The effects of data properties on local, piecewise, global, mixture of experts, and boundary-optimized classifiers for medical decision making

Abstract

This paper investigates the issues of data properties with various local, piecewise, global, mixture of experts (ME) and boundary-optimized classifiers in medical decision making cases. A local k-nearest neighbor (k-NN), piecewise decision tree C4.5 and CART algorithms, global multilayer perceptron (MLP), mixture of experts (ME) algorithm based on normalized radial basis function (RBF) net and boundary-optimized support vector machines (SVM) algorithm are applied to three cases with different data sizes: A stroke risk factors discrimination case with a small data size N, an antenatal hypoxia discrimination case with a medium data size N and an intranatal hypoxia monitoring case with a reasonably large data size individual classification cases. Normalized RBF, MLP classifiers give good results in the studied decision making cases. The parameter setting of SVM is adjustable to various receiver operating characteristics (ROC).