Deep-learning/transfer-learning based Overall Survival prediction conditional on Progression-Free Interval with TCGA RNA-seq expression and KEGG-pathways.

We developed a novel deep-learning network that integrated Progression-Free Interval (PFI) as a condition based on transfer learning and fine tuning, utilizing pathway-gene (KEGG) relationship to predict Overall Survival (OS) for The Cancer Genome Atlas (TCGA) RNA-seq expression data. In practice, the acquisition of clinical and experimental data can be costly, which generally results in a scarcity of data. To accommodate scenarios with low sample sizes for a tumor type, we pre-trained the network with the data from thirty-one other tumor types. We evaluated the performance of our method for ten tumor types in TCGA, conducting a comparison analysis against several existing methods, including meta-learning, Cox-lasso, Cox-ridge, Cox-ElasticNet, DeepSurv, DeepHit, Nnet-survival, Cox-Time, Cox-CC and PASNet. For nine of ten tumor types (excluding LGG), our method significantly outperformed the other methods in terms of C-index and Integrated Brier Score. For LGG, our method demonstrated a comparable performance. Furthermore, our results could be used to investigate the temporal changes for the association between KEGG pathways and OS. Many well-known pathways, such as circadian rhythm and DNA replication pathways, were commonly identified.