USING CORRELATION COEFFICIENT IN ECG WAVEFORM FOR ARRHYTHMIA DETECTION

BIOMEDICAL ENGINEERING- APPLICATIONS, BASIS & COMMUNICATIONS USING CORRELATION COEFFICIENT IN ECG WAVEFORM FOR ARRHYTHMIA DETECTION 147 CHUANG-CHIEN CHIU 1,2, TONG-HONG LIN 1 AND BEN-YI LIAU 2 1 Institute of Automatic Control Engineering, Feng Chia University, 2 Graduate Institute of Electrical and Communciations Engineering, Feng Chia University, Taichung, Taiwan Biomed. Eng. Appl. Basis Commun. 2005.17:147-152. Downloaded from www.worldscientific.com 1. INTRODUCTION Disturbances of impulse formation and conduction may occur anywhere in the human heart. The most common cardiac arrhythmia is the ventricular ABSTRACT Arrhythmia is one kind of diseases that gives rise to the death and possibly forms the immedicable danger. The most common cardiac arrhythmia is the ventricular premature beat. The main purpose of this study is to develop an efficient arrhythmia detection algorithm based on the morphology characteristics of arrhythmias using correlation coefficient in ECG signal. Subjects for experiments included normal subjects, patients with atrial premature contraction (APC), and patients with ventricular premature contraction (PVC). So and Chan's algorithm was used to find the locations of QRS complexes. When the QRS complexes were detected, the correlation coefficient and RR-interval were utilized to calculate the similarity of arrhythmias. The algorithm was tested using MIT-BIH arrhythmia database and every QRS complex was classified in the database. The total number of test data was 538, 9 and 24 for normal beats, APCs and PVCs, respectively. The results are presented in terms of, performance, positive predication and sensitivity. High overall performance (99.3%) for the classification of the different categories of arrhythmic beats was achieved. The positive prediction results of the system reach 99.44%, 100% and 95.35% for normal beats, APCs and PVCs, respectively. The sensitivity results of the system are 99.81%, 81.82% and 95.83% for normal beats, APCs and PVCs, respectively. Results revealed that the system is accurate and efficient to classify arrhythmias resulted from APC or PVC. The proposed arrhythmia detection algorithm is therefore helpful to the clinical diagnosis.. Biomed Eng Appl Basis Comm, 2005(June); 17: 147-152. Keywords: Arrhythmia; ECG; Correlation coefficient; Atrial premature contraction (APC); Ventricular premature contraction (PVC) Received: Sep 3, 2004; Accepted: April 30, 2005 Correspondence: Chuang-Chien Chiu, Ph.D. Institute of Automatic Control Engineering, Feng Chia University, Taichung, Taiwan E-mail: chiuc@auto.fcu.edu.tw premature beat. Premature beats may be atrial, atrioventricular (A-V) junctional or ventricular in origin, and may be found in apparently healthy individuals. However, premature beats may become clinically significant if they occur too frequently, originate from multiple foci or are found in individuals with proven heart disease. The most common cause of cardiac arrhythmias is the coronary artery disease. It has been shown that 90 to 95 percents of patients with acute myocardial infarction have some associated cardiac arrhythmia. Two most common cardiac arrhythmias during an acute myocardial infarction are premature ventricular contractions and 37

148 sinus tachycardia. The analysis of the electrocardiogram (ECG) signal is the most readily available method for diagnosing cardiac arrhythmias. Arrhythmia is one kind of diseases that gives rise to the death and abnormal beats may momentarily paralyze blood pressure. Cardiac arrhythmias are dysfunctions or disturbances in the behavior of the heart. These disturbances produce abnormal in rate, rhythm and the site of impulse formation; factors that may in turn alter the normal sequence of atrial and ventricular activation. In electrocardiograms, such arrhythmias manifest themselves as deformations in the observed waveform. Such deformations, as associated with a diagnosed arrhythmia, occur with a consistency and morphological similarity that they may be looked upon as a waveform pattern in the temporal domain. New techniques of arrhythmia detection using morphology of the waveform have shown promising results in correctly detecting fatal arrhythmias [1-4]. In general, ECG arrhythmia waveforms can be subdivided into two classes; (1) events: which are single ectopic occurrences and (2) rhythms: which are a continuous series of one or more event types. In this study, an arrhythmia classifier based on correlation coefficient has been presented to identify normal beats, abnormal premature ventricular contraction (PVC) beats and atrial premature contraction in ECG. The occurrence of an arrhythmia is unpredictable. The purpose of this study is to develop a method to distinguish healthy and abnormal subjects using the correlation coefficients of ECG waveforms. Two kinds of cardiac arrhythmia, PVC (premature ventricular contractions) and APC (atrial premature constructions) will be discussed. They are the most common types of cardiac arrhythmias in ECG monitoring. In this paper, we describe the development of system for arrhythmia diagnosis. The beat of ventricular premature contraction will be classified on the basis of beat morphology. The classification of more complex arrhythmias will depend on a combination of rhythm and morphology analysis. In brief, the proposed system of arrhythmia detection includes the modules of beat recognition and rhythm analysis. 2. MATERIAL AND METHODS In general, the normal ECG rhythm means that there is a regular rhythm and waveform. However, the ECG rhythm of the patient with arrhythmia will not be regular in certain QRS complex. We utilize the different characteristics that arrhythmias exhibit to detect the abnormal ECG waveform. We first need to find the location of every QRS complex. The ECG signal is formed of P wave, QRS complex, and T Vol. 17 No. 3 June 2005 wave. The locations of QRS complex have the maximum variation in the slopes. This property was used to detect the location of QRS complex. The method of So and Chan was adopted to detect the location of QRS complex [5]. The property of variation in slopes and an adaptive threshold was applied to detect the R point. The detection of QRS complex The So and Chan QRS detection method [5] is trended to implement on the ambulatory ECG monitor. The computational requirement is kept at a reasonable level while without compromising its accuracy. Therefore, the approach of first derivative was selected and made substantial improvement to that technique. First, let x(n) represents the amplitude of the ECG data at a discrete time n. The slope of the ECG wave is calculated by equation 1. 2....(1) The slope_threshold is computed using equation When two consecutive ECG data satisfy the condition that slope (n) > slope_threshold, the onset of the QRS complex is detected. According to the suggestion given in [6], the parameter theresh_param can be set as 2, 4, 8 or 16 and the filter_parameter can be set as 2, 4, 8 or 16. After the detection of the onset of QRS complex, we shift the appropriate samples to detect the maximum point (maxi) and take as the R point. The maxi is then updated by equation 3. The first_maxi is defined by equation 4....(2)...(3) first_maxi=height of QRS onset-height of R point...(4) The initial maxi is the maximum slope of first 250 points. The appropriate threshold_param is 8 and the filter_parameter is 16 in [6]. Minimum error could be made if we set the parameters as above. However, the R waves of some arrhythmias were not able to detect by setting the threshold parameter being 8. The ventricular premature contraction (PVC) has smooth variation of slopes in the location of QRS complex. If 38

BIOMEDICAL ENGINEERING- APPLICATIONS, BASIS & COMMUNICATIONS the value of threshold_param is too high, it will not satisfy the situation that two consecutive ECG data conform the condition that slope (n) > slope_threshold. In such case, the QRS complex of PVC will not be detected accurately as shown in Figure 1. In order to solve the problem, we lowered the threshold_param to be 5. The QRS complex of PVC by setting the threshold-param being 5 is detected correctly as shown in Figure 2. 149 (1) Normal beat (N) (2) Ventricular Premature Contraction (V) (3) Atrial Premature Contraction (A) For each beat type, templates are selected from the MIT-BIH arrhythmia database. Template matching was performed using a normalized correlation coefficient [8] defined as equation 5...(5) Biomed. Eng. Appl. Basis Commun. 2005.17:147-152. Downloaded from www.worldscientific.com Fig.1. The incorrect detection of QRS complex. Fig.2. The correct detection of QRS complex. In order to evaluate the influence that the threshold_param changed, we tested the performance by detecting the R wave on ECG from the MIT-BIH normal sinus rhythm [7]. The MIT-BIH normal sinus rhythm ECG testing data are sampled at 128 Hz. Each data file contains two channels of ECG signal. Each data file is accompanied with an annotation of normal beats. The results of QRS detection of the So and Chan that lower the threshold_param to be 5 by using 10 MIT-BIH normal sinus rhythm ECG data files are summarized in experimental results. In experimental results, we will discuss the influence of parameter variation. Correlation coefficient ECG classification is carried out based on the correlation coefficient approach. A window is positioned at the QRS complex for the matching of the constituent elements. The window only spans the duration of the QRS complex rather the whole cardiac cycle. The beneficial property helps to minimize the computation in order to achieve classification quickly. In this study, the classifier is trained to recognize three types of beats Where xy = the correlation coefficient, N = the number of template points, x(n) = the template points, y(n) = the signal points under analysis, x = the average of the template points, y = the average of the signal points, and k is the time index of the signal y(n) at which the template is placed. The correlation coefficient falls within a range 1< xy <1, where +1 indicates a perfectly matched between signal and template. The classifier based on correlation coefficient is an effective way. Arrhythmias have different morphologies. For example, a premature ventricular contraction (PVC) beat is an ectopic beat originating in either the left or the right ventricular. It comes early in the cardiac cycle, before the next expected beat, and it occurs in the presence of an underlying rhythm, usually a sinus rhythm. The ectopic pacemaker of the PVC locates on the cell of ventricular muscle. The speed of the transmission of muscle is much slower than that of the nerve. So the PVC appears as a premature beat with a wide QRS and a long pause. Therefore, the PVC has a specific waveform. We will take advantage of the property of waveform to classify it by using the correlation coefficient. Flow chart of the detection algorithm First, we extract ECG data of five seconds and use the method of So and Chan to extract the features from ECG data. At the same time, we will aim at the similarity of waveform to recognize. We use the correlation coefficient to evaluate the similarity between typical normal beats and the test beat. The total number of computational points is only twenty points. If it is similar to the ventricular premature contraction (PVC) beat, our system will categorize it as PVC. If it is similar to the normal beat, our system will categorize it as normal or APC. Because the waveform of normal and APC are close to each other, we will analyze further if the RR-interval duration is normal or not. If the RR-interval duration is normal, the beat will be normal. Otherwise, it will be classified as the atrial 39

150 premature contraction (APC). After finishing the classification of the five seconds, the algorithm will read the next five seconds ECG data and repeat the same processes. The detailed flow chart of the detection algorithm is shown in Figure 3. 3. EXPERIMENTAL RESULTS Evaluation of QRS complex detection Generally, the premature ventricular contraction (PVC) beat has the smooth slopes on the location of QRS complex. Therefore, the R point was difficult to be detected by the original setting of threshold_param. We altered the setting of threshold_param in order to detect every location of QRS complex correctly. The MIT-BIH normal sinus rhythm database was taken as the test data in order to evaluate the effects of changing the parameter setting by So and Chan method. There are ten test data files and the sampling frequency is 128 Hz. The QRS complexes of these test data files were classified as normal in the MIT-BIH normal sinus rhythm database. The results were shown in Table 1. Fig.3. Flow chart of the detection algorithm. Vol. 17 No. 3 June 2005 Results of arrhythmia detection The detection of QRS complex has achieved good accuracy as shown in Table 1. Thereafter, two important characteristics including the morphology of QRS complex and the RR-interval duration were applied to detect the arrhythmic beats. The arrhythmic beats of APC and PVC have different kinds of RRinterval duration and morphology of QRS complex. For example, the atrial premature contraction (APC) beat has the normal morphology of QRS complex, but it does not have the normal RR-interval duration. The ventricular premature contraction (PVC) beat has abnormal RR-interval duration and abnormal morphology of QRS complex. We chose another ten test files from MIT-BIH arrhythmia database [9], and each data set includes some ventricular premature contraction (PVC) beats or atrial premature contraction (APC) beats, and normal beats. The sampling frequency of the data was 360 Hz. Each beats of QRS complex was classified and defined in the MIT-BIH arrhythmia database. The detection results using our detection algorithm are listed in Table 2. The sensitivity of our proposed method almost reaches 100% except for files 119 and 221. The results of classification for every beat of ECG data are shown in Table 3. Element (i, j) in the Table 3 represents the total number of beats annotated from the database as category j and classified from the classification algorithm as category i. Table 1. Using So and Chan method to detect the locations of QRS complex by setting the threshold_ param to 5. File True False False Positive Positive Negative Sensitivity 16265 194 0 0 100% 16272 125 1 1 99.21% 16273 182 0 0 100% 16420 185 0 0 100% 16483 184 1 2 98.98% 16539 163 0 0 100% 16786 142 0 0 100% 17052 147 0 0 100% 17453 168 0 0 100% 18184 175 0 0 100% Average 99.82% Note: True positive is that tests positive for a condition and is positive. (i.e., have the condition)false positive is that tests positive but is negative.false negative is that tests negative but is positive. 40

BIOMEDICAL ENGINEERING- APPLICATIONS, BASIS & COMMUNICATIONS Table 2. Results of arrhythmia detection. File True False False Sensitivity Positive Positive Negative 100 146 0 0 100% 209 94 0 0 100% 220 70 0 0 100% 116 34 0 0 100% 119 28 1 0 96.55% 202 24 0 0 100% 208 48 0 1 100% 221 33 2 0 94.28% 103 49 0 0 100% 101 49 0 0 100% In Table 3, the total number of QRS complex is 571. The number of normal beats defined by database annotation is 538. Two normal beats are classified as atrial premature contraction beats and one normal beat is classified as ventricular premature contraction beat. The number of abnormal beats (atrial premature contraction, APC) defined by database annotation is 9. All APCs defined by database annotation were classified as APC correctly by the classification algorithm. The accuracy of APC classification achieves 100%. The number of abnormal beats (ventricular premature contraction, PVC) defined by database annotation is 24. Only one ventricular premature contraction (PVC) beat is classified as normal beat. Table 3. Results from the classification algorithm. Classification Database Annotation...(6)...(7) Normal APC PVC Total Normal 535 0 1 536 APC 2 9 0 11 PVC 1 0 23 24 Total 538 9 24 571 151 Table 4 shows the sensitivity (equation 6) and positive prediction (equation 7) of the classification algorithm. The sensitivity of APC is low because our system classifies some normal beats as APCs. The total number of APC defined by database annotation is small, it might cause the sensitivity of APC being low. The positive prediction of normal beats achieves 99.44%. The positive prediction of APC achieves 100%. The positive prediction of PVC achieves 95.83%. Total performance of our system is 99.3% by equations 8 and 9. By the illustration, we know that our system has the perfect efficiency on the recognition of arrhythmia detection. Table 4. Sensitivity and positive prediction for each beat category. Category Sensitivity Positive Prediction Normal 99.81% 99.44% APC 81.82% 100% PVC 95.83% 95.83% 4. CONCLUSIONS...(8)...(9) It is clearly shown in Table 4 that the proposed classification algorithm base on correlation coefficient and RR interval is very effective on classifying arrhythmic beats. However, in our selected dataset, the number of normal beats (535 normal beats, almost 94.22% of the total number of beats) was very large compared to the other categories (atrial premature contraction 1.58% and ventricular premature contraction 4.2%). The total performance is high (99.3%) due to the achieved sensitivity and positive prediction are high for the normal beats (99.81% and 99.44% respectively), compared with to results for the results of the atrial premature contraction and for the results of ventricular premature contraction (81.82% sensitivity and 100% positive prediction for the atrial premature contraction and 95.83% sensitivity and 95.83% positive prediction for the ventricular premature contraction). In addition, the percentage of beats misclassified as normal beats in these categories was 0% for the atrial premature contraction and 4.35% for the ventricular premature contraction. Moreover, the false alarms were three normal beats being classified as arrhythmic. The main advantage of the 41

152 system is that it uses the property for rhythm and morphology and does not waste any other time consuming and the system can be very efficient. Some other advantages of our arrhythmic detection system can be mentioned. First, we detect the location of QRS complexes only based on slope. Thus, the speed of QRS complex detection could be more quickly than another method of Pan and Tompkins. Also, So and Chan method has higher accuracy demonstrated by Tan et al. in 2000 [5]. Moreover, the detection of arrhythmia detection does not rely on P wave since the P wave is hard to detect and usually exists in nosing signal level. Another advantage of the proposed system can be mentioned is to utilize the method of correlation coefficient. The method of correlation coefficient mainly recognizes the degree of similarity between certain typical waveform and arrhythmia waveform. We could recognize three kinds of beats including normal beat, atrial premature contraction (APC) beat, and ventricular premature contraction (PVC) beat. A window only spans the duration of the QRS complex for the matching of the constituent elements rather than the whole cardiac cycle. Total points of computation for QRS complex similarity are relatively small which minimizes the time of computation in order to achieve the efficient classification. In the future, if we can find more different beats for arrhythmia, the proposed arrhythmia detection system can recognize some other different arrhythmic beats. In conclusion, our system has many advantages including efficiency, accuracy, and simplicity. We believe that it is very suitable to arrhythmic detection in clinical practice. ACKNOWLEDGEMENT The authors would like to thank the National Science Council, Taiwan, R.O.C., for supporting this research under Contract No. NSC 91-2320-B-035-001. REFERENCES 1. Kumar VK : A novel approach to pattern recognition in real-time arrhythmia detection. Engineering in medicine and biology society. Proceedings of the annual international conference of the IEEE 1988; 1: 7-8. 2. Giraldo BF, Marrugat J, and Carninalti P : Design of an expert system for arrhythmia diagnosis. Engineering in medicine and biology society. Proceedings of the annual international conference of the IEEE 1992; 3, 1255-1256. Vol. 17 No. 3 June 2005 3. Giraldo BF, Binia M, Marrugat J and Caminal P : Arrhythmia diagnosis system: validation methodology. Engineering in medicine and biology society. IEEE 17th annual conference 1995; 1: 737-738. 4. Dickhaus H, Gittinger J and Maier C : Classification of QRS morphology in Holter monitoring. Engineering in medicine and biology. 21st annual conference and the annual fall meeting of the biomedical engineering society 1999; 1: 270. 5. Tan KF, Chan KL and Choi K : Detection of the QRS complex, P wave and T wave in lectrocardiogram. Advances in medical signal and information processing 2000; 41-47. 6. So HH and Chan KL : Development of QRS detection method for real-time ambulatory cardiac monitor. Engineering in Medicine and Biology society. Proceedings of the 19th Annual International Conference of the IEEE 1997; 1: 289-292. 7. MIT-BIH normal sinus rhythm database, third edition, May1997.(http://www.physionet.org/physiobank/dat abase/nsrdb) 8. Rangayyan RM. Biomedical signal analysis, John wiley & sons, Inc, New York, USA, 2002, 93-99. 9. MIT-BIH arrhythmia database, third edition, May 1997.(http://www.physionet.org/physiobank/databas e/mitdb/) 42