Classification of ECG arrhythms using time-frequency based features

Early detection of heart diseases can prevent permanent damage or even death that may occur in the future. In this thesis, it is aimed to determine Ecg Arrhythms with various machine learning algorithms. This is very important for clinical decision support systems. The thesis study used a large database of R-R intervals selected from the MIT-BIH Arrhythmia database. The study consists of two phases. In the first stage heart diseases are divided into two classes to be healthy and arrhythmia ECG records. At the first stage, it is aimed to determine whether there is any heart disease in the person. In the second phase, the ECG records are divided into 9 classes, and the aim is to find out what is the arrhythmia.The thesis study used a large database of R-R intervals selected from the MIT-BIH Arrhythmia database. Time-frequency transformation methods are used in the feature extraction phase. These methods are Smoothed Pseudo Wigner-Ville, Choi-Williams, Born-Jordan, Bessel and Zhao-Atlas-Mark Transforms. In the classification phase, Support Vector Machines, K Nearest Neighborhood algorithms, Ensemble classifiers, Tree methods and Discriminant Analysis were used. It has been shown that the proposed algorithm has a fairly short computation time and adaptability to real-time systems. The study will be part of the Decision Support System in the Telemedicine system being developed. The performance results of study based on classification for two-class are obtained as respectively 93.88%, 92.14% , 95.62% , 95.46% for accuracy sensitivity, specificity and positive predictive values, and results of study based on classification for nine-class are obtained as respectively 98.56%, 71.69%, 98.48%, 76.71%, 89.31%, 98.91% for accuracy sensitivity, specificity, F score, positive predictive and negative predictive values.