Contribution of modelling jointly progression-free survival and biomarker longitudinal data for therapeutic evaluation in oncology
Antoine Pitoy (1,2,3), Solène Desmée (2), Hoai-Thu Thai (1), Marc Cerou (1), Dorothee Semiond (4), Christine Veyrat-Follet (1), Julie Bertrand (3)
(1) Translational Disease Modelling Oncology, Sanofi, 91380 Chilly-Mazarin, France. (2) Université de Tours, Université de Nantes, UMR 1246 SPHERE, INSERM, 37000 Tours, France. (3) Université de Paris, UMR 1137 IAME, INSERM, F-75018 Paris, France. (4) Sanofi Translational Medicine & Early Development, Cambridge, MA, USA.
Objectives:
Joint modeling has been increasingly used for therapeutic evaluation in oncology, as it allows the simultaneous fit of longitudinal and time-to-event (TTE) data to characterize and quantify the association between biomarker dynamics and risk of event. Unlike linear models [1] initially used for biomarker kinetics, nonlinear joint models are implemented with more mechanistic models [2], such as tumor growth inhibition models. At the individual level, nonlinear joint models have been shown to improve patient follow-up by providing dynamic predictions [3]. However, the benefit of joint modeling to inform and support decision-making at the population level remains to be assessed. Such models, when developed in early phases of clinical trials, could potentially contribute to an estimate of Phase III clinical trial outcomes at interim analyses, thus enabling earlier decision-making, especially for trials with long duration with TTE as the primary endpoint. The objective of the present simulation study was to evaluate the performance of an approach based on a nonlinear joint model to estimate the primary endpoint at the interim and final analyses of a Phase III oncology clinical trial, compared with a Cox and a survival parametric model.
Methods:
Our simulation setting is inspired by the Phase III ICARIA-MM clinical trial (NCT02990338) that compared progression-free survival (PFS) of isatuximab (anti-CD38 monoclonal antibody) plus pomalidomide/dexamethasone versus pomalidomide/dexamethasone alone in patients with relapsed and refractory multiple myeloma who had received at least two prior lines of treatment [4]. We used a Claret Tumor Growth Inhibition (TGI) model for the biomarker longitudinal data and a log-logistic proportional hazard model for the PFS with slope of the biomarker as link function. We simulated 100 two-arm Phase III clinical trials under the null hypothesis where the treatment arm, “TEST,” did not improve the survival outcome compared with the standard control arm, “REF”. Interim analyses were performed when 33% and 65% of the expected PFS events occurred. The nominal level α was set to 0.025 at the final unilateral analysis (superiority of TEST over REF) and according to three alpha spending functions (Pocock, O’Brien and Fleming, Haybittle-Peto) at each interim analysis [5,6].
Hazard ratios (HRs) and their 1-α% confidence intervals (CIs) were calculated using (i) Cox, (ii) parametric survival model, and (iii) nonlinear joint survival model. For the approach based on nonlinear joint model, the HRs and their 1-α% CIs are obtained through simulating 1000 clinical trials until study completion, accounting for population parameter estimate uncertainty and fitting the latter using Cox regressions. The type I error rates of each method and their 95% CIs were derived at each interim and final analysis.
Results:
For the nonlinear joint survival model, we observed a systematic overestimation of the Claret TGI treatment resistance apparition rate parameter (relative bias of about 23%) at both interim and final analyses; all parameters were estimated with relative bias of <10% and relative root mean square error of <30%. The type one error rates were 0 [0.000-0.036] at the first and second analysis for all approaches, and 0.02 [0.002-0.070], 0.01 [0.00025-0.054] and 0.02 [0.002-0.070] at the final analysis for Cox, parametric survival and nonlinear joint models, respectively. These estimates were not significantly different from the nominal level of the test, regardless of the alpha spending function.
Conclusions:
In this simulation study, we showed that using a nonlinear joint survival model is feasible for an earlier assessment of the outcome of a Phase III clinical trial while controlling for inflation of the type I error. We are currently assessing the power of all approaches to demonstrate the benefit of using biomarker longitudinal data for decision-making.
Funding: Sanofi
References:
[1] Rizopoulos D. Joint Models for longitudinal and time-to-event data: With applications in R. (1st ed.). 2012; Chapman and Hall/CRC.
[2] Desmée S et al. Using the SAEM algorithm for mechanistic joint models characterizing the relationship between nonlinear PSA kinetics and survival in prostate cancer patients. Biometrics. 2016; 73(1):305–12.
[3] Tardivon C et al. Association between tumor size kinetics and survival in patients with urothelial carcinoma treated with atezolizumab: Implication for patient follow-up. Clin Pharmacol Ther. 2019; 106(4):810–820.
[4] Attal M et al. Isatuximab plus pomalidomide and low-dose dexamethasone versus pomalidomide and low-dose dexamethasone in patients with relapsed and refractory multiple myeloma (ICARIA-MM): a randomised, multicentre, open-label, phase 3 study. Lancet. 2019; 394 (10214):2096–2107.
[5] DeMets DL and Lan KK. Interim analysis: the alpha spending function approach. Stat Med. 1994; 13 (13-14):1341–1352; discussion 1353-1356.
[6] Peto R et al. Design and analysis of randomized clinical trials requiring prolonged observation of each patient. I. Introduction and design. Br J Cancer. 1976; 34(6):585–612.