ANALYSIS AND INTERPRETATION OF ECG SIGNAL USING WAVELET TRANSFORM

Abstract

Electrocardiogram (ECG) is a graphical representation of the electromechanical activity of the cardiac system. It provides a fast and reliable information to the expert cardiologist with respect to the functional aspects of the heart. The ECG is recorded in many situations, viz., to know the state of a patient under medical diagnosis and treatment, to keep a watch on the state of patients in the intensive cardiac care units, to know the response of the patient under medical treatment, to know the condition of cardiac system under stressed conditions, and to monitor the state of the ambulatory patients. The number of cardiac patients are increasing at an alarming rate and it is not possible for the existing number of cardiologists to take care of all the cardiac patients under all the conditions. This problem was realized about four decades back and a large number of individuals and groups started work on the computer aided analysis and interpretation of ECG signal throughout the world. One of the first attempts to automate the ECG analysis was made in 1957 by Pipberger and his group in USA. The ECG data acquisition and preprocessing; detection of waves, peaks, and interwave segments; feature extraction of all waves, peaks and segments; and usage of these in disease classification and diagnostics are the important stages in computer aided ECG analysis and interpretation. After carrying out the detailed survey about four years back by the author, it was found that there is a gap between what is ideally required and what has been achieved so far in the area of computer aided ECG analysis and interpretation. The gaps were identified and the work has been carried out in this thesis to bridge the gaps in the automated analysis and interpretation of the ECG signal by making effective use of the wavelet transform. The work covers the ECG feature detection & extraction, and cardiac disease diagnostics using multilead ECGs, disease diagnostics using rhythm analysis and data compression using non-redundant template and wavelet transform. After dealing with the general introduction and the brief outline of the work, the first stage of the work deals with the QRS detection using wavelet transform. In this study, the detection of QRS complexes by different wavelets, using standard CSE and MIT/BIH data bases has been carried out. Looking at the potentials of various approaches, it is very difficult to claim that a particular method (syntactic, non-syntactic, hybrid or transformative type) is always suitable for QRS detection. Due to the physiological variability of the QRS wave and the presence of noise and artifacts in the ECG signal, none of the technique has claimed an Xll accuracy of 100% in QRS detection. In recent times, the use of wavelet transform in QRS detection has shown upper edge in terms of no-need of preprocessing, accuracy of detection, and simplicity in calculations. And it has been found that even with the wide variations in ECG morphologies, QRS detection is more accurate by wavelet. In the present work, five existing wavelets (WT1, WT2, WT3, WT4, WT5) and a new wavelet (WT6) developed by the author have been used for the QRS detection. Their performance is checked using the first 25 records of the CSE database. The WT1, having high and low magnitude filter coefficients, gives QRS detection rate of99.11%. The performance of WT2 is similar to WT1 and gives QRS detection rate of 99.16% . These long filter length wavelets usually fail to detect QRS complex, if it present at the start or end of data. The compact wavelet VVT3 is computationally simple, faces less difficulty in detecting the end point QRS complexes, and gives high detection rate of 99.91%. The WT4 faces difficulty in detecting QRS complex when the signal amplitude is low, and gives QRS detection rate of 88.18%. The WT5 is computationally suitable for detection of end point QRS complexes, and gives detection rate of 99.70%. After the performance evaluation ofWT1-WT5 and also by keeping in view the selection guidelines for wavelets, a new wavelet (WT6) has been constructed. The WT6 gives QRS detection rate of 99.89%. The quadratic spline wavelet (QSWT i.e.,WT3) and WT6 are found to be the most suitable for QRS detection as they detect QRS complexes even with the wide variety of ECG morphologies, noise and/ or artifacts. The developed software gives the QRS detection rate of 99.866% . The QSWT, when tested on all 48 records of the MIT/BIH, gives the QRS detection rate of 99.806%. 'fhis performance of the developed software proves the utility of the wavelets for the detection of QRS complexes. In the second stage of the work, all the fundamental ECG parameters have been detected and measured by using the QRS location as a time reference. From these parameters, diagnostically important features namely, heart rate, P amplitude, P duration, PR interval, QRS interval, QRS peak-to-peak amplitude, QT interval, VAT and Tamplitude are obtained. The comparison of five wave fiducial points shows that most of the values are well within the tolerance limits suggested by the CSE working party and the overall accuracy in the measurement is about 91.00 %. Out of a total of 125 fiducial location estimates in 25 records, 11 estimates deviate from the tolerance. The software has also been tested using the ECGs recorded in the laboratory, and 5 beats have been selected for feature extraction and the resulting diagnostic parameters have been extracted. After confirmation of the reliability of software using CSE DS-3 and the ECG records ofthis lab, diagnostic dataset DS-5 has been used for the analysis and disease diagnosis. As there are no measurement results published by xiii the CSE for dataset DS-5, statistical analysis has been used to see the distribution of program estimates around a mean value. From a record, five regular beats per lead are selected and 29 parameters per beat per lead are extracted. Therefore, the software extracts the parameters from five such beats in 12 standard leads and stores the feature extraction data. Statistical parameters are used to see measurement performance of the developed software. With respect to interval and amplitude measurements of various components of the QRS complex as well as of the P wave and ST-T complex, a median value of the results from 12 standard leads (SL) has been used as a reference. The median value derived from 12 SL measurements proved to correlate in the best way with the results of the visual analysis. The statistical parameters namely, variance and standard deviation (SD) derived from 12 SL measurements show the best performance in the ECG analysis. As the ultimate aim of the ECG analysis is disease diagnosis, therefore, in the third stage of the work, the diseases namely, Left Ventricular Hypertrophy (LVH), Right Ventricular Hypertrophy (RVH), Myocardial Infarction (MI) (anterior, lateral, and inferior), Right Bundle Branch Block (RBBB), Left Bundle Branch Block (LBBB), Tachycardia and Bradycardia are considered for disease classification. The disease classification is based on the main features of the ECG. The testing of disease diagnostic has been carried out on all 125 cases of CSE DS-3 database. For this data two diagnostic criteria have been used to enable the validation of the disease classification. It is observed that the final diagnostic results obtained according to the score and the thresholds by both the criteria arc the same. Alter the testing of software on CSE DS-3, the disease diagnosis of CSE DS-5 records and the ECGs recorded from different subjects in the laboratory have been carried out. The combined results of existing scoring schemes for LVH, RVH and MI diseases are used to give an overall diagnostic statement, which is a resultant statement based on the results obtained by the existing three LVH, three RVH and two MI scoring criteria. A typical record D-00008.DCD from the CSE DS-5 has been used and the five data files of ECG parameters and the corresponding disease interpretation by existing as well as modified scoring schemes have been carried out. The ultimate result is obtained from the net results of five interpretations given by modified scoring schemes. Using the strategy of disease diagnostics from the features of five ECG beats, the validation of the software has been carried out using the CSE database. The results of this evaluation are compared with the results obtained by the existing scoring schemes and also with the diagnostic truth given by the CSE Committee. The CSE Working Party has considered the case as normal even if the XIV record shows minor abnormalities such as non-specific ST-T changes, incomplete right or led BBB, left anterior fascicular block, minor intraventricular conduction defects (QRS < 120ms) or even myocardial ischemia, as a single statement without making reference to any of the seven primary categories, namely, Normal, LVH, RVH, BVH AMI, IMI, and MIX Ml. To compare the results of existing and modified scoring schemes with the CSE results, the diseases Bradycardia, Tachycardia, RBBB, and LBBB are not considered. From the comparison with the CSE diagnostic results of first 10 records of DS-5, the diagnostic interpretation performed by existing scoring criteria matches up to 60% with the truth and by the modified criteria up to 80% , thereby resulting a gain of 20% . The gain is due to three factors: i) use of combined wavelets for feature extraction, ii) use of five beats in place of one beat for analysis, and iii) the use of modified scoring scheme. In addition to the disease diagnosis using multilead ECGs, rhythm analysis using single lead recording has been carried out to study heart rate variability (HRV) in the next stage of work. This includes the use of the WT for QRS detection and removal of ectopic beats and artifacts, determination of spectral and non-spectral indices and displaying of HRV related plots, namely R-R interval and PSD curves. The system performance has been evaluated by using the standard MIT/BIH database, because this database has long records and the database created by on-line recording from different subjects in the laboratory itself. TheWThas been used to detect the R-R normal intervals from the ectopic beats and artifacts. This is due to splitting the ECG signal in different band of frequencies and the use of frequency band containing the QRS complexes. On the basis of these results, it can be stated that the HRV spectral and non-spectral indices are less prone to fluctuations in heart rate due to autonomic imbalance than the fluctuations due to improper and incorrect detection of a single heart beat. This false detection gives substantial rise to the values of heart rate (HR), standard deviation of normal to normal R-R intervals (SDNN), low frequency power (LFP) and high frequency power (HFP) parameters. Thus, HRV analysis needs accurate detection of normal R-R intervals. The second set of database obtained from the subjects in the laboratory characterises the dynamic response of the heart to the vagus nerve i.e. during slow, comfortable and fast paced respiration. Three different rates, 12, 19, and 24 breaths/min, were chosen to represent slow, comfortable, and fast pacing rates, respectively. Quantization of the data in terms of their relative spectral and non-spectral indices for different respiration phases illustrates the influence on the vagal activity, HRV and corresponding PSD curves. For slow respiration, the HR gradually changes and this change is in the range of 80 to 100 BPM. So far as normal breathing is concern, the HR change is fast and varies from 80 to 120 BPM. For fast respiration, the HR variations are less, the change is in the range from 90 to 110 BPM. There is a negative relationship between the respiratory rate and the spectrum measures of parasympathetic activity (vagal power). For slow respiration, high power peak emerges around the frequency of 0.3 Hz. Alow power peak emerges in case of fast respiration. This indicates the influence of vagal control on the heart activity. The test results are consistent and reliable and show high promise for the effective use ofWT based R-R detection technique for the HRV study and analysis. To handle the large volume ofECG data without losing the diagnostic information, it is necessary to use faithful data compression techniques. Thus, there is a need to develop such techniques which have better performance in comparison to existing techniques. In the present work a simple technique has been introduced named as non-redundant-template direct data compression (NRT-DDC) technique. It performs the compression by downsampling the ECG signal in steps. The removed data samples in the process ofdownsampling are stored in a data array as non-redundant template. The signal is compressed by a factor of 8 for the ECG signal sampled at 500 Hz or 16 for the signals sampled at 1000 Hz. The performance evaluation ofthe non-redundant template NRT-DDC has been carried out using the CSE data sets-3 and -5. The reconstruction accuracy even in the low frequency and low amplitude (baseline) region is within the tolerance limits, which helps in accurate detection of onsets and offsets of the ECG waves. Aspect of holding the information of the ECG locations is being carried out by storing the compressed signal. Hence, this helps to retain 100% accurate diagnostic information. To compare the performance of this method with the existing techniques, we have selected the reported data as well as the techniques reported with the performance evaluation based on compression ratio (CR) and percent root mean square difference (PRD) for data sampled with 500Hz. The CR by the existing techniques ranges from 7 to 10 in most of the cases (except in two, having high CR of about 16) and PRD from 3to 28. It means that only CR or PRD does not give proper scale for comparison. Therefore, the only way is to see whether the clinical information is being retained or can be retrieved or not from the reconstructed signal. This aspect has been considered to evaluate the performance ofthe developed methods in this work and the comparison of the diagnostically important parameters measured in original and reconstructed signal is carried out with the CSE results. In addition to this data compression technique, an algorithm for WT based data compression has also been developed. Cardiologists suggest that the clinically useful information present in original ECG signals is preserved by 8:1 compression, and in most xvi cases 16:1 compressed ECGs are clinically useful. Considering this, the data compression has been carried out using the WT technique to provide the CR of 8:1 and 16:1. Finally it can be stated that the work contributes significantly to the area of computer aided analysis and interpretation of ECG signal. It also raises number of questions for carrying out further work. The overall work done in this thesis may be considered a positive and significant contribution in this field.