In-depth analysis of neural network ensembles for early detection method of diabetes disease

Abstract

Lifestyle-driven disease such as diabetes mellitus has become a serious health problem worldwide. We propose the fusion of neural network-based classifiers, i.e., neural network and support vector machine to assist in early detection of diabetes mellitus. These classifiers are combined to produce the final prediction. However, when considering a number of classifiers in the pool, the selection of combination rule is not easy to understand. The aim of this paper is to investigate the performance of different combination rules, including several single classifiers involved in the ensemble. We use various performance metrics and validation tests to assess the performance of these classifiers using a real-world dataset. Finally, among the classifiers we evaluate their performance differences using statistical significant test. The experimental results indicate that combination rule with average voting scheme is the best performer compared with other combination rules and single classifiers in the ensemble.