Using a Markov cohort model, the experience of patients with R/R PTCL receiving pralatrexate (treatment arm) or CC (comparator arm) was simulated. The cycle length of model, in which a health state transition can occur, was set as a week. The simulation began at 48 years of age, reflecting the mean age of Korean patients with R/R PTCL , and continued for 15 years. The model consisted of five health states; initial treatment, treatment pause, subsequent treatment, stem cell transplantation (SCT) success, and death (Figure 1). All four alive health states can transition to death throughout the simulation. Except “SCT success” and “death” states, each health state was composed of four nested states; complete response (CR), partial response (PR), stable disease (SD), and progressive disease (PD). According to the PROPEL study, the proportion of patients achieving CR, PR, SD, or PD as a result of initial treatment with pralatrexate is 12.6%, 21.2%, 22.1%, and 44.2%, respectively. Among the Korean patients included in the CMCA study , 19.8%, 21.4%, 3.2%, or 55.6% achieve CR, PR, SD, or PD following initial treatment with CC, respectively.
Patients from each arm began the simulation at the “initial treatment” state. It was assumed that the initial treatment state lasts for 14 weeks because the drug regimen for pralatrexate and CC is two treatment cycles with a seven-week cycle  and 4.67 cycles with a three-week cycle (according to a local clinician survey), respectively.
After completing the initial treatment, patients moved to the “treatment pause” stage, if they survived, for eight weeks until subsequent treatment began. It was assumed that the response rate during the treatment pause state was the same as that during the initial treatment state. Among patients achieving CR or PR, those receiving SCT with a successful outcome moved to the “SCT success state” and it was presumed that they will have the same survival rate and quality of life as the general population. Therefore, the age-specific mortality rate of the general population provided by the Korean National Statistics Office was applied to them. Regardless of the type of drug treatments provided, the probability of receiving SCT in the treatment pause state following initial treatment was assumed to be 0.529 based on an expert panel survey of six clinicians who treat most R/R PTCL patients in Korea. The probability of success among those receiving SCT, defined as survival for at least one year, was 0.663 based on a local study . In general, patients with R/R PTCL receive SCT when they achieve the best response . According to the clinician panel survey, the median time of Korean patients with R/R PTCL achieving the best response is 20 weeks following initial treatment. Therefore, we assumed that SCT is performed at the 20th week from the start of initial treatment. Patients with CR not receiving SCT or receiving SCT with failure remained in the “treatment pause state” for the rest of the model simulation. The other patients (i.e., patients with PR not receiving SCT or receiving SCT with failure, and patients with SD or PD) transitioned to the “subsequent treatment” state at the 23rd week from the start of model simulation.
In the subsequent treatment state, all patients from either the treatment or comparator arm received CC for 4.33 cycles with a three-week cycle (according to a local clinician survey). After completing subsequent therapy, patients moved to the “treatment pause” state again on the 36th week. Among the Korean patients included in the CMCA study , the proportion of patients achieving CR, PR, SD, or PD following subsequent treatment was 12.3%, 15.8%, 3.5%, or 68.4%, respectively. Among patients with CR or PR in the treatment pause state, those receiving SCT with successful outcomes moved to the “SCT success state.” The probability of receiving SCT in the treatment pause state following subsequent treatment was assumed to be 0.458 based on the local clinician survey. The same probability of SCT success (0.663) was used. Applying local practice patterns, it was assumed that patients receive SCT at the 42nd week, which is 20 weeks after the subsequent therapy begins. The other patients (i.e., patients with CR or PR not receiving SCT or receiving SCT with failure, and patients with SD or PD) remained in the “treatment pause state” for the rest of the model simulation.
Patients assigned to the comparator arm were assumed to receive CC, such as DHAP (dexamethasone, cisplatin, and cytarabine), ESHAP (etoposide, methylprednisolone, cytarabine, and cisplatin), or ICE (ifosfamide, carboplatin, and etoposide), according to recommendations by the National Comprehensive Cancer Network guidelines and reimbursement list of the Korean National Health Insurance (NHI). The proportion of R/R PTCL patients receiving DHAP, ESHAP, or ICE in Korea was obtained from a clinician panel survey.
Input parameter and data sources
Table 1 presents model parameters and their data sources.
1) Comparative effectiveness
The comparative effectiveness of pralatrexate versus CC was derived from a CMCA study . Patients treated with pralatrexate had a median OS of 15.24 months, while those in the control group had 4.07 months (hazard ratio (HR): 0.432, 95% confidence interval (CI): 0.298 to 0.626).
Transition probabilities were derived from the CMCA study by digitizing OS rates from Kaplan-Meier curves. Since the time horizon for the cost-effectiveness analysis was greater than the duration of the CMCA study (approximately 52 months), extrapolation of survival rates was required to provide the long-term survival benefits of pralatrexate. Since access to individual patient data (IPD) was not possible, it was reconstructed from the CMCA study, so that parametric survival analysis could be conducted. The Kaplan-Meier survival curves for both treatment groups were digitized using DigitizeIt software. The method developed by Guyot et al. was applied to reconstruct IPD from the extracted coordinates of the published Kaplan-Meier survival curves using R software . The estimated median OS and HR were similar to those of the CMCA study, demonstrating the validity of the reconstructed data . Parametric survival analyses were conducted corresponding to guidance from the National Institute for Health and Care Excellence . The proportional hazards assumption did not hold according to goodness-of-fit test (p-value < 0.05), indicating that fitting parametric models separately to each treatment arm may be preferred. Parametric survival curves were fitted to the reconstructed IPD using various distribution functions. This approach was based on the NICE technical support document 14: survival analysis for economic evaluations alongside clinical trials-extrapolation with patient-level data, which says “There are a wide range of parametric models available, and each have their own characteristics which make them suitable for different data sets. Exponential, Weibull, Gompertz, log-logistic, log normal and Generalised Gamma parametric models should all be considered.” . In order to identify the most appropriate distribution function, Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) as well as visual inspection were assessed. For the pralatrexate arm, a generalized gamma model was selected as the best fit by visual inspection, as well as the lowest AIC(434.1) and BIC(431.4) (Figure 2). In contrast, the Gompertz model was selected as the most appropriate for the comparator group (AIC: 436.3, BIC: 433.7) (Figure 2). Transition probabilities of death beyond the duration of the CMCA study were then calculated from survival curves in each arm.
Costs were estimated from a societal perspective, including medical costs, transportation costs to visit healthcare institutions, and caregiver’s costs for hospitalization. All costs are presented as US dollars in 2019 value (1 USD approximately equal to 1,100 Korean won).
Cost of initial treatment
Initial treatment is provided with pralatrexate or comparator drugs. For each treatment, we calculated drug costs, drug administration costs, costs to treat adverse events (AEs), and monitoring costs. The cost of the comparator arm was calculated as a weighted average based on the usage proportion of DHAP (15.3%), ESHAP (9.9%), and ICE (74.8%) obtained through clinical expert consultation.
Based on the unit price, and doses recommended by the NCCN guidelines, weekly drug costs were calculated as $2,279 for pralatrexate, $54 for DHAP, $64 for ESHAP, and $100 for ICE. Drug costs of essential concomitant drugs were included: oral folic acid and vitamin B12 for pralatrexate, and pegfilgrastim, pegteograstim, tripegfilgrastim, and lipegfilgrastim for comparator drugs. Drug administration costs included outpatient visits, hospitalization, medication management, aseptic preparation, and injections. Unit costs for each item were determined using the Maximum Payable Amount Table of Korean NHI.
The common AEs among patients treated with pralatrexate or comparator drugs include anemia, mucositis, nausea/vomiting, neutropenia, peripheral neuropathy, and thrombocytopenia. The level of AE for all of these is grade 3 or higher, which indicates AEs requiring hospitalization . Therefore, the average cost per hospitalization for each condition reported by the Korean NHI statistics was used as the cost of treatment for the AE. Probabilities of AEs in each arm were obtained from literature [6, 15-20], and applied during the initial and subsequent treatment state.
Type and frequency of monitoring tests routinely performed for patients with R/R PTCL were identified by a clinician panel survey; positron emission tomography-torso; computed tomography for chest, abdomen, and neck; complete blood cell count; and general chemical tests.
Cost of subsequent treatment
In the subsequent treatment state, patients with R/R PTCL from both arms receive CC and best supportive care (BSC). Drug costs, drug administration costs, costs to treat AEs, and monitoring costs were calculated as above. Medical costs of the BSC were derived from a local study . Costs of subsequent treatments were calculated as a weighted average based on the usage proportion of DHAP (33.9%), ESAHP (16.2%), ICE (45.5%), and BSC (4.4%).
Cost of SCT
From Health Insurance Review and Assessment Service-National Patients Sample (HIRA-NPS) data from 2011 to 2016, patients with PTCL, defined as those having at least one claim record with a diagnosis of PTCL (ICD-10 codes: C84.4, C84.6, C84.7, C86.0~C86.3, C86.5, C86. 6, or C91.5), hospitalized for hematopoietic stem cell injection (X5131~X5136) were identified. The average hospitalization cost was considered as cost per autologous SCT. Unlike autologous SCT patients using their own stem cells, allogeneic SCT patients use hematopoietic stem cells from other healthy individuals. Therefore, cost per allogenic SCT was computed as cost per autologous SCT plus cost of obtaining matched hematopoietic stem cells, estimated using the data from the Korean Hematopoietic Stem Cell Transplantation Association and the Korea Marrow Donor Program. Using the proportion of autologous and allogeneic SCT performed among patients with PTCL in HIRA-NPS data from 2011 to 2016, the weighted average cost of SCT among patients with PTCL was calculated.
3) Quality of life
Utility values for each health state and disutility values for AEs were obtained from a local study  and foreign literature [23, 24]. Utility decrement due to AEs was applied during the initial and subsequent treatment state. Utility during the SCT success state was assumed to be the same as that of the general population aged 45 to 65 years .
During the modeled simulation of 15 years from the beginning at the “initial treatment” state, the average cost that a patient with R/R PTCL treated with pralatrexate or CC would have, was calculated for each Markov cycle by multiplying the probability of transition to each of the five health states during the cycle and their corresponding costs, and calculating the sum of the costs of all five health states. Then, we calculated the sum of the average costs of all Markov cycles (i.e., from the first cycle (1st week) through the last cycle (782nd week (=365 days/7 days × 15 years) of the Markov model) for each treatment arm, considering them as the expected costs that a patient with R/R PTCL would have for 15 years from the start of the simulation. The expected treatment effectiveness, defined as life-years (LYs) or quality-adjusted life-years (QALYs) gained as a result of treatment with pralatrexate or CC, was estimated in the same manner as the costs. For each Markov cycle, the average LYs were calculated as a product of the “probability of transition to each of the five health states during the cycle” and “1/52.1 or 0 LYs (= “1 week/52.1 weeks” if alive or “0 week/52.1 weeks” if dead).” Detailed explanations for how the expected costs and treatment effectiveness were computed as a result of the modeled simulation are presented in Appendix 1.
Incremental cost-effectiveness ratios (ICERs) were computed for the base case by dividing the incremental costs associated with providing pralatrexate therapy versus CC (i.e. the expected costs of pralatrexate arm less the expected costs of CC arm) by incremental effectiveness measured in LYs and QALYs gained, respectively. An annual discounted rate of 5% was used for both cost and effectiveness.
One-way sensitivity analyses were conducted to confirm the robustness of the ICER of the base case. The impact of the uncertainty in the extrapolation of the OS benefit of pralatrexate beyond the observation period of the CMCA was examined by applying different distribution functions. For utility for the health states of CR, PR, SD, and PD, upper and lower values of 95% CI  were used to re-calculate the ICER. The success rate of SCT was varied by 10% from the base case. Assuming that dose intensity is correlated with survival gain in lymphoma , a 20% decrease was applied to the dose of comparator drugs due to AEs, and the corresponding costs were adjusted accordingly. Based on the time horizon used in the earlier studies, the time horizon of the modelled simulation was varied to 10 and 30 years.