ELECTROCARDIOGRAM SIGNAL AUTOMATIC ANALYSIS

Modern engineering offers sophisticated means for monitoring diagnosis and therapy of cardiac diseases, the most common cause of death in the western world. There has been a rapid expansion and an increased utilization of computerized electrocardiography in recent years. The number of different ECG computer programs has steadily increased. Each of these programs has basically two distinct parts, a measurement and an interpretation part. In the measurement part various algorithms are used for wave recognition, beat selection and parameter extraction. In the interpretation phase different criteria and classification schemes are applied on the computed measurements in order to reach diagnostic ECG statements. The result is an improved quality of life as well as an increased life expectancy for many individuals suffering from cardiac diseases. Several approaches to investigating cardiac activity are today available, the 12-lead surface electrocardiogram (ECG) being by far the most widely used. The ECG provides a global description of the heart electrical activity and it has been found helpful in diagnosing a multitude of cardiac diseases, by considering rate and morphology of the ECG signal. The three-lead vector cardiogram (VCG) also provides information of the electrical activity; being an orthogonal description, the VCG is useful for studying the electrical activity in the three spatial dimensions. The endocardial electrogram (EGM) is used to study local activity and it is used in implantable devices designed for cardiac rhythm management (CRM) where surface electrodes are inappropriate. The recognition and analysis of the ECG signals is a very important task. This could be difficult because the size and form of these signals may change eventually and can be noised. Many tools, methods and algorithms from signal processing theory have been proposed, described and implemented in the literature. Greater part of them was developed under MATLAB, using Signal processing and Wavelet Toolboxes. This doctoral thesis contributes to the field of cardiology by presenting several algorithms for improved analysis and is to put forward solutions to certain problems encountered in the processing of ECG. 1- First chapter is theoretical and is focused on heart anatomy and physiology in order to describe the ECG signal origin and typical signal parameters in normal cases. It concludes with the description of the manifestation of two types of pathologies recognizable by the signal: arrhythmia and ischemia. 2- Second chapter includes ECG theoretical study representing different examples of abnormal traces. 3- Third chapter discusses distortion factors in an EGC signal explaining different noise types overlapped upon the signal itself especially in the frequency domain. 4- Fourth chapter discusses the different pre-processing methods studied in the literature in order to filter the signal from the types of noises of the previous chapter. We tried to choose some methods to create a filtering system that can reduce the noise with the smallest lost in the signal. 5- In the fifth chapter we start with the first ECG pattern recognition passage, the QRS-detection. We discuss the performance of four different QRS-detection algorithms set up in literature and then we develop a new one. We finish with a comparison between the several studied algorithms and the developed one. In this study we took the data from the database 48 ECG traces MIT-BIH. This algorithm realized an improvement in the obtained outcomes comparing with the other studied ones. Such development is composed by two main phases: pre-filtering and thresholding. 6- Sixth chapter continues the previous subject and is about the automatic recognition of the other two ECG components: the P and T waves. ELECTROCARDIOGRAM SIGNAL AUTOMATIC ANALYSIS G. KARRAZ, 2003-2006 6. In this chapter, as in the fifth one, we study four different algorithms, already studied in literature, and then we develop a new one comparing the several algorithm outcomes in the EGC trace. 7-In seventh chapter we explain the neural networks theory we’ll use later for the ECG signal classification. This chapter comprehends the study of forward and backward neural networks in different topologies in one and multi hidden layers. Moreover, we studied some neural network training algorithms we used later in this thesis. 8- In eighth chapter we propose a new ECG filtering system going by the QRS detection algorithm and the back-propagation method of a neural network. 9- In chapter nine we studied Heart Rate Variability (HRV) and ischemic episodes in long term ECG signal. Firstly, we discussed the HRV parameters meaning both in time plane and in frequency plane using several methods proposed by literature to calculate these parameters; secondly, we developed two new algorithms for premature cardiac beat classification and ischemic episodes automatic recognition, the first one uses the relative ratio between two consecutive points in the HRV time series. Where the second algorithm is based on the ST segment morphologic study because of its importance in such cases and the C-mean Fuzzy logic algorithm. Outcomes have been shown on two leads long term traces together with HRV parameters. We noted that, in some cases of cardiac arteries diseases (CAD), the rise of an ischemia can be perceived thanks to HRV parameters. 10-In chapter ten we create a new method to classify five different types of arrhythmia based on the use of neural networks with two hidden layers and trained by three different algorithms (Bayesian, Back-propagation, and Levenberg Marequardt). We noted that Bayesian classifier has the best performance compared to the other developed ones. In chapter nine we studied Heart Rate Variability (HRV) and ischemic episodes in long term ECG signal. Firstly we discussed the HRV parameters meaning both in time plane and in frequency plane using several methods proposed by literature to calculate these parameters; secondly we developed two new algorithms for premature cardiac beat classification and ischemic episodes automatic recognition, the first one uses the relative ratio between two consecutive points in the HRV time series. Where the second algorithm is based on the ST segment morphologic study because of its importance in such cases and the C-mean Fuzzy logic algorithm. Outcomes have been shown on two leads long term traces together with HRV parameters. We noted that, in some cases of cardiac arteries diseases (CAD), the rise of an ischemia can be perceived thanks to HRV parameters. In chapter ten we create a new method to classify five different types of arrhythmia based on the use of neural networks with two hidden layers and trained by three different algorithms (Bayesian, Back-propagation, and Levenberg Marquardt). We noted that Bayesian classifier has the best performance compared to the other developed ones.