A Cox-optimized survival model based on GrowNet

Abstract

In this study, we propose an improved Cox optimization model – GrowSurv, based on the Cox proportional hazards model (Cox model) and the GrowNet framework. The newly proposed GrowSurv model extends the traditional linear Cox model to study nonlinear problems. When solving the objective function, the new GrowSurv model is less prone to be trapped in a local minimum and requires a much smaller number of hyper-parameters to be adjusted. Furthermore, the new GrowSurv model performs well in high-dimensional datasets. We extensively tested our GrowSurv algorithm on simulated and real datasets and estimated the model performance by five-fold cross-validation. By comparing with other survival models under two metrics the Concordance index (C-index) and Integrated Brier Score (IBS), experimental results show that GrowSurv exhibits good prediction performance through controlled experiments on datasets under different dimensions.

Keywords:

• Survival analysis
• Cox proportional hazards model
• GrowNet
• GrowSurv
• Multiple impulation method

Acknowledgments

The authors would like to thank the editor-in chief and the referees for their careful reading and constructive suggestions to improve the paper.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

All 11 real survival datasets used in our study are available in the R packages and python packages. In detail, Veteran dataset, Gbsg dataset and Mgus* dataset are available in the R package ‘survival’, while Pbc* dataset, PeakVO2 dataset are available in the R package ‘randomForestSRC’, Beer dataset is available in the R package ‘pensim’, while EMTAB386 dataset, GSE32062 dataset and GSE49997 dataset are available in the R package ‘curatedOvarianData’, Metabric dataset is available in the python package ‘pycox’, Breast Cancer dataset can be downloaded from https://github.com/Sage-Bionetworks/predictiveModeling/.

Additional information

Funding

The project is supported by the National Natural Science Foundation of China [grant number 11971214].