2.1 General Workflow and Baseline Dataset

A visual representation of the workflow is displayed in Fig. 2. An overview of the ADEMP structure used in this work can be found in Table S1 in Online Resource 1. The methods were designed to align with the CHEERS (Consolidated Health Economic Evaluation Reporting Standards) guidelines to ensure transparency and comprehensiveness. However, the primary aim of this work was to compare pharmacoeconomic modeling methods, not to determine the cost–utility of sunitinib treatment for a specific country. This focus required certain methodological choices to facilitate a robust comparison between different modeling approaches, rather than strict adherence to the guidelines.

Fig. 2

Overview of the workflow. ICUR incremental cost-utility ratio, PE traditional pharmacoeconomic, PMX pharmacometric-based pharmacoeconomic, SLD sum longest diameter, TDM therapeutic drug monitoring, QALY quality-adjusted life years

As a first step, “true” data from a clinical trial was simulated using an earlier developed pharmacometric model framework for sunitinib in patients with GIST [22]. The model framework consisted of four models describing the time course of adverse events [hypertension, neutropenia, hand-food syndrome (HFS), and fatigue], soluble Vascular Endothelial Growth Factor Receptor-3 (sVEGFR-3) concentration, tumor growth, and OS (time-to-event Weibull model) (Fig. 1). Additionally, a separate model describing the time course of the adverse event thrombocytopenia, similar to the neutropenia model, was developed and added to the framework. A more detailed description of each model is provided in Online Resource 1. Patients were solely censored at death or when study end was reached.

A target population (Npatient = 1000) representing patients metastatic and/or unresectable GIST was generated using the distribution of original patient demographics [weight (normal distribution): mean = 73.5, standard deviation (sd) = 18.7, interval 36–185 kg; baseline tumor size (log-normal distribution): mean = 182.7, sd = 134.2, interval 29–822 mm) [22, 24, 25] (Fig. 2: “Dataset generation”). The selection of 1000 patients for the initial simulation was made to represent data from a large clinical trial. Using this population in combination with the pharmacometric framework, two study arms were simulated to represent (1) one arm with a continuous sunitinib dosing regimen of 37.5 mg daily and (2) one arm without sunitinib. Both study arms ran over a period of 104 weeks. The simulated clinical outcomes of the model included tumor progression (more than 20% growth from nadir), death, and the incidence of adverse events over time.

Following accepted clinical practices and the prescribing information [27], dose reductions (possible doses: 0 mg, 12.5 mg, 25 mg, and 37.5 mg, based on available tablet sizes) were implemented in the case of unacceptable adverse events, with the included adverse events variables being closely monitored. In the simulations, fatigue and HFS were monitored daily, while neutrophil count (ANC), diastolic blood pressure (dBP), and platelet count were evaluated every 6 weeks in accordance with established clinical protocols. For the first occurrence of grade 3 adverse events (or grade 2 for HFS and fatigue), the patient dose was reduced to 0 mg/day until the toxicity improved to ≤ grade 1, at which point the original dose was reinstated. In instances of grade 4 AEs (or grade 3 for HFS and fatigue), as well as for recurring adverse events [i.e., more than one instance of grade 3 AEs (or grade 2 for HFS and fatigue)], subsequent dosing was resumed at a reduced level (by 12.5 mg) following resolution of the toxicity under 0 mg/day [27].

The simulated toxicity data was transformed into a binary format denoting the presence (yes) or absence (no) of unacceptable toxicity, aggregated for every 6-week cycle. The criteria for defining unacceptable adverse events were as follows: neutropenia—an ANC < 1 × 109/L, thrombocytopenia—a platelet count < 50 × 109/L, diastolic hypertension—a dBP ≥ 110 mmHg, HFS ≥ grade 3 and fatigue ≥ grade 3. If a patient experienced unacceptable adverse events, they were classified as experiencing an adverse event related to that specific toxicity variable for the duration of the treatment cycle in question. Patients continued therapy and were monitored for survival throughout the study period, regardless of whether they experienced disease progression.

All data visualization and data management were conducted in RStudio in R (version 4.2.0). Model simulations were conducted with the package mrgsolve (version 1.0.3), datasets were generated using dmutate (version 0.1.3) and dplyr (1.0.9). Figures were generated using the package ggplot2 (version 3.4.3).

2.2 Traditional Pharmacoeconomic Model Estimations

All traditional pharmacoeconomic model parameters were estimated in NONMEM version 7.5.0 on the basis of the simulated data (Fig. 2: “Estimation of PE models”). The pharmacometric-based models were not re-estimated, except for the thrombocytopenia model, which was newly developed (see Online Supplementary 1 for further details). The traditional pharmacoeconomic model structures were selected on the basis of four distinct mathematical models that were utilized to perform a CUA of sunitinib [6,7,8,9]. In the recreation of the four traditional pharmacoeconomic models describing progression-free survival (PFS) and OS, the indicator function f(θ,x) was employed to indicate the comparators of interest, i.e., whether the patient received sunitinib treatment by multiplying an estimated model parameter (θDRUG) by the binary indicator variable Xdrug [Xdrug = 0: absence of sunitinib treatment (study arm without sunitinib), Xdrug = 1: presence of sunitinib treatment (study arm with continuous sunitinib dosing regimen at starting dose 37.5 mg daily and potential dose adjustments)].

A more detailed discussion of the model structures and the equations underlying each separate model can be found in Online Resource 1.

2.3 Generation of the Model Frameworks

The final model structures and corresponding parameter estimates were translated into mrgsolve for simulations. In addition to the existing pharmacometric-based pharmacoeconomic model framework, the four traditional pharmacoeconomic model frameworks were generated by merging the corresponding PFS, OS (separated for the TTE models, combined for the Markov Models), and toxicity models (five logistic regression models) (Fig. 2: ‘Creation of simulation frameworks’).

Individual utility values (UVi) were determined at each occasion (UVij) ranging from 0 (lowest quality of life) to 1 (perfect quality of life) following the EuroQol-5 Dimension (EQ-5D) questionnaires [28], using a multiplicative function (Online Resource 1) [29]. The decision to confine utility values between 0 and 1 was motivated by the relatively high quality of life observed in patients with GIST, even at advanced stages of the disease [30]. The specific UVs for each condition were based on literature (UVbaseline: 0.712, UVneutropenia: 0.9777 (≥ grade 2), UVthrombocytopenia: 0.9895 (≥ grade 3) or 0.892 (≥ grade 4), UVHFS: 0.8813 (≥ grade 3), UVfatigue: 0.999 (≥ grade 1), or 0.9107 (≥ grade 3), UVprogression: 0.577) [18, 31,32,33], with the baseline UV accounting for the reduced quality in life owing to presence of GIST in all patients.

The economic evaluation was conducted from the Dutch healthcare system perspective, with costs reported in 2023 (Euros). A time horizon of 2 years (104 weeks) was selected as most adverse events occur within this period, and the median survival is approximately 2 years (around 100 weeks). This time span allowed for a more focused analysis of model performance by concentrating on the critical period during which the most significant events and costs occur. The total costs were calculated per patient over the span of 104 weeks or until the time of death. Information on the drug costs of sunitinib was gathered from the Dutch National Health Care Institute [34]. The costs of regular follow-up consisted of imaging, laboratory costs and medical visits declared via the Dutch Healthcare Authority [35] where visits were considered to occur every 16 weeks (as reported in 2023 (Euros)) [18]. Costs for adverse events were based on those estimated in a previous study in a German population (as reported in 2019 (Euros)) [36, 37]. The costs for adverse events were based on data from Germany owing to the unavailability of equivalent Dutch data. Given the similarities between the Dutch and German healthcare systems [36], German data was deemed to serve as a suitable proxy. A summary of all costs is provided in Table S3 (Online Resource 1). Both costs and life-year outcomes were discounted by 4% per year, following Dutch guidelines [38]. No adjustments for inflation were made as the cost data were directly taken from recent studies.

2.4 Simulations of Treatment Outcomes and Costs

The final models were utilized to conduct a CUA comparing sunitinib treatment with no treatment in patients with GIST, assessing treatment outcomes and the ICUR. For each of the five model frameworks, a single virtual patient cohort (Npatient = 10,000) was created over a 2-year period (Fig. 2: ‘CUA’). The selection of 10,000 patients for the virtual cohort was made to ensure a robust analysis, as a larger sample size reduces noise owing to random variability and helps to ensure that observed differences between models are owing to structural factors rather than random chance.

Two simulations were performed for each model framework: one with sunitinib treatment (Xdrug = 1) and one without (Xdrug = 0), both using the same underlying patient population. The ICUR was calculated by comparing sunitinib treatment with no sunitinib treatment.

Additionally, the precision of the outcomes for each model (OS, QALYs, and ICUR) was evaluated by repeating the simulations (Nsimulation = 100). In each iteration, a new virtual population was generated. The outcomes of the pharmacoeconomic model frameworks were compared to the outcomes of the pharmacometric model framework. For each traditional pharmacoeconomic model framework (‘PE’), the mean of the 100 simulations was quantified as the relative deviation from the base (RDB), i.e., the relative deviation from the pharmacometric-based pharmacoeconomic model framework (‘PMX’) (Eq. 2).

$$\text= \frac}_}-}_}}}_}} \times 100\%$$

(2)

The impact of reducing drug prices following the expiration of patents was additionally explored. As the median reduction in the Netherlands was 41% [39], a 40% reduction scenario was considered alongside two more extreme scenarios of 60%, and 80% to the drug price. For each discount scenario, the total costs and the ICUR were based on the mean of 100 simulations for each model framework.

2.5 Therapeutic Drug Monitoring

A scenario with Ctrough-based dose adjustments was evaluated where the developed frameworks were used for CUA of sunitinib in TDM (Fig. 2: TDM workflow). Firstly, it was assessed whether each of 1000 simulated patients from the sunitinib treatment arm (Sect. 2.1 Baseline dataset) had a simulated Ctrough value below or above the target exposure at day 57, for both efficacy (> 37.5 ng/ml) and safety (< 75 ng/ml) targets separately [40] (Fig. 2: ‘Dataset generation’ TDM workflow). This procedure was similar to a previous pharmacoeconomic study of TDM for sunitinib [36]. Differences in PFS and OS between patients above and below the efficacy threshold of 37.5 ng/ml were estimated using Eq. 1. Similarly, differences in toxicity between patients above and below the safety threshold of 75 ng/ml were estimated, using Eq. 1 to estimate the difference between patients. In addition, four pharmacoeconomic model frameworks were generated, where the influence of sunitinib exposure (Ctrough above or below safety and efficacy thresholds) on clinical outcomes was assessed (Fig. 2: ‘Estimation PE models and Creation of simulation frameworks’ TDM workflow).

Thereafter, using the model frameworks, the impact of TDM, with accompanying dose adjustments, was explored for a virtual patient population (Npatient = 10,000) over a 2-year time horizon (Fig. 2: ‘CUA’). A Ctrough above 37.5 ng/ml was associated with efficacy and a Ctrough of 75 ng/ml to toxicity. Two variables were created per patient (X37.5 and X75) indicating whether the patient was below, within or above these limits. Patients were randomly assigned into three groups according to the estimated underlying probability distribution based on the 1000 simulated patients from the sunitinib treatment arm: sub-therapeutic (< 37.5 ng/ml, p = 0.557: X37.5 = 0, X75 = 0), therapeutic (37.5–75 ng/ml, p = 0.414: X37.5 = 1, X75 = 0) and supra-therapeutic (> 75 ng/ml, p = 0.029: X37.5 = 1, X75 = 1) [41].

In the traditional pharmacoeconomic model frameworks, TDM intervention was simulated following a previous publication [36] to occur after cycle 1 of therapy (day 42). Here, all patients were assigned into the therapeutic group unless toxicity had occurred during cycle 1, thereby mimicking immediate correct dose adjustment following TDM. In the following cycles (every 42 days), patients remained in the assigned group or transited to the sub-therapeutic group, in case of toxicity (i.e., through dose reduction).

Within the pharmacometric-based pharmacoeconomic model framework, simulations were not treatment cycle-dependent and sunitinib monitoring was conducted on day 15, 29, and 57, adhering to a recommended TDM schedule [42]. Dose adjustments were individualized based on a predetermined range of available dosages (0, 12.5, 25, 37.5, 50, 62.5, and 75 mg). Doses were increased by increments of 12.5 mg for exposures below the target, while exposures above the target led to a decrease by the same amount. Not all patients achieved the therapeutic window following a single dose adjustment; occasionally, multiple dose alterations were required. The protocol for dose modifications in response to toxicities was consistent with the procedures outlined in Sect. 2.1.