Predicting survival in non-small cell lung cancer by preceding tumor assessment data

Abstract

There has been research evaluating the relationship and predictability of progression-free survival (PFS) on overall survival (OS) for treatments of various cancers. Most studies reported low to moderate correlation between PFS and OS. One possible reason for this may be that PFS is time to a categorical outcome (derived by dichotomizing change in tumor size) and doesn't fully capture the correlation between tumor assessment over time as a continuous variable and OS. In this paper, we developed a data-driven model to predict future survival status at time t using both pre-treatment covariates and tumor assessment data (tumor size change from pre-treatment assessment, disease progression status, appearance of new lesions) up to prior time t∗ ( <t ). The method is illustrated with data from a phase III non-small cell lung cancer (NSCLC) trial. In oncology drug development, the primary objective of a typical phase II clinical trial is to evaluate anti-tumor activity of an experimental anti-cancer therapy by measuring whether and to what degree the therapy can lead to reduction of tumor or delay of disease progression (defined as certain degree of tumor growth or appearance of new lesions). To this end, patients enrolled in the trial undergo tumor assessments prior to initiation of treatment and a sequence of scheduled evaluation visits during the trial. Some pre-specified criteria, such as RECIST for solid types of tumor, will then be applied to determine tumor responses by tumor size change over time. If sufficient anti-tumor activity is shown, as demonstrated by a sufficient number of patients experiencing certain percentage of tumor shrinkage or delay of disease progression, then a phase III clinical trial is likely to be conducted to investigate whether the therapy can improve overall survival. This line of drug development relies on an unstated assumption that the anti-tumor effect of a therapy will lead to prolongation of patient's survival time. Whiletherationalebehindthisunderlyingassumptionisthebiologicalmechanismoftumorgrowth,itisofclinical interest to quantitatively assess whether and to what degree the anti-tumor effect is correlated with and predicts overall survival (OS). Research has been done on this topic in various tumor types, mostly focused on relationship between progression-free survival (PFS, defined as survival time without disease progression) and OS. See Buyse et al. (2), Ballman et al. (5), Mandrekar et al. (7). Most of these studies found only low to moderate correlation between PFS and OS. Ballman et al. (5) studied the relationship between PFS rate at 6 months and OS rate at 12 months and also concluded that the relationship was moderate. These somewhat weak results could have been due to the fact that treatment effect on OS is often confounded by post study anti-cancer treatment (PST) whereas treatment effect on PFS is not. Hence the more effective the PST is on OS, the less PFS can be expected to correlate with and predict OS. Another possible reason is that PFS is a dichotomized version of tumor size change over time and hence may not fully capture the correlation between tumor assessment over time as a continuous variable and