An Ensemble Approach to Predicting Overall Survival (OS) Benefit Jointly From Multiple Surrogate Endpoints (SE) in Oncology: A Case Study in Previously Untreated Metastatic Melanoma

OBJECTIVES: In absence of OS data from randomized controlled trials (RCTs), SE-based predictions for OS benefit facilitate early assessment of treatments which may accelerate drug development. Surrogacy studies conventionally evaluate a single SE. We devised an approach to blend in the available predictive information from multiple SEs to improve the accuracy in prediction of OS benefit.

METHODS: The evidence base consisted of 24 RCTs in previously untreated metastatic melanoma identified by a systematic review. SEs were progression-free survival (PFS), objective response rate (ORR), and complete response rate (CRR). Given predictions on OS hazard-ratios obtained from bivariate random effects meta-analyses individually for each SE in each RCT, we developed an optimization model to elicit the weights to be assigned to the predictions of each SE. In the model, weights assigned to the predictions of SEs indicated their exclusive probability of correctness relative to the predictions from other SEs, and the objective was to minimize the gap between the observed and predicted OS hazard-ratios across all RCTs using Euclidean and supremum norms. Improvements in maximum standard error around the predicted log-transformed OS hazard-ratios by the ensemble approach versus individual SEs were reported as performance measures.

RESULTS: The optimized weights for the predictions from the hazard-ratios of PFS, and the odds-ratios of ORR and CRR were 0.254, 0.446, and 0.3, respectively under the Euclidean norm, and 0.606, 0, and 0.394, respectively under the supremum norm. Average absolute improvement in the maximum standard error was 0.039 and 0.044, respectively, under the Euclidean and supremum norms which respectively corresponded to 14.5% and 18.6% relative improvements in the maximum standard error.

CONCLUSIONS: Our tractable and scalable ensemble approach can consolidate the predictions from multiple SEs to maximize the utilization of available intermediate endpoint data while reducing uncertainty around the predicted OS benefit borne by choosing a single SE.