Performance of SVM Classifiers in Predicting Diabetes

This paper investigates the ability of several models of Support Vector Machines (SVMs) with alternate kernel functions to predict the probability of occurrence of Diabetes (DT) in a mixed patient population. To do this a SVM was trained with 13 inputs (symptoms) from the medical dataset obtained from a university hospital. Different kernel functions, such as Linear, Quadratic, Polyorder (order three), Multi Layer Perceptron (MLP) and Radial Basis Function kernel (RBF) were coded and tested to build the medical diagnosis system (MDS). A detailed database, comprising of healthy and diabetic patients from a university hospital was used for training the SVM for prediction of diabetes. All kernel functions for SVM models showed reasonably good accuracy in prediction of disease (s), with linear kernel structure showing best prediction in 3 out of 4 datasets and Polyorder in one database. Thus the best choice appears to be situation specific.