Using stacked generalization and complementary neural networks to predict Parkinson's disease

Abstract

This paper proposes the integration between stacked generalization and complementary neural networks to diagnose Parkinson's disease. The Parkinson speech dataset acquired from the UCI machine learning repository is used in our study. Complementary neural networks compose of the truth and the falsity neural networks which are trained to predict the truth output and the falsity output. Stacked generalization consists of two levels. They are level 0 and 1. Ten-fold cross validation is used for training complementary neural networks created in level 0. All outputs produced from each fold are merged to create new input feature. Five sets of machines are trained to create five features which are used as input used to train complementary neural networks created in level 1 of stacked generalization. It is found that the combination between stacked generalization and complementary neural networks provides better performance than using only the traditional stacked generalization or neural network in the prediction of Parkinson's disease.