Performance Evaluation of Extreme Learning Machines Classification Algorithm for Medical Datasets

Abstract

The choice of efficient algorithms is a critical issue in the classification of medical datasets. This requires the consideration of a number of measures to ensure reliable results. In this study, the robustness of Extreme Learning Machine (ELM) and some state-of-arts classifiers were investigated on six (6) different (complete and incomplete) medical datasets. Multiple imputation technique with 5-fold-iteration was used to address the issue of missing data points in datasets with holes. The technique regenerated the missing values 100% in all the datasets. The performance of ELM was compared with Support Vector Machine (SVM), k-Nearest Neighbour (KNN) and Classification and Regression Tree (CART) on the complete and imputed datasets. The evaluations were based on classification accuracy, computational time and stability of the algorithms. ELM has 83.33% overall best accuracy, and 100% best computational time of the simulations. However, the stability of ELM is subject to further improvement, which is an area of further research.