Predicting Clinical Outcomes of Ovarian Cancer Patients: Deep Survival Models and Transfer Learning

Abstract

With the advent of high-throughput sequencing technologies, the genomic platforms generate a vast amount of high dimensional genomic profiles. One of the fundamental challenges of genomic medicine is the accurate prediction of clinical outcomes from these data. Gene expression profiles are established to be associated with overall survival in cancer patients, and this perspective the univariate Cox regression analysis was widely used as primary approach to develop the outcome predictors from high dimensional transcriptomic data for ovarian cancer patient stratification. Recently, the classical Cox proportional hazards model was adapted to the artificial neural network implementation and was tested with The Cancer Genome Atlas (TCGA) ovarian cancer transcriptomic data but did not result in satisfactory improvement, possibly due to the lack of datasets of sufficient size. Nevertheless, this methodology still outperforms more traditional approaches, like regularized Cox model, moreover, deep survival models could successfully transfer information across diseases to improve prognostic accuracy. We aim to extend the transfer learning framework to “pan - gyn” cancers as these gynecologic and breast cancers share a variety of characteristics being female hormone-driven cancers and could therefore share common mechanisms of progression. Our first results using transfer learning show that deep survival models could benefit from training with multi-cancer datasets in the high-dimensional transcriptomic profiles.