Our study was conducted based on a validated Chinese diabetes-related health outcomes model called the Chinese Hong Kong Integrated Modeling and Evaluation (CHIME) simulation model (9). The CHIME model is the first validated tool for predicting outcomes in Chinese patients with prediabetes and type 2 diabetes that was developed using Chinese data. The CHIME simulation model comprises 13 risk equations to predict all-cause mortality, diabetes-related macrovascular events (myocardial infarction, ischemic heart disease, heart failure, and cerebrovascular disease), microvascular events (peripheral vascular disease, neuropathy, amputation, ulcer of the skin, renal failure, cataracts, and retinopathy), and development of diabetes status. More details about the development of the CHIME model have been reported elsewhere (9). A CHIME-based discrete event micro-simulation cost-effectiveness model (CHIME-CE) was built to simulate the costs and outcomes for different treatments and long-term mortality was adjusted for diabetes all-cause mortality risk using population natural mortality (10). Natural mortality was extracted from China’s 6th National Census (11), the model structure is shown in Figure 1. A cohort of 30,000 patients was simulated and the cycle length was 1 year. This cost-effectiveness analysis was conducted from the perspective of the Chinese healthcare system, and a lifetime horizon was applied, with 99% of patients dying. All costs and health outputs were discounted at a rate of 5%. Our economic evaluation adhered to the Consolidated Health Economic Evaluation Reporting Standards 2022 (CHEERS 2022).

The target population was in line with the population analyzed in the trials, that is, Chinese adults, aged between 50 and 60, with T2DM which was previously inadequately controlled on metformin and received GLP-1RAs plus metformin as second-line therapy. The patient characteristic profiles were assumed to be the means reported in the original RCTs. When data pertaining to a specific parameter that was used for estimating the complications, such as the usage of antihypertensives or insulin, were not available, information from the Chinese Clinical Management System was used as a references (9, 12). More details are presented in Tables 1, 2.

A total of six GLP-1RAs were considered in our study, namely, exenatide (5 μg/bid for the first 4 weeks and then 10 μg/bid), liraglutide (0.6 mg/day in the first week, 1.2 mg/day in the second week, and 1.8 mg/day in the third week and beyond), loxenatide (5 μg/week), dulaglutide (1.5 mg/week), semaglutide (increase from a starting dose of 0.25 mg weekly to 1 mg, doubling the dose every 4 weeks), and lixisenatide (10 μg/day in the first week; 15 μg/day for the second week; and 20 μg/day in the third week and beyond). We did not consider benaglutide due to there being no relevant Phase 3 RCTs. All the patients were switched to insulin glargine U100 after 5 years, as their T2DM progressed with deterioration of their beta-cell function (37).

To inform the clinical inputs for the model, a systematic literature review was conducted in July 2021 to identify randomized controlled trials of relevant GLP-1RAs in the treatment of T2DM patients (38). A total of nine phase III RCTs with 3,850 patients were identified in the systematic literature review: Exendin-4 (13), Gao-2020 (14), SUSTAIN-7 (15), LEAD-2 (16), AWARD-5 (17), AWARD-6 (18), GetGoal-X (19), GetGoal-M (20), and GetGoal-F1 (21). The baseline characteristics of the patients in these trials (such as age, physical status, glycosylated hemoglobin (HbA1c), duration of diabetes, and fasting plasma glucose) were similar and therefore comparable. Additionally, a considerable proportion of Asian populations have been included in the global trials, therefore, they can serve as the source of clinical data for this study population. More details are provided in Table 2.

Network meta-analyses (NMAs) were then performed to assess the comparative efficacy of these six GLP-1RAs combined with metformin in the second-line treatment of T2DM. Four indicators were included in the analysis, namely the changing rates of glycated hemoglobin, body mass index (BMI), systolic blood pressure (SBP), and diastolic blood pressure (DBP) during the first 6 months. Considering that none of the included trials addressed the effect of lixisenatide on blood pressure, we used data from GetGoal-M-Asia (22) instead, as GetGoal-M and GetGoal-M-Asia were similar in design and had consistent results (e.g., 0.4% change in glycated hemoglobin with loxenatide compared to placebo in both studies during the first 6 months). The rates of change of the four indicators are presented in Supplementary Tables S1–S4. Random-effect Bayesian models estimated the risk ratios (RRs) of the changing rates between the six treatments via Markov chain Monte Carlo algorithms. The reference for glycated hemoglobin and BMI was set as lixisenatide in GetGoal-X, which was the trial with the largest sample size in our network, and semaglutide in SUSTAIN-7 was used as the reference for SBP and DBP for the same reason. Noninformative priors were used to allow the observed trial data to explain the effect estimates (39). We used the gemtc package (40) in R, version 4.1.0 (41) with four parallel Markov chains consisting of 50,000 samples after a 10,000 sample burn-in. The convergence of the Markov chains was checked by trace plots and Gelman-Rubin diagnostic statistics. The design-by-treatment approach was used to check the consistency in the entire NMA and the global inconsistency of the study was assessed with the I² (chi-square test) statistic. The local inconsistency was evaluated by node-splitting analysis, which was used to assess the inconsistency of the model by separating evidence on a particular comparison into direct and indirect evidence. The significance level was α = 0.05 for statistical tests.

Only the direct costs of implementing each treatment were included considering the Chinese healthcare system. All cost data were converted to US dollars using the exchange rate from 2022 (1 USD = 6.693 CNY). The prices of the GLP-1RAs were obtained from the access price of the national basic medical insurance drug catalog in 2021, and the prices of insulin and metformin were derived from the median price of national centralized drug procurement in 2021 (26). The cost of antidiabetic treatment and blood glucose test strips related to T2DM was collected from a large screening study based on the Chinese population (27). We considered only adverse events (AEs) with rates >5% reported in the RCTs, and related treatment costs and duration of diabetes were from published articles. Other potential health resource consumption, such as outpatient treatment costs, hospital expenses related to diabetes complications, and end-of-life costs, were extracted from published cost-effectiveness research based on Chinese patients (28–33, 35). More details about the cost-related parameters are listed in Table 1.

Health state utilities were collected from a report of 12,583 Chinese patients with T2DM, and a validated Chinese EQ-5D-5L instrument was used to investigate the utility for diabetes mellitus and cardio-cerebrovascular disease without complications (42). Other utilities that were not reported in that study, such as ESRD and minor and major amputations, were retrieved from published studies based on Chinese patients (29, 30, 34, 35). Disutilities of AEs were from a Chinese-based CEA (36). All the utility-related parameters are shown in Table 1.

The short-term costs and utilities were calculated based on the microvascular or macrovascular events and AEs reported in the RCTs, and the long-term costs and utilities were predicted by the CHIME-CE model.

The primary outputs of the model included costs, life years, quality-adjusted life years (QALY), net monetary benefit (NMB), incremental cost-effectiveness ratio (ICER), and incremental net monetary benefit (INMB). Considering the fact that the effectiveness of various treatments in our model is very similar, ICER was not used as the indicator in our sensitivity analysis. Instead, INMB was adopted. China’s per capita GDP in 2021 was used as the willingness-to-pay threshold (USD 12,728) in this study as recommended by the 2020 China Pharmacoeconomics Guideline (25).

Sensitivity analyses were performed to address the uncertainties in parameter values. We performed the one-way sensitivity analysis to test the sensitivity of results to changes in treatment effects-, costs-, and utilities-related parameters. Tornado graphs were plotted to visualize the parameters that had meaningful association with the results, and the INMB was used as a measure of cost-effectiveness. An INMB over 0 means the treatment was more cost-effective. A Monte Carlo simulation was performed for 50,000 iterations and the probabilistic sensitivity analysis was conducted. The gamma distribution was selected for cost; the beta distribution for probability, proportion, and utilities; and the normal distribution for the RRs between the treatments. All the parameters were adjusted within the reported 95% confidence intervals (CI) or assuming reasonable ranges of the base-case values. Cost-effectiveness acceptability curves were used to analyze the cost-effectiveness of each regimen with various Willingness-to-pay (WTP) thresholds.

Considering the impacts of research time limits, longer simulation time frames, while closer to patients’ lifetime costs and outcomes, also introduced more uncertainty. Therefore, we conducted a scenario analysis, comparing the cost, utilities, and NMB of each treatment when the research time limit was 10, 20, 30, and 40 years.

Combined with the results of base-case analysis, we calculated the cost and negative effect of diabetes complications caused by different treatments. In addition, we used a generalized linear model (GLM) (43) with HbA1c and BMI as independent variables, costs and disutilities as dependent variables, and all other disease characteristics of patients as covariates (such as age, gender, blood pressure, disease history). According to the values, BMI and HbA1c were both converted into binary variables according to the relevant standards of the World Health Organization (WHO). Namely, HbA1c greater than 7% was regarded as insufficient glycated hemoglobin control (44), and BMI greater than 25 kg/m² was considered to be ineffective in weight control (45). Through the GLM, the impacts of effective blood glucose or weight control on disease burden were analyzed.

Network plots are provided in Supplementary Figure S1. The RRs between the 6 treatments for glycated hemoglobin, BMI, SBP, and DBP are provided in Table 1. The results of global inconsistency and local inconsistency suggested that there was no inconsistency in all networks for the four indicators, and no significant differences were found between direct and indirect comparisons. More details can be found in Supplementary Figure S2.

The model was validated against the China Health and Retirement Longitudinal Study (CHARLS) cohort which was a nationally representative longitudinal cohort of Chinese residents aged 45 and older and other outcomes models such as the United Kingdom Prospective Diabetes Study Outcomes Model 2 (UKPDS-OM2) and the Risk Equations for Complications of Type 2 Diabetes (RECODe). More details have been reported elsewhere (9). We summarise the main process for model validation in eMethods 1, 2 in the Supplementary material.

The results of the base-case analysis are shown in Table 3. The mean QALYs (95% CI) for patients who received dulaglutide, lixisenatide, exenatide, semaglutide, liraglutide, and loxenatide combined with metformin were 12.50 (12.46–12.54), 12.52 (12.48–12.56), 12.56 (12.51–12.60), 12.56 (12.51–12.61), 12.58 (12.54–12.62), and 12.65 (12.61–12.69), respectively, ranked from least to most effective. The mean costs (95% CI) for patients who received loxenatide, exenatide, dulaglutide, lixisenatide, semaglutide, and liraglutide combined with metformin were USD 42,092 (41,765–42,377), 42,114 (41,796–42,444), 42,763 (42,479–43,066), 44,016 (43,081–43,737), 46,414 (46,103–46,734), and 47,026 (46,685–47,341), respectively, ranked from least to most costly. The mean INMB (95% CI) for patients who received loxenatide, exenatide, dulaglutide, lixisenatide, semaglutide, and liraglutide combined with metformin was USD 118,865 (118,268–119,417), 117,736 (117,159–118,330), 116,325 (115,685–116,908), 116,029 (115,455–116,641), 113,442 (112,857–114,029), and 113,069 (112,496–113,653), respectively, ranked from highest to lowest. Compared with the INMB of exenatide combined with metformin, the difference in loxenatide, dulaglutide, lixisenatide, semaglutide, liraglutide was USD (95% CI) 1,124 (326–1,988), −1,418 (−2,209–594), −1,713 (−2,569–−815), −4,298 (−5,121–−3,465), and −4,672 (−5,519–−3,835), respectively, ranked from highest to lowest. Loxenatide was identified as the most cost-effective choice. The ICERs of each treatment compared to exenatide are presented in Supplementary Table S8. A breakdown of diabetes complication-related costs and utilities for all treatments is presented in Table 4.

According to Table 4, the diabetes complication-related costs (USD 9256) and disutilities (0.39) of semaglutide combined with metformin were the lowest and were combined with the best controlling effects for glycosylated hemoglobin, BMI, and blood pressure from among the six treatments, followed by dulaglutide (cost, USD 9741; disutilities, 0.42), loxenatide (cost, USD 9980; disutilities, 0.41), lixisenatide (cost, USD 10326; disutilities, 0.44), liraglutide (cost, USD 10467; disutilities, 0.41), exenatide (cost, USD 10607; disutilities, 0.44). In the lifetime limits, compared to effective glycated hemoglobin control (HbA1c < 7%), patients taking medications with poor glycated hemoglobin control would cost $667 more, together with 0.013 QALYs lost. Compared to effective weight control (BMI < 25 kg/m²), medications that are effective in weight control would save patients $706, together with 0.017 QALYs gained. More details are provided in Supplementary Table S7.

Compared with previous similar studies (2, 53, 54), our study has the following advantages. First, we are the first to systematically compare the efficacy and cost-effectiveness of GLP-1RAs approved in China, thereby providing comprehensive evidential support for clinical rational drug use and medical service decision-making. Second, we adopted the CHIME-CE model to predict outcomes which were developed based on a validated tool for predicting the progression of diabetes and its outcomes among Chinese patients. Compared with other models, CHIME-CE is the first model based on the Chinese population, which means that our model was more accurate in predicting the outcomes of Chinese T2DM patients. Third, compared with other studies discussing the cost-effectiveness of GLP-1RAs, we not only considered the effects of GLP-1RAs on HbA1c but also other indicators that affect the risk of comorbidities. Fourth, we considered the impact of short-term AEs on costs and utilities, which had been overlooked in previous studies (2, 28–30, 35, 37, 53, 54). Finally, we quantified the impacts of effective glycated hemoglobin control and weight control on disease burden in patients with diabetes.

This study also has the following limitations. There were no head-to-head studies that directly compared all the glucose-lowering treatments against each other which meant that we had to perform an NMA. Although the articles included in this study were of high quality, indirect comparisons still brought some uncertainty. More head-to-head RCTs are needed to validate the findings of our study. Due to the limitation of data, the cost of retinopathy per event was sourced from Taiwan. However, the sensitivity analysis results indicate that the uncertainty of this parameter did not affect the conclusions of this study. Then, we only considered partial transition probability-related parameters such as RRs to simplify model calculations in the one-way sensitivity analysis, which may have led to an excessive agreement with base-case analysis results. Furthermore, the sources of cost and utility parameters were not from the same patients, which may have led to bias in economic evaluation. Due to the limited information reported by the RCTs, in the subsequent years, HbA1c, BMI, SBP, and DBP were predicted by the CHIME model. However, our study population and the CHIME population somewhat differed, predicted results may be biased, and future studies should pay more attention to this issue. In addition, some clinical parameters were sourced from global RCTs, and although this was necessary, there may still be potential biases to a certain extent. We only considered AEs with an occurrence rate of 5% or higher, this may lead to a certain degree of underestimation of costs and overestimation of utility in the base-case analysis and scenario analysis results. Finally, under the current circumstances, conducting a subgroup-based cost-effectiveness analysis is challenging, and we did not study the availability and affordability of GLP-1RAs, implying that further research is required.