Department of Electronic Engineering, Chang Gung University, Kwei-Shan, Tao-Yuan 333, Taiwan [email protected]
Abstract
In this paper, a pattern recognition based algorithm developed for the discrimination of ventricular fibrillation (VF) and ventricular tachycardia (VT) is presented. The method jointly utilized temporal and spectral features, in conjunction with a linear discriminant function (LDF), for the task of VF/VT classification. Tests conducted using 70 ECG episodes collected from the MIT-BIH database were performed. The numerical results showed that the novel method achieved correct detection rates of 96.00% and 95.56% for VF and VT, respectively, permitting 95.71% overall detection accuracy.
1. Introduction
A number of previous literature/reports have indicated that cardiac arrhythmias account for over 50% of all deaths due to heart disease.
Prevention of such cardiac deaths requires rapid and accurate identification of lethal arrhythmias , such as ventricular fibrillation (VF) and ventricular tachycardia (VT), either from the surface or intracardiac electrocardiogram (ECG) tracings . Since VF and VT are required to be effectively treated by delivering different electrical energy shock at early stage, development of built-in algorithms f o r “smart”
defibrillators that are capable of differentiating VF from VT as well as making quick and accurate shock decisions is thus essential to reducing mortality from cardiac deaths. There were a number of detection algorithms for cardiac arrhythmias having been proposed
previously [1], [2], [3].
In this paper, a novel method based on time and frequency analysis is introduced. Two features, dubbed threshold crossing interval (TCI) and frequency spread, were defined in our study.
In general, after obtaining both features respectively by applying time and frequency analyses to the original ECG recordings, an optimal linear mapping procedure was then applied. Such a linear mapping is adopted to reduce a 2-D feature vector to a 1-D scalar so that the detection of VF and VT can be achieved simply by comparing the sole feature value with a pre-determined threshold that well separates VF from VT. Descriptions of the method and performance evaluation are given in the subsequent sections.
2. Method
A block diagram of the overall process flow is depicted in Fig. 1. Initially, an ECG signal is preprocessed by a bandpass filter with the passband of [2,20] Hz to remove the baseline drift, the motion artifacts and the 60Hz power line interference.
A. Feature Extraction
A time-domain feature was defined as follows.
First, a 5 s ECG signal was converted into a binary sequence by comparing the ECG sample value with an adaptive threshold. If a sample value was greater than or equal to the threshold, then a digit “1” was assigned to represent that sample; otherwise, a digit “0” resulted. Note that the threshold was selected as 20% of the peak
22
Figure 1.Block diagram of the process flow
value of a 1 s ECG segment and was updated every 1 s, thus producing an adaptive threshold. As a result, a number of pulses were generated. Such a thresholding scheme can reduce low-level noise and does not affect the heart signal. Moreover, adaptive thresholding leads to immunity from significant changes in the signal amplitude. Next, the number of pulses, denoted as N, was counted. The mean pulse interval, referred to as the threshold crossing intervals (TCI’s) corresponding to a binary sequence resulting from the original 5 s ECG signal can be estimated by TCI=5000/N (ms). TCI was then considered as the temporal feature in our method.
As for the spectral feature, it was defined in the following manners. The fast Fourier transform (FFT) was first computed for each original 5 s ECG signal.
Fig. 2 gives a typical plot of the magnitude frequency response obtained from a VF ECG recording. We here took the 50% of the maximum magnitude as the
“clip” and denoted the largest and the smallest frequencies whose corresponding spectral magnitudes were greater than the clip value as fh and fl, respectively. A feature referred to as frequency spread, denoted by S, was then defined as
S= fh - fl,
where S has the unit in Hz.
Figure 2. Spectrum of a VF ECG signal
B. Pattern Classification
To perform the task of classification, two features were jointly applied. A linear mapping derived based on a likelihood ratio test (LRT) was introduced for this purpose and formulated as:
=
S A TCI
y
T ,where A is a 2×1 linear mapping matrix used to transform from a 2-D feature vector to a smaller 1-D scalar y that retains the information relevant for classification [4]. In our study, the
matrix A was derived from an existing training database. As a result, the decision rule can be expressed as
If y≧T, => VF;
If y﹤T, => VT,
where T represented a threshold that can well differentiate VF from VT groups.
3. Results and Discussion
A subset of the MIT-BIH database consisting of 70 recordings (VF: episode numbers 1-25; VT: episode numbers 26-70) was adopted in our analysis. Each recording was 10 s in length (the sampling frequency fs = 250Hz). Consequently, the optimal results achieved by our algorithm were 96.00% (24 out of 25) for VF detection and 95.56% (43 out of 45) for VT detection, producing 95.71% overall detection accuracy, as highlighted in Table 1. A scattering plot of the classification results was also given in Fig. 3.
Table.1 Performance with different threshold values.
The optimal result is highlighted.
Finally, several points are addressed as follows.
First, when either time or frequency feature was solely applied, the performances resulting from both cases were both lower than that obtained from a joint use of both features, implying that the information
utilized for classification in time and frequency domain should be independent each other. Secondly, it should be noted that one limitation of our study is that the algorithm performance was tested on the same database for training process. Therefore, a larger database would be essentially required for prospectively evaluating the pattern classification performance for our method. Moreover, since the novel algorithm only required a short ECG recording (say, 5 s, for example), thus it can meet the requirements, such as rapid and accurate shock decision, for implementation in a practical automated external defibrillator
(AED)
Figure 3. Plot of classification (VF:x; VT:o)
4. References
[1] N. V. Thakor, Y. -s. Zhu, and K.-Y. Pan,
“Ventricular tachycardia and fibrillation detection by a sequential hypothesis testing algotithm,” IEEE Tran. Biomed. Eng., vol. 37, pp. 837-843, 1990.
[2] S.-W. Chen, “A two-stage Discrimi- nation of Cardiac Arrhythmias Using a Total Least Squares-Based Prony Modeling Algorithm,”
IEEE Tran. Biomed. Eng., vol. 47, pp.
1317-1327,Oct. 2000
[3] M. E. Cain, H. D. Ambos, J. Markham, B. D.
Lindsay, and Arthur, “Diagnostic implications of spectral and temporal analysis of the entire cardiac cycle in patients with ventricular tachycardia,” Circ., vol. 83, no.5, pp. 1637-1648, May 1991.
[4] K. Fukunaga, Introduction to Statistical Pattern Recognition. New York:
Academic, 1990.
T Detection Accuracy -5.6~-5.7 VF: 84.00%
VT: 95.56%
-5.8~-5.9 VF: 92.00%
VT: 95.56%
-6.0~-6.8 VF: 96.00%
VT: 95.56%
-6.9~-7.5 VF: 96.00%
VT: 93.33%
0 10 20 30 40 50 60 70
-20 -15 -10 -5 0 5
x--VF o--VT
24