Model overview and structure

A three-state PSM was developed in R (version 4.4) using the flexsurv package [11], following ISPOR-SMDM modeling good research practices [12] and reported per CHEERS 2022 [13]. Health states comprised PFS, post-progression survival (PPS), and dead. State membership at each 4-week cycle was defined as: P_PFS(t) = S_rPFS(t); P_Dead(t) = 1 minus S_OS(t); P_PPS(t) = S_OS(t) minus S_rPFS(t), floored at zero. Where S_rPFS(t) exceeded S_OS(t) at any cycle, S_rPFS(t) was set equal to S_OS(t). A half-cycle correction was applied throughout. Time horizon: 40 years (lifetime). Annual discount rate: 3% for both costs and outcomes, per Second Panel on Cost-Effectiveness recommendations [14]. Perspective: US third-party payer (healthcare costs only; no productivity or indirect costs). The analysis was preregistered on OSF prior to any cost-effectiveness calculations (https://doi.org/10.17605/OSF.IO/68R2J). State-occupancy traces for each arm are shown in Fig 1.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Partitioned survival model state-trace plots for capivasertib arm (Panel A) and placebo arm (Panel B) over a 40-year lifetime horizon.

PFS = progression-free survival; PPS = post-progression survival.

https://doi.org/10.1371/journal.pone.0353373.g001

Clinical inputs and survival modeling

Efficacy data were sourced from CAPItello-281 (NCT04493853), published by Fizazi et al. [5]. Individual patient-level survival data were reconstructed fromthe published CAPItello-281 Kaplan-Meier curves for rPFS and OS (Fizazi et al. [5], Figs 2A and 3A) using the Guyot algorithm [15], implemented via the IPDfromKM package in R. Curve coordinates were obtained by automated color-channel (HSV) segmentation of the published figure images using OpenCV, with axis calibration derived from the printed axis ticks, and numbers at risk were transcribed from the published at-risk tables. The full image-to-coordinate-to-pseudo-IPD pipeline is deposited in the OSF project repository (https://doi.org/10.17605/OSF.IO/EWXQT) and reproduces the coordinates used here from the figure images. Reconstruction recovered the published arm sizes exactly (507 and 505). For each curve, the root-mean-square deviation between the reconstructed Kaplan-Meier step function and the digitized input coordinates was 0.01 or less (maximum point deviation 0.10), confirming faithful inversion; reconstructed medians (rPFS 33.3 and 25.7 months) corresponded closely to the published medians (33.2 and 25.7 months), with a maximum deviation of 0.15 months.

Six parametric distributions were fitted by maximum likelihood to each of the four reconstructed datasets (rPFS capivasertib, rPFS placebo, OS capivasertib, OS placebo): exponential, Weibull, log-normal, log-logistic, Gompertz, and generalized gamma. Selection criteria were AIC (primary), BIC (secondary), and biological plausibility, defined as S(40 years) < 0.01. For rPFS, per-arm best-AIC selection was applied; tail plausibility was assessed, but because rPFS curves were subsequently bounded by OS within the partitioned survival model (S_rPFS capped at S_OS each cycle), the best-AIC rPFS capivasertib fit (log-logistic) was retained despite an unconstrained S(40 years) of 0.0243, as its tail is structurally bounded and cannot contribute implausible residual survival. Background mortality was applied as a floor using sex-specific 2022 US male CDC life tables [16], implemented at each cycle as the maximum of the parametric hazard and the age-matched general-population hazard. The modeled cohort entered at the CAPItello-281 median age of 68 years; age advanced by cycle, and the same life-table floor was applied identically to both arms via the OS function. Reconstructed Kaplan-Meier curves with the selected parametric extrapolations overlaid are shown in Fig 2.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2.

Reconstructed Kaplan-Meier curves with parametric extrapolations for rPFS (Panel A) and OS under the proportional hazards base case (Panel B). Capivasertib, green; placebo, red. Solid lines, reconstructed Kaplan-Meier estimates; dashed lines, selected parametric extrapolations to 40 years; dotted vertical line, end of the 60-month observed period.

https://doi.org/10.1371/journal.pone.0353373.g002

For OS, independent per-arm best-AIC fitting produced a clinically implausible inversion driven by a noisy terminal Kaplan-Meier step (n approximately 7 in the capivasertib arm), yielding the capivasertib arm (Gompertz, 40-year restricted mean survival 54.6 months) extrapolating to worse survival than placebo (log-logistic, 40-year restricted mean survival 84.4 months). AIC differences across all OS distributions were less than 1 point (exponential 1404.7, Weibull 1404.4, Gompertz 1404.2), providing no statistical basis for preferring any single family. A proportional hazards base case was pre-specified in the OSF preregistration prior to any ICER calculations: the best-AIC plausible-tail distribution was fitted to the placebo OS arm, and the capivasertib OS curve was derived by applying S_capi(t) = S_placebo(t)^0.90. This approach preserves the trial OS point estimate under a proportional hazards assumption while avoiding the unstable independent-arm tail artifact documented in the preregistration [17,18]. All six single-distribution per-arm OS scenarios plus the per-arm best-AIC artifact scenario are reported for transparency in Table 1.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. OS Extrapolation Scenario Analyses.

https://doi.org/10.1371/journal.pone.0353373.t001

Treatment stopping rules: capivasertib and abiraterone were assumed to be discontinued at rPFS progression. Full dose intensity was assumed for both arms throughout PFS. CAPItello-281 reported that 58% of capivasertib-arm patients remained on treatment for 12 months or longer and 35% for more than 2 years, supporting sustained exposure during PFS; however, dose reductions (32%) and interruptions (65%) were common, so the full-dose assumption is conservative against capivasertib and may overestimate its drug acquisition cost relative to real-world exposure. Patients in the PPS state received post-progression costs only; no active capivasertib was modeled post-progression.

Costs

All costs are expressed in 2025 US dollars. Capivasertib list price was $23,762 per 28-day cycle (64 tablets; 4-days-on/3-days-off schedule), sourced from a publicly available price database [10]. This is a list-price proxy; actual net payer costs may differ due to confidential rebates, which cannot be quantified in the absence of disclosed net price. Generic abiraterone cost was $101 per 28-day cycle (CMS NADAC, April 2026) [19], applied to both arms during PFS. Prednisone 5 mg once daily: $15 per 28-day cycle. Leuprolide acetate 7.5 mg (1-month depot) (ADT backbone, both arms): $176.45 per month (Medicare Part B ASP, HCPCS J9217, Q1 2026) [20], applied to both arms during PFS. The VENTANA PTEN (SP218) RxDx Assay companion diagnostic: one-time cost of $107.55 (CPT 88360, 2026 Medicare Physician Fee Schedule) [21], capivasertib arm only at cycle 1. Post-progression costs for the PPS state (both arms): $9,652 per 28-day cycle ($10,492 per month), derived from the published mean all-cause monthly cost of metastatic castration-resistant prostate cancer of $9,717 per patient-month (2022 US dollars) reported by Kaye et al. [22], inflation-adjusted to 2025 US dollars using the US Bureau of Economic Analysis Personal Consumption Expenditures health care price index (factor 1.0798). Fasting glucose (CPT 82947) and HbA1c (CPT 83036) monitoring was included as a capivasertib-arm cycle-level cost during PFS, reflecting the prescribing information schedule (fasting glucose at weeks 1, 2, 4, 6, and 8 then monthly, and HbA1c every 3 months) [6]; unit costs were 2026 Medicare Clinical Laboratory Fee Schedule allowables (CPT 82947 $3.93; CPT 83036 $9.71; effective 2026-01-01) [21]. Adverse event management costs for grade 3 or higher hyperglycemia, diarrhea, and cutaneous reactions were applied to the capivasertib arm only at incremental rates derived from CAPItello-281 safety data [5] and conservative one-time event-management assumptions, varied in OWSA and PSA; because these costs were small relative to drug acquisition costs, they did not materially affect conclusions.

Health state utilities

Health state utilities were sourced from the EQ-5D-5L analysis of the TITAN trial (apalutamide in metastatic castration-sensitive prostate cancer) by Agarwal et al. [23], a contemporary mHSPC population with comparable disease characteristics. The PFS state utility (0.80; sensitivity range 0.72 to 0.85) was taken as the on-treatment progression-free value reported in TITAN, and the PPS state utility (0.63; range 0.58 to 0.75) as the post-progression (progressed-disease) value reported in TITAN; both were applied equally to both arms in the base case. Adverse event disutilities for grade 3 or higher hyperglycemia, diarrhea, and cutaneous reactions were applied to the capivasertib arm as one-cycle decrements drawn from the standard oncology toxicity-disutility literature (Lloyd et al. [24]; Nafees et al. [25]). A pre-specified exploratory scenario (S2) applied an additional PFS utility decrement of 0.04 to the capivasertib arm to reflect the numerically faster time to clinically meaningful FACT-P physical well-being decline reported in the CAPItello-281 PRO analysis by George et al. [26]; this decrement was not derived from a formal EQ-5D mapping and is labeled exploratory accordingly.

Model validation

Internal validity was confirmed by verifying that all state-occupancy probabilities summed to 1.0 at every cycle, that the PPS fraction remained non-negative after applying the rPFS floor, and that survival functions were monotonically non-increasing. Face validity was assessed by comparing reconstructed rPFS medians with published CAPItello-281 values; reconstructed medians (33.3 and 25.7 months) corresponded to the published values (33.2 and 25.7 months) within 0.15 months. Extreme-value tests were performed by setting all cost parameters to zero and confirming that the model produced zero costs with nonzero QALYs, and by setting OS equal to rPFS and confirming zero PPS occupancy. Background mortality was validated by confirming that projected survival curves did not exceed age-matched US general population survival at any cycle.

Ethics statement

Ethics approval and informed consent were not applicable. This study used only previously published aggregate trial data and publicly available cost data, and no human subjects were involved.

Use of artificial intelligence tools

Two Anthropic tools were used in the preparation of this work: Claude (Opus-class large language model) and Claude Code. Claude Code assisted the author in writing and debugging the R analysis pipeline that the author specified: reconstruction of individual patient data from published Kaplan-Meier curves (IPDfromKM, flexsurv), the partitioned survival model, the probabilistic and one-way sensitivity analyses, the scenario analyses, and figure generation. Claude assisted with manuscript structuring, language editing, consistency checking against CHEERS 2022 reporting items, and reconciliation of arithmetic between the model outputs, tables, and text. The author defined the model structure, all parameter values and their sources, and every analytical decision; no clinical or economic input value was generated or retrieved from the open web by an AI tool. All numerical results were produced by the analysis code, not by a language model, and the author verified every reported figure against the model engine output and every cost input against its primary source. The author takes full responsibility for the content, accuracy, and integrity of the work. The analysis code is openly available in the OSF repository.

Sensitivity and scenario analyses

OWSA varied each parameter individually across its pre-specified range while holding all others at base-case values. A tornado diagram was generated for the 12 parameters with greatest influence on net monetary benefit (NMB) at $150,000 per QALY; NMB was used rather than the ICER because incremental QALYs approach and cross zero under some parameter values, rendering ICERs unstable. Eight pre-specified OS extrapolation scenarios were run in the full PSM (Table 1). Four additional scenario analyses were conducted: S1, brand abiraterone list price (e.g., Zytiga or Yonsa) instead of generic NADAC; S2, exploratory PRO-informed PFS utility decrement of −0.04 in the capivasertib arm; S3, 10-year time horizon; S4, WTP threshold of $200,000 per QALY. Uncertainty in the OS hazard ratio itself was characterized through the eight OS extrapolation scenarios and through lognormal sampling of the trial OS HR (point estimate and 95% CI) in PSA, rather than as a separate one-way scenario. PSA used 10,000 Monte Carlo iterations: drug costs sampled from gamma distributions (CV 0.10); utilities from beta distributions (CV 0.10); AE rates from beta distributions derived from CAPItello-281 confidence intervals; AE costs from gamma distributions (CV 0.20); survival model parameters from multivariate normal distributions on log-scale flexsurv coefficients using the full variance-covariance matrix (preserving within-model parameter correlations); and the OS hazard ratio sampled lognormally from its trial point estimate and 95% CI. rPFS and OS parameter draws were sampled independently across endpoints. The CEAC was generated across a WTP range of $0 to $500,000 per QALY in $10,000 increments. The summary probabilistic ICER is reported as the ratio of mean incremental cost to mean incremental QALY across all iterations, not as the mean of individual iteration ICER ratios. OWSA ranges were chosen to span the plausible uncertainty for each parameter: cost parameters were varied by plus or minus 20% to 50% (wider for less certain inputs such as generic drug cost, narrower for CMS-scheduled costs), utilities across their published confidence limits, the OS hazard ratio across its trial 95% confidence interval (0.71 to 1.15), and the time horizon and discount rate across conventional reference-case bounds (10 years to lifetime; 0% to 5%). A coefficient of variation of 0.10 was applied to cost parameters as a conventional, moderately informative choice reflecting that the central cost inputs are known with relatively high precision; because the incremental cost exceeds the maximum cost-effective cost by more than an order of magnitude, a wider cost distribution does not change the probabilistic result.

Model inputs

Model input parameters are summarized in Table 2. The placebo OS arm was best fit by generalized gamma among plausible-tail distributions after excluding log-normal and log-logistic for implausible 40-year tails; the AIC advantage over Weibull was small (1465.70 vs 1466.75) and BIC differences were negligible, so this residual structural uncertainty was explored directly in the OS extrapolation scenarios (Table 1). AIC and BIC values for all six OS distributions in both arms, and the selected rPFS distributions per arm, are presented in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Model Input Parameters (Base Case).

https://doi.org/10.1371/journal.pone.0353373.t002

Base case

In the base case (Table 3), the capivasertib arm accrued 4.103 discounted QALYs at a total discounted cost of $1,455,507, compared with 3.703 discounted QALYs at $207,230 for the abiraterone comparator. The incremental cost was $1,248,277, driven primarily by capivasertib acquisition costs during PFS. Incremental QALYs were 0.400, yielding an ICER of $3,120,720 per QALY. Net monetary benefit at a WTP threshold of $150,000 per QALY was approximately negative $1,188,000.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Base-Case Cost-Effectiveness Results.

https://doi.org/10.1371/journal.pone.0353373.t003

One-way sensitivity analysis

The tornado diagram (Fig 3) plots incremental NMB at $150,000 per QALY and shows that capivasertib list price was by far the most influential parameter, followed by the discount rate, time horizon, and the OS hazard ratio. Incremental NMB remained substantially negative (below approximately negative $800,000) across the full plausible range of every parameter, including the most favorable value of capivasertib price tested (negative 30%). No single parameter, varied across its full range, produced positive NMB at any threshold up to $500,000 per QALY.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. One-way sensitivity analysis tornado diagram.

Bars show incremental net monetary benefit (NMB) at a willingness-to-pay of $150,000 per QALY as each parameter is varied across its plausible range (parameter ranges in parentheses); the dashed line marks the base-case NMB. NMB is plotted rather than the ICER because incremental QALYs cross zero under some parameter values, making ICERs unstable. Incremental NMB remained negative across the full range of every parameter.

https://doi.org/10.1371/journal.pone.0353373.g003

Overall survival extrapolation scenarios

Across the eight pre-specified OS extrapolation scenarios, capivasertib was either dominated or produced an ICER exceeding $1.35 million per QALY (Table 1). The most favorable scenario was log-normal independent arms (ICER $1,359,445 per QALY). Two scenarios produced negative incremental QALYs, in which capivasertib was dominated: the generalized gamma independent-arms fit and the per-arm best-AIC artifact scenario (Gompertz capivasertib, log-logistic placebo), the latter reflecting the distribution-mixing tail artifact documented in the preregistration and reported here for transparency. The PH base-case ICER of $3,120,720 per QALY was intermediate among the scenarios yielding positive incremental QALYs, consistent with its preregistered role as the methodological central estimate rather than a result-optimizing choice.

Probabilistic sensitivity analysis

In PSA (n = 10,000; Fig 4), the ratio of mean incremental cost to mean incremental QALY was $3,148,744 per QALY. The probability of cost-effectiveness was 0% at WTP thresholds of $100,000, $150,000, $200,000, and $500,000 per QALY (Fig 5). Of the 10,000 iterations, 8,646 (86.46%) fell in the northeast quadrant of the cost-effectiveness plane (higher cost, positive incremental QALY) and 1,354 (13.54%) in the northwest quadrant (higher cost, negative incremental QALY). Because a subset of iterations yielded a negative incremental QALY, the distribution of individual iteration ICER ratios is uninformative, and the cost-effectiveness acceptability curve and plane provide the appropriate characterization of decision uncertainty.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Probabilistic sensitivity analysis cost-effectiveness plane (n = 10,000 iterations).

Incremental cost (y-axis) versus incremental QALYs (x-axis); willingness-to-pay reference lines at $100,000 (dotted orange) and $150,000 (dashed red) per QALY. All iterations lie above the WTP lines (higher cost without cost-effective benefit); approximately 13.54% fall left of the vertical axis, reflecting iterations with negative incremental QALYs under immature-OS sampling.

https://doi.org/10.1371/journal.pone.0353373.g004

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Cost-effectiveness acceptability curve (CEAC) showing the probability that each strategy is cost-effective across willingness-to-pay thresholds from $0 to $500,000 per QALY (n = 10,000 PSA iterations).

Capivasertib had a 0% probability of being cost-effective across the entire range, so its curve lies on the x-axis and the comparator curve lies at 1.0; this reflects that no PSA iteration favored capivasertib at any threshold up to $500,000 per QALY, not a plotting error.

https://doi.org/10.1371/journal.pone.0353373.g005

Scenario analyses

Full scenario analysis results are presented in Table 4. Scenario S1 (brand abiraterone list price instead of generic) produced a minimally different ICER ($3,249,560 per QALY), reflecting the symmetry of abiraterone costs across both arms. Scenario S2 (exploratory PRO-informed PFS utility decrement of −0.04 in the capivasertib arm) worsened the ICER to $5,316,451 per QALY by reducing the incremental QALY gain. Scenario S3 (10-year time horizon) yielded an ICER of $4,426,676 per QALY. Under scenario S4 (WTP threshold of $200,000 per QALY), net monetary benefit remained negative (approximately negative $1.17 million). No scenario brought capivasertib near cost-effectiveness.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Pre-Specified Scenario Analyses. ICERs are calculated from unrounded incremental cost and QALY values; recomputation from the rounded values displayed may differ by up to a few thousand dollars per QALY.

https://doi.org/10.1371/journal.pone.0353373.t004