Cost-effectiveness of pembrolizumab for previously treated MSI-H/dMMR solid tumours in the UK

Abstract

Objectives

Patients with previously treated microsatellite instability-high (MSI-H)/mismatch repair deficient (dMMR) tumours have limited chemotherapeutic treatment options. Pembrolizumab received approval from the EMA in 2022 for the treatment of colorectal, endometrial, gastric, small intestine, and biliary MSI-H/dMMR tumour types. This approval was supported by data from the KEYNOTE-164 and KEYNOTE-158 clinical trials. This study evaluated the cost-effectiveness of pembrolizumab compared with standard of care (SoC) for previously treated MSI-H/dMMR solid tumours in line with the approved EMA label from a UK healthcare payer perspective.

Methods

A multi-tumour partitioned survival model was built consisting of pre-progression, progressed disease, and dead health states. Pembrolizumab survival outcomes were extrapolated using Bayesian hierarchical models (BHMs) fitted to pooled data from KEYNOTE-164 and KEYNOTE-158. Comparator outcomes were informed by published sources. Tumour sites were modelled independently and then combined, weighted by tumour site distribution. A SoC comparator was used to formulate the overall cost-effectiveness result with pembrolizumab as the intervention. SoC comprised a weighted average of the comparators by tumour site based on market share. Drug acquisition, administration, adverse events, monitoring, subsequent treatment, end-of-life costs, and testing costs were included. Sensitivity and scenario analyses were performed, including modelling pembrolizumab efficacy using standard parametric survival models.

Results

Pembrolizumab, at list price, was associated with £129,469 in total costs, 8.30 LYs, and 3.88 QALYs across the pooled tumour sites. SoC was associated with £28,222 in total costs, 1.14 LYs, and 0.72 QALYs across the pooled tumour sites. This yields an incremental cost-effectiveness ratio (ICER) of £32,085 per QALY. Results were robust to sensitivity and scenario analyses.

Conclusions

This model demonstrates pembrolizumab provides a valuable new alternative therapy for UK patients with MSH-H/dMMR cancer at the cost of £32,085 per QALY, with confidential discounts anticipated to improve cost-effectiveness further.

Keywords:

• Microsatellite instability
• mismatch repair deficient
• economic analysis
• cost-effectiveness
• pembrolizumab
• multi-tumour
• histology-independent

JEL CLASSIFICATION CODES:

• D61
• D6
• D
• H51
• H5
• H

Introduction

Tumours with mismatch repair deficiency (dMMR) can be found in many different sites throughout the body. While they may arise because of inherited conditions, such as Lynch syndrome, they are more often attributed to somatic mutations¹. MMR genes support a process that helps to maintain genomic stability by repairing the base–base mismatches and changes in DNA that are generated during replication and recombination. Mutated or non-functioning DNA MMR proteins can cause high microsatellite instability (MSI-H), which are mutations in the short, tandem sequences of mononucleotide, dinucleotide, or higher-order nucleotide repeats known as microsatellites that are scattered throughout the human genome². MSI-H caused by dMMR can subsequently lead to tumour development, and dMMR tumours are more susceptible to tumour mutations than MMR-proficient tumours³.

The prevalence of MSI-H/dMMR varies by tumour site and disease stage. It is found in approximately 25% of patients with endometrial cancer, more than 10% of patients with colorectal or gastric malignancies, and only 0–3% of patients with biliary cancer^4,5. A study by Le et al.⁶ suggested that the prevalence of MSI-H in late-stage disease was lower than in earlier stages. Separately, studies have also shown that dMMR is prognostic of worse survival outcomes than in patients with MMR-proficient tumours⁷. Patients with previously treated MSI-H/dMMR tumours have limited chemotherapeutic treatment options. While treatment pathways vary by tumour site, most comprise either multi- or single-agent chemotherapy regimens. For some histologies, such as colorectal cancer (CRC), the availability of existing targeted therapies has resulted in genomic testing becoming part of standard clinical practice; for cancer in other sites, testing is far more limited^8–13.

Cells from dMMR tumours often exhibit increased expression of the immune checkpoint protein programmed death ligand 1 (PD-L1). Furthermore, infiltrating lymphocytes display high levels of checkpoint proteins. This combination has been suggested to render these tumours vulnerable to treatment with immune checkpoint inhibitors such as pembrolizumab.

Pembrolizumab is a programmed cell death protein 1 (PD-1) inhibitor immunotherapy that originally received accelerated approval by the US Food and Drug Administration (FDA) in May 2017 for the treatment of adult and paediatric patients with unresectable or metastatic MSI-H/dMMR solid tumours that have progressed after prior standard treatment and have no satisfactory alternative treatment options¹⁴. In April 2022, pembrolizumab received similar approval from the European Medicines Agency (EMA) for previously treated MSI-H/dMMR tumours in adults with unresectable or metastatic CRC after previous fluoropyrimidine-based combination therapy; for adults with advanced or recurrent endometrial carcinoma who have disease progression on, or following prior treatment with, a platinum-containing therapy in any setting and who are not candidates for curative surgery or radiation; and for patients with unresectable or metastatic gastric, small intestine or biliary cancer who have disease progression on or following at least one prior therapy.

The approval in Europe for pembrolizumab in treating these five MSI-H/dMMR tumour types was supported by data from the KEYNOTE-164 (KN-164; NCT02460198) and KEYNOTE-158 (KN-158; NCT02628067) clinical trials^1,15. KN-164 is a Phase II single-arm trial of patients with previously treated, unresectable, locally advanced or metastatic MSI-H/dMMR CRC. The KN-158 trial enrolled patients with unresectable or metastatic MSI-H/dMMR solid tumours from multiple tumour sites, including endometrial, gastric, small intestine or biliary cancers. The primary endpoint of both trials was objective response rate (ORR)¹⁶. Secondary endpoints for both trials included duration of response, progression-free survival (PFS), overall survival (OS), safety and tolerability, and health-related quality-of-life (HRQL) – the latter of which was only collected in KN-158.

Examples of previous economic evaluations for histology-independent therapies (HITs) such as pembrolizumab can be found for larotrectinib and entrectinib, both treatments of NTRK fusion-positive solid tumours^17,18. In the absence of direct comparative evidence, two methods have been explored to derive relative effectiveness estimates: using data from published studies for comparator therapies and using a within-trial analysis of intervention outcomes compared with outcomes from the prior line of therapy.

To model outcomes for HITs, studies have typically pooled data from patients with multiple tumour types collected from a basket trial, assuming homogeneity in patient outcomes between different tumour sites^17–20. However, guidance published by Murphy et al. suggests that if the assumption of homogeneous clinical outcomes across histologies fails to hold, this could have consequences for both cost and health outcomes, which could influence the results of the cost-effectiveness analysis and subsequent health technology assessment (HTA) decision-making²¹. The authors propose a Bayesian hierarchal modelling (BHM) framework to capture heterogeneity in outcomes by tumour site. Consequently, when reviewing their methods for HTA, the UK’s National Institute for Health and Care Excellence (NICE) now recommends that “when heterogeneity between groups within a population is a concern, any assumptions about homogeneity or heterogeneity and generalizability to clinical practice must be clearly presented, tested and fully explored”²².

A review of the literature confirmed that, while studies have been published in individual tumour sites for CRC and endometrial cancer^23,24, to our knowledge no economic evaluations of pembrolizumab for previously treated MSI-H/dMMR solid tumours in multiple sites have been published. The objective of this study was to assess the cost-effectiveness of pembrolizumab for patients with previously treated MSI-H/dMMR solid tumours from a UK healthcare payer perspective. The patient population included the five tumour sites within the approved EMA label.

Methods

Decision problem

The modelled patient population included adult patients with previously treated MSI-H/dMMR solid tumours, as per the KN-164 and KN-158 clinical trials and in line with the approved EMA label. The five tumour sites considered in the cost-effectiveness analysis were colorectal, endometrial, gastric, small intestine, and cholangiocarcinoma (biliary).

The intervention in the model was pembrolizumab 200 mg, given intravenously (IV) every 3 weeks for up to 35 cycles or until progression^1,15. In the clinical trials, patients who achieved a complete response could stop study treatment after receiving at least eight administrations of pembrolizumab, which was reflected in the analyses of time on treatment data from the trial informing the cost-effectiveness analysis. Patients who had confirmed disease progression but still experienced clinical benefits without any additional increase in tumour burden could continue pembrolizumab therapy^1,15.

Comparators by tumour site were identified using European and National Comprehensive Cancer Network (NCCN) guidelines^9–13, and were validated by UK clinical experts who confirmed which treatments are currently used in UK clinical practice⁹. These comparators were then included in an accompanying clinical systematic literature review (SLR) to identify published evidence to inform the economic model. Regimens of a similar class were grouped where clinical experts suggested that there would be negligible difference in outcomes. A standard of care (SoC) comparator was used to formulate the overall cost-effectiveness result. SoC comprised a weighted average of the comparators by tumour site based on market share. Comparators and their respective market shares are included in Supplementary Materials 1.

Model structure

A cost-effectiveness model (CEM) was built in Microsoft Excel® with a UK healthcare payer perspective and adhered to best practice guidance from NICE. The time horizon was set to 40 years (equivalent to lifetime) and a model cycle length of 1 week applied to capture treatment dosing schedules. Cost and QALY outcomes were discounted at 3.5% per year²⁵.

A partitioned survival modelling approach was utilized with three mutually exclusive health states: pre-progression, progressed disease (PD), and death. A partitioned survival analysis structure provides the flexibility to extrapolate immature data, such as from clinical trials. PFS and OS were also readily available from the published evidence, which was critical to generate comparator survival outcomes given that both KN-164 and KN-158 are single-arm trials.

Alive states were further separated into on- and off-treatment. Patients may discontinue treatment before or after progression, meaning that they enter either the pre-progression off-treatment or PD off-treatment states.

Each tumour site and comparator were modelled independently to address the potential heterogeneity associated with pooling treatments and tumour sites. The model then combined the individual tumour site results by applying the tumour site prevalence distribution to create a weighted average of costs, QALYs, and LYs. The overall incremental cost-effectiveness ratio (ICER) was then calculated as the weighted average of incremental costs across all tumour sites divided by the weighted average of incremental QALYs across all tumour sites. Tumour site distribution was determined using a real-world analysis of available UK epidemiological data for the included tumour sites. The calculations to determine the eligible population were specific to each tumour site and structural assumptions were applied in line with previous health technology appraisals in similar indications or based on other publicly available sources. Incidence estimates per tumour site were sourced from Cancer Research UK^26–29 and a NICE technology appraisal (TA762)³⁰. Each tumour site was associated with the following weights used in the model: CRC, 57.9%; endometrial, 14.8%; gastric, 19.1%; small intestine, 5.3%; and cholangiocarcinoma, 2.9%. Additionally, all non-pembrolizumab treatment arms were combined, weighted by market share data elicited from clinical experts³¹, to create the SoC comparator for an overall cost-effectiveness result with pembrolizumab as the intervention.

Model parameters

Baseline patient characteristics

Baseline patient characteristics were used to calculate several inputs, including the age- and sex-dependent general population utility and mortality, as well as dosage derivation for some interventions. These were sourced from the KN-158 and KN-164 trials (Supplementary Materials 2).

Efficacy

Intervention

In the base case, the BHM framework recommended by Murphy et al.²¹ was extended to the extrapolation of time-to-event outcomes through a Bayesian formulation of standard parametric survival models. BHMs were fitted to pooled data from KN-164 and KN-158 and fitted to OS and PFS outcomes. For each model, the goodness of fit to the observed data was first assessed visually by overlaying the parametric curve on the observed Kaplan–Meier curve, then assessed statistically via model comparison statistics (deviance information criterion [DIC] for Bayesian models, and Akaike information criterion [AIC] and Bayesian information criterion [BIC] for frequentist models).

The hierarchical nature of BHMs means that parametric distributions fitted have both shared (fixed-effects) parameters and tumour-site-dependent parameters. Fixed-effects parameters are shared by all tumour sites, while an exchangeable (random-effects) parameter that is unique to each tumour site captures the heterogeneity of outcomes observed. Estimates of the level of heterogeneity across sites and of the pooled treatment effects for each tumour were produced. The BHMs included covariates considering age (normalized), sex, Eastern Cooperative Oncology Group (ECOG) score, cancer stage, and number of prior lines of therapy. The mean of the posterior distribution was used for the deterministic analysis and individual iterations were randomly sampled from the posterior distribution in each run of the probabilistic sensitivity analysis (PSA), allowing for simultaneous sampling across all tumour sites.

Standard one-piece parametric survival models were also fitted separately for each tumour site. The best-fitting models were selected considering a mixture of the AIC and the BIC, as well as visual inspection of fit to available Kaplan–Meier data. Survival outcomes using standard parametric survival models (PSMs) fitted separately for each tumour site were explored in scenario analyses. Given the heterogeneity observed in survival outcomes, as shown in Figure 1, it was not appropriate to pool outcomes from multiple tumour sites and fit a single survival model.

Figure 1. Kaplan–Meier plots of survival outcomes from KEYNOTE-158 and KEYNOTE-164, pembrolizumab (a, OS; b, PFS). Abbreviations. CRC, colorectal cancer; OS, overall survival; PFS, progression-free survival.

Modelled pembrolizumab time to treatment discontinuation (TTD) was assumed to be equivalent to PFS. A stopping rule at 105 weeks was applied to pembrolizumab in each tumour site, consistent with the approved drug label and trial design^1,15,32.

The survival outcomes for patients treated with pembrolizumab in KN-164 and KN-158 are presented in Figure 1 for OS (a) and PFS (b). A summary of the OS and PFS standard PSM distributions with AIC and BIC are presented in Supplementary Materials 3 and 4, respectively.

Comparators

Several approaches were explored to inform outcomes for comparators within each of the different tumour sites. In the base case, published clinical data identified by the SLR were extracted, then digitized using methods previously described by Guyot et al.³³ PSMs were fitted to the generated pseudo-patient-level data (pseudo-PLD). Where multiple studies were identified for a single comparator regimen, data were pooled in order to maximize the available data using the methods described in Combescure et al.³⁴ Naïve indirect treatment comparison and matching-adjusted indirect comparison methods were also explored. However, since the proportional hazards assumption did not hold and adjustments for observed confounders had a negligible effect, these analyses were not considered further. The fitted curves were therefore used directly.

For TTD, there were limited published data for comparators, either because data within HTA submissions were redacted or because TTD Kaplan–Meier data were not reported. Therefore, where clinical experts suggested that it was appropriate to do so, TTD was assumed to be equivalent to PFS. Where this was not appropriate and summary statistics were available, an exponential curve fitted to median time on treatment data from summary of product characteristics (SmPC) drug labels or respective clinical trials was used. A summary of the base case selections is provided in Supplementary Materials 5.

Overall survival

Model predictions for OS are presented in Figure 2 for each tumour site. Each pane includes the base case pembrolizumab (BHM) and comparator (naïve PSM) OS predictions compared to the Kaplan–Meier curve.

Figure 2. Pembrolizumab and comparator OS base case selections versus Kaplan–Meier (a, CRC; b, endometrial; c, gastric; d, small intestine; e, cholangiocarcinoma). Abbreviations. CRC, colorectal cancer; OS, overall survival.

Progression-free survival

Model predictions for PFS are presented in Figure 3 for each tumour site (a–e). Each pane includes the base case pembrolizumab and comparator PFS predictions compared with the Kaplan–Meier curve.

Figure 3. Pembrolizumab and comparator PFS base case selections versus Kaplan–Meier (in order of upper left to bottom right: CRC, endometrial, gastric, small intestine, cholangiocarcinoma). Abbreviations. CRC, colorectal cancer; PFS, progression-free survival.

Safety

Grade 3+ adverse events (AEs) with an incidence of ≥1% of patients in the pembrolizumab arm (KN-164 and KN-158) and an incidence of ≥3% of patients in the comparator arms were considered in the analysis^1,15. Including a higher incidence threshold for comparator AEs was considered a conservative but pragmatic assumption to avoid the total number of AEs included in the model being excessively large. For treatment arms not evaluated in KN-164 or KN-158, incidence rates were informed by literature gathered as part of a targeted literature review (TLR) of trial data and published studies. Costs associated with AEs were applied as a one-off cost in the first model cycle.

Health-related quality-of-life

HRQL data were collected in the KN-158 trial using EQ-5D-3L questionnaires and the UK value set applied to derive utility values. The base case method derived values based on progression status and tumour site, therefore capturing heterogeneity between patient HRQL observed between different histological sites. Health state utilities were adjusted to account for age-matched general population using a utility multiplier derived from Hernández Alava et al. 2022³⁵.

Utilities were assumed to be the same across treatment regimens, thereby assuming that patients treated with comparator therapies would have the same HRQL as patients treated with pembrolizumab.

Utilities specific to the CRC tumour site were obtained from Grothey et al.³⁶ as no HRQL data were collected in KN-164. In the model, EQ-5D utility values of 0.73, 0.74, and 0.59 were used for pre-progression on-treatment, pre-progression off-treatment, and PD states, respectively. Health state utility values by progression status and tumour site are presented in Table 1.

Table 1. Health state utility values by progression status and tumour site.

Tumour site	Pre-progression utility		Progressed utility	Source
CRC	0.73*	0.74**	0.59	Grothey et al.^36
Endometrial	0.72		0.67	KN-158
Gastric	0.71		0.65	KN-158
Small intestine	0.81		0.74	KN-158
Cholangiocarcinoma	0.80		0.70	KN-158

Abbreviation. CRC, colorectal cancer.

*on-treatment, **off-treatment.

AE disutilities were not included in the base case as it was assumed the impact of AEs on patient HRQL would be captured by the health state utility values, thereby avoiding any potential double counting.

Costs

Drug acquisition and administration costs

The dosing schedule for pembrolizumab was based on dosages observed in KN-164 and KN-158^1,15. Comparator dosing schedules were identified from SmPC labels where possible or from the literature. Unit costs were identified from the Monthly Index of Medical Specialities (MIMS)³⁷ and the UK Government’s drugs and pharmaceutical electronic market information tool eMIT³⁸. All drugs were included at list price. A summary of dosing schedules and unit costs is provided in Supplementary Materials 6.

The costs associated with treatments administered by IV injection were determined by whether the drug was associated with a simple or complex first administration. Subsequent administrations were associated with their own respective administration cost. It was assumed for all IV-administered treatments that vials are shared between patients when necessary. This method is a conservative approach as pembrolizumab has a fixed dose, while the majority of IV administered comparators are based on weight or body surface area (BSA) and would potentially be subject to drug wastage (thus increasing their costs). The administration cost of orally administered therapies was assumed to be zero.

Adverse event costs

The cost of managing Grade 3+ AEs were applied as a one-off cost for patients entering the model. Costs of each AE are provided in Supplementary Materials 7 and were derived from the National Health Service (NHS) reference costs and Personal Social Services Research Unit (PSSRU) costs where possible^37,38. Where no relevant code could be identified, values were taken from published literature and previous NICE technology appraisals.

Testing costs

Testing costs to identify MSI-H/dMMR tumours in patients were applied in the first model cycle to patients in the pembrolizumab treatment arm. Testing costs for polymerase chain reaction (PCR) and immunohistochemistry (IHC) tests were applied for the gastric, small intestine and cholangiocarcinoma tumour sites, assuming 50% of patients currently undergo testing for microsatellite instability or mismatch repair in current clinical practice. NICE currently recommends all patients with CRC receive an IHC test or microsatellite instability test upon diagnosis, and that all patients with endometrial cancer receive an IHC test upon diagnosis. Therefore, additional testing costs for these tumour sites were not incurred.

Testing costs for pembrolizumab per tumour site (Table 2) were calculated as the proportion of patients receiving each test multiplied by the PCR and IHC unit costs sourced from NICE DG27⁴¹. Testing costs were adjusted for the prevalence of MSI-H/dMMR in each tumour site to effectively calculate the unit cost of identifying one additional eligible patient^42,43. Unit costs and the proportion of patients who are tested in current clinical practice are presented in Supplementary Materials 8.

Table 2. Testing costs by tumour site, per treated patient.

Tumour site	Testing costs for pembrolizumab
CRC	£0.00
Endometrial	£0.00
Gastric	£1,028
Small intestine	£1,589
Cholangiocarcinoma	£4,236

Abbreviations. CRC, colorectal cancer; IHC, immunohistochemistry; PCR, polymerase chain reaction.

Note. Unit costs for PCR and IHC testing are £202 and £210, respectively. Costs were inflated to 2021 figures and are presented in Supplementary Materials 8.

Monitoring costs

Monitoring costs were included to reflect the change in resources used by patients on- and off-treatment across model health states by tumour site and treatment arm. These included one-off costs upon entering a health state, and recurring costs applied per cycle. Resource use included initial and follow-up consultations along with monitoring tests, such as computed tomography (CT) scans and blood tests. Healthcare resource use estimates were sourced from previous NICE appraisals using the NHS reference costs, NICE technology appraisals, and PSSRU costs. A summary of healthcare unit costs and sources are provided in Supplementary Materials 9.

Subsequent treatment costs

Subsequent treatment costs were applied as a one-off cost to patients transitioning to the PD health state. The proportion of patients receiving one or more subsequent therapies was informed by the KN-164 and KN-158 trials, as were the duration and distribution of treatments^1,15. Subsequent therapies used in clinical trials that were not reflective of UK clinical practice were identified by clinicians and excluded. It was assumed that the same proportion of patients, regardless of initial line of therapy, would receive subsequent treatment. Total treatment cost for each subsequent treatment was calculated by summing the drug acquisition and administration cost across each subsequent treatment’s duration and dosing schedule. Subsequent treatment distributions and costs are presented in Supplementary Materials 6.

End-of-life costs

The model incorporated the cost of end-of-life (EOL) care, which was applied as a one-off cost to patients upon death. Tumour-site-specific sources were identified where possible. The unit cost applied for CRC, gastric cancer, and cholangiocarcinoma was sourced from Round et al., a 2015 study that estimated the cost of EOL care for patients with lung, breast, colorectal, or prostate cancer, and used data from published literature and publicly available data sets from the UK⁴⁴. The cost of EOL care for endometrial cancer was taken from an economic evaluation of previously treated patients with metastatic MSI-H/dMMR endometrial cancer that cited a US study²⁴. EOL costs for small intestine cancer were found using a study referred by the 2017 NICE TA488 for the submission of regorafenib in previously treated unresectable or metastatic gastrointestinal stromal tumours⁴⁵. EOL costs are presented in Supplementary Materials 11.

Analyses

The model utilizes deterministic and probabilistic analyses. Model results are presented for the entire indication using results for pembrolizumab versus a weighted SoC in each tumour site, weighted by the distribution of patients between the different tumour sites. One-way sensitivity analyses and scenario analyses were conducted to examine the influence of specific inputs and assumptions on cost-effectiveness results.

PSA was conducted with 1,000 iterations to estimate the probability of each treatment being cost-effective under different willingness-to-pay thresholds. In each iteration, inputs were randomly drawn from their specified distributions. Uncertainty parameters or variance–covariance matrices of the selected distributions were based on original data sources; if unavailable, the standard error was assumed to be 10% of the mean value.

A quality-adjusted life year (QALY) weight above 1 would be expected to be applied when disease severity is great enough, as per the latest NICE guidance²⁵. However, our analysis conservatively does not apply a QALY modifier. The impact of this assumption is explored in the Discussion.

Results

Base case results

Pembrolizumab was associated with an increase in both life years (LYs) and QALYs compared with a weighted SoC. Pembrolizumab, at list price, was associated with £129,469 in total costs, 8.30 LYs, and 3.88 QALYs across the pooled tumour sites. SoC was associated with £28,222 in total costs, 1.14 LYs, and 0.72 QALYs across the pooled tumour sites. This yields incremental differences of £101,248 in total costs, 7.16 LYs, and 3.16 QALYs, with an ICER of £32,085 per QALY.

Disaggregated total costs by cost category and disaggregated total QALYs by health state are presented for pembrolizumab and SoC in Tables 3 and 4, respectively.

Table 3. Disaggregated costs by cost category (discounted).

Cost group	Pembrolizumab	SoC	Increment
Treatment costs	£90,978.53	£6,169.48	£84,809.05
Administration costs	£7,399.36	£4,256.31	£3,143.05
Subsequent treatment costs	£4,524.81	£4,932.42	−£407.61
Subsequent administration costs	£1,466.71	£1,602.57	−£135.86
Testing costs	£404.29	£0.00	£404.29
Healthcare resource use costs	£18,401.88	£3,243.66	£15,158.22
End-of-life costs	£6,174.08	£7,558.30	−£1,384.22
Adverse event costs	£119.63	£459.04	−£339.42
Total	£129,469.29	£28,221.78	£101,247.51

Abbreviation. SoC, standard of care.

Table 4. Disaggregated LYs (undiscounted) and QALYs (discounted) by health state.

Health state	Pembrolizumab	SoC	Increment
LYs
Progression-free	3.63	0.49	3.14
Progressed	4.67	0.65	4.02
Total	8.30	1.14	7.16
QALYs
Progression-free	2.00	0.35	1.65
Progressed	1.88	0.37	1.51
Total	3.88	0.72	3.16

Abbreviations. LY, life year; QALY, quality-adjusted life year.

Probabilistic sensitivity analysis results

The probabilistic sensitivity analysis is presented as a cost-effectiveness plane in Figure 4. Uncertainty is presented as 95% confidence ellipses for the overall population and by histology. The probabilistic ICER of pembrolizumab versus the weighted SoC of £31,995 per QALY was very similar to the deterministic ICER.

Figure 4. PSA cost-effectiveness plane with 95% confidence ellipses – pembrolizumab vs SoC. Abbreviations. CRC, colorectal; PSA, probabilistic sensitivity analysis; QALY, quality-adjusted life year; SoC, standard of care.

Small intestine cancer is associated with wide 95% confidence ellipses reflecting the small sample size and immaturity of the observed survival data. Cholangiocarcinoma is similarly associated with uncertainty due to the small sample size; however, observed survival outcomes were relatively more mature, yielding an overall smaller confidence ellipse.

Deterministic sensitivity analysis results

Figure 5 presents the most influential univariate parameters on the ICER of pembrolizumab versus weighted SoC. The parameter that was associated with the greatest uncertainty is the utility value for the progressed disease off-treatment health state for the CRC tumour site from Grothey et al.³⁶ The plausible ICERs ranged from £31,489 to £32,715 per QALY when applying utilities of 0.55 and 0.63, respectively. The upper and lower bounds are each within £630 of the base case ICER. The one-way sensitivity analysis results are also presented in tabular form in the Supplementary Materials 12.

Figure 5. One-way sensitivity analysis tornado plot – pembrolizumab vs SoC, ICER. Abbreviations. CRC, colorectal cancer; HCRU, healthcare resource use; ICER, incremental cost-effectiveness ratio; IV, intravascular; SoC, standard of care.

Scenario analysis results

The scenarios with the greatest impact on the ICER were a time horizon of 10 years, and undiscounted QALYs and costs. Changing these base case values yielded a £15,234 and −£7,324 change in the base case ICER, respectively. The scenario analysis results are presented as a tornado diagram in Figure 6 and in tabular form in Supplementary Materials 13.

Figure 6. Scenario analysis results – pembrolizumab vs SoC, ICER. Abbreviations. AE, adverse event; BHM, Bayesian hierarchical model; ICER, incremental cost-effectiveness ratio; OS, overall survival; PFS, progression-free survival; PSM, parametric survival model; QALY, quality-adjusted life year; SoC, standard of care.

Discussion

This study investigated the cost-effectiveness of pembrolizumab versus therapies currently used in the UK for the treatment of MSI-H/dMMR solid tumours. The model base case suggests that pembrolizumab is associated with an increase in LYs, QALYs, and costs versus SoC, which yields an estimated ICER of £32,085 per QALY gained. A number of important assumptions are applied throughout the analysis, which are likely to impact the resulting cost-effectiveness. The average QALYs for patients receiving SoC across tumour sites (0.72) compared with the expected QALYs for the general population (12.9) results in a proportional QALY shortfall calculation of 94.4%⁴⁶. In line with current NICE methods, this would result in a severity modifier (QALY weight) of 1.2 and would lower the ICER to £26,738. Furthermore, this analysis was conducted considering the publicly available list prices of drugs – that is, without considering any confidential commercial arrangements that may be in place between manufacturers and payers. The results may therefore not be reflective of the true cost-effectiveness of the products to payers. Given pembrolizumab has been approved on multiple occasions by NICE, it is anticipated that confidential pricing arrangements would improve the cost-effectiveness outcomes.

It is important to note that a key challenge of this analysis is that, although pembrolizumab treats a clinically distinct population of patients characterized by expression of MSI-H/dMMR, heterogeneity in health outcomes is observed between patients with tumours at different sites. Furthermore, due to the uncontrolled nature of the KN-158 and KN-164 trials, there is difficulty deriving measures of relative effectiveness.

An additional challenge arises from the limited number of patients observed within the small intestine and cholangiocarcinoma tumour sites. Due to the small sample size, there is a greater degree of uncertainty, therefore potentially making it more difficult to generalize the findings from these populations to a broader demographic. The BHM approach can be used to reduce this uncertainty by effectively allowing for the “borrowing” of information between the included tumour sites; however, it cannot eliminate it entirely. Considering the inherent uncertainties associated with rare histological sites or those with a low prevalence of MSI-H/dMMR, there exists a justification for healthcare decision-makers to exhibit a greater tolerance of uncertainty compared with more prevalent tumour sites. It is also important to note that, when observing outcomes and results for the overall pan-tumour population, small intestine and cholangiocarcinoma sites have a smaller impact compared with other tumour sites, due to the relative proportion within the overall indication.

When developing the structure of the model, two prior NICE technology appraisals were identified of tumour-agnostic therapies, larotrectinib and entrectinib, which both used three-state partitioned survival models^17,18. Following this, Bellone et al. published a cost-effectiveness analysis of entrectinib from the perspective of the Italian National Health Service (Servizio Sanitario Nazionale)⁴⁷. Michels et al. also adapted and published the NICE HTA model for larotrectinib for the Netherlands⁴⁸. Both therapies for NTRK fusion-positive solid tumours were studied in basket trials (i.e. across multiple tumour sites). A key feature of their economic analysis was that data were pooled across tumour sites and analysed as one overall population, failing to evaluate or capture any possible heterogeneity in outcomes between different histological sites. We sought to improve upon this limitation by expanding on recommendations published by Murphy et al.²¹ and developing an approach to extrapolate OS and PFS outcomes within a BHM framework. This approach allowed sharing of information between baskets while generating extrapolated curves that varied by histological site, thus capturing the observed heterogeneity in pembrolizumab survival outcomes. To generate a single cost-effectiveness estimate, tumour-site-specific outcomes were then weighted by the distribution of patients between histological sites to generate an overall weighted ICER for pembrolizumab versus an overall weighted SoC.

Various assumptions may be made about heterogeneity, or the lack thereof, in outcomes between the different tumour types that are represented. There may be complete homogeneity in survival outcomes where it is assumed that outcomes in different tumour sites are equal or that the differences in outcomes between tumour sites may be negligible. Under this assumption, data can be pooled and analysed together, as was previously done in appraisals of larotrectinib and entrectinib^17,18. In our analysis, we assumed that outcomes, or the efficacy of the intervention, are similar (but not equal) across different tumour sites, and that the different histologies do not determine a particular ordering of effectiveness a priori (i.e. the baskets are exchangeable)²¹. Therefore, using a BHM approach was deemed an effective and reasonable approach in this scenario, as such a model is able to capture heterogeneity in survival outcomes between tumour sites and allow information to be borrowed between groups or “baskets” through the use of shared parameters. This method is thought to increase the precision of estimates when compared with analysing individual baskets separately, while also reducing the chances of obtaining implausible estimates for tumour sites represented by few patients^49,50. To our knowledge, this is the first use of BHMs to extrapolate time-to-event outcomes to inform a cost-effectiveness analysis.

An important limitation to address is the challenge of measuring comparator outcomes and the availability of sufficient data necessary for this analysis. The only alternative to the use of surrogacy assumptions was using naïve indirect comparisons to inform comparator health outcomes. Where possible, a matching-adjusted indirect comparison was used to adjust for observed potential confounders and effect modifiers. However, in some tumour sites, the limited patient numbers meant population-adjustment methods were not feasible. Both unadjusted and adjusted indirect treatment comparisons require the assumption of proportional hazards to hold, which was not found to be the case for most comparators in this analysis. Therefore, independent parametric curves were also fitted to the available comparator data. Methods allowing for time-varying hazard ratios were not feasible, given the small patient numbers for each tumour site.

The paucity of data for comparators, especially those evaluated in MSI-H/dMMR populations, and the unclear prognostic effect or level of bias is another limitation worth noting. With the exception of a few studies, SoC outcome data were sourced from patients unselected for MSI-H/dMMR. The impact of this as well as other unknown confounders are likely to bias the results of this evaluation, although the direction of bias remains unknown. In line with this, due to the paucity of TTD data for comparators, TTD was assumed to be equivalent to PFS.

Parameter and structural uncertainty were explored through PSA, univariate one-way sensitivity analysis, and scenario analysis. Cost-effectiveness results were shown to be most sensitive to the utility values by health state for the CRC tumour site and undiscounted costs and QALYs. Overall, the sensitivity and scenario analyses found that, under a range of assumptions, the ICER for pembrolizumab compared with SoC is robust.

Other scenarios including two-piece BHMs and exclusion of testing costs resulted in small reductions in the ICER. In addition, the impact of using standard PSMs to extrapolate pembrolizumab OS and PFS fitted independently by tumour site slightly decreased the ICER.

Several uncertainties remain that were not possible to address within this analysis. Most can either be directly or indirectly attributed to uncertainty and the potential risk of bias associated with the need to derive estimates of relative efficacy using unanchored indirect treatment comparisons (required due to the uncontrolled design of the KN-158 and KN-164 trials). Regulatory agencies and HTA bodies will need to consider such challenges when considering reimbursement strategies for tumour-agnostic therapies.

Conclusion

This economic evaluation is the first to evaluate the cost-effectiveness results of pembrolizumab in this indication. To our knowledge, it is also the first to explicitly capture heterogeneity in survival outcomes between histological sites by extrapolating time-to-event outcomes using parametric survival distributions applied within a BHM framework. Pembrolizumab substantially improves health outcomes at an increased cost, and using the current base case settings, this model suggests that pembrolizumab provides a valuable new alternative therapy for patients with MSH-H/dMMR cancer at the cost of £32,085 per QALY, with confidential pricing arrangements anticipated to improve cost-effectiveness further. The analysis was informed by the best available clinical, HRQL, and cost data for both pembrolizumab and relevant comparators.

Transparency

Declaration of funding

This study was supported by funding from Merck Sharp & Dohme Corp., a subsidiary of Merck & Co., Inc., Kenilworth, NJ, USA.

Declaration of financial/other interests

KY, RX, and MA are employees of Merck Sharp & Dohme LLC, a subsidiary of Merck & Co., Inc., Rahway, NJ, USA. GM and CSH are employees of MSD (UK) Ltd., London, UK. MMW and EB are employees of Lumanity, UK, which received consultancy fees from Merck Sharp & Dohme LLC, a subsidiary of Merck & Co., Inc., Rahway, NJ, USA in connection with this study. KM was an employee of Lumanity, UK at the time of the study. TM is an employee of Lumanity, US, which received consultancy fees from Merck Sharp & Dohme LLC, a subsidiary of Merck & Co., Inc., Rahway, NJ, USA in connection with this study. JM was an employee of MSD, Switzerland, at the time of the study. RAI is an employee of Merck Canada Inc., Kirkland, QC, Canada.

Author contributions

GM, KY, MMW, and TM contributed substantially to the concept and design of the economic evaluation, data analysis and interpretation, drafting of the manuscript, and critical revision of the manuscript for important intellectual content. EB and KM contributed substantially to data analysis and interpretation, drafting of the manuscript, and critical revision of the manuscript for important intellectual content. JM contributed substantially to the statistical analysis and interpretation of the data, and the critical revision of the manuscript for important intellectual content. RX, CSH, and MA contributed substantially to the interpretation of the data, and critical revision of the manuscript for important intellectual content. RA contributed substantially to the concept and design of the economic evaluation, and critical revision of the manuscript for important intellectual content.

Research sponsor

Merck Sharp & Dohme LLC, a subsidiary of Merck & Co., Inc., Rahway, NJ, USA (MSD).

Reviewer disclosures

Peer reviewers on this manuscript have no relevant financial or other relationships to disclose.

Acknowledgements

The authors gratefully acknowledge the support of Jean Muller and Mitashri Chaudhuri of MSD for their support with statistical programming. Howard Thom kindly provided external validation and feedback on the proposed modelling approach and model structure. Gianluca Baio provided technical and methodological input to the design of the Bayesian hierarchical models that informed the cost-effectiveness analysis.