(156) A Longitudinal Early-Indicator of Overall Survival Based on Prognostic Scores

Publication Number: 1201

Background: In oncology, the standard efficacy endpoint is Overall Survival (OS). However, in early clinical trial phases, due to limited follow-up time and cohort size, surrogate endpoints (such as Progression-Free Survival) are used as estimates of OS. In many cases though, these surrogate endpoints do not correlate well with OS.

Objectives: To create an early indicator of OS that can reliably estimate efficacy in early clinical trial phases.

Methods: We propose a new method to estimate treatment benefit based on the ROPRO, a state-of-the-art pan-cancer prognostic score, and its machine-learning twin DeepROPRO. These prognostic scores estimate the mortality risk in such a way that a higher risk is correlated with a lower time-to-event. Thus, by measuring the patient's mortality risk over time, we can get an estimate of their reaction to medication. To extensively validate our new method, we extracted a large list of advanced non-small-cell lung cancer (aNSCLC) clinical trials from ClinicalTrials.gov. We emulated these clinical trials using patients from the nationwide Flatiron Health (FH) electronic health record (EHR)-derived de-identified database. From FH, we extracted information on the first line of therapy of 34,061 advanced non-small-cell lung cancer patients, including treatment details, date of death and biomarker measurements. To evaluate the performance of our new methodology, we analyze: 1) the concordance between our new method and OS benefit, 2) whether our method can identify treatment benefit earlier than the log-rank test of OS, and 3) if at early time-points our method can predict the final OS.

Results: We extracted an initial list of 184 clinical trials from ClinicalTrials.gov. From this initial list we excluded 171 trials based on study type (we only included comparative studies of new medications) and population (the population must be similar to that of FH). The final list of 13 recent clinical trials includes multiple types of medication, such as immunotherapy, tyrosine kinase (TK)-inhibitor and platinum-doublet chemotherapy. We emulated these clinical trials using a cohort of aNSCLC patients from FH, obtaining equivalent OS results to the original clinical trials. Our initial analyses showed a high correlation of our new method and OS.

Conclusions: Accurate early OS estimators are warranted, particularly as established endpoints may not always correlate with OS. Using this comprehensive experimental setup, we will present results showing that our new method is a good early-OS estimator. Still, additional analyses are required to evaluate the performance in other cancer indications.