Neural network models of survival analysis data

Abstract

This thesis trains, tests and compares the traditional Cox Proportional Hazard model (Cox PH) and a Cox proportional hazards deep neural network model (DeepSurv) on simulated and real survival data sets. For simulated survival data, two experiments are performed: one with a linear risk function and one with a nonlinear (square) risk function. For real survival data, We use two data sets: 1. an Employee data set with 10,499 observations and a response that indicates the time that an employee spends with a company and 2. a Churn data set with 2000 observations and a response that indicates the number of months until a customer stops doing business with a company. We demonstrate that DeepSurv performs as well or better than the tradi-tional Cox PH model on right-censored survival data with both linear and nonlinear risk functions. We use the concordance index (C-index) and an integrated Brier score to compare the predictive power of Cox PH method and DeepSurv methods on test sets.