Two novel methods for multiclass ECG arrhythmias classification based on PCA, fuzzy support vector machine and unbalanced clustering

In this paper we propose two novel methods of ECG classification to discriminate five heart beat types. The first approach combines principal component analysis (PCA) and modified fuzzy one-against-one (MFOAO) method for multiclass categorization. The fuzzy one-against-one method (FOAO) converts the n-class problem of classification to n(n-1)/2 two-class problems, and performs the binary classification with SVM. It was introduced to solve the problem of the unclassified regions induced by the classical pairwise classification one-against-one. Our novel modified algorithm of FOAO uses fuzzy support vector machine (FSVM) for the binary classification in order to discard outliers. The second approach integrates PCA, unbalanced clustering (UC) and FOAO algorithms. PCA is used to extract the principal characteristics of the signal and reduce its dimension. UC algorithm is used to discard outliers, and reduce the training set by replacing samples with prototypes. The first goal of this work is to compare the ability of the two novel methods to discard outliers and enhance the performance of the classification with PCA and FOAO; the second one is to highlight the efficiency of the combined method PCA-UC-FOAO in the classification of long term ECG records.