Abstract— The automatic processing of ECG for classification of heartbeat is presented in this paper. This work gives ability to the classifier to classify the beats to one of the four classes as recommended by ANSI/AAMI EC57:1998 standard. The beats are normal, ventricular, supraventricular and fusion. The data obtained from MIT-BIH database. Six hundred beats have chosen from each class. The accuracy, sensitivity, specificity and predictivity for supraventricular beats are 96%, 90.8%, 97.7% and 92.8% respectively. For ventricular beats they are 97%, 92.8%, 98.3% and 95.1%. For normal beats they are 98%, 96.7%, 98.6% and 95.9%. For fusion beats 95%, 92%, 96.1% and 88.7% respectively. These results obtained are an improvement on previously reported results.
Keywords— ANN:artificial neural network; AAMI: american association of medical instruments; MLP:multilayer perceptron;
I. INTRODUCTION
Heart diseases are one of the most life threatening disease in the world now a days. The common method for diagnosing cardiac diseases is ECG. ECG is simple, non-invasive and cost effective tool for diagnosing purpose. It is important to classify the beats properly to find out the arrhythmia. Classification of heart beat is tedious and time consuming job therefore there is need of automatic classification which assist cardiologist.
Automated classification of heartbeats for diagnosing of heart beat has been done using a large number of features. In 2004 philip de chazel et. al. uses structural(morphological) and time interval features. This classification supports AAMI standard. It is one of the earliest paper published which support AAMI standard [1]. In 2006 they use same feature and form patient adaptive classifier [2]. Some author’s proposed other features such as frequency component with time interval feature [5], Higher order cumulate features [6] and Karhunen–Loeve expansion of ECG morphology [7]. Tanis mar et. al uses large feature set and apply feature selection method to choose more appropriate feature. Whereas in this work simple morphological and statistical features are used and results obtained are very appreciable and comparable to previous proposed methods.
II. ECG DATA PREPARATION
In this study, all the ECG data used have been obtained from the MIT-BIH Arrhythmia Database (MITDB) [8]. Actually, more than 100,000 beats are present in database which are individually belonging to one of 15 possible beat types. However, it is necessary to follow the classification rules as proposed by AAMI for the evaluation of ECG classifiers [9], which recommends to group all present morphologies in six classes according to their physiological origin which is described in table no.1. When evaluating classifiers it is standard recommendation to ignore records containing paced beats from these six classes, and some authors have proposed to ignore unknown beats too, for being too poorly represented and of no help for further classification purposes [3].
Applied to MIT-BIH data base 2400 beats are taken from different classes.600 beats are taken from normal class, 600 beats are taken from supraventicular class, 600 beats are from ventricular and rest 600 beats are from fusion group. Out of 2400 beats 50%beats are used for training , 15% are used for validation and 35% are using for testing.
III. METHOD FOR AUTOMATIC HEARTBEAT CLASSIFICATION
Fig no.1
Fig.1 shows the stages of an automated system suitable for heartbeat classification which is used in this study. It consists of three stages: a pre-processing stage, a processing stage, and a classification stage. In pre-processing stage the digitized ECG is applied at the input of pre-processing stage. In this stage a filtering unit is present to remove artefact signals from the ECG signal. These signals include baseline wander, power
AAMI heart beat class N S V F Q
description
Any heart beat not in the S, V, F, Q classes Supraventricular ectopic beat
Ventricular ectopic beat
Fusion beat
Unknown beat
MIT-BIH heartbeat types
Normal beat(NOR)
Atrial premature beat
Premature ventricular contraction (PVC) Fusion of ventricular and normal beat (FVN) Paced beat (p)
Left bundle branch block(LBBB)
Aberrated atrial premature beat(AAP) Ventricular escape beat (VE) Fusion of paced and normal beat (PN)
Right bundle branch block(LBBB)
Nodal (junctional) premature beat(NP)
Unclassified beat(U)
Atrial escape beat(AE)
Supraventricular premature beat (SP)
Nodal (junctional) escape beat (NE)
line interference, and high-frequency noise. The processing stage consists of heartbeat detection and feature extraction.
The heartbeat detection module attempts to locate all heartbeats. The feature extraction module is concerned with forming a vector of measurements (feature vector) from each heartbeat that are processed by the classifier stage. The feature extraction modules are required because, although it is possible for the classification stage to process the ECG samples directly, greater classification performance is often achieved if a smaller number of discriminating features (than the number of ECG samples) are first extracted from the ECG. The classifier units normally contain parameters which are set during the system development to optimize the classification performance. These stages are discussed in more detail below.
A. ECG Filtering
The All ECG signals were filtered with FIR band pass filter to remove the baseline wander. The cut-off frequencies of FIR filter is 5-15Hz (i.e lower cut-off frequency is 5Hz and upper cut-off frequency is 15Hz)[5]. Due to filtering power line interference and high-frequency noise were removed from ECG signal. Now the signal may be use for processing stages.
B. Heartbeat Detection
This study did not investigate the problem of heartbeat detection from the ECG; instead we have utilized the heartbeat fiducial point times provided with the MIT-BIH arrhythmia database. The provided fiducial points occur at the instant of the major local extremum of a QRS-complex (i.e., either the time of the R-wave maximum or S wave minimum). These fiducial points were obtained manually on a beat-by-beat basis. Four types of beats i.e normal, supraventricular, ventricular and fusion is shown in fig no.2. in this study we have taken 180 sample from each beat by taking 90 samples on left and 90 sample on right side of the peak.
C. Feature extraction
Morphological and statistical features are used to form feature vector. Features included namely area, energy, maximum amplitude, minimum amplitude, ratio of maximum amplitude and minimum amplitude, mean, variance, kurtosis and skewness.
Fig. no.2
IV. RESULTS
The confusion matrix is shown in fig no.3 which shows the accuracy and sensitivity of different classes. ROC of MLP classifier is given in next figure. Table no.2 summarized the results of different classes of classifier.
ROC of given approaches to ideal condition except for fusion class. The performance measure parameter for different classes is shown in table no.2. as we can see from table no.2 the accuracy for each class is above 95%.
Table no.2
Performance measure of MLP for given FS
clas
s Accuracy (%) Sensitivity (%) Specificity (%) Predictivity (%)
Supraventricular 96 90.8 97.7 92.8
Ventricular 97 92.8 98.3 95.1
Normal 98 96.7 98.6 95.9
Fusion 95 92.0 96.1 88.7
V. DISSCUSSION
ANN based classifier are good tool for classification of ECG beats. MLP is one of the most simple and robust technique. In this study MLP classifier gives overall accuracy of this classifier is 96.5% and other parameter sensitivity, specificity and predictivity are 93%, 97.7% and 93.1% respectively. When we compare this work with previously done work we find MLP classifier have good classification performance in less computational load. A comparison table is given in table no.3.
Table no.3
literature classifier AAMI support classes accuracy
Chazel et al (2004) Linear discriminant yes 5 86.2%
Tanis mar et al LDA Yes 4 84.6%
MLP Yes 4 89.0%
Chazel et al (2006) Adaptable linear discriminent yes 5 94.0%
This work MLP yes 4 96.5%
VI. REFRENCES
[1] P. de Chazal, M. O’Dwyer, and R. Reilly, “Automatic classification of heartbeats using ECG morphology and heartbeat interval features,” IEEE Trans. Biomed. Eng., vol. 51, no. 7, pp. 1196–1206, Jul. 2004.
[2] P. de Chazal and R. Reilly, A patient-adapting heartbeat classifier using morphology and heartbeat interval features,” IEEE Trans. Biomed. Eng., vol. 53, no. 12, pp. 2535–2543, December 2006.
[3] Tanis mar, Sebastian zaunseder, juan Pablo, mariano llamedo and rudiger poll, “optimisation of ECG classification by means of feature selection,” IEEE Trans. Biomed. Eng., vol. 58, no. 8, pp. 2168–2176, August 2011.
[4] Jiapu pan and willis j. Tompkins, “A real-time QRS detection algorithm,” IEEE Trans. Biomed. Eng., vol. 32, no. 3, pp. 230–236, December 1985.
[5] M. Llamedo and J. P. Martinez, “Heartbeat classifier using feature selection with database generalization criteria,” IEEE Trans. Biomed. Eng., vol. 58, no. 3, pp. 616–625, Mar. 2011.
[6] K. Lai and K. Chan, “Real-time classification of electrocardiogram based on fractal and correlation analyses,” in Proc. 20th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., Hong Kong, 1998, pp. 119–122.
[7] M. I. Owis, A. H. Abou-Zied, A.-B. M. Youssef, and Y.M. Kadah, “Study of features based on nonlinear dynamical modeling in ECG arrhythmia detection and classification,” IEEE Trans. Biomed. Eng., vol. 49, no. 7, pp. 733–736, Jul. 2002.
[8] R mark and g.moody MIT-BIH database [online].
[9] Testing and reporting performance results of cardiac rhythm and ST segment measurement algorithms, 1998. Association for advancement of medical instrumentation (AAMI).