Prediction of Bladder Cancer Prognosis by Deep Cox Proportional Hazards Model Based on Adversarial Autoencoder

Due to the highly aggressive and heterogeneity of bladder cancer, it is important to predict bladder cancer prognosis accurately and identify the survival-related biomarkers. Recent studies have increasingly concentrated on exploring the influence of genes on cancer prognosis. Research indicates that how to deal with the obstacles caused by the redundant genes in genetic data for prognosis remains a challenge. To address this problem, in this study we propose a deep network (AAE-Cox) to accurately predict bladder cancer prognosis, which is designed by combining an improved adversarial autoencoder (AAE) with a deep Cox proportional network. Specifically, on the basis of normal AAE model, we maintain adversarial training while removing the decoder, to use the improved model to better gene extraction in the code space. The improved AAE model was then combined with a deep Cox network to achieve higher prognostic accuracy. Comprehensive experiments reveal the success of our approach in significantly enhancing the prognosis prediction accuracy for bladder cancer, and ablation experiment verifies the effectiveness of the improved model. Moreover, our method outperforms the existing advanced techniques in this domain.