Data collection
12 lead ECGs of the 90 patients without structural cardiac disease, in whom PVC or VT had been ablated successfully, were used. The data was collected in Tehran Arrhythmia Clinic. The PVC foci were confirmed during the ablation procedure and attached as a label of data. Based on EPS reports, foci were divided into six subgroups including: LVOT (anterior & posterior), RVOT septum, basal RV, RVOT free-wall, Aortic cusp (Left Coronary cusp, Right Coronary cusp & Non-Coronary cusp) and LV body.
According to the EPS reports, 8 out of all PVCs were originated from the LVOT. Of the 52 PVCs originating from the RVOT, 15 were grouped as of the free-wall origin, 37 of the septum zone. 18 of PVCs had originated from the aortic cusp. 3 of PVCs were arisen from the LV body. The origins of remained PVCs were identified as the basal RV. Due to the lack of patients whose PVCs originating in LV body (there were only 3 patients), this group was omitted and study was done with remaining 87 patients.
Figure 1 shows a sample of 12-lead ECG of patients with PVC originating from RVOT septum.
PVC detection
First of all, PVCs are recognized and distinguished from the normal beats. Because of their greatness in height, depth and length, PVCs could easily be detected. Since 12 leads ECG signals were recorded simultaneously, one lead was used to detect duration of the PVC beat and finally the PVC beats were extracted in all leads. Because PVC was more obvious than normal one in lead III, this lead was used for detecting PVC heartbeats.
In order to attenuate high frequency components, the ECG signal was filtered using a low pass equiripple finite-duration impulse response (FIR) filter with cut-off frequency at 25 Hz implemented with MATLAB 7.8.0 (R2009a) (The Mathworks, Inc.). In this way, considering lower frequency content of PVCs, beats would become more obvious than normal one and would be detected easier.
The standard parameters of the ECG waveform can be determined with high accuracy using wavelet transforms (Sumathi & Sanavullah 2009; Bensegueni & Bennia 2012). The wavelet transform (WT) provides a representation of the signal in time-scale domain, allowing representation of the temporal features of a signal at different resolutions. Detection of these main points is based on maximum absolute values and zero crossing of WTs at specific scales (Chang et al. 2013).
The Haar wavelet was used in our study at scale 21 for PVC detection (Sumathi & Sanavullah 2009). R peak in PVC beat is detected by marking the zero crossing of the WT between positive maximum-negative minimum pair (Martínez et al. 2004). Ascendant edge of the wave is relevant to negative minimum and decreasing edge of wave is relevant to positive maximum.
First we chose a peak greater than the threshold (0.1) and then absolute of WT signal was calculated. Then we determined the nearest peaks before and after the chosen peak and finally selected the greater one. The first minimum before the initial peak and the first minimum after the second peak were specified as onset and offset of PVC beat. Because of simultaneous recording of ECG leads, duration of onset and offset of PVC beat was assigned to other leads and PVCs were determined in all leads.
Features extraction
In this study we extracted the morphological, frequency and spectrogram features after PVCs detection to classify their five foci. Since a physician classifies arrhythmia with the information of rhythm and morphology, an input vector can consist of features that illustrate the rhythm and morphology properly (Song et al. 2005).
One of the morphological features is polarity of each-lead ECG signal. The positive, biphasic or negative polarity was considered with 1, 0, or −1 respectively. Existence or absence of notching (Yamashina et al. 2011) in signal was shown with 1 or 0. Notching in QRS complex is determined as a tri-phasic R or Q wave with an interval greater than 40 m-sec between the first and second peak of the QRS complex. Existence of notching is considered when notching is observed in more than three of the six limb leads. As it was obtained, notching was observed more often in the PVCs arising from the free-wall rather than in the PVCs originating from the septal region (Tada et al. 2007). Figure 2 shows a sample of notched QRS complex.
Spectrum content of ECG signal in various leads is another feature that is used in classification of PVCs in this study. Frequency band of P and T waves in lead II, the most applicable lead, is approximately 0.5 to 10 Hz and QRS complex has frequencies ranging approximately from 3 to 40 Hz. Heart rate dependencies of waves in ECG must be taken into consideration. Heart rate changes will cause changes in waveforms and frequency contents. So some frequency components of ECG signal can be defined as classifying features and Fast Fourier Transform (FFT) was used for this purpose. Fourier transform of a signal and its inverse are calculated by following equations (Subha et al. 2010).
(1)
(2)
In these equations, t represents time and f denotes frequency, x(t) is original signal in time domain and X(f) represents its Fourier transform in frequency domain.
The Short-time Fourier transform (STFT) is obtained from the Fourier transform by multiplying the time signal x(t) by a window function (Tokmakei & Erdogan 2009; Hardalac et al. 2007).
(3)
Spectrogram, magnitude of STFT, shows frequency characteristics of signals in the time domain. The average of spectrogram greater than the 70 percent (Sheikkani et al. 2012) was computed for all leads and used as a feature for separating the groups.
The other spectral features are as follows: Maximum amplitude of spectrum signals, variance of signals’ spectrum, average of FFT of signal, power spectrum and power spectrum bandwidth of signals, the average of amplitudes of spectrum signals greater than the 70 percent and the average of Haar wavelet coefficients at level 1 and 2.
Statistical analysis and classification
Since the assumptions of normal distribution and similarities were valid, statistical analyses of One-Way ANOVA were performed in SPSS 17.0 software to evaluate differences of features. The p< 0.05 was considered to show significant differences. Mahalanobis distance (Yeh 2009) was used for classifying the five groups by considering morphological, spectrogram and frequency features.
To determine the separation ability of pairwise groups, we used SVM classifier in MATLAB. Support Vector Machines (SVM) are supervised learning models with associated learning algorithm that analyze data and recognize patterns, used for classification. The basic SVM takes a set of input data and predicts, for each given input, which of two possible classes forms the output. It is more considerable using SVM classifier than Mahalanobis distance to discriminate two groups. We set data and their label of each group in separate matrixes, defined train and test data, then classified two groups each time and computed errors and compared them.