ResSurv: Cancer Survival Analysis Prediction Model Based on Residual Networks

TL;DR

ResSurv is a deep residual network-based model for cancer survival prediction that effectively handles high-dimensional genomics data, overcoming overfitting and degradation issues common in deep Cox models, achieving state-of-the-art results.

Contribution

The paper introduces ResSurv, a novel deep residual learning framework combining Cox proportional hazards with residual networks for improved survival analysis.

Findings

ResSurv effectively extracts high-dimensional features.

ResSurv outperforms existing models in survival prediction accuracy.

Adding normalization layers prevents gradient issues in deep networks.

Abstract

Survival prediction is an important branch of cancer prognosis analysis. The model that predicts survival risk through TCGA genomics data can discover genes related to cancer and provide diagnosis and treatment recommendations based on patient characteristics. We found that deep learning models based on Cox proportional hazards often suffer from overfitting when dealing with high-throughput data. Moreover, we found that as the number of network layers increases, the experimental results will not get better, and network degradation will occur. Based on this problem, we propose a new framework based on Deep Residual Learning. Combine the ideas of Cox proportional hazards and Residual. And name it ResSurv. First, ResSurv is a feed-forward deep learning network stacked by multiple basic ResNet Blocks. In each ResNet Block, we add a Normalization Layer to prevent gradient disappearance and gradient explosion. Secondly, for the loss function of the neural network, we inherited the Cox proportional hazards methods, applied the semi-parametric of the CPH model to the neural network, combined with the partial likelihood model, established the loss function, and performed backpropagation and gradient update. Finally, we compared ResSurv networks of different depths and found that we can effectively extract high-dimensional features. Ablation experiments and comparative experiments prove that our model has reached SOTA(state of the art) in the field of deep learning, and our network can effectively extract deep information.