This study was based on previously collected, anonymized clinical trial data from KEYNOTE-826 provided by the study sponsor; as such, it was exempt from institutional review board or ethical approval. The KEYNOTE-826 trial was approved by the appropriate ethics body at each participating center [6, 7].

We built a cost-effectiveness model in Microsoft Excel with a 1-week cycle length and 50-year time horizon. The model captures lifetime effects and costs, given an average age of 51.0 years at diagnosis. The cycle length allowed accurate accounting for the dosing schedules of included interventions and avoided material half-cycle correction. We adopted a US healthcare perspective and followed relevant guidelines, including the Consolidated Health Economic Evaluation Reporting Standards (CHEERS) 2022 statement [10,11,12,13].

In KEYNOTE-826, progression-free survival was seen to plateau for patients receiving pembrolizumab + chemotherapy. This resulted in scenarios where the extrapolations of progression-free survival and overall survival crossed in the partitioned survival model, as this approach includes no structural link between these endpoints. As the model structure was anticipated to create uncertainty around the results, we followed best practice and developed both a state-transition and partitioned survival analysis approach [14]. We used the state-transition approach in our base case, leveraging KEYNOTE-826 individual patient-level data to derive time-to-progression and post-progression survival estimates to inform transition, alongside progression-free and overall survival (Fig. 1). The partitioned survival approach is considered to quantify structural uncertainty [15].

Three-health state semi-Markov state-transition model. Numbers in square brackets indicate transition probabilities, defined in the figure. PFS(t, arm), TTP(t, arm), and PPS(t, arm) are the survivor functions for progression-free survival, time to progression, and post-progression survival, respectively, at time t and the given treatment arm. PFS progression-free survival, PPS post-progression survival, TTP time to progression

The population modeled was patients with advanced cervical cancer and a combined positive score ≥ 1 [16]. This population accounted for 89% of the patients in KEYNOTE-826 [6]. KEYNOTE-826 was designed to evaluate pembrolizumab + chemotherapy (paclitaxel + cisplatin or paclitaxel + carboplatin) ± bevacizumab versus chemotherapy ± bevacizumab in women with cervical cancer [6].

Within the model, dosing for pembrolizumab + chemotherapy is aligned with KEYNOTE-826 and the FDA license [16]. Pembrolizumab is given until progression for up to 2 years or 35 treatment cycles; hence, a 105-week stopping rule was applied. Except for bevacizumab, which has no stopping rule, chemotherapy in both arms is given until progression for a maximum of six treatment cycles; hence, an 18-week stopping rule was applied for the concerned regimens.

Subsequent treatment was modeled as per KEYNOTE-826: 48% of patients who progressed in KEYNOTE-826 received second-line chemotherapy, which aligns with US real-world data [3]. For simplification, the model only included second-line treatment costs administered to ≥ 3% of patients in either arm of KEYNOTE-826.

Cost-effectiveness was assessed in terms of costs per life-year and quality-adjusted life-year (QALY) gained. Consistent with the current Institute for Clinical and Economic Review Value Assessment Framework, a US $150,000/QALY willingness-to-pay threshold was used [10].

The median follow-up period in KEYNOTE-826 was 28.6 months; therefore, it was necessary to perform extrapolations to estimate long-term outcomes for a lifetime time horizon of up to 50 years. Standard parametric survival models (exponential, Weibull, log-normal, log-logistic, Gompertz, and generalized gamma) were fit to the full individual patient-level data separately for pembrolizumab + chemotherapy and chemotherapy (Supplementary Material 1). More flexible methods were explored if one-piece models did not provide a reasonable fit to the patient-level data. These methods included “two-piece” models with the Kaplan–Meier curve followed by a parametric survival model fit to the data from a certain time point onwards, and Royston–Parmar spline models with up to three knots [17]. Curve selection was based on visual fit to patient-level data, the clinical plausibility of long-term extrapolations and hazard functions, and the statistical fit to the patient-level data. Data from the Gynecologic Oncology Group (GOG)-240 trial were used to validate extrapolations [18].

Time-to-progression and progression-free and post-progression survival inputs were informed by data from the primary KEYNOTE-826 analysis. In the base case, three-knot spline models were used for time to progression and progression-free survival. Deaths were recorded as an event for progression-free survival but as a censoring event for time to progression. Therefore, progression-free survival is the time to a composite endpoint of progression or death, whereas time to progression is based on disease progression alone. The flexible spline model follows the change in hazard, meaning it captured the plateau in pembrolizumab patient survival well. The model assumptions and selection process are described in Supplementary Materials 1 and 2, respectively.

As the number of deaths in KEYNOTE-826 before progression was small, pre-progression mortality was estimated from time to progression and progression-free survival using the formula

$$}\left( }t}t + 1} \right) = S_}}} \left( \right)/S_}}} \left( t \right) \, \, S_}}} \left( \right)/S_}}} \left( t \right),$$

where Sx(t) is the survival function for endpoint(x) at time(t).

To avoid clinically implausible pre-progression mortality estimates in the cost-effectiveness model, the same modeling approach and specification were selected for progression-free survival as for time to progression (Fig. 2).

Kaplan–Meier curves and long-term extrapolations of survival outcomes from KEYNOTE-826. OS overall survival, PEM pembrolizumab, PFS progression-free survival, PPS post-progression survival, pSoC pembrolizumab + standard-of-care chemotherapy, SoC standard of care, TTP time to progression

In the base-case analysis, post-progression survival data from KEYNOTE-826 were extrapolated using generalized gamma models fit to all Kaplan–Meier data (Supplementary Material 1).

When using the partitioned survival model approach, progression-free survival was modeled using the same three-knot spline as the state-transition model base case. Similarly, overall survival was modeled using three-knot spline models.

In the base case, we included grade ≥ 3 drug-related adverse events that occurred in ≥ 5% of patients in either arm, including anemia, neutropenia, urinary tract infection, hypertension, thrombocytopenia, febrile neutropenia, and decreased platelet, white blood cell, and neutrophil counts. The mean duration of adverse events was obtained from KEYNOTE-826.

The European Quality of Life 5-Dimension 5-Level version instrument was used to measure generic health status in patients enrolled in KEYNOTE-826. We used patient-level data from KEYNOTE-826 and the US-specific value set to calculate utilities [19].

We computed utilities by time to death to capture how pembrolizumab impacts quality of life (QoL) in patients with advanced cervical cancer. Time-to-death utilities were deemed more appropriate than utility inputs stratified by progression status because “pseudoprogression” issues can be encountered with immuno-oncology drugs, where the action of treatment may be mistaken for disease progression. They also provide finer gradations in utility across patients compared with progression-based utilities [20]. In the base case, we applied a regression model with time-to-death categories and grade ≥ 3 adverse events to predict treatment-independent utilities for use in the analysis (Supplementary Material 1). Utilities in the model were capped using US-specific data on general population utility [21].

The analysis considered drug acquisition (wholesale acquisition cost) [22], treatment administration, PD-L1 marker testing, adverse events, resource use, and end-of-life costs. Costs were adjusted to 2022 US dollars using the US Bureau of Labor Statistics, Consumer Price Index, and Medical Care inflation calculator. Costs and QALYs were discounted at 3% per year.

Time on treatment was based on KEYNOTE-826 patient-level data. Extrapolation was not needed for pembrolizumab, paclitaxel, cisplatin, or carboplatin, as the clinical trial covered the maximum duration these treatments were given in both arms. For bevacizumab, time-on-treatment patient-level data were extrapolated using a two-piece log-logistic model that was fit from 46 weeks onward (Supplementary Material 1). We sourced wholesale acquisition cost prices from the AnalySource database [22]. Treatment administration, PD-L1 testing, and other resource use costs were sourced from the Centers for Medicare and Medicaid Services [22, 23]. The cost of PD-L1 testing reflected the cost per PD-L1-positive patient identified rather than the cost per test, with an average of 1.1 tests required per positive patient identified. Clinical experts informed resource use frequencies in the pre- and post-progression health states, assuming no differences between treatment arms. Adverse events were costed using Medicare severity diagnosis-related groups [10]. We retrieved end-of-life costs from the literature [24]. The tested scenarios, model assumptions, and inputs are reported in Supplementary Material 1.

We performed one-way sensitivity analyses to explore the impact of uncertainty in individuals’ input parameters on results and a probabilistic sensitivity analysis of 5000 iterations to explore the combined uncertainty of all inputs on results.

Multiple scenario analyses were conducted to investigate various modeling assumptions (Supplementary Material 3).

A key scenario was assessing the structural uncertainty introduced by the choice of modeling approach—state-transition or partitioned survival. The partitioned survival model was parameterized using the same inputs as the state-transition model in every plausible instance. The only difference was modeling survival using an extrapolation of overall survival data from KEYNOTE-826. Overall survival was extrapolated using a Royston–Parmar spline model that had three knots.

Treatment options in second line are rapidly evolving, and real-world data suggest that tisotumab vedotin and pembrolizumab are widely administered at second line in this US subpopulation [25]. Therefore, we performed a scenario analysis where tisotumab vedotin and pembrolizumab were added to subsequent treatment options. Inputs for this scenario were clinically informed and resulted in the following assumptions: in the pembrolizumab arm, 60% of progressed patients receive tisotumab vedotin, 20% pembrolizumab, and 20% chemotherapy. In second-line chemotherapy, we assumed that 75% received tisotumab vedotin and 25% pembrolizumab. We evaluated results separately in patients who received bevacizumab and those who did not using a 90-day friction-cost method [26].

We also explored pembrolizumab’s societal impact by including productivity gains. These were calculated using inputs from Supplementary Material 1 and assume that patients with progressed disease no longer work, while those who are progression-free work part-time.