Heartbeats Arrhythmia Classification Using Probabilistic Multi-Class Support Vector Machines: A Comparative Study

The support vector machines were originally created to classify binary problems. Their extension for multiclass problems was the subject of several researches. Usually, a multiclass classifier is obtained by combining several binary classifiers. During the last years, the attention is focused on four main models of Multi-class Support Vector Machines (M-SVM), which consider all classes simultaneously: Weston and Watkins (WW) [1] model, Crammer and Singer (CS) [2] model, Lee, Lin and Wahba (LLW) [3] model and the quadratic loss multi-class support vector machines (MSVM) [4] model introduced by Guermeur and Monfrini. This study aims to develop a new method based on four M-SVMs models to classify seven different arrhythmia, plus normal ECG obtained from the Physiobank database [5]. The four models work separately, each one aims to output class posterior probability estimates. The results indicate that models achieved an average accuracy between (95.81%) and (98.42%), however the generalization ability of M-SVM is better than the other models. This results show that the M-SVMs can be useful classifier for the automatic detection of heart diseases.