ECG arrhythmia classification using simple reconstructed phase space approach

ECG Arrhythmia Classification Using Simple Reconstructed Phase Space Approach AS Al-Fahoum1, AM Qasaimeh2 1 Biomedical Centre of Excellence, Yarmouk University, Irbid 21163, Jordan 2 Jordan University of Science and Technology, Irbid, Jordan important tool in ICU and CCU that enables online Abstract monitoring of the cardiac activities that require special ECG arrhythmias such as ventricular and atrial algorithms for detection and prediction [2]. There are arrhythmias are one of the common causes of death. several methods used to detect cardiac arrhythmia These abnormalities of the heart activity may cause an depending on heart rate variability [3], spectral analysis, immediate death or cause a damage of the heart. In this time-frequency distribution [2], [4-6], and nonlinear paper, an arrhythmia classification algorithm is signal processing techniques [3], [7-11] presented. The proposed method uses the nonlinear R Acharya, et al. [3] presents a classification algorithm dynamical signal processing techniques to analyze the of heart arrhythmias depending on the heart rate ECG signal in time domain. The classification algorithm variability (HRV) signals. Linear and nonlinear is based upon the distribution of the attractor in the parameters are calculated to differentiate the different reconstructed phase space (RPS). The behavior of the types of arrhythmias; the linear parameters are estimated ECG signal in the reconstructed phase space is used to in the time domain and frequency domain; the nonlinear determine the classification features of the whole parameters also are calculated to specify correlation classifier. To evaluate the performance of the presented dimension (CD), largest Lyapunov exponent (LLE) and classification algorithm, data sets are selected from the approximation entropy (ApEn). Also the phase space MIT database. Two groups of data, learning and testing plots are obtained to differentiate the different types of datasets, are used to design and test the proposed arrhythmias from the shape (behavior) of the phase space. algorithm. A classification sensitivity and specificity of L. Khadra, et al.[2] present a high order spectral analysis 100% are used to fine tune the parameters of the selected algorithm to classify different types of heart arrhythmias features using the learning dataset. Forty five signals are such as AF, VT and VF. The method used was bispectral used to test the proposed approach resulting in 85.7- analysis technique. F.M.Roberts, et al. [7] obtain the 100% sensitivity and 86.7-100% specificity are obtained reconstructed phase space for using the ECG leads II and respectively. VI by plotting them against each other for four different types of arrhythmias using 100 features extracted from 1. Introduction the RPS and used as input to artificial neural network THE ELECTROCARDIOGRAM (ECG) is the graphical (ANN) classifiers. R.J.Povinelli, et al.[8] use the representation of the electrical activity generated by the reconstructed phase space for different types of heart [1]. The first stage of the heart beat begins when the arrhythmias by using ECG leads II and VI for different sino-atrial node (SA) depolarizes. SA node, located in the segment time intervals from 0.5 -3.0 seconds using 101 right atrium, is the pacemaker of the heart, depolarizing in features (attractors) extracted from the reconstructed regular time interval to ensure proper pacing. Then the phase space. R.J.Povinelli, et al. [9] use distribution electrical signal moves rapidly through the heart muscle models as statistical representations over multi- with normal rhythmicity. If the electrical system of the dimensional reconstructed phase space both heart does not properly function, the hart's rhythm nonparametric distributions based on binning and becomes abnormal due to the firing in the SA node or the occurrence counts and parametric distributions based on transmission of the signal throughout the heart muscle. Gaussian Mixture Model (GMM) are used. F. M. Roberts, These abnormalities can be monitored using changes in et al. [10] filter different types of heart arrhythmias into the ECG recording whether in its behavior or rate. four sub-bands: 0.5-5, 5-10, 10-20 and 20-32 Hz. A phase Reconstructed phase space can easily differentiate these space is constructed with embedding dimension of three behavior differences depending on the signal distribution. and a time lag of 20. R. J. Povinelli , et al. [11] use the Quantitative classification of cardiac arrhythmia is an phase space to classify four different ECG rhythms by ISSN 0276−6547 757 Computers in Cardiology 2006;33:757−760. using the global false nearest-neighbor technique to available at PhysioNet website [12]. Different datasets are calculate RPS dimension and build Gaussian Mixture selected, each dataset represents different type of heart Models (GMM) for each signal class from the arrhythmia; MIT-BIH Normal Sinus Rhythm Database reconstructed phase spaceType the introduction here. (NSRDB) for normal sinus rhythm, CU Ventricular Tachyarrhythmia Database (CUDB) for ventricular 2. Methods tachycardia, AF Termination Challenge Database This work is dealing with classification problem of (AFTDB) and MIT-BIH Atrial Fibrillation Database four different types of heart arrhythmias; normal sinus (AFDB) for atrial fibrillation, and MIT-BIH Malignant rhythm (NSR), atrial fibrillation (AF), ventricular Ventricular Ectopy Database (VFDB) for ventricular fibrillation (VF) and ventricular tachycardia (VT), fibrillation arrhythmia. In order to create a database for depending on their distribution in the reconstructed phase the classifier implemented in this research, the records are space. In previous work [7]-[11], as mentioned above, the divides into two groups: RPS was used in cardiac arrhythmias classification, at Group I (learning set) contains 40 records, these least 100 features were extracted from the RPS to records are divided into four categories, each category has distingue each rhythm, while only three classification 10 records which represents one type of heart arrhythmia. parameters are used in this proposed classification Each record has 5 seconds time duration. This group of algorithm. data is used for building the classifier. A two–dimensional phase space plot may explain the Group II (test set) contains 45 records; 14 records for structure which is hidden in the dynamics. In such a plot normal sinus rhythm (NSR), 15 records for atrial each data point is plotted versus the value sampled at a fibrillation (AF), 8 records for ventricular tachycardia chosen fixed time delay earlier. The formal basis of this (VT) and 8 records for ventricular fibrillation (VF). This simple tool lies in the concept of phase space dataset is used for testing the classification process. reconstruction. Each point in the RPS is calculated as follows The proposed algorithm uses small segments of waveform, each waveform consists of 5 seconds tracing. xn =[xn−(d−1)τ ..... xn−τ xn ] A 250 Hz sampling frequency is used for all the datasets, so the data is re-sampled to 250 Hz. From the nature of For n = (1 + ( d − 1)τ ).....N the ECG signals coming from different individuals and Where N is the dimension of the time series, τ is the different recording periods for the same individuals, it is delay time and d is the embedding dimension. Then the obvious that these signals have different amplitudes and entire phase space is generated by base lines. These variations are due to the muscle artifacts  x1 ⋯ x1+( d −1)τ  and the power line interferences. In order to correct these  x ⋯ x2+(d −1)τ  effects, the data must be adjusted to have standard x1+τ statistical parameters. The mean of the signal is calculated X= and subtracted so that the signal is zero meaned. Next, the x2+τ  ⋮ ⋮  2 signal is divided by the standard deviation to give a unit   ⋱ variance. xN −(d −1)τ xN  2.1.2. Phase space reconstruction xN −(d −2)τ ⋯ To reconstruct the phase space for the data, the time lag and the embedding dimension should be determined. In phase space reconstruction, different time The time lag can be determined by using the first series fill different subset in the phase space; this subset is minimum of the automutual information function method, called an attractor. In other words, there are different the first zero crossing of the autocorrelation [13], or patterns or trajectories in the RPS that are produced by empirically to obtain maximum classification accuracy. the trajectory matrix, these different geometrical The dimension can be selected using the false nearest distributions are used to characterize different time series neighbors, Cao’s method [14], or empirically. To in the RPS. reconstruct the phase space for the ECG signals the 2.1. Analysis of ECG signal following steps are performed The first minimum of the automutual information function is calculated for each time series. 2.1.1. Data and pre-processing A histogram of the calculated time lags is drawn, and The datasets are chosen from MIT-database which is the peak value is chosen as the time lag. An embedding dimension of 2 is chosen empirically to 758 achieve maximum classification accuracy. The classification algorithm is tested using different A two-dimensional RPS is plotted for all arrhythmias embedding dimension, lag time and different signal using the obtained time lag. duration. For different time lag it concludes that the over The set shows three peaks at 19, 21 and 22. Therefore, all accuracy of the classification algorithm remains a time lag of 19 is chosen. approximately constant over the region with time lag of 9-19 units, which is 91.1%. This behavior is obtained 2.1.3. Features extraction because the attractors in the reconstructed phase space Features extracted from the reconstructed phase space reserve there dynamical behavior in this region, while the are depending on the distribution of the data in the phase overall accuracy below and above this range is space to follow the transmission of the ECG signal in the decreasing. heart muscle, in other words, these features are depending The results for time duration less than and equal to 5 on the geometric structure of the attractor in the seconds show that the sensitivity and specificity increase reconstructed phase space. Three boxes are chosen in the while increasing the signal duration, however, in some RPS to extract such features as follows: types of arrhythmias sensitivity and specificity remain Three distinguished boxes (a, b and c) are determined constant regardless of the time duration which indicates depending on data density distribution in the phase space, that the classification process do not completely depend where on the time duration of the classified signal. This result a is the box in the RPS centered at zero with –0.5 and indicates that the nonlinear dynamical characteristics of 0.5 edges, the ECG signal do not completely lost when reducing the b is the box in the RPS bounded by time duration. This conclusion goes in line with what was { X ≤ −1.2,−0.5 ≤ X + τ ≤ 0 } obtained in [8]. Table 2 shows the sensitivity and specificity for the test dataset versus different signal and duration. The results of the classification accuracy for test c is the box bounded by { X + τ ≤ −0.2 } data versus the embedding dimension show that the The percentages of the number of points bounded by overall accuracy remains constant for embedding the three areas (P(a),P(b), and P(c)) are calculated with dimension of 2 to 6 which is 91.1%. Using other values respect to the whole number of points in the RPS. will decrease the obtained accuracy. It can be concluded A classification rules are generated depending upon that the reconstructed phase space is to be topologically the distribution of the arrhythmias bounded by these equivalent to the original state space of the system when boxes. the embedding dimension is suitable, in our experiments The reconstructed phase spaces were initially dimension of 2-6 seems suitable. created for the learning dataset. Each type of ECG arrhythmia is found to occupy a distinguished geometrical 4. Discussion and conclusions distribution in the RPSDescribe your methods here. The method presented in this paper deals with the 3. Results nonlinear dynamical behavior of the ECG arrhythmias, which is used to identify the cardiac arrhythmias. The Figure 1 shows the overall classification algorithm method used here is different from the previous based on the selected threshold values of the predefined approaches that used the reconstructed phase space in classification parameters. The threshold values of arrhythmias classification which used many classification 31.27%, 0.63% and 54.22% for P(a), P(b) and P(c) parameters [7-11]. Since heartbeats depend on other respectively are chosen for the whole classification bodily events such as hormone and chemical levels, it can process. The false positive (FP) is defined as the number be modeled as a nonlinear system. The nonlinearity in the of misclassified signals, and the false negative (FN) is the behavior of such a system can be captured by the RPS number of signals that are classified as a part of group that contains state variables and relationships between when they are not. Sensitivity is defined as the ability of state variables that provide greater differentiability across the classifier to classify a certain signal as being a part of classes than the original state variable by itself [2, 9]. the group it actually belongs to. Specificity refers to the Comparison between different methods that were used ability of the classifier to correctly rule out signals that do in cardiac arrhythmia classifications and the proposed not belong to the group [2]. Table 1 shows the results of approach shows that the sensitivity and specificity for the the sensitivity and specificity of the classification proposed algorithm are within the range of 85.7-100% algorithm using data in group II which contains 45 and 86.7-100% respectively. These results outperform waveforms. While for the learning data the sensitivity and those results that are provided in [7-8]. The classification specificity were 100%. accuracy is 100% for VF arrhythmia which is the most dangerous type among other arrhythmias. Other research 759 groups achieved a 100% accuracy only for the normal reconstructed phase spaces. IEEE Transactions On case, and the classification accuracy provided by other Knowledge And Data Engineering 2004; 16: 779-783. researchers for classifying VF arrhythmia were 91.7% [12] The research resource for complex physiologic signals, [2], 96.5% [9], 88% [8] and 95.1% [7]. PhysioNet. From the Web site www.physionet.org [13] Kantz H. and Schrreiber T., Non Linear time series The simplicity of the algorithm can be helpful for the analysis. Cambridge: Cambridge University Press 1997. real time implementation of the classification algorithm to [14] Cao L., Practical method for determining the minimum decrease the time needed both in classification itself and embedding dimension of a scalar time series. Physica D in providing the suitable therapy to the patients. Future 1997; 110: 43-50 work is needed to increase the classification accuracy of [15] Frye S. J., Cardiac Rhythm Disorders: An introduction the proposed algorithm; this may be done by combining using the nursing process. Williams & Wilkins 1988, this method with other classification methods. Baltimore, MD, USA Acknowledgements Authors would like to thank Yarmouk University for its continuous support. References [1] Grauer K. and Curry R., clinical Electrocardiography: A primary care approach. Blackwell Scientific Publication,1992, Boston, USA [2] Khadra L., Al-Fahoum A., and Binajjaj S., A Quantitative Analysis Approach for Cardiac Arrhythmia Classification Using Higher Order Spectral Techniques. IEEE Transactions on Biomedical Engineering 2005; 45: 1878- 1885. [3] Acharya R.U., Kannathal N. and Krishnan S.M., Comprehensive analysis of cardiac health using heart rate signals. Physiol.Meas.2004;25: 1139-1151 [4] Addison P.S., Watson J.N., Clegg G.R., Holzer M., Sterz Fig. 1 The overall classification algorithm F., Robertson C.E. , Evaluating arrhythmias in ECG signals using wavelet transforms. IEEE Engineering In Medicine And Biology September/October 2000: 104-109. [5] Sun Y., Chan K. and Krishnan S., Life-threatening ventricular arrhythmia recognition by nonlinear descriptor. Table 1: Sensitivity and specificity of the classification BioMedical Engineering OnLine 2005; 4:1-11. algorithm for the test group [6] Afonso V. and Tompkins W., Detecting Ventricular Fibrillation. IEEE Engineering In Medicine And Biology Sensitivity 1995; 14: 152-159. FP FN % Specificity % [7] Roberts F., Povinelli R. and Ropella K., Identification of NSR 2 1 85.7 92.9 ECG Arrhythmias using phase space reconstruction. AF 1 2 93.3 86.7 Proceedings of principles and practice of knowledge discovery in database (PKDD'01) 2001; Freibureg , VT 1 0 87.5 100 Germany, 411-423. [8] Povinelli R., Roberts F., Johnson M., and Ropella K, Are VF 0 1 100 87.5 nonlinear ventricular arrhythmia characteristics lost as Table 2: Sensitivity and specificity of the classification signal duration decreases? Computers in Cardiology 2002 algorithm versus signal duration (test dataset) ;29: 221-224 [9] Povinelli R., Johnson M., Lindgren A., Roberts F. and Ye Duration (s) Sensitivity % Specificity % J., Statistical Models of Reconstructed Phase Space for signal classification. IEEE Transactions on Signal 1 57.1-100 73.3-100 processing 2006; 54:2178-2186. 2 50-100 66.7-92.9 [10] Roberts F., Povinelli R. and Ropella K., Rhythm 3 64.3-100 75-92.9 classification using reconstructed phase space of signal 4 71.4-100 75-92.9 frequency sub-bands. Computers in Cardiology 2003; 30: 61-64 [11] Povinelli R., Johnson M., Lindgren A. and Ye J., Time Amjed Al-Fahoum, Hijjawi Faculty for Eng. Technol., Yarmouk series classification using Gaussian mixture models of University, Irbid 21163, Jordan. (afahoum@yu.edu.jo) 760