Deep survival forests with feature screening

Abstract

Deep forest has recently received much attention, mainly from the machine learning community, because it has fewer hyper-parameters, better performance, and is easier to construct than deep neural networks. This approach, as well as its variations, are commonly used for image classification, face recognition, music classification, and other related tasks. Its potential in survival analysis, on the other hand, has yet to be determined. In this paper, we propose a deep survival forests framework for modeling right-censored data with high and ultra-high dimensional covariates. We construct the proposed framework by merging the cascade survival forest structure with a feature screening mechanism. Experimental and statistical analysis results on well-known ultra-high and high-dimensional datasets show that the proposed methodology outperforms popular machine learning-based methods such as deep neural survival networks, random survival forests and their several variants in terms of prediction accuracy.