Patient Specific Electrocardiogram Beat Classification using Expectation Maximization Algorithm

The abnormalities of human heart are usually diagnosed from a biological signal known as the Electrocardiogram (ECG). It gives the electrical variations caused to the body due to the depolarization and repolarization cycles as part of the blood circulation system. Various arrhythmias can be identified using the morphological and temporal changes occurred to ECG signals by cardiologists. The process of identifying various arrhythmias from the changes in ECG signal is a tedious and complex process even for the experts. There exists a requirement for the automatic arrhythmia classification systems. Most of the automatic arrhythmia classification systems explained in literature are of supervised in nature. Inter patient variability of ECG signal causes problem to the performance of these supervised classifiers. MIT_BIH arrhythmia database is used for this work. R peak locations in the annotation file provided along with each ECG signal of the database is used to identify the QRS complex. A feature window of 256 samples around R peak is considered for every QRS complex identified and extracted a set of 32 features from each of them. Each feature set includes 24 Dual Tree Complex Wavelet Transform (DTCWT) features as well as other 8 statistical and temporal features. This work used Expectation Maximization (EM) algorithm for the clustering of data which is an unsupervised method. The data model used is Gaussian Mixture Model (GMM). The labelling of clusters is done according to the recommendations provided by Association for the Advancement of Medical Instrumentation (AAMI) with the help of the database. The performance of the work carried out is evaluated using commonly used measures. Overall sensitivity of 70.43% is obtained and it is a promising result for an unsupervised classification method