In this paper, we present our winning method for survival time prediction in the 2015 Prostate Cancer DREAM Challenge, a recent crowdsourced competition focused on risk and survival time predictions for patients with metastatic castration-resistant prostate cancer (mCRPC). We are interested in using a patient's covariates to predict his or her time until death after initiating standard therapy. We propose an iterative algorithm to multiply impute right-censored survival times and use ensemble learning methods to characterize the dependence of these imputed survival times on possibly many covariates. We show that by iterating over imputation and ensemble learning steps, we guide imputation with patient covariates and, subsequently, optimize the accuracy of survival time prediction. This method is generally applicable to time-to-event prediction problems in the presence of right-censoring. We demonstrate the proposed method's performance with training and validation results from the DREAM Challenge and compare its accuracy with existing methods.
Ensemble learning of Multi-View CNN models for survival time prediction of brain tumor patients using multi-modal MRI scans