Deep Neural Network for Cure Fraction Survival Analysis Using Pseudo Values

Abstract

Introduction: The hidden assumption in most of survival analysis models is the occurrence of the event of interest for all study units. The violation of this assumption occurs in several situations. For example, in medicine, some patients may never have cancer, and some may never face Alzheimer. Ignoring such information and analyzing the data with traditional survival models may lead to misleading results. Analyzing long term survivals can be performed using both traditional and neural networks. There has been an increasing interest in modeling lifetime data using neural network due to its ability to handle complex covariates if any. Also, in several numerical results it provides a better prediction. However, for long-term survivors only one neural network was introduced to estimate the uncured proportion together with the EM algorithm to account for the latency part. Neural network in survival analysis requires special cost function to account for censoring.

Methods: In this paper, we extend the neural network using pseudo values to analyze cure fraction model. It neither requires the use of special cost function nor the EM algorithm

Results: The network is applied on both synthetic and Melanoma real datasets to evaluate its performance. We compared the results using goodness of fit methods in both datasets with cox proportional model using EM algorithm.

Conclusion: The proposed neural network has the flexibility of analyzing data without parametric assumption or special cost function. Also, it has the advantage of analyzing the data without the need of EM algorithm. Comparing the results with cox proportional model using EM algorithm, the proposed neural network performed better