Trial-level prediction of long-term outcome based on pathologic complete response (pCR) after neoadjuvant chemotherapy for early-stage breast cancer (EBC)

Background The US Food and Drug Administration and European Medicines Agency have published guidance for industry on the use of pathologic complete response (pCR) as a surrogate endpoint to accelerate the regulatory approval of neoadjuvant agents in high-risk early-stage breast cancer (EBC). Three meta-analyses, the CTNeoBC consortium (Cortazar 2014), Berruti (2014), and Korn (2016), evaluated the association between the pCR odds ratio and the event hazards ratio but did not identify strong trial-level associations. Thus, uncertainties remain with respect to whether the magnitude of effect-size increase in pCR reasonably predicts long-term clinical benefit. Findings Trial-level data from CTNeoBC were used as the training data set to derive an empirical nonlinear model for predicting long-term outcomes based on pCR results. A Cox regression model was used to evaluate the relationship among treatments, event hazards, and pCR as joint covariates. The trial-level association between treatment and event hazard was derived and then linked with pCR rates. Magnitude of the patient-level association was also included in the analysis. Additional published trials were used to validate the predictive model. Numerical differences between the perfect surrogate prediction and observed effect followed normal distribution based on the Kolmogorov-Smirnov test. For event-free survival (EFS), the Student t-test P value of 0.02 suggested a statistically significant nonzero difference, with a mean value of −0.163 (SE 0.058). For overall survival (OS), the Student t-test P value of 0.0027 suggested a statistically significant nonzero difference, with a mean value of −0.153 (SE 0.038). Three studies, including GeparSixto, BOOG, and Neo-tAnGo, were used for validation. The F test suggested the model fit the test data well. Implications The observed hazard ratios fit well with the predicted hazard ratios for both EFS and OS and suggest plausible trial-level associations with the new predictor. Major findings Our model predicted the correlation between pCR and EFS as well as OS. This model could be used as a supporting tool to help interpret positive pCR results in neoadjuvant clinical studies in patients with high-risk EBC.