Dynamic predictions of patient survival using longitudinal tumor size in non-small cell lung cancer: Approach towards personalized medicine.

Abstract

e20606 Background: Tumor burden has long been used for the clinical diagnosis, staging, prognosis and treatment of non-small cell lung cancer (NSCLC), as described, for example, in the 7th edition of the AJCC/UICC NSCLC staging guidelines. Previous longitudinal tumor size approaches have used fixed tumor kinetic parameters or tumor shrinkage at a given timepoint, to correlate PFS and OS in a stepwise fashion. Here we describe a joint modeling approach which allows for individual, patient-level predictions of survival during NSCLC treatment. Joint modeling simultaneously fits OS and tumor size dynamics, converting full information from individual tumor assessments into a personalized prediction of survival - thereby avoiding dichotomization of response measure in a patient. Methods: Clinical data from IPASS Phase 3 study of Iressa (gefitinib) in NSCLC were used to fit a joint model of OS and tumor size. The data from a follow-up study (IFUM, Phase 4) for the same drug in a narrower population were used to validate the model on an independent set of subjects. This part included simulating clinical trials from the model and comparing the simulated survival with the observed data. The survival estimation method for individual patients followed from a Bayesian formulation and was implemented in R packages JM and JMbayes. Results: A joint model for overall survival and tumor size was developed and validated using clinical trial simulations. Individual survival estimates were obtained for subjects in a subsequent study based on early data cut-off for tumor assessments. Patient-level predictions were shown to be accurate as well as study-level survival estimates. The model was able to update individual survival predictions in real time. Conclusions: Joint tumor size / survival modeling provides a promising area of investigation for prediction of survival in individual patients. It can be used as a quantitative tool for estimating time-evolving risk of death based on early tumor size measurements. A clinically validated version of such a tool may allow physicians to better choose between treatment continuation and change, following tumor size measurements from standard clinical care.