Comparison of artificial neural network and Cox regression models in survival prediction of gastric cancer patients

authors:

avatar Akbar Biglarian 1 , avatar Ebrahim HajiZadeh ORCID 1 , * , avatar Anoshirvan Kazemnejad ORCID 2

Dept. of Biostatistics, Faculty of medical science, Tarbiat Modares University, Tehran, Iran.
Dept. of Biostatistics, Faculty of medical science, Tarbiat Modares University, Tehran, Iran

how to cite: Biglarian A, HajiZadeh E, Kazemnejad A. Comparison of artificial neural network and Cox regression models in survival prediction of gastric cancer patients. koomesh. 2010;11(3):e153802. https://doi.org/10.5812/koomesh-153802.

Abstract

Introduction: Cox regression model is one of the statistical methods in survival analysis. Proportionality of hazard rate is an assumption of this model. In the recent decades, artificial neural network (ANN) model has increasingly used in survival prediction. This study aimed to predict the survival probability of Gastric cancer patients using Cox regression and ANN models. Materials and Methods: In this historical-cohort study, information of total of 436 gastric cancer patients with adenocarcinomas pathology who underwent surgery at the Taleghani hospital of Tehran between 2002 and 2007 were included. Data were divided to training and testing (or validation) groups, randomly. The Cox regression model (semi-parametric model) and a three layer ANN model were used for analyzing of database. Furthermore, the area under receiver operating characteristic curve (AUROC) and classification accuracy were used to compare these models. Results: Prediction accuracy of ANN and Cox regression models were 81.51% and 72.60%, respectively. In addition, AUROC of ANN and Cox regression models were 0.826 and 0.754, respectively. Conclusions: ANN was better than Cox regression model in terms of AUROC and accuracy of prediction. Therefore, ANN model is recommended for prediction of survival probability. These finding are very important in health research, particularly in allocation of medical resources for patients who predicted as high-risks