Comparison of Artificial Neural Network and Logistic Regression in Predicting of Binary Response for Medical Data: the stage of disease in Gastric Cancer

Logistic regression is a general model to determine the relationship between covariates and binary response variables. Artificial neural network model is an alternative flexible model which can be used in these cases, too. This study aimed to make a comparison between the predictions of ANN and logistic regression model for binary outcome of medical data.
Material & Data gathered from 639 registered gastric cancer patients between January 2002 and October 2007 at the Research Center for Gastroenterology and Liver Diseases of Shahid Beheshti University of Medical Sciences, Tehran, Iran. Stage of disease was considered as the dependent variable. Network performance was assessed by using of least square error of prediction and then concordance indexes and area under receiver operative characteristic curves (AUROC) were used to comparison of neural network and logistic regression models. Data analysis was performed by R 2.12 software. Results showed that the concordance index of ANN and LR for drug treatment was calculated as 0.771 and 0.710; respectively. In addition AUROC for ANN and LR models were 0.725 and 0.699, respectively. The difference between the values of observed and predicted of the dependent variable by two models was significant (P=0.002). As a result, the total accuracy prediction of the ANN model is better than LR model, so this model is suggested to predict the stage of gastric cancer disease and also diagnostic goals.