Comparing relative weight reduction and principal component analysis methods applied to perinatal outcome classification ANN methods

Abstract

This thesis compares relative weight reduction (RWR) and principal component analysis (PCA) variable selection methods applied to perinatal outcome classification artificial neural network (ANN) models. The two perinatal outcomes classified using ANN models are Apgar score at 5 minutes after baby birth and vaginal or cesarean section delivery type from Niday Perinatal Database. The important factors of determining the perinatal outcomes are found separately using RWR and PCA methods applied to ANN tools. It is found that RWR performs better than PCA on specificity measures and it has the advantage of keeping individual indicator data information. The PCA method performs better on sensitivity measures and it is suitable for large input data noise reduction. The Niday 2001 models are also verified in a five member committee of classifiers using Niday 2004 unseen data. The performance measures show that the Niday 2001 models are sufficient in classifying the desired outcomes.