Abstract
INTRODUCTION: Longitudinal tumor measurements (TM) are commonly recorded in cancer clinical trials of solid tumors. To define patient response to treatment, the Response Evaluation Criteria in Solid Tumors (RECIST) categorizes the otherwise continuous measurements, which results in substantial information loss. We investigated two modeling approaches to incorporate all available cycle-by-cycle (continuous) TM to predict overall survival (OS) and compare the predictive accuracy of these two approaches to RECIST. MATERIAL AND METHODS: Joint modeling (JM) for longitudinal TM and OS and two-stage modeling with potential time-varying coefficients were utilized to predict OS using data from three trials with cycle-by-cycle TM. The JM approach incorporates TM data collected throughout the course of the clinical trial. The two-stage modeling approach incorporates information from early assessments (before 12 weeks) to predict subsequent OS outcome. The predictive accuracy was quantified by c-indices. RESULTS: Data from 577, 337, and 126 patients were included for the analysis (from two stage IV colorectal cancer trials (N9741, N9841) and an advanced non-small cell lung cancer trial (N0026), respectively). Both the JM and two-stage modeling reached a similar conclusion, i.e. the baseline covariates (age, gender, and race) were mostly not predictive of OS (p-value > 0.05). Quantities derived from TM were strong predictors of OS in the two colorectal cancer trials (p < 0.001 for both association in JM and two-stage modeling parameters); but less so in the lung cancer trial (p = 0.053 for association in JM and p = 0.024 and 0.160 for two-stage modeling parameters). The c-indices from the two-stage modeling were higher than those from a model using RECIST (range: 0.611-0.633 versus 0.586-0.590). The dynamic c-indices from the JM were in the range of 0.627-0.683 indicating good predictive accuracy. CONCLUSION: Both modeling approaches provide highly interpretable and clinical meaningful results; the improved predictive performance compared with RECIST indicates the possibility of deriving better trial endpoints from these approaches.