ECG ANALYSIS AND HEART DISEASE CLASSIFICATION

Abstract

The heart is a vital organ in humans and animals which circulates blood through blood vessels in the whole body. The blood provides nutrients and oxygen to all cells, and removes wastes of metabolism. Heart disease is the main killer of men and women in the world; particularly in the United States, Canada, England, and Asian countries. The common diseases of the heart areheart attack (Myocardial Infarction), arrhythmias and sudden cardiac arrest. Therefore, early detection and accurate diagnosis are important issues in clinical practice of cardiologists and physicians. The cardiologist or physician identifies heart diseases on the basis of ECG signals. The electrocardiogram (ECG) is a noninvasive method for detecting heart diseases. The ECG signal basically represents normal or abnormal functioning of heart activity. The normal ECG signal generally consists of P-wave, QRS-complex and T wave. The cardiologists and physicians have defined rules and definitions for visual ECG analysis, which may have subjectivity and are not uniform, so computer based interpretation is the need of this day. The ECG signal analysis and classification involve acquisition of ECG data, pre-processing of ECG signals, detection of ECG wave complexes such QRS complex, P & T-wave, etc. Next, their wave boundaries such as Ponset-Poffset, QRSonset-QRSoffset and Tend are found and then clinical relevant intervals, such as P duration, duration of QRS complexes, ST-T interval and QT interval along with morphologies of each wave are determined. On the basis of ECG wave complexes, amplitude and wave interval features are extracted for classification. The aim of the present work is to do ECG analysis using simple approaches and then improve diagnostic performance. Recorded ECG signals contain noises and artifacts such as power line noise, baseline wander, motion artifacts, etc. In this work, removal of baseline wander and motion artifacts has been implemented for detection of QRS complexes. The second stage is detection of QRS complexes and wave components. Here, QRS complex detection of single and multilead ECG signals has been done by proposing a new algorithm. The detection of P and T wave has also been implemented in multilead ECG signals. The clinical parameters in ECG signal are calculated on the basis of boundary marking of P-QRS-T complexes, so marking boundaries of ECG wave complexes required accurate and reliable method. This has been achieved by proposed new algorithm. Diagnosis and classification of ECG signal required accurate and reliable method. Here classification of Myocardial Infarction, Cardiomyopathy, and bundle branch block has been done using the detected diagnostic parameters. Description of research work: The acquired raw ECG signal contains noises and artifacts. In this work we have developed a two stage median filter to remove baseline ii wander using window width size fs/2 and fs for stage first and second in terms of sampling frequency (fs). Single lead QRS detection: A simple and efficient new method for QRS detection in the ECG is proposed in this research work. The initial data is preprocessed using two stage median filter for removing baseline drift. The second stage, enhances the peaks of ECG wave components by using the sixth power of the signal. The next stage identifies the QRS complex by taking a variable window size. The performance of the new algorithm is evaluated against the CSE, MIT/BIH AD, ESC ST-T and QT databases. These four standard databases were used to perform QRS detection and 368 cases were considered which were, tested on 10,06,168 beats and achieved overall average sensitivity 99.52% and positive predictivity of 99.69%. The QRS detection was also performed on 12 datasets of noisy, full length signals (118e24 to 118e_06 and 119e24 to 119e_06) from MIT–BIH Noise Stress Test Database and obtained performance is higher and comparable to other algorithms in literature. Multilead QRS detection: QRS detection in 12-Lead Electrocardiogram (ECG) using composite lead and peak enhancement method is proposed in this thesis. Initially raw signals of 12-Lead electrocardiogram having a sampling frequency fs are pre-processed for baseline wander removal using a two stage median filter with window widths of fs/2 and fs respectively. The point by point average of the preprocessed signals corresponding to 12-Leads is taken to generate a composite lead. In order to obtain a variable size search window for QRS detection, the composite lead is enhanced by the sixth power of the signal and its mean value is determined. The maximum value of the search space defined by the search window was mapped on the composite lead and other 12 ECG leads of 12-lead ECG individually for QRS detection. The performance of the algorithm is evaluated against the CSE multilead measurement database and St. Petersburg Institute of Cardiological Technic’s 12-lead Arrhythmia Database and PTB Database. The overall performance of the proposed method, using different standard multilead databases, such as CSE, PTB and St-Petersburg multilead Arrhythmia with different cases and, total 2,55,925 beat was analyzed. The overall average sensitivity of 99.24% and positive predictivity of 99.90% was achieved considering all different standard databases. Boundary marking of ECG wave components and diagnostic parameter detection: Boundary point’s detection in simultaneously recorded 12-Lead ECG signal using a composite lead is proposed in this work. The complexes of the composite lead are better enhanced and noise free compared to others in any of the 12 lead signals. After detection of the QRS location of composite lead QRSonset and QRSoffset were determined by using the standard deviation method. Detection of P-wave location and onset-offset was carried out by using the standard deviation method and similarly T wave location was determined, and Tend iii was marked. The performance of the algorithm is evaluated against the CSE multilead database, the main boundary marking of Ponset, Poffset, QRSonset, QRSoffset and Tend estimated are within limits recommended by the CSE working party. In this software we obtained unbiased measurement within specified limits. For automatic ECG analysis and diagnosis system a dominant beat is required for measurements and classification. The onsets of P, QRS and offsets of P, QRS and T wave are detected on the composite beat and boundary values of the composite beat were mapped in all the average beats of 12-Leads. After determination of P duration, QRS complex duration, ST-T complex interval and QT interval, and other parameters such as peak to peak amplitude, area, mean, standard deviation, skewness and kurtosis of P duration, QRS duration and ST-T complex interval of all average beats of 12-lead ECG are calculated. In this work disease diagnosis and classification were performed using different ECG lead arrangements with SVM and ANN classifiers. Detection of myocardial infarction has been performed using composite lead parameters and all 12 lead parameters with SVM and ANN classifiers and it is observed that ANN classifier obtained maximum accuracy in composite lead and all 12 lead systems. Also, it is observed that after reduction in dimensionality using PCA, obtained classification accuracy is 100% in both lead systems. Thus, it can be concluded that the composite lead system performed comparable and significant MI detection. Detection of cardiomyopathy has been performed using twenty two features from composite lead and 220 features from all 12 lead with SVM and ANN classifier. In this case performance of cardiomyopathy detection is higher in ANN classifier for both composite lead and all 12 lead system and it is observed that after reduction of dimensionality using PCA, performance of SVM and ANN classifiers decreased in both lead systems. Thus, it can be finally concluded that cardiomyopathy detection using ANN classifier with a composite lead system performs better than SVM. Detection of bundle branch block has been performed using extracted features of composite lead and all 12 Lead systems with SVM and ANN classifiers. In this study ANN classifier (accuracy with PCA: 80%, accuracy without PCA: 100%) performed better than SVM (accuracy with PCA: 68.75%, accuracy without PCA: 68.75%) with all 12 lead systems. The Computer Assisted ECG Analysis and Classification system is designed for healthy, myocardial infarction, cardiomyopathy and bundle branch block with ANN classifier using composite lead and all 12 lead features. In this case, classification accuracy obtained is 100% with the composite lead system using PTB annotated database. Thus, it can be finally concluded that the composite lead system contributes significantly for ECG analysis and classification systems. The overall work done in this thesis may be considered a positive and significant contribution in this field.