Comparison of a new ad-hoc classification method with Support Vector Machine and ensemble classifiers for the diagnosis of Meniere's disease using EVestG signals

In this paper, we compared the performance of our previously designed ad-hoc classifier with Support Vector Machine (SVM) and a family of ensemble learners on classification of patients with Meniere's disease (MD) based on Electrovestibulography (EVestG) signals. The ad-hoc classifier was developed based on the average vote of classifiers each built for a single feature using Linear Discriminant analysis (LDA). The training and test datasets included EVestG signals recorded from 14 MD patients and 16 age-matched healthy controls for training and 9 MD patients and 10 age-matched controls for test dataset. The feature space was built based on the EVestG characteristic signals produced in response to side tilt stimulation. The most discriminative features of the training set were selected using the minimum-redundancy-maximum-relevancy (mRMR) algorithm following one-way analysis of variance (ANOVA). SVM and three ensemble methods, including Bagging, Adaptive Boosting (AdaBoost) and Random Subspace methods were used for comparing the classification performance with that of our ad-hoc voting classifier. The classification results on the test data set showed that the ad-hoc voting classifier outperformed the competitor algorithms in terms of sensitivity, specificity and overall accuracy. The implications of the results are discussed.