DEVELOPMENT OF PATTERN RECOGNITION TECHNIQUES FOR THE ANALYSIS OF ECG WAVEFORM

Abstract

Electrocardiogram (ECG) is an important bio-electrical signal which is extensively used by the cardiologists for diagnosing the state of cardiac system. Electrocardiogram is the graphical representation of the electrical activity of the heart. It has also very good correlation with the mechanical activity of each chamber of the heart. An ECG signal recording can be decomposed into ECG cycles, and every ECG cycle can further be decomposed into simpler subpatterns like P, Q, R, S, T peaks and PR and ST interwave segments. These peaks and interwave segments form the pattern vector of the ECG. The pattern vector of the ECG is then used for classification as normal and abnormal ECG. Although a good amount of work has been done so far in this field, but still no method has reached the stage of perfection. There is utmost necessity of making perfect the existing methods or developing new methods which can be acceptable to experts (cardiologists) for computer aided interpretation of ECG signal. As another step in this direction, the present work has been done on the development of pattern recognition techniques for the analysis of electrocardiograms. The work presented in the thesis meets out the objectives of analysis and pattern recognition of ECG signal through several steps of processing and identification. As a first step, the raw ECG signal is preprocessed to remove baseline wander and noise. The baseline wander is removed using adaptive filters. The noisy peak patterns are recognized using amplitude and duration criteria. This step is named as first stage of filtering. After first stage of filtering, the foremost step is the identification of QRS complex as it is the most prominent feature of the ECG signal. A set of consecutive peaks without interwave segments in between and satisfying a set of criteria are clubbed together to form the complex. Using K-means method of clustering and the incremental energy, the complexes are clustered into two classes. The complex (iv) belonging to the cluster with higher value of mean Incremental energy is identified as QRS complex. After recognition of QRS complex, the P and T peaks are identified by selecting a search interval between two consecutive QRS complexes. Candidate peaks are selected from the background of noisy peaks (which could not be removed during first stage of filtering) using peak-to-peak amplitude criteria. Two prominent peaks in the first half search interval with highest average value of the peak-to-peak amplitude are selected and a fuzzy classifier is used to identify T-peak. Similarly, P-peak is identified in the second half search interval. The peaks other than the identified real peaks are rejected as noisy peaks during this second stage of filtering. After completing both the stages of filtering, the measurements of attributes like amplitude, duration, boundaries of the peaks and duration and shape of interwave segments are carried out. Besides this, other features like RR-interval, PR-interval, ventricular activation time (VAT), QRS-duration, and ST-segment elevation or depression are also measured. All these attributes along with their normal permissible values are given in the tabular form in leadwise output. The deviation of the measured attributes from its normal permissible value is highlighted in the output indicating the particular feature of the ECG that should be specially looked into during diagnosis. A new method for frontal plane axis (FPA) angle calculation has been developed. From the FPA angle of any subpattern, the angle between two subpatterns can be found. The FPA angle of QRS complex is found and the FPA deviation from the normal permissible range, e.g. left axis deviation (LAD) and right axis deviation (RAD) is also indicated in the output. If the attributes of the various subpatterns of the ECG are within normal permissible range, the ECG is classified to be normal and, if not, the ECG is said to be abnormal. In case the category comes out

as abnormal, knowledge base for classification of ECG for category of disease to which it belongs, is invoked. The diseases which have clear, precise and unambiguous diagnostic rules have been taken into account in the present work, namely, Bradycardia, Tachycardia, Sinus Arrhythmia, Left ventricular hypertrophy and Right ventricular hypertrophy. A point scoring method for detecting the left and right ventricular hypertrophy has been developed which gave satisfactory diagnosis when tested on 12-lead CSE-ECG database. In the output, the diagnosis of the ECG is also given. All the work on pattern recognition and analysis has been carried out on a personal computer. Implementation on personal computer demonstrates suitability of the technique developed for its use in bedside on-line ECG analyzer in hospitals. The algorithm have been developed for each stage of processing, identification and classification of ECG signals. The effectiveness and efficiency of each algorithm has been tested by using the ECG database of CSE-library. Although the algorithms have .been tested on large number of cases, only a few representative cases are given in the thesis with each algorithm as illustrative examples. The general conclusions in respect of the work are given at the end of the thesis as a comprehensive view of the overall results obtained in the present work. Suggestions for carrying out the further work are also made in this chapter.