.03/QALY and $29,788.05/QALY for TLE-CHM compared to PLB-CHM, respectively. In scenario 3, when the WTP threshold is adjusted to $19,067/QALY, TLE-CHM has only a 0.6% probability of being cost-effective compared to PLB-CHM. In scenario 3, when the survival distributions were adjusted to Weibull, log-normal, and exponential, the ICERs of TLE-CHM compared to PLB-CHM were $35,739.34/QALY, $29,234.05/QALY, and $37,433.23/QALY, respectively. All of these values were above the predefined WTP threshold. As shown in Supplementary Figure B, the results of scenario 4 indicate that when the Dirichlet distribution is employed to model the probability distribution of survival states within each cycle, the probability of TLE-CHM being cost-effective is 28.10%.