Examples of the ECG signals used: (a) Typical normal sinus rhythm (N) waveform; (b) typical ventricular tachycardia (VT) waveform; (c) typical ventricular fibrillation (VF) waveform 

Context in source publication

Context 1

... tachycardia (VT) and ventricular fibrillation (VF) are potentially life-threatening forms of cardiac arrhythmia. Fast and accurate detection of these conditions can save lives. We used semantic mining to characterize VT and VF episodes by extracting three significant parameters (frequency, damping coefficient and input signal) from electrocardiogram (ECG) signal. This method was used to analyze four-second ECG signals from a widely recognized database at the Massachusetts Institute of Technology (MIT). The method achieved a high sensitivity and specificity of 96.7% and 98.3%, respectively, and was capable of detecting normal sinus rhythm (N) from VT and VF signals without false detection, with a sensitivity of 100%. VT and VF signals were recognized from each other, with a recognition sensitivity of 96% and 94%, respectively. This newly proposed method using semantic mining shows strong potential for clinical applications because it is able to recognize VT and VF signals with higher accuracy and faster recognition times compare to existing methods. Keywords: ventricular fibrillation, ventricular tachycardia, ecg, ventiruclar arrhythmias, life threatening arrhythmias, semantic mining Ventricular fibrillation (VF) is a fatal form of arrhythmia that is considered to be life threatening. VF is potrayed by irregular contractions of the heart’s ventricles, resulting in a failure to pump blood and death within minutes unless corrective measures are taken promptly. An electric defibrillator can be used to return the heart to the normal sinus rhythm (N). However, if the combination of a normal sinus rhythm and ventricular tachycardia (VT) is misinterpreted as a false positive (FP) for VF, the patient receives a needless shock that could deteriorate the heart function and cause serious consequences to the patient. Conversely, an incorrect diagnosis for VF can result in a false negative (FN) and is life threatening. Therefore, the accurate and early detection of VT or VF is extremely important. Automated systems for the detection of VF using various detection methods, such as time- and frequency-domain analysis, neural networks, wavelet and nonlinear analysis, have been described previously by other researchers. These automated systems can recognize the VF signal from either a NSR or a VT signal. Most of these methods used a time-domain analysis approach because of the advantage of using simple algorithms and real-time analysis via a computer. The time-domain analysis includes an autocorrelation function (AFC) (Chen et al., 1987; Clayton et al., 1993), a threshold-crossing interval (TCI) (Clayton et al., 1993), time-delay methods (Amann et al., 2007), and threshold-crossing sample counts (TCSC) (Arafat et al., 2009). Frequency-domain analysis is usually performed using power spectrum analysis, whereby the VF signal is characterized by observing the power spectrum of the signal (the VF power signal is reported to vary from 4 to 7 Hz) (Clayton et al., 1993). Other methods using frequency-domain analysis include modified amplitude distribution analysis (MADA) (Fokkenrood et al., 2007) and multifractal singularity spectrum analysis (Wang et al., 2007). When artifial neural network (ANN) techniques have been proliferated in late 1990’s, most researchers opt for this technique because it has the ability to deal with nonlinear discrimination between classes (Clayton et al., 1994; Minami et al., 1999). In addition, wavelet transforms (Khadra et al., 1997; Abbas et al., 2004; Nawarvar et al., 2004) and nonlinear analysis (Jekova et al., 2002; Sun et al., 2005; Daoming et al., 2007) were proposed as detection techniques for the classification of VF signals. Despite the previous proposed methods shown promising results in terms of sensitivity and specificity, a few drawbacks have been identified to improve the accuracy of detection. Some of the proposed techniques are too complicated to be materialized in real time. For the discrimination of VT and VF signals, the detection time becomes the key factor to determine the patient’s fate. According to Minami et al. 1999, the time delay due to the detection of VT/VF signals must be as short as possible; otherwise, the patient is at risk of death. If the proposed technique is too complex, the processing time will be too long to be effective. Additionally, some of the previous methods only discriminate between N and VF signals, without testing for VT (Amann et al., 2007; Clayton et al., 1994; Jekova et al., 2002; Arafat et al., 2009). As mentioned above, VF must be accurately identified and discriminated from VT and N signals to prevent any unnecessary shock to the patient that could damage the heart. The accuracy of the detection technique must be as high as possible to prevent misinterpretation or a false negative, which would have fatal consequences for the patient. Previously, we have proposed a semantic mining technique that can characterize ventricular arrhythmia (V) from normal sinus rhythm (Othman et al., 2010; Othman et al., 2012). However, the previous study only concentrated on discriminating between N and V. Because of this reason, in this study, we expand the technique instead of only characterize V, we try to characterize VT and VF as well. This method is essentially different from other approaches because this algorithm can portray the heart’s oscillatory behavior by classifying behavior patterns using a semantic concept. This algorithm extracts the significant characteristics of the ECG signal (the frequency, damping coefficient and input signal) and classifies them into three types of rhythms: normal sinus rhythm (N), ventricular tachycardia (VT) and ventricular fibrillation (VF). The ECG data were acquired from the PhysioBank database (Goldbeger et al., 2000), which is a large and growing online database of well-characterized digital physiological signals and related data for use by the biomedical research community. In this study, the MIT ‘nsrdb’ (normal sinus rhythm database) with ECG recordings of people that only contains recordings of normal sinus rhythms and the MIT ‘cudb’ (The Creighton University ventricular tachyarrhythmia database) with ECG recordings of patients with sustained ventricular tachycardia and ventricular fibrillation were used. Figure 1 shows an example of the signal epochs used. The ECG signal for a normal sinus rhythm consists of identical PQRST waves. Typically, in a normal sinus rhythm ECG (Figure 1a), each P wave is followed by a QRS complex and then a T wave, and the rhythm is regular. In a ventricular tachycardia ECG (Figure 1b), no P waves are found, and the QRS complexes are wide and abnormal, with a rhythm that is sometimes regular and sometimes irregular. In Figure 1c, an ECG of ventricular fibrillation is shown; P waves and QRS complexes are absent, and the rhythm is chaotic. Usually, a raw ECG signal contains noise, which affects the ability of certain ECG recognition systems to recognize patterns in ECG signals (Arafat et al., 2009; Jekova et al., 2002; Pan & Tomkins, 1985). Examples of noise includes electrode motion artifacts, baseline wander and power line interference. Thus, a bandpass Butterworth was applied to ECG signals in order to reduce noise. A pass band from 1–30 Hz was chosen because the power spectrum of N, VT and VF signals are reported to be within this range (Chen et al., 1987; Clayton et al., 1993; Minamin et al., 1999). The transfer function of second order Butterworth bandpass filter ...