Use of crowdsourced research to develop a prognostic model for first-line metastatic castrate resistant prostate cancer (mCRPC).

Abstract

180 Background: Project Data Sphere, LLC (PDS) and Sage Bionetworks/DREAM have completed the “Prostate Cancer DREAM Challenge” (Challenge), a crowdsourced competition, using historical prostate cancer clinical trial data from PDS. The Challenge aimed to improve prognostic models for overall survival (OS) and to explore predictive models for treatment toxicity in mCRPC patients. Methods: Control arms of 4 randomized phase III trials (total 2,070 patients) were used as training and validation data sets for the Challenge: ASCENT2, MAINSAIL, VENICE and ENTHUSE33. All subjects were first line mCRPC patients receiving docetaxel treatment. Curated baseline clinical covariates (demographics, comorbidity, prior treatment, laboratory, lesion and vital signs) were modeled along with raw clinical data tables. The primary purpose of the Challenge was to develop a prognostic model for OS (SubChallenge 1). The models were scored using concordance index and integrated area under receiver operator curve (iAUC) from 6-30 months. The published mCRPC OS model of Halabi, et al., JCO, 2014, was used as the benchmark. Results: The Challenge attracted over 160 active participants who formed 50 teams that submitted final models for SubChallenge 1. Median iAUC was 0.76 (0.67-0.78) with a maximum score of 0.792. Over half (n = 35) of these models exceeded the published benchmark (0.743 iAUC). Teams explored new methodologies such as model-based imputation and machine learning techniques to develop the best performing models. Many leveraged raw clinical data sets to create their own covariates and expanded beyond existing prognostic models. Conclusions: The Challenge externally validated Halabi’s first line prognostic model. New prognostic models were proposed and validated with significant improvements over the benchmark. Further analyses are needed to examine the winning models for new prognostic factors and to validate them using additional trial data from PDS. The Challenge drove interest from cross-disciplinary teams of global experts to explore and enhance their technical abilities using real clinical data whilst serving as a vehicle to accelerate medical innovation.