A framework for prediction of progression free survival based on modelling of sub-endpoints
Sreenath M. Krishnan (1), Lena E. Friberg (1), Joakim Nyberg (1), Celine Sarr (1), Ronald Niebecker (2), Shaonan Wang (3), and Alejandro Perez Pitarch (4)
(1) Pharmetheus AB, Uppsala, Sweden (2) Boehringer Ingelheim Pharma GmbH & Co.KG, Biberach, Germany (3) Boehringer Ingelheim (China) Investment Co., Ltd. , Shanghai, China (4) Boehringer Ingelheim Pharma GmbH & Co.KG, Ingelheim, Germany
Introduction: Compared to overall survival, progression free survival (PFS) requires a shorter follow-up and can analyze the efficacy of therapy under investigation without any confounding effects of subsequent lines of treatment. Therefore, it has been increasingly used as a primary endpoint in clinical development to accelerate the time to drug approvals.
The assessment of PFS in solid tumors is based on Response Evaluation Criteria in Solid Tumors (RECIST)[1], which evaluate the following sub-endpoints - the change in the sum of longest diameter (SLD) of target lesions, appearance of new lesion (NL), and progression of non-target lesion (NTR) and death events before imaging-based tumor progression. Traditionally, time to event (TTE) models has been used in describing PFS. TTE models’ development and its application is less efficient, especially in early phases as only a few events are observed. An alternative, potentially more informative and powerful way was suggested by Yu et. al. (2020)[2], which was based on the models for target lesion (SLD) and non-target progressions where, non-target progression merges the data on non-target lesion progression (NTR), new lesion (NL), and death into a single endpoint. The approach by Yu et. al. has the potential to provide predictions of PFS based on the data collected at early stages of drug development.
Aim: The aim of the current study was to develop a modelling and simulation framework based on sub-endpoints, that predicts long-term typical clinical readouts such as PFS, objective response rate (ORR), duration of response (DOR) or best objective response (BOR).
Methods: The data for the model development was available from three Phase III studies in which the safety and efficacy of (a) Afatinib versus chemotherapy in non-small cell lung cancer (NSCLC) [ClinicalTrials.gov Identifier: NCT00949650], (b) Nintedanib along with standard Docetaxel therapy in NSCLC [Identifier: NCT00805194] and (c) Nintedanib [Identifier: NCT01015118] in combination with carboplatin and paclitaxel in advanced ovarian cancer were investigated. The models for sub-endpoints were developed separately before they were combined into a joint model. For the longitudinal SLD data, a tumor growth inhibition (TGI) model or a bi-exponential model was considered while binary NTR and NL data collected at each tumor scan were modeled using a logistic regression model estimating the probability an event at the time of the scan. In case of estimation issues with the logistic regression models, a TTE models were considered for modeling NTR and NL. The death events as well as the events related to dropout from study due to other reasons than progression could happen at any time during the study, hence a TTE model was used in the modeling. In the joint modeling framework, the clinical endpoints including PFS were simulated based on the events in the sub-endpoints’ models. The model evaluations of sub-endpoints as well as the joint model were based on visual predictive checks (VPCs). NONMEM 7.4 software was used for model development. [3]
Results: Population models for sub-endpoints were developed and jointly they described the progression free survival events. A TGI model with an exponential tumor growth function and an arm-dependent shrinkage rate along with the resistance function described the SLD data adequately in all three studies. The NTR and NL data were best described by logistic regression models in study (a) and (b) while in study (c) TTE models gave good fit to the data. The death events were modelled using TTE models with an exponential (study a), Gompertz (study b) or Weibull (study c) distribution of events and the dropouts were described by a Weibull function in all three studies. The VPCs showed a good predictive performance of the models for the sub-endpoints as well as PFS derived using the models for the sub-endpoints. The model framework could also predict ORR, DOR and BOR despite the diversity of responses observed in the 3 Phase III trials.
Conclusion: A framework for prediction of progression free survival based on modelling of sub-endpoints was successfully developed based on Phase III data. When further investigated in Phase I/II settings, this approach may be used for predicting PFS before the PFS data are mature and may provide support for early decision making for indications where PFS is an important clinical endpoint.
References:
[1] Eisenhauer et al., Eur J Cancer. 2009.
[2] Yu et al., CPT Pharmacometrics Syst. Pharmacol. 2020.
[3] NONMEM Users Guides, (1989-2018). Icon Development Solutions, USA.
