Abstract
Tumor size-overall survival (TS-OS) models can support decision-making in oncology drug development by predicting long-term OS based on TS data from early data cuts and baseline patient factors. The current work describes the development of a TS-OS framework capable of predicting OS across a variety of treatment modalities and mechanisms of action in patients with non-small cell lung cancer from seven clinical studies. The presented framework jointly models TS with a bi-exponential Stein model and OS with an accelerated failure time log-normal survival model. In the corresponding link function between TS and OS, the most significant predictor of OS was the tumor growth rate (k(g)), applied via an Emax function. Time to tumor growth and baseline TS were additional TS predictors informing OS. Albumin, total protein, and neutrophil-to-lymphocyte ratio were selected from the tested baseline factors as the most significant predictors of OS. Significant baseline covariates for the TS model included number of target lesions on baseline TS, tumor PD-L1 expression on tumor shrinkage rate, and lactate dehydrogenase levels on k(g). The TS-OS framework model adequately describes the OS distributions within this specific set of treatment modalities-chemotherapies, immuno-oncology treatments, and combinations thereof-using a single treatment-independent link function, supporting the use of the framework to support evaluation and design of future studies. Our findings contribute to a body of literature exploring and qualifying TS-OS modeling as a methodology capable of supporting and accelerating oncology drug development.