2.1 Overview

To estimate the effectiveness and cost effectiveness of fremanezumab in migraine patients, a decision analytic model is needed. The effectiveness of fremanezumab and placebo on the number of MMDs and HRQOL in EM and CM patients was captured in a previous study by Wang et al. using regression models [25]. For the current research, these regression models were implemented in a probabilistic Markov model that was developed in Microsoft Excel. This model was further populated with healthcare resource use (HCRU) and cost data and used to estimate the cost effectiveness of fremanezumab compared with SOC. An adjusted Japanese public healthcare perspective was adopted in which productivity losses were included, and the main model outcomes were life-years (LY), QALYs, costs in Japanese Yen (¥) inflated to 2022, and an incremental cost-effectiveness ratio (ICER). The current cost-effectiveness model (CEM) performed separate analyses for the EM and CM patients and merged the results for combined analyses.

2.2 Model Population

The model population was based on a cohort of Japanese EM (n = 357) and CM (n = 571) patients whose data was obtained from Japanese-Korean clinical trials [19, 20, 26]. The population consisted of patients who had a history of migraine for ≥ 12 months prior to trial admission. EM patients had 6–14 headache days with ≥ 4 migraine days (MDs), and CM patients ≥ 15 headache days with ≥ 8 MDs per 28 days. For the base-case analysis, only data from the Japanese patients was used. More details on the clinical trials are listed in the electronic supplementary material (ESM).

2.3 Treatments for Episodic and Chronic Migraine

In the clinical trials [19, 20, 26] that were used to estimate MMDs and HRQOL [25], patients were treated with fremanezumab 675 mg in the first month, followed by two doses of placebo in the second and third months. Alternatively, patients could also be treated with monthly administered fremanezumab 275 mg. Patients who were not treated with fremanezumab were treated with monthly administered placebo. Patients that were treated with both dosing schedules for fremanezumab were in the same intervention arm of the current CEM. The patients who received placebo were considered as the control arm (i.e., SOC) in the current CEM.

2.4 Model Structure

The disease course was simulated in the Markov model with three health states: alive (on treatment), alive (off treatment), and death (Fig. 1). Within each ‘alive’ health state, patients were partitioned over 29 MMDs per month (assuming 28-day months and 0 MMDs included) using the regression models from the previous modeling study [25]. Patients begin treatment, and after 3 months, either respond to and remain on treatment, or do not respond and discontinue (i.e. off treatment). ‘No response’ was defined as patients who did not achieve a 30% reduction in mean MMDs for CM and 50% for EM at 3 months. These definitions for no treatment response are considered clinically meaningful in the respective patient populations and were validated with Japanese clinical experts [27, 28]. The model population was simulated with 28-day cycles over a 25-years’ time horizon. Migraine is most common in younger and middle-aged people aged 25 years and older, and predominately effects women [2]. Since the course of the disease changes significantly after menopause, which on average occurs around the age of 50 years, the 25-year time horizon was deemed appropriate [29]. The Markov model was developed in Microsoft Excel.

Fig. 1

Model structure. Patients could either be on treatment, off treatment or die. The arrow between on and off treatment represented the discontinuation rate and the arrows to the death state the general mortality rate. In each ‘alive’ health state the distribution of patients across based on mean MMDs was determined, which was different for patients that were treated with fremanezumab and SOC (i.e., off treatment). MMD monthly migraine day, SOC standard of care

2.5 Model Input2.5.1 Treatment Efficacy

Model input parameters are listed in Table 1. Results from the previous modeling study by Wang et al. [25] were used to inform treatment efficacy and utility data (for utilities see Sect. 2.5.3) in the current model. Parametric regression models were used to model mean MMD distributions for EM and CM patients treated with fremanezumab or SOC over a period of 3 months to inform health state transitions. For the base case, the zero-inflated beta binomial (ZIBB) distribution was used with only trial data from the Japanese participants [25]. This approach effectively modeled all 29 potential health states based on the number of MMDs during each model cycle. Due to a lack of long-term efficacy data, the model assumed that the treatment effect at 3 months was maintained as long as patients remained on treatment. This assumption was validated by clinical experts within the field of neurology (TT and FS). More details on estimating the MMDs can be found in the previous modeling study [25].

Table 1 Model input parameters

Patients discontinued treatment when they did not adhere to the response criteria for fremanezumab in EM and CM and, therefore, a ‘negative’ stopping rule was applied. This concerned 50% of the EM and 25% of the CM patients that were treated with fremanezumab, which was based on clinical experts’ opinions (TT and FS). For each non-responder to fremanezumab, the treatment effect obtained by fremanezumab was replaced with the treatment effect for SOC. More details on treatment discontinuation can be found in the ESM.

Patients could also discontinue treatment for any reason [19, 20, 26]. An exponential parametric model was fitted to the observed data to estimate the discontinuation rate per model cycle (Table 1).

Migraine-related mortality was not modeled. However, given the 25-year time horizon, age-related mortality was included and modelled separately, using life tables from the Ministry of Health, Labor and Welfare in Japan [30].

2.5.2 Costs and Resource Use

Costs (2022 Japanese Yen [¥]), HCRU and discount rates were based on the Japanese C2H guidelines, clinical experts’ opinion and real-world patient survey data [31, 32]. Model costs and HCRU were chosen if they were expected to introduce any incremental differences between fremanezumab and SOC. The following cost categories were included in the model: treatment acquisition and administration costs, generalist visits, acute treatment costs, and productivity losses.

Treatment acquisition costs were applied to all patients that were on treatment with fremanezumab. The SOC drug acquisition costs were captured under resource use costs for patients in the fremanezumab arm that were off treatment and in the SOC arm. The SOC drug acquisition costs were calculated based on the JMDC database from Jan 2021 to Dec 2021 [33]. The ten most-sold medicines in the database were averaged (weighted by usage) and it was assumed that patients used these drugs once every migraine day in the model. An overview of these ten most-sold medicines is provided in the ESM. Treatment required a trained specialist to perform each administration. It was assumed that all patients had the injections administered in the base case. As such, all patients incurred an additional cost of ¥200 per administration [32]. Cost and HCRU for generalists visits were obtained from the literature [32]. The HCRU was obtained for 1 MMD and extrapolated linearly as the number of MMDs increased. Acute treatment costs were equal to SOC acquisition costs and applied to the fremanezumab arm. The HCRU of acute treatment costs was assumed to be 1 per MMD and also extrapolated linearly per MMD. Finally, productivity losses were included in the base-case analysis of the model because of the strong relationship between migraine and productivity [8]. The productivity losses were calculated using the annual salary for men and women in Japan multiplied with the percentage of work time missed (absenteeism) and impairment while working (presenteeism). The absenteeism and presenteeism differed between the number of MMDs based on the literature [34]. This calculation resulted in the annual productivity loss which was recalculated to monthly costs and applied to each model cycle. Costs and effects were discounted with an annual discount rate of 2% [31]. More details on productivity losses are listed in the ESM.

2.5.3 Utilities

MMD-specific utility values were used that were also estimated in the previous modeling study [25]. In that study, utility values were obtained from the Japanese-Korean clinical trials [19, 20, 26], with the migraine-specific quality-of-life (MSQ) questionnaire and mapped to EuroQol–five dimension–three level (EQ-5D-3L) values [25]. The MSQ-derived EQ-5D-3L values were modelled as a function of MMDs using a linear regression model, resulting in specific values for each health state in the current Markov model [25]. Different from the previous modeling study, treatment-specific utilities were assumed in the base case by incorporating treatment as a covariate in the regression model. This was advised by clinical experts within the field of neurology (TT and FS) and health technology assessment (AI) as they believed that the number of MMDs should not be the only factor that impacts HRQOL, but also factors such as the length and severity of migraine. More details on the utilities can be found in the ESM.

2.6 Analyses2.6.1 Base-Case Analysis

A deterministic base-case analysis was performed using the point estimates of each model parameter. Incremental LY, QALY and costs were calculated between fremanezumab and SOC. The ICER was calculated using the incremental QALY and costs but there was no specific WTP threshold in Japan as a reference.

2.6.2 Sensitivity Analyses

Several sensitivity analyses were performed to test the robustness and uncertainty of the model. Deterministic one-way sensitivity analyses (OWSA) were performed to test the influence of each individual parameter on the model outcomes. Parameter values were varied on their standard error, which was estimated based on the input’s source, or set to 20% of the mean if not available.

Probabilistic sensitivity analyses were performed in Microsoft Excel Visual Basic for Applications (VBA) to evaluate uncertainty of all model parameters simultaneously. This was done by drawing random values for each model parameter from prespecified distributions using Monte Carlo simulations with 2000 iterations. Model convergence was tested to determine if this was enough iterations for the PSA to produce robust outcomes. More details on model convergence can be found in the ESM.

Scenario analyses were performed to explore the impact of certain model settings and assumptions on the results. Scenario 1 included both the Japanese and Korean clinical trial data to model the effectiveness of fremanezumab and SOC (both on mean MMDs and utilities) and explore the difference compared with using Japanese-specific data only. Scenario 2 removed productivity losses from the analysis, since the Japanese healthcare perspective does not typically include indirect costs. In Scenario 3 and Scenario 4, the zero-inflated negative-binomial (ZINBI) and zero-inflated gamma (ZAGA) models were used to model the mean MMD distributions over time, respectively, and explore the impact of the selected parametric model distribution. Scenario 5 and Scenario 6 were a combination of Scenario 2 and Scenario 3 and Scenario 2 and Scenario 4, respectively. In Scenario 7, a time horizon of 5 years was adopted to estimate the short-term impact of fremanezumab on costs and effects. Scenario 8 included a positive stopping rule for patients to discontinue treatment after 1 year if they have not experienced loss of response. It was assumed that this concerned 10% of the EM and CM patients and that the treatment effect linearly waned back to baseline (i.e., treatment effect at month 0) over a period of 1 year.

2.7 Model Validation

The model was checked for face validity and assumptions were validated by experts within the field of neurology (KY, FS) and health economic outcomes research (AI). Model input parameters, outcomes and scenarios analyses were also validated by the clinical physicians (KY, FS, AI). The model was also checked for validity by one of the authors (MS) using the Assessment of the Validation Status of Health Economic decision models (AdViSHE) checklist [37]. The completed checklist can be found in the ESM.