Comparison of Cox regression and Artificial Neural Network models in prediction of survival in acute leukemia patients

Cox regression model is one of the most common methods of survival analysis for whose application an assumption of proportional hazards needs to be established. Recently، neural network models without having certain assumptions have been shown to be suitable alternatives in predicting survival. This study aims to compare Cox regression and Artificial Neural Network (ANN) models to predict survival in acute leukemia patients. In the present retrospective study، the information on 197 patients with acute leukemia in Sayyed-O-Shohada Hospital was collected using a checklist. Firstly، the assumption of proportional hazards was tested; Cox regression model was fitted to the observations. To select an efficient ANN to compare with Cox regression model، the number of hidden layer neurons was changed. The prediction accuracy of the two models was compared using receiver operating characteristic (ROC) curve and kappa. Data analysis was performed using SPSS 19، Splus2000، and MatlabR 2009 software packages. Out of 9 ANN models with one hidden layer and 4 to 12 neurons، an ANN with 5 neurons in hidden layer was a superior model compared with Cox regression model. The areas under ROC curve for ANN model and Cox model were estimated to be 0. 0709 and 0. 458، respectively. The accuracies of prediction of survival for ANN model and Cox model were estimated as 78. 9% and 50. 3%، respectively. Due to the high predicting accuracy of ANN models، the use of different models of ANN and their development in various fields of medical science are recommended.