BroadSurv: A Novel Broad Learning System-based Approach for Survival Analysis

Abstract

Survival analysis (time-to-event analysis) is a set of statistic methods to analyze time-to-event data and is widely used in many fields such as economics, finance and medicine. One of the fundamental problems in survival analysis is to explore the relationship between the covariates and the survival time. Recently, with the development of deep learning-based techniques, various approaches have been proposed for survival analysis. To better handle the censoring, special cost functions or sophisticated network structures are usually designed for these methods. In this paper, a novel two-stage method is proposed to model the survival data. In the first stage, pseudo conditional probabilities are computed, which can act as the quantitative response variables in regression problems. In the second stage, with these pseudo values, a complicated survival analysis problem is transformed into a regression problem that can be effectively solved by broad learning system. The experimental results show that, with a flexible structure and a simple cost function, our proposed method has a better performance in handling the censored problems.