Conditional overall survival (OS) simulations with a joint tumor size (TS)-OS model to support oncology development decision making
Sebastiaan C. Goulooze, PhD (1), Morris Muliaditan, PhD (1), Alejandro Mantero, PhD (2), Sandra A. G. Visser, PhD (3), Teun M. Post, PharmD, PhD (1), Chetan Rathi, PhD (3), Herbert Struemper, PhD (4)
(1) LAP&P Consultants, Leiden, The Netherlands; (2) Disease Area Strategy, Oncology Biostatistics, GSK, Waltham, MA, USA; (3) Clinical Pharmacology Modeling & Simulation, GSK, Upper Providence, PA, USA; (4) Clinical Pharmacology Modeling & Simulation, GSK, Durham, NC, USA
Objectives:
TS-OS models characterize the relationship between TS dynamics and OS in oncology [1,2] to support drug development decision making. These models can be used to generate conditional simulations of long-term OS, fully accounting for and extrapolating early TS and OS data. Here we retrospectively investigated the impact of data cut-off time (i.e., interim data length of follow-up time to inform the conditional simulations) on the quality of conditionally simulated OS. This research can inform clinical development to allow robust OS predictions that can be used for decision making.
Methods:
This post hoc analysis used the final dataset for the single-arm GARNET study (NCT02715284) with ~4-year follow-up of dostarlimab in non-small cell lung cancer patients. Hypothetical ‘interim datasets’ were created with varying follow-up times by censoring at cut-off dates (1, 1.25, 1.5 or 2 years after study start). Cut-off dates <1 year after first patient first visit were not explored because enrollment lasted ~1 year. To each artificial ‘interim’ GARNET dataset, a TS-OS model was fitted in which TS parameters were estimated based on the interim data, while OS parameters (including the impact of TS parameters on OS) were fixed on the estimated value of the TS-OS model developed on feladilimab data [3]. This corresponds to a prospective application in which immature OS data at hand are not sufficient to re-estimate OS parameters. The TS-OS model consisted of a log-normal OS distribution, with a Stein model to characterize the TS dynamics over time [3,4].
Conditional simulations of the long-term OS
For each patient and simulation repetition, samples were generated from the patient’s individual uncertainty in post hoc TS parameters. Based on these conditionally sampled TS parameters, the individual OS distribution was obtained. A sample from the individual OS distribution was taken, conditional on the available interim OS data that the TS-OS model was based on: patients that had already experienced the OS event in the interim dataset were not re-simulated, and OS events would be sampled only in the future (i.e., if an individual already survived for 200 days in the interim dataset, the simulated OS event for that individual always occurred after 200 days). Simulated OS events after 4 years were treated as censored at 4 years in line with GARNET follow-up. The 95% prediction interval from each conditional simulation was visually compared to the observed 4-year OS data using a Kaplan-Meier survival plot.
Results:
Compared to a model estimated on the full GARNET dataset, the TS-OS models fitted to the ‘interim’ datasets had higher estimates for the TS parameters tumor growth rate (kg) (+14–39%) and tumor shrinkage rate (ks) (+28–65%) for Dostarlimab. This was believed to be related to the behavior of the Stein model, where long-term stable TS levels can be ‘fitted’ only with equally low values of kg and ks.
Despite the difference in TS parameters when ‘censoring’ the TS and OS data, the conditional simulations well predicted long-term OS for all cut-off dates (ranging from 1–2 years after study start for a 4-year total study duration). This is likely because both kg and ks were included as predictors of OS (with opposite direction of effect) in the TS-OS model, resulting in a similar prediction of OS with dostarlimab despite lower kg and ks. Removing ks from a TS-OS model applied to the full GARNET dataset causes over-estimation of the predicted OS over time.
Conclusions:
Conditional simulations using a TS-OS model can leverage early TS and OS data from an ongoing study to forecast OS at the study end. We obtained robust OS predictions at 1 year post study start (i.e., around the time of last patient first visit) in GARNET. We identified a possible dependence of the Stein model on the follow-up time, likely related to parameter values that were needed to fit the long-term stable TS profiles observed in some patients in GARNET. Although kg was the most significant predictor of OS, the inclusion of ks as a predictor for OS appeared necessary to maintain the long-term predictive performance of the TS-OS model for early data cuts of the GARNET study. Taking study-specific enrollment rates into account, the presented simulation framework could be extended to support minimal follow-up required to obtain OS predictions.
Funding: GSK [NCT02715284]. Editorial support from Fishawack Health, funded by GSK.
References:
[1] Chan et al. CPT Pharmacometrics Syst Pharmacol (2021) 10(10):1171–1182.
[2] Desmée et al. AAPS J (2015) 17(3):691–699.
[3] Struemper et al. Submitted to PAGE 2023.
[4] Stein et al. Clin Cancer Res (2011) 17(4):907–917.
