Development of a joint tumor size (TS)–overall survival (OS) modelling and simulation (M&S) framework supporting oncology development decision making
Herbert Struemper, PhD (1), Chetan Rathi, PhD (2), Morris Muliaditan, PhD (3), Sebastiaan C. Goulooze, PhD (3), Alejandro Mantero, PhD (4), Teun M. Post, PharmD, PhD (3), Sandra A. G. Visser, PhD (2)
(1) Clinical Pharmacology Modeling & Simulation, GSK, Durham, NC, USA; (2) Clinical Pharmacology Modeling & Simulation, GSK, Upper Providence, PA, USA; (3) LAP&P Consultants, Leiden, The Netherlands; (4) Disease Area Strategy, Oncology Biostatistics, GSK, Waltham, MA, USA
Objectives:
Early oncology development programs are often limited to base decision making on overall response rate endpoints due to immature OS data, despite the critical importance of the OS endpoint for regulatory pathways. Joint M&S of the individual longitudinal TS and OS has the potential to bridge this gap by establishing the predictive relationship of TS parameters to OS, quantifying the direct impact of patient characteristics in often heterogeneous patient populations on OS, and simulating OS comparisons to control arms or historic controls [1,2,3]. Here we describe the development of a joint TS-OS M&S framework and its evaluation based on immuno-oncology (IO) and chemotherapy responses in patients with non-small cell lung cancer (NSCLC).
Methods:
In this post hoc analysis, the proposed TS-OS model characterizes TS dynamics via a Stein model [4]—with parameters baseline TS (TSb), tumor shrinkage rate (ks) and tumor growth rate (kg)—and links the resulting TS parameters and patient baseline characteristics to an accelerated failure time (AFT) log-normal survival model. Simultaneous joint estimation of TS and OS model parameters was performed in NONMEM 7.5. Various (combinations of) TS link parameters and baseline characteristics were tested in a covariate analysis using stepwise forward inclusion (a=0.01) followed by backward elimination (a=0.001). Goodness of fit plots and visual predictive checks were used to evaluate the TS-OS model.
Results:
The TS-OS dataset included 254 second-line+ NSCLC participants (187 from the feladilimab [FELAD] studies NCT02723955, NCT03693612, and NCT03739710; 67 from the dostarlimab [DOSTAR] study NCT02715284) and 1163 TS observations, including chemotherapy (CHEMO), immune agonist, and checkpoint inhibitor monotherapy and combination arms. The model was iteratively developed with data from the FELAD program and tested on DOSTAR data, followed by a final model update using the combined TS-OS dataset. The TS model rate constants ks and kg were estimated separately for five treatment categories: FELAD, CHEMO, FELAD+CHEMO, DOSTAR, IO-COMBO.
In the final model, the covariates on TS were number of target lesions and hemoglobin at baseline on TSb and alkaline phosphatase at baseline on kg. For OS, the covariates were kg, ks, TSb, age at baseline, albumin at baseline, number of prior lines of therapy, and neutrophil-to-lymphocyte ratio at baseline. The statistically most significant predictor of OS was kg. Any treatment differences in OS were driven by differences in treatment-specific TS parameters (kg and ks); no TS-independent treatment effect on OS was needed.
As part of the model development, we compared our TS-OS M&S framework with a TS-OS modelling framework based on a large set of NSCLC data described in Chan et al [2] and noted specific similarities and differences. Both frameworks use the Stein TS model, which allowed us to test the kg to OS link parameter from Chan et al on our data. Fixing the OS parameters based on the estimates from Chan et al while estimating the TS parameters with a joint TS-OS approach resulted in an adequate fit of the FELAD data and confirmed the applicability of the TS-OS approach across development programs. However, when modelling the combined FELAD and DOSTAR data, we noticed that the inclusion of ks as a TS-OS link parameter notably improved the fits.
The relevance of using a joint TS-OS modelling approach in avoiding bias from informative dropout was confirmed by the two-stage approach under-estimating the kg parameter (between -18% and -57%). In a separate investigation, this model has been applied to extrapolate TS responses to OS with conditional simulations to investigate the impact of various data cut-offs on prediction precision [5].
Conclusions:
The TS-OS model characterized differences in OS across five different treatment categories using the same link functions for the influence of TS parameters and OS in patients with NSCLC, with kg as the most significant predictor. Our work further supports the potential of TS-OS modelling as a tool to leverage early TS data to obtain insights on treatment effects on OS. The TS-OS M&S framework will be applied to predict first-line NSCLC data from the PERLA study (NCT04581824; head-to-head DOSTAR+CHEMO vs pembrolizumab+CHEMO), with plans to further validate and expand the framework with additional data sources.
Funding: GSK [NCT02715284]. Editorial support from Fishawack Health, funded by GSK.
References:
[1] Bruno et al. Clin Cancer Res (2020) 26(8):1787–1795.
[2] Chan et al. CPT Pharmacometrics Syst Pharmacol (2021) 10(10):1171–1182.
[3] Desmée et al. AAPS J (2015) 17(3):691–699.
[4] Stein et al. Clin Cancer Res (2011) 17(4):907–917.
[5] Goulooze et al. Submitted to PAGE 2023.
