A comparison of two stage and joint tumor growth inhibition-progression free survival modeling approach to simulate clinical outcome in oncology
Laurent Claret (1), Rene Bruno (1), Bei Wang (2), Mathilde Marchand (1), Chunze Li (2), Sandhya Girish (2), Jin Jin (2), Angelica Quartino (2)
(1) Pharsight Consulting Services, Pharsight, a CertaraTM Company, Marseille, France (2) Genentech Research and Early Development (gRED), Roche, South San Francisco, CA
Objectives: In anticancer drug development, tumor growth inhibition (TGI) metrics have been increasingly used to predict clinical outcomes: overall survival (OS) or progression free survival (PFS) [1]. This approach involves two step modeling (TSM): 1) a longitudinal TGI model to estimate TGI metrics and 2) a time to event model to link TGI to clinical outcome. This TSM approach has been criticized [2] because the uncertainty of the individual TGI metrics estimations is not carried forward in the time to event model likelihood, and it is proposed to use a joint model (JM) fitting simultaneously both dependent variables. If JM cannot be easily applied for OS as longitudinal and time to event data are not recorded simultaneously and TGI models cannot be extrapolated after treatment stop, it can be evaluated on PFS.
Methods: TSM and JM were evaluated on a study comparing two treatments based on PFS hazard ratio (HR) in cancer patients. TSM and JM were implemented in NONMEM 7.2. TGI data are described by a bi-exponential model [3] implemented in a non-linear mixed effect approach and PFS data by a Weibull distribution. PFS models are evaluated in their abilities to simulate distributions and treatment HR.
Results: In the TSM approach the Weibull PFS distribution is defined by patient characteristics and TGI model parameters [3]. The model is qualified to predict PFS in patient sub-groups and treatment HR. The JM model development revealed that TGI metrics provided better fit than hazard as function of longitudinal baseline adjusted tumor size prediction and a Weibull distribution of the intercept was better than a constant one. The final JM got the same structure. Model parameter point estimates and posterior predictive distributions of JM and TSM were similar. PFS model parameters were estimated slightly more precisely with JM.
Conclusions: This comparison, based on an analysis of single study, shows that despite a slightly better precision of JM model parameter estimates, structural models and prediction performances were similar for both approaches. TGI metrics were better than tumor size as function of time to drive PFS hazard function. JM model can be used to predict PFS based on TGI profiles. Ultimately, the goal of the model will have to be accounted for e.g. develop the best model to simulate alternative clinical trial designs, assess an early marker of effect to support decisions or predict individual patient outcomes to support therapeutic decisions.
References:
[1] Bruno R et al. Clin Pharmacol Ther. 95, 386-93, 2014
[2] Mansmann UR et al. J Clin Oncol. 31, 4373, 2013
[3] Stein WD et al. Oncologist 13, 1046–1054, 2008
