The value of surrogate endpoints for predicting real-world survival across five cancer types

Objective It is unclear how well different outcome measures in randomized controlled trials (RCTs) perform in predicting real-world cancer survival. We assess the ability of RCT overall survival (OS) and surrogate endpoints – progression-free survival (PFS) and time to progression (TTP) – to predict real-world OS across five cancers.

Methods We identified 20 treatments and 31 indications for breast, colorectal, lung, ovarian, and pancreatic cancer that had a phase III RCT reporting median OS and median PFS or TTP. Median real-world OS was determined using a Kaplan–Meier estimator applied to patients in the Surveillance and Epidemiology End Results (SEER)-Medicare database (1991–2010). Performance of RCT OS and PFS/TTP in predicting real-world OS was measured using t-tests, median absolute prediction error, and R² from linear regressions.

Results Among 72,600 SEER-Medicare patients similar to RCT participants, median survival was 5.9 months for trial surrogates, 14.1 months for trial OS, and 13.4 months for real-world OS. For this sample, regression models using clinical trial OS and trial surrogates as independent variables predicted real-world OS significantly better than models using surrogates alone (P?=?0.026). Among all real-world patients using sample treatments (N?=?309,182), however, adding trial OS did not improve predictive power over predictions based on surrogates alone (P?=?0.194). Results were qualitatively similar using median absolute prediction error and R² metrics.

Conclusions Among the five tumor types investigated, trial OS and surrogates were each independently valuable in predicting real-world OS outcomes for patients similar to trial participants. In broader real-world populations, however, trial OS added little incremental value over surrogates alone.