Cost-effectiveness analysis of radiotherapy versus surgery for esophageal squamous cell carcinoma in China: a Markov model study based on real-world data

Abstract

Background

Studies on cost-effectiveness analysis of esophageal cancer are relatively common. However, there are few cost-effectiveness analyses that take a comparative approach between radiotherapy and surgery for esophageal squamous cell carcinoma (ESCC). A cost-effectiveness Markov model was constructed to explore the relationship between the benefits and costs of radiotherapy versus surgery for ESCC in China.

Methods

According to the observational data (patients’ medical records), 196 patients with ESCC enrolled for data were classified into surgery and radiotherapy two groups, both groups received preoperative and postoperative chemoimmunotherapy. A Markov model was constructed using TreeAge Pro Healthcare software to simulate the diseases progression after treatment. The model cost parameters were derived from the average treatment expenditures by real-world patients. The survival formula, fitted to real-world patient data in R, was used to calculate model transition probabilities.​ The utility value parameters were obtained by reviewing literature. Base-case, one-way deterministic sensitivity, and probabilistic sensitivity analyses were performed. The results were evaluated against willingness-to-pay (WTP) thresholds to determine the cost-effectiveness of treatments.

Results

Among 196 patients (131 males; 114 surgery group), median overall survival was 41.3 vs. 30.4 months and progression-free survival 28.0 vs. 20.6 months for surgery versus radiotherapy groups.​ Compared to radiotherapy, surgery had an incremental cost of 411,574.32 US dollars (USD), an incremental utility of 11.85 quality-adjusted life years (QALYs), and an incremental cost-effectiveness ratio of 34,744.52 USD/QALY, reaching 90% of the upper bound of WTP (three times the gross domestic product [GDP] per capita in China). One-way deterministic sensitivity analysis revealed that the costs of adverse event management and immunotherapy were key drivers affecting the incremental cost-effectiveness ratio. In probabilistic sensitivity analysis, at a WTP of 12,741.11 USD/QALY and 38,223.34 USD/QALY (1 to 3 times China’s per capita GDP), the probability of surgery being cost-effective was 10.1% and 62.5%, respectively. At a WTP of 33,080.09 USD/QALY, the probability was equal.

Conclusions

In China, Surgery was marginally cost-effective compared to radiotherapy for ESCC.​​ WTP thresholds, adverse reactions costs and immunotherapy costs constituted the primary factors influencing health economic evaluation outcomes in ESCC surgery versus radiotherapy.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12913-026-14060-w.

Background

China has a high incidence of esophageal cancer, which poses a significant threat to public health [1]. Unlike Western countries, esophageal cancer in China is predominantly squamous cell carcinoma [2], and the complexity of cancer treatments​​ lies in the fact that a one-size-fits-all approach cannot be mechanically applied to every patient. Instead, therapeutic strategies must be tailored to the individual, based on their specific disease characteristics and tolerance to side effects. This involves selecting the appropriate surgical procedure; specific drugs and their dosages; radiotherapy targets and dosing; and the sequence of different modalities (e.g., chemotherapy followed by surgery versus surgery followed by chemotherapy).​ Although the various treatment methods have shown promising efficacy, questions remain regarding how to coordinate these methods and determine the optimal treatment duration. In addition to these treatment concerns, patients must consider treatment costs alongside efficacy. Improved survival is usually associated with increased costs, which is an unavoidable dilemma. Physicians, medical institutions, and medical decision-making departments need to pay special attention to the cost-effectiveness of treatments. This does not represent the first cost-effectiveness study on ESCC, Research on cost-effectiveness analysis for esophageal cancer predominantly focuses on pharmaceutical interventions [3–6], with relatively less attention given to non-pharmacological treatment approaches such as surgery and radiotherapy. Although our study is not the first cost-effectiveness analysis of esophageal squamous cell carcinoma (ESCC) in China, it analyzes ESCC from a new perspective—comparing surgery with radiotherapy.

In China, an individual patient’s medical expenses are covered by two or three sources: the government basic medical insurance covering approximately 95% of the population, out-of-pocket payments by the patient, and commercial health insurance purchased by a minority. Consequently, both government/commercial medical insurance decision-making departments and patients themselves need to carefully consider the cost-effectiveness of treatments.​.

Therefore, the principles of pharmacoeconomics and single-center retrospective case data were used to construct a cost-effectiveness model based on real-world treatment outcomes and costs. In this study, we aimed to explore the relationship between the benefits and costs of surgery and radiotherapy for esophageal squamous cell carcinoma (ESCC), thereby providing a basis for clinical decision-making.

Methods

Experimental design

Constructed​​ a standard Markov model that ​​included​​ two groups with distinct radical strategies: One group ​​underwent​​ surgery while the other group ​​received​​ radiotherapy. Both groups ​​received​​ two cycles of preoperative neoadjuvant chemoimmunotherapy before and two cycles of postoperative consolidation chemoimmunotherapy following surgery/radiotherapy. This Markov model ​​was iterated​​ in TreeAge to simulate the disease progression after treatment. Upon importing required parameters, the program ​​output​​ the cost-effectiveness analysis result of the comparison between the two groups. The flowchart ​​was​​ shown in Fig. 1.

Fig. 1.

Experimental design flowchart

To obtain parameters for model simulation from real-world patients, medical records of ESCC patients treated at the Department of Thoracic Surgery and the Department of Radiotherapy of our center between January 2019 and February 2022 were collected and analyzed based on observational data. In alignment with the model settings, the patients were classified into two groups based on the radical treatment method: the surgery group, which received preoperative neoadjuvant chemoimmunotherapy + surgery + postoperative consolidation chemoimmunotherapy; and the radiotherapy group, which received pre-radiotherapy neoadjuvant chemoimmunotherapy + radiotherapy + post-radiotherapy consolidation chemoimmunotherapy.

It should be noted that this retrospective analysis used non-randomized data with simulated randomized treatment, which may bias the model. These limitations will be addressed in future research.​.

The inclusion criteria were as follows: (1) patients with pathologically confirmed ESCC; (2) completion of radical surgery or radical radiotherapy; (3) at least two cycles of neoadjuvant chemotherapy and two cycles of neoadjuvant immunotherapy before radical treatment; (4) at least two cycles of consolidation chemotherapy and two cycles of consolidation immunotherapy after radical treatment; and (5) complete data on treatment costs.

The exclusion criteria were as follows: (1) patients without a diagnosis of ESCC; (2) patients who did not undergo radical treatment; (3) patients who did not receive complete radical, neoadjuvant, or consolidation treatment; and [4] incomplete data on treatment costs.

The real-world survival data of patients were imported into R 4.41 (R Core Team, Vienna, Austria) to plot overall survival (OS) and progression-free survival (PFS) curves, followed by modeling and fitting with probability density distributions to obtain the formulas of survival functions and transition probability tables. These formulas can calculate the transition probabilities and counts of patients in a specified state at any given time. The Markov model was constructed in TreeAge Pro Healthcare 2022 R1.2 (TreeAge Software, LLC., Williamstown, Massachusetts, USA) to simulate the diseases progression after treatment. The calculated transition probabilities, along with the mean hospitalization costs of each major treatment item per patient and utilities, were imported into the TreeAge software for cost-effectiveness analysis.

Markov model

Model structure

The Markov model used in this study included three mutually exclusive health states: progression free survival (PFS), progressive disease (PD), and death. All patients started from the PFS state and then transitioned to one of these three progressive and irreversible states. The transition probabilities between these states were calculated using the method described below. The states in the model are illustrated in Fig. 2.

Fig. 2.

Schematic diagram of the Markov model structure

To simplify the model, the cycle length was assumed to be 1 month. Surviving patients automatically entered the next cycle at the end of each cycle, and the cycle terminated when 99% of the patients died (i.e., OS < 0.01).

The model generated cost outputs in units of US dollars (USD) and utility outputs in units of quality-adjusted life years (QALY) [7]. The discount rate was set at 5% for cost and utility outputs [8, 9]. The incremental cost-effectiveness ratio (ICER) was calculated using TreeAge software. To determine the cost-effectiveness of a treatment, the ICER was compared with a predefined willingness-to-pay (WTP) threshold set between one and three times the gross domestic product (GDP) per capita of China in 2022 (85,698–257,094 Chinese Yuan [CNY]/QALY) [9–11]. Based on the average exchange rate of 6.7261 CNY to 1 USD in 2022, the WTP threshold was 12,741.11–38,223.34 USD/QALY.

Treatment regimens in the model

To simplify the model, the simulated treatment regimens adopted in this study were as follows. Patients first received two cycles of first-line chemotherapy and two cycles of immune checkpoint inhibitor therapy (immunotherapy) and were then randomly assigned to the surgery or radiotherapy group for radical surgery or radical radiotherapy. After local treatment, patients received two cycles of first-line chemotherapy and two cycles of immunotherapy and then entered the Markov model. If the patient remained in the PFS state, immunotherapy was administered monthly until the patient entered the PD state. If the patient entered the PD state, second-line chemotherapy or best supportive care was administered monthly until death. Second-line chemotherapy options included, but were not limited to, the following regimens: Irinotecan plus S-1, docetaxel monotherapy, paclitaxel monotherapy, and irinotecan monotherapy, etc. These strategies are consistent with clinical practice guidelines.

Given that this was a retrospective study aiming to closely reflect real-world conditions, various surgery, radiotherapy, chemotherapy, and immunotherapy regimens were included. First-line chemotherapy, second-line chemotherapy, and best supportive care were considered without specifying the regimens used in this model.

General equations for transition probabilities

Let N_PFS(t) be the number of patients with PFS in cycle (t), then, N_PFS(t+1) is the number of patients who remain progression-free in the subsequent cycle. Let pFtF(t) be the probability of remaining in the PFS state at cycle t. Thus, we obtain the following formula:

1

Based on the definition of the transition probability from PFS to death, the direct transition from PFS to death was not caused by esophageal cancer, and the mortality rate was approximated as the natural mortality rate. Mortality rises with age. However, due to the unavailability of finer age-stratified natural mortality data and the inability to precisely calculate the natural mortality rate for this cohort, we had to adopt a simplified approach by using annual natural mortality rate in the calculations. In 2022, the annual natural mortality rate in China was 7.37‰ [11], and the monthly natural mortality rate was calculated to be approximately 0.0006, then:

2

According to the model assumptions, all patients with PFS had one of three outcomes: PFS, PD, and death, and the sum of their probabilities equals 1. Here, set pFtP(t) represents the probability of state PFS to PD at cycle (t), pFtF(t) represents the probability of state PFS to PFS at cycle (t), and pFtD represents the probability of state PFS to death at cycle (t), then:

3

In cycle (t), the number of patients in the PD state was calculated as the sum of the number of patients who transitioned from PFS to PD and the number of patients who remained in the PD state. Set N_PD(t+1) represents the number of patients with PD in cycle (t + 1), N_PFS→PD(t) represents the number of patients from PFS to PD in cycle (t), N_PD→PD(t) represents the number of patients remained PD in cycle (t), then:

4

Set N_PFS(t) represents the number of patients with PFS in cycle (t), pFtP_(t) represents the transition probability from PFS to PD in cycle (t), N_PD(t) represents the number of patients with PD in cycle (t), pPtP_(t) represents the transition probability from PFS to PFS in cycle (t), then:

5

6

Equations (5) and (6) are substituted into Eq. (4) to obtain the following:

7

Based on Eq. (7), we have:

8

Set N_OS(t) represents the number of overall survival in cycle (t), N_PFS(t) represents the number of progression-free survival in cycle (t), we have:

9

Equation (9) is substituted into Eq. (8) to obtain the following:

10

Since patients with PD transition to either PD or death, and the sum of these two probabilities equals 1, we have:

11

Since the number can be proportionally converted to proportion probability, the N_PFS(t) in Eqs. (1), (2), (3), (10) and (11) can be replaced by the PFS in cycle (t). Similarly, N_PFS(t+1), N_OS(t), and N_OS(t+1) can be replaced.

Fitting survival curves

Distributions and parameters of curve fitting functions

How to select survival distributions depends on the plausibility of the extrapolation. We considered the following parts of the plausibility: Considering the survival function is monotonically decreasing, therefore we adopted standard parametric models. Because DIC is more suitable for Bayesian models, while Markov models mostly choose AIC/BIC, therefore the evaluation of goodness-of-fit combines AIC/BIC with visual inspection. Based on the likelihood, distribution characteristics, AIC/BIC values, and visual inspection, the formulas and parameters of the survival functions were determined. These functions were used to calculate the time-dependent transition probability tables for the surgery and radiotherapy groups.

Multiple probability density distributions can be used to describe survival functions; the common distributions [12] are shown in Supplementary Material Attachment 2.

The transition probabilities between states as basic parameters were obtained using a code (shown in Supplementary Material Attachment 1) written in R based on the survival package [13, 14], survminer package [15], and survHE package [16]. The code was executed to generate the Akaike information criterion (AIC), Bayesian information criterion (BIC), and deviance information criterion (DIC). Considering their respective applicability conditions [17], AIC and BIC were selected in this study to evaluate the fitting performance, with a small AIC/BIC indicating high goodness of fit.

Time-dependent transition probability tables

Since OS and PFS from follow-up did not provide transition probabilities for all cycles and survival data beyond the follow-up period, it is necessary to fit and extrapolate these curves. This allows for the calculation of OS and PFS for cycles in the Markov model, as well as the transition probabilities of these cycles to generate a time-dependent transition probability table that covering various states in model. The calculations were conducted using R, and the results would be used for calculation in TreeAge.

Costs and utilities

The cost data were derived from the actual medical expenses of the enrolled patients. The utility data were collected from the literature. We combined keywords such as ‘ESCC’, ‘esophagus cancer’, ‘esophageal gastric junction cancer’, ‘cost-effectiveness’, and ‘utility’ to search databases including PubMed, ScienceDirect, Ovid, and CNKI, retrieving 53 relevant articles. Initially, we excluded 33 articles unrelated to esophageal cancer or esophageal-gastric junction cancer based on their titles and abstracts. For the remaining 20 articles, we conducted a full-text review and retained those reporting utility values for surgery/radiotherapy in esophageal cancer/esophageal-gastric junction cancer, summarized and synthesized these utility values. The detailed treatment costs and distributions, detailed utilities and distributions are shown in Table 1.

Table 1.

Costs, utilities and distributions

Item	Mean (USD)	Minimum (USD)	Maximum (USD)	Distribution	Source
Surgery group costs
Surgery costs	10,026.30	4,872.83	16,103.09	Gamma	Enrolled study patients
Immunotherapy costs	3,588.77	1,737.41	6,239.21	Gamma	Enrolled study patients
First-line chemotherapy costs	955.77	463.38	1,665.20	Gamma	Enrolled study patients
Second-line chemotherapy/best supportive care costs	805.58	313.19	1,515.02	Gamma	Enrolled study patients
Hospitalization costs	381.83	84.00	817.26	Gamma	Enrolled study patients
Laboratory test costs	356.56	65.42	2,441.53	Gamma	Enrolled study patients
Examination costs	789.75	508.47	1,714.96	Gamma	Enrolled study patients
Adverse event management costs	2,386.05	553.59	7,871.93	Gamma	Enrolled study patients
Radiotherapy group costs					Enrolled study patients
Radiotherapy costs	4,774.36	1,693.41	6,581.51	Gamma	Enrolled study patients
Immunotherapy costs	1,987.02	248.12	5,215.57	Gamma	Enrolled study patients
First-line chemotherapy costs	528.34	3.68	4,311.30	Gamma	Enrolled study patients
Second-line chemotherapy/best supportive care costs	439.86	14.16	4,190.27	Gamma	Enrolled study patients
Hospitalization costs	145.18	7.43	1,344.02	Gamma	Enrolled study patients
Laboratory test costs	236.85	3.57	1,599.22	Gamma	Enrolled study patients
Examination costs	252.27	2.97	1,575.71	Gamma	Enrolled study patients
Adverse event management costs	644.75	0.68	4,752.45	Gamma	Enrolled study patients
Death costs△	5,556.71	5,001.04	6,112.38	Gamma	[18]
Utilities
Radiotherapy△□	0.770	0.69	0.85	Beta	[19–21]
Surgery△□	0.300	0.27	0.33	Beta	[19–21]
Pre-treatment	0.820	0.75	0.89	Beta	[22]
Progression-free survival	0.797	0.64	0.96	Beta	[23, 24]
Progressive survival	0.577	0.46	0.69	Beta	[23, 24]
Death	0	0	0	--	--

Note: △ The maximum and minimum values were unavailable and were set at the mean ± 10%; □ Remained for approximately 1.5 months (equivalent to 1.5 cycle lengths in the model)

Sensitivity analyses

Sensitivity analyses, also known as uncertainty analyses, were performed by varying parameters within their ranges to avoid bias caused by the base-case analysis of deterministic mean values.

In one-way deterministic sensitivity analysis (DSA), the ICER was calculated using the upper and lower values of parameters to assess their impact on the ICER. The results are presented as a tornado diagram.

In probabilistic sensitivity analysis, second-order Monte Carlo simulations were performed using 1,000 iterations based on random sampling from the parameter distributions. The results are presented as an incremental cost-effectiveness scatter plot (ICESP) and a cost-effectiveness acceptability curve (CEAC). The parameters used for the sensitivity analyses are also listed in Table 1.

Statistical analysis

Frequencies, rates, and row × column variables were tested using Fisher’s exact test. The Kaplan–Meier method was used to plot survival curves, and the log-rank test was used to compare differences in survival between groups. A two-sided P value < 0.05 (P < 0.05) was considered statistically significant. All statistical analyses were performed using R.

Results

Patient characteristics

A retrospective analysis was conducted on patients with esophageal cancer who were treated at our hospital between January 2019 and February 2022. According to the inclusion and exclusion criteria, 196 patients were included in the study and classified into the surgery and radiotherapy groups based on the radical treatment method. The surgery and radiotherapy groups comprised 114 and 82 patients, respectively. Among these patients, 131 were male and 65 were female. Patients in the radiotherapy group did not undergo surgery and were staged according to the China 2009 staging system for patients with non-surgical esophageal cancer [25]. There were 2, 21, 63, 77, 19, and 14 patients diagnosed with stage tumor in situ (Tis), I, II, III, IVa, and undetermined esophageal cancer, respectively. All stage IVa patients were classified as locally advanced without evidence of distant metastasis. Patients in the surgical group who met the criteria for surgery following downstaging with neoadjuvant therapy. The detailed data are presented in Table 2. By Fisher’s exact test, there was no significant difference (all P-values are greater than 0.05) in the composition of sex, age, and TNM stage between the two groups of patients, indicating comparability.

Table 2.

Comparison of baseline characteristics

Item	Surgery	Radiotherapy	P-value
Sex
Male	77	54	0.878
Female	37	28
Age
< 60 years	33	26	0.7527
≥ 60 years	81	56
Stage
Tis	2	0	0.05922
I	12	9
II	34	29
III	41	36
IVa	12	7
NA	13	1

Follow-up analysis

As for the follow-up endpoint, the median follow-up time for all patients was 38.5 months, with 14 patients lost to follow-up, representing a loss to follow-up rate of 7.14%. The median OS was 41.3 (95% CI: 27.9–NA) vs. 30.4 (95% CI: 21.8–NA) months for the surgery and radiotherapy groups. The median PFS was 28.0 (95% CI: 19.7–NA) vs. 20.6 (95% CI: 15.1–NA) months in the surgery and radiotherapy groups. The OS and PFS curves are shown in Fig. 3. By log-rank test, both OS and PFS curves showed no statistically significant difference (OS: P = 0.12; PFS: P = 0.14) between the two groups.

Fig. 3.

Comparison of survival curves among patients treated with surgery and radiotherapy in the real world. (A) Overall survival. (B) Progression-free survival

Survival curve fitting

Using the method described in the Section “Distributions and parameters of curve fitting functions” of Method, calculations were performed in R, and the AIC and BIC values of various distributions were obtained in Supplementary Material Attachment 3 Table 1.

According to the principle of minimum AIC (Surgery OS: 489.4697, Surgery PFS: 577.4106, Radiotherapy OS: 260.0655, Radiotherapy PFS: 353.8332), likelihood, distribution characteristics, and visual inspection, a Lognormal distribution was selected to fit both the OS and PFS curves. The µ and σ were derived from R. As a result: surgery group: µ_OS = 3.643621, σ_OS = 0.888692, µ_PFS = 3.29926, σ_PFS = 1.0777. RT group: µ_OS = 3.38177, σ_OS = 0.911483, µ_PFS = 3.0264, σ_PFS = 1.16487.

We simulated overlaps between various distributions and the actual survival curve in R, where the Lognormal distribution exhibited the smallest AIC value and demonstrated the closest alignment with the actual survival curve.

Based on the formulas of the fitted curves, the OS and PFS were calculated in R. The comparison of curves between the Markov model and real-world is shown in Fig. 4.

Fig. 4.

Survival curves in the Markov model and the real-world

Transition probability tables

The transition probabilities between states were calculated by substituting the OS and PFS at each time point t into Eqs. (1, 2, 3, 10, 11). These transition probabilities are summarized in a transition probability table. In the model, ​​the progression rate for the surgery group during the initial Markov cycle​​ (simulating a one-month period, as detailed in Section “Model structure” of the Methods) ​​had been previously established as pFtD [1] = 1 − pFtF [1] − pFtP [1] (derived from Eq. 3) = 1 − 0.9933 − 0.0011 = 0.0056, with the same applying to the radiotherapy group.

In the first and second cycles, negative pPtP values were calculated, which are impossible in real-world scenarios. This may be explained by the curve-fitting deviations and the fact that the pPtP graph does not intersect the origin (see discussion for details). Therefore, negative values were treated as 0.

The data for the first 10 months are listed in Supplementary Material Attachment 3 Tables 2 and 3.

Table 3.

Base-case analysis

Strategy	Total cost (USD)	Total utility (QALYs)	Incremental cost (USD)	Incremental utility (QALYs)	ICER (USD/QALY)
Radiotherapy	233,619.63	57.20	--	--	--
Surgery	645,193.80	69.04	411,574.32	11.85	34,744.52

Note: ICER: Incremental cost-effectiveness ratio, QALY: Quality-adjusted life-year

Base-case analysis

The OS in the transition probability tables indicates that 99% of the patients were in the death state at cycle 245 of the radiotherapy group and cycle 303 of the surgery group. Therefore, the termination criteria for the Markov model were 245 cycles for the radiotherapy group and 303 cycles for the surgery group. The cost data, utility data, and transition probabilities were incorporated into TreeAge for model construction and cost-effectiveness analysis. The results of the base-case analysis are presented in Table 3.

The results suggest that compared to radiotherapy, surgery patients had an incremental cost of 411,574.32 USD, an incremental utility of 11.85 QALYs, and an ICER of 34,744.52 USD/QALY. The ICER was lower than, but reached 90% of, the three times WTP threshold of 38,223.34 USD/QALY. Surgery was marginally more cost-effective compared to radiotherapy.

One-way deterministic sensitivity analysis

The results of the one-way deterministic sensitivity analysis are presented in a tornado diagram in Fig. 5.

Fig. 5.

Tornado diagram of the one-way deterministic sensitivity analysis

The diagram shows that the top five factors influencing the ICER were costs of adverse event management in the surgical group, costs of adverse event management in the radiotherapy group, immunotherapy costs in the surgery group, immunotherapy costs in the radiotherapy group, and test costs in the surgery group. The costs of death, costs of first-line chemotherapy in both groups, costs of radical treatments in both groups, utilities of radical treatments in both groups, and utilities of therapies before radical treatments in both groups had little impact on the ICER. Other variables have mild to moderate effects on the ICER. The results of one-way deterministic sensitivity analysis suggest that elevated costs of specific items reduced the cost-effectiveness of surgery and may imply that surgery is no longer cost-effective compared to radiotherapy.

Probabilistic sensitivity analysis

Following 1,000 iterations of second-order Monte Carlo simulations, a cost-effectiveness acceptability curve was generated (Fig. 6) to compare surgery and radiotherapy. An incremental cost-effectiveness scatter plot was also generated (Supplementary Material Attachment 3 Fig. 1).

Fig. 6.

Cost-effectiveness acceptability curve

As shown in Fig. 6, when the WTP was 33,080.09 USD/QALY, the probability of cost-effectiveness was equal for radiotherapy and surgery. When the WTP threshold was set at the GDP per capita of China (12,741.11 USD/QALY), the probability of surgery being cost-effective was 10.1%. When the WTP threshold was set at three times the GDP per capita of China (38,223.34 USD/QALY), the probability of surgery being cost-effective was 62.5%. When the WTP was 62,443.32 USD/QALY, the probability of surgery being cost-effective was 82.7%. These results are consistent with the base-case analysis.

Discussion

China has a high incidence of esophageal cancer [1], accounting for a significant proportion of malignant tumors. Compared to Western countries, esophageal cancer in China differs in pathological characteristics, treatment strategies, and socioeconomic features, with diverse and combined treatment approaches.

The various treatment modalities have demonstrated encouraging therapeutic effects. For treating esophageal cancer, the primary emphasis has been on the efficacy of interventions. But considering only the efficacy and ignoring the cost factor is far from enough. A study reported that the average expenditure per Chinese esophagus cancer patient had increased from 25,111 Chinese Yuan (CNY) in 2002 to 46,124 CNY in 2011. Especially, the expenditure on surgery (38,492 CNY) was significantly higher than that on radiotherapy (27,933 CNY) [26]. From the patient perspective, it is essential to consider not only the efficacy of treatments but also their associated costs. It is natural to question whether treatment benefits justify associated costs. In our study, surgery is marginally cost-effective compared to radiotherapy given a willingness to pay threshold of 38,223.34 USD/QALY. In general, extended survival is associated with increased expenditure. Like weights placed on either side of a balance scale, this presents an inevitable dilemma that requires careful consideration. Physicians, medical institutions, and decision-making departments must pay special attention to the cost-effectiveness of treatment strategies.

In China, an individual patient’s medical expenses are covered by two or three sources: first, the government basic medical insurance that covers approximately 95% of the population; second, the patient’s personal contribution; and third, commercial medical insurance purchased by some individuals. In recent years, at the point of hospital discharge, the government immediately covers the portion covered by basic medical insurance. Patients only need to pay their out-of-pocket share, while for those with commercial medical insurance, part or all of their personal expenses may be reimbursed by their commercial insurance.

CROSS [27], NEOCRTEC5010 [28], PERFECT [29], CheckMate577 [30], and the study by Park et al. [4] have confirmed acceptable efficacy and safety for neoadjuvant chemotherapy, neoadjuvant immunotherapy, consolidation chemotherapy, and consolidation immunotherapy. Currently, neoadjuvant immunotherapy combined with chemotherapy shows promising efficacy in resectable locally advanced esophageal cancer. Moreover, single-arm phase I/II clinical trials have indicated that chemoradiotherapy combined with immunotherapy is effective and well tolerated. Postoperative immunotherapy with an anti-programmed death 1 monoclonal antibody as a maintenance treatment is recommended for patients with locally advanced resectable esophageal cancer who have undergone neoadjuvant chemoradiotherapy without achieving pathological complete response [31]. A great many studies have already verified the efficacy and safety of neoadjuvant chemotherapy plus immunotherapy and postoperative consolidation chemotherapy plus immunotherapy. However, these studies themselves are mostly concentrated on prognostic indicators such as overall survival, progression-free survival, local control, and adverse reactions.

Our model employed a “perioperative” or “peri-radiotherapy” approach as well as these established treatment strategies. This involves the administration of preoperative or pre-radiotherapy neoadjuvant chemotherapy and immunotherapy, as well as postoperative or post-radiotherapy consolidation chemotherapy and immunotherapy.

Many studies have focused on the cost-effectiveness of esophageal squamous cell carcinoma (ESCC) treatment, particularly following the randomized controlled trials of immunotherapy. A Chinese cost-effectiveness analysis compared tislelizumab and camrelizumab as second-line ESCC treatment. Tislelizumab showed lower lifetime costs, higher QALYs, and was dominant over camrelizumab under China’s healthcare system perspective [3]. Another study evaluated adding durvalumab/tremelimumab to concurrent chemoradiotherapy (CCRT) for locally advanced ESCC. Survival outcomes significantly improved vs. CCRT alone historically, with PD-L1 positive status strongly predicting benefit [4]. Ran Qi etc. evaluation of pembrolizumab + chemo vs. chemo as first-line esophageal cancer therapy found it cost-effective only in the PD-L1 CPS ≥ 10 subgroup, not in ESCC or intention-to-treat (ITT) groups, at standard WTP thresholds [5]. Fenghao Shi etc. found a camrelizumab cost-effective second-line treatment for advanced/metastatic ESCC compared to chemotherapy, with an ICER below China’s typical WTP thresholds [6]. These studies demonstrated the advances of new drugs in the treatment of esophageal cancer. Numerous cost-effectiveness second-analyses have emerged based on published clinical trials for drugs. Although these studies on systemic medications exemplify classic applications of pharmacoeconomics, they are limited by designs that deviate from real-world conditions and fail to consider combination therapies and the complexion of actual treatment. Recognizing that both local and systemic treatments are indispensable, we summarized and extracted data from a single-center retrospective cohort, and substituted it into a Markov model to compare two common local treatment modalities for esophageal cancer in combination with systemic medication. We aimed to explore the relationship between patient outcomes and treatment costs using real-world data.

The radiotherapy of esophageal cancer, as an auxiliary or salvage treatment [32], has few studies on the economics compared with surgery. Salcedo et al. conducted a cost-effectiveness analysis comparing chemoradiotherapy followed by esophagectomy with chemoradiotherapy alone for ESCC and concluded that chemoradiotherapy alone was cost effective most of the time [33]. However, the authors noted that the risk of surgical complications in patients with esophageal cancer was insufficient to significantly affect costs and utility. Additionally, a 2022 Chinese study evaluated the cost-effectiveness of neoadjuvant therapy followed by surgery versus surgery alone for stage III esophageal cancer [34]. The authors determined that the neoadjuvant approach demonstrated superior efficacy and cost-effectiveness relative to upfront surgery. Notably, this study focused on establishing the necessity of neoadjuvant therapy rather than directly comparing surgery against radiotherapy. Collectively, these findings align with our results regarding the economic profiles of distinct treatment modalities. In our study, adverse event management was the most significant influencing factor for both surgery and radiotherapy. This discrepancy may be attributed to the lack of consolidation immunotherapy in the study by Salcedo et al. Adverse events compromise treatment outcomes by increasing costs, reducing patient utility, and diminishing therapeutic efficacy. In severe cases, they may interrupt treatment or cause disability/death.​​ This quantitative analysis highlighted the relatively mild adverse reactions and low radiotherapy costs.

With the promotion of immunotherapy, pharmacoeconomic studies on immunotherapy for esophageal cancer have also begun to increase. A cost-effectiveness analysis based on KEYNOTE-590 trial compared pembrolizumab + chemotherapy to chemotherapy alone for the first-line treatment of ESCC [5]. The current study revealed that the combination of pembrolizumab and chemotherapy as a first-line treatment for advanced or metastatic esophageal cancer was cost-effective only in some subgroup. This may be attributed to the high cost of pembrolizumab, which was offset by the substantial benefits observed for this subgroup. In our study, although surgery demonstrated marginal improvements in OS and PFS compared to radiotherapy, the increase in complication costs resulted in an insignificant enhancement in cost-effectiveness, as indicated by the ICER.

In our model, immunotherapy accounts for the majority of time during the most phase of esophageal cancer treatment, exerted a substantial influence on outcomes, second only to adverse event management. Similar results were observed in cost-effectiveness analyses of clinical trials comparing immunotherapy and chemotherapy. In a cost-effectiveness analysis comparing tislelizumab and camrelizumab for the second-line treatment of metastatic ESCC, Chen et al. identified the cost of immunotherapeutic agents as the most critical factor [3]. As a long-term strategy for esophageal cancer, chemotherapy combined with surgery or radiotherapy combined with chemotherapy is applied in the initial phase of treatment. Immunotherapy is then continued until disease progression. In this approach, immunotherapy often constitutes the largest proportion of treatment modalities, and its costs may significantly influence the overall cost-effectiveness of the therapy.​.

In the selection of treatment strategies for esophageal cancer, the main controversy between surgeons and radiation oncologists centers on the approach for locally advanced stages. Some thoracic surgeons believe that patients who have received neoadjuvant radiotherapy and chemotherapy have much more surgical difficulties and postoperative risks, especially anastomotic fistula and pulmonary infection. However, compared to esophagectomy, neoadjuvant concurrent chemoradiotherapy may improve survival rates and potentially be more cost-effective [35].

It should be also noted that there were some parts worth discussing in the model simulation process and statistical methods: In the present study, the Markov model was constructed using survival data derived from the fitted curve formulas rather than from real-world data. The reason lies that low-frequency follow-up in real-world studies may result in survival curves remaining flat across sequential Markov cycles, preventing estimation of cycle-specific transition probabilities.​.

During data processing, we considered extrapolating beyond the last real-world data point by fitting OS/PFS curves, i.e., building the model using a combination of real-world and extrapolated data. However, one issue was that ​​integrating empirical and fitted data introduced uncontrolled heterogeneity​​, complicating the model with excessive influencing factors and undermining simplification goals. Additionally, real-world OS/PFS curves are non-smooth, making it impossible to calculate transfer probabilities at arbitrary timepoints. Crucially, the transfer probability ​​between two adjacent points on the empirical curve could not be determined​​: Should it be linear? Should it follow a specific function? Would correction factors be needed? To streamline the model, we ultimately utilized ​​transfer probability matrices derived from fitted survival curves​​.

Intriguingly, the pPtP (probability of PD to PD) values for the first and second cycles were negative during the calculation of transition probabilities, which is impossible in real-world scenarios. Equation of lognormal distribution (Supplementary Material Attachment 2) shows that as x approaches 0 from positive values, log(x) tends toward negative infinity. Consequently, there must be a value of (t), such that as (t) approaches this value, pPtP tends toward negative infinity. Moreover, the curve fitting discrepancies resulted in the pPtP graph not intersecting the origin, leading to negative pPtP values in the initial cycles. These negative values were treated as 0.

The survHE package offers three curve-fitting methods: maximum likelihood estimation (MLE), Hamiltonian Monte Carlo (HMC), and integrated nested Laplace approximation (INLA). The MLE method is computationally simple and suitable for datasets with small sample sizes; thus, the MLE method was deemed appropriate for our study. Survival distribution selection depends on extrapolation plausibility. Given the monotonically decreasing nature of survival functions, we adopted standard parametric models. Based on likelihood, distribution characteristics, AIC/BIC values, and visual inspection, we determined survival function formulas and parameters. Regarding the criteria for assessing goodness-of-fit, AIC, BIC, and DIC have distinct applications [17]. ​​These functions calculated time-dependent transition probabilities for surgery and radiotherapy groups, with lower AIC/BIC values indicating better fit.​.

The limitations of our study were as follows:

1. This study is a retrospective analysis rather than a prospective randomized controlled trial. Although the data source involved non-random allocation, the simulated treatment was randomized, which could introduce potential bias. However, Table 2 demonstrates that the two real-world patient groups showed no difference in composition regarding gender, age, and tumor TNM stage, thereby minimizing such bias. These factors may introduce bias into the model, and these limitations will be addressed in future research.

2. Although major data used in the model were sourced from real-world evidence, the modeling process and complexity were inadequate to fully capture real-world scenarios. These complexities include, but are not limited to: heterogeneity in treatment regimens and drug dosages across different patients, the presence of various underlying diseases, differing tolerance levels to treatment complications, and economic constraints that may limit the implementation of subsequent therapies, etc.

3. The utility data were obtained from published literature rather than from the enrolled patients in present observational study, and the cost data derived from a single-center cohort may be outdated and may not accurately reflect resource use across China.

4. Additionally, we had to set values to zero on occasions to avoid impossible negative values, as mentioned earlier in the Discussion.

Conclusions

1. Surgery is marginally cost-effective compared to radiotherapy for ESCC in China; however, this advantage might be offset at relatively low WTP thresholds.​.

2. WTP thresholds, adverse reactions costs and immunotherapy costs constituted the primary factors influencing health economic evaluation outcomes in ESCC surgery versus radiotherapy in China.​.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AIC

Akaike information criterion

BIC

Bayesian information criterion

CCRT

Concurrent chemoradiotherapy

CEAC

Cost-effectiveness acceptability curve

CNY

Chinese Yuan

DIC

Deviance information criterion

DSA

Deterministic sensitivity analysis

ESCC

Esophageal squamous cell carcinoma

GDP

Gross domestic product

HMC

Hamiltonian Monte Carlo

ICER

Incremental cost-effectiveness ratio

ICESP

Incremental cost-effectiveness scatter plot

INLA

Integrated nested Laplace approximation

ITT

Intention-to-treat

MLE

Maximum likelihood estimation

OS

Overall survival

PD

Progressive disease

PFS

Progression-free survival

pFtD

Probability of PFS to death

pFtF

Probability of PFS to PFS

pFtP

Probability of PFS to PD

pPtD

Probability of PD to death

pPtP

Probability of PD to PD

QALY

Quality-adjusted life year

RT

Radiotherapy

Tis

Tumor in situ

USD

US dollars

WTP

Willingness-to-pay

Author contributions

Li’ang Xu designed project, collected, analyzed and interpreted data, and was a major contributor in writing the manuscript. Rong Liu collected, analyzed and interpreted data, and was another major contributor in writing the manuscript. Xiaoxi Chen, Lihong Liu contributed to data collection and collation. Lan Wang reviewed and polished the manuscript. Chun Han was the corresponding author of the manuscript and polished it. All authors read and approved the final manuscript.

Funding

This study was supported by Medical Science Research Project of Hebei, funding No. 20242007.

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

The experimental protocol was established, according to the ethical guidelines of the Helsinki Declaration and was approved by the Ethics Committee of Fourth hospital of Hebei Medical University. Written informed consent was obtained from individual or guardian participants. All authors have participated in the work and have reviewed and agree with the content of the article.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.