Determination of The Neural Network Performances In The Medical Prognosis By Roc Analysis

Abstract

Recently, artificial neural networks are widely used in medical prognosis. The goal of this work is to predict whether a patient will live at least one year after a heart attack by using neural networks as an example of prognosis. With this aim, Multi Layer Perceptrons (MLP), Radial Basis Function Networks (RBF), Probabilistic Neural Networks (PNN), Generalized Regression Neural Networks (GRNN) and Learning Vector Quantization Networks (LVQ) are used. To demonstrate the real performances of the networks, not only classification accuracies but also Receiver Operation Characteristics (ROC) analysis must be investigated. For this purpose, both sensitivity-specificity values and ROC curves are evaluated for all networks.