This study proposes a method for ventricular late potentials (VLPs) detection using time-frequency representation and wavelet denoising in high-resolution electrocardiography (HRECG). The analysis is performed both with the signal averaged electrocardiography (SAECG) and in real time. A comparison between the temporal and the time-frequency analysis is also reported. In the first analysis the standard parameters QRSd, LAS40, and RMS40 were used; in the second normalized energy in time-frequency domain was calculated. The algorithm was tested adding artificial VLPs to real ECGs.
The ventricular late potentials (VLP) are high-frequency (in relation to the bandwidth of the electrocardiographic signal) and very-low-intensity signals. The presence of VLPs in the electrocardiographic signal has been associated with damages in the ventricular myocardial tissues. The necrosis or ischemic death of myocardial cells causes the formation of high-resistivity areas, where the propagation of cardiac action potential is delayed. This phenomenon affects the electrocardiographic signal with the presence of electrical activity, although of low intensity, between the end of the QRS complex and the initial part of the ST segment, where it should not be (Figure
Schematic ECG without (a) and with very exalted VLPs (b). Some characteristic points are also marked.
VLPs are localized at the end of QRS complex and in the initial part of the ST segment. Their intensity is at least two orders of magnitude smaller than the electrocardiographic signal, so they are usually “hidden” below the noise produced by the acquisition hardware and the electrical activity not related to the heart. For these reasons VLPs are not easily visible on the ECG.
Several statistical studies demonstrated a correlation between the presence of VLPs and the possibility of sudden cardiac death due to arrhythmia, often tachycardia. Patients with previous ischemic events are the most at risk. A correct VLPs detection makes the prevention of this serious malignant arrhythmias possible.
A classic electrocardiographic signal has amplitude of the order of a few mV and it contains most of the information at frequencies below 100 Hz. VLPs, if present, are considered nonstationary and non-Gaussian signals with an amplitude between 1 and 20
Table
Ventricular late potentials (VLPs) | |
---|---|
Causes | Areas of myocardium with reduced conductivity |
Effects | Onset of serious arrhythmias |
Bandwidth | 40–300 Hz |
Time location | End of QRS complex and initial part of ST segment |
Duration | <50 msec |
Amplitude | 1–20 |
The low-amplitude and high-frequency dispersion makes VLPs detection very difficult, often the signal is dominated by the noise. It is therefore necessary to process the signal to drastically reduce the noise level.
In this paper we propose a method to detect VLPs in two different conditions: analyzing a few minutes of prerecorded ECG with the use of signal averaged electrocardiography (SAECG) and examining the ECG during its acquisition. The first analysis allows the identification of weaker VLPs, while the second can be used for real-time diagnostic purposes.
Another technique of denoising, the wavelet denoising, was used in order to obtain a better detection and to optimize the real-time analysis. The simultaneous use of wavelet denoising and SAECG, in postacquisition processing, yielded good results also with short-term ECG. However, the most innovative feature of the work regards the search for parameters, in the time-frequency domain, that confer robustness to the method, as the introduction of a parameter not influenced from the J point location (ENEND). Finally, a bivariable separation between the time-frequency parameters improved the effectiveness of the method.
Although VLPs can be of great importance in the arrhythmic risk prevention, their detection is not yet widespread, due to the lack of suitable equipment. Our work is an attempt to solve this problem by providing a robust and reliable algorithm.
For clarity, we divided the algorithm into two phases: the preprocessing and the detection phase (Section
Real electrocardiographic signals, provided by “PhysioNet” database, were used to develop the algorithm. We chose “PTB Diagnostic ECG Database,” a collection of real ECGs acquired by the Physikalisch Technische Bundesanstalt (PTB), the German national metrology institute [
The aim of the signal pre-processing is the reduction of the noise level in the ECG record. For this purpose two denoising techniques will be used: the wavelet denoising and SAECG. The end result will be the vector magnitude (VM), which contains the information of all leads in a single signal. Figure
Steps of the pre-processing phase.
The first operation consists in filtering the signal. The filter performs two functions simultaneously: the first is to remove both the DC component and the low-frequency oscillations (which can be due for example to breathing) and the second is to limit the bandwidth to the component of interest, in order to limit the noise energy.
The studies carried out by Jane Raimon, Pablo Laguna, and Pere Caminal have demonstrated the superiority of nonlinear phase filters for this application. In particular the filter which gives the best results is a band-pass Butterworth filter of the fourth order, with cut-off frequencies at 25 Hz and 300 Hz. Figure
Effect of a fourth-order Butterworth filter on an ECG signal.
In some signals we found the presence of peaks in frequency, located at 50 Hz and its harmonics, which are most likely due to a poor filtering of the power supply in the acquisition phase. So it was necessary to remove these components with notch filters. We used a second-order filter, with attenuated band of 5 Hz around the detected peak. The effect of the filter is shown in Figure
Effect of a second-order Butterworth notch filter on ECG and its frequency spectrum.
This operation can be performed either in the acquisition phase or after the calculation of the SAECG.
One of the first methods developed to reduce the noise power in an electrocardiographic signal is the signal averaged electrocardiography (SAECG). This method is highly effective, and it is used in the preliminary analysis of all algorithms for VLPs detection. It is based on the principle according to which, averaging
The SAECG is, therefore, obtained by averaging each of these individual beats, to obtain a signal with a very low noise level, suitable for the detection of small-intensity signals, as the VLP. The voltage noise, however, is proportional to the standard deviation, which is reduced by a factor
Several techniques have been developed for the beats alignment. In particular the authors Jane et al. have distinguished three main variants that include all the methods used in VLP detection algorithms [ Double-level method: an amplitude threshold is set in such a way that the signal exceeds it only in correspondence of the QRS complexes. The alignment is determined averaging the instant at which the threshold is exceeded by the signal and the instant in which the level returns below it. Normalized integrals method: each beat, called Matched filtering method: this is a classical method for the detection of a known signal in the presence of additive white Gaussian noise. It consists of a linear time-invariant system with impulse response
LTI system of a matched filter.
The latter is the most recent and effective method. In the algorithm we use a variant of it, introduced by CD Woody [
Cross-correlation between template and signal. The arrows show the peaks to be considered for the alignment.
The relative maximum of the cross-correlation, marked in Figure
The SAECG is obtained by aligning the individual beats and averaging them, repeating this process for each lead available (Figure
Example of SAECG.
The effectiveness of the SAECG increases with the number of beats analyzed. However, it is not always possible to operate on long time ECG; in these cases it is therefore necessary to use alternative denoising methods. For this purpose we have chosen, for its direct effect on individual beats, the wavelet denoising. Details of this technique can be found in [
Wavelet denoising general scheme.
For decomposition and reconstruction we used the family of wavelets “Coiflets 5,” which gives the best results, decomposing the signal into five levels. The choice of the threshold value is very important, we do not recommend to set this value as a constant. Better results are obtained by setting a threshold proportional to the residual noise. The noise level is measured in the second level of decomposition, calculating the standard deviation of the signal in a section in which there is only noise.
The vector magnitude quantifies the energy measured by the three bipolar leads:
The result of the pre-processing phase is shown in Figure
Example of a VM relative to an ECG with VLPs.
The aim of this phase is the measurement of parameters that are directly influenced by the presence of VLPs. To make a comparison we performed both time and time-frequency analysis. Figure
Steps of the detection phase.
The J point marks the end of the QRS complex. In a healthy ECG, it separates a section characterized by a wide signal from a section without electrical activity. In the presence of VLPs this separation is not so sharp, and it is difficult to locate the point. We propose a method for J point location based on energy comparison. The principle was introduced by Legarreta et al. [
The first step is the estimation of residual noise level. For this purpose the VM is binned with time steps of 10 msec. Then the energy in each trait is evaluated, and the minimum energy is associated with the noise energy (Figure
VM division for residual noise estimation.
An energy threshold, proportional to the noise energy, is then set. Starting from the R wave peak, which is the absolute maximum of the VM, the threshold is compared with the energy evaluated in a 10 ms interval after each ECG sample. The first point that provides an energy below the threshold is considered as the J point.
The same method is then used to locate another reference point on the ECG, we called it “QRSoff.” It represents the end of the QRS complex and has to be independent of the possible presence of VLPs. This means that the energy threshold used for its identification must be significantly greater than the one used for J point detection. This point is useful in time-frequency analysis in case of incorrect location of the J point. The best results were obtained with thresholds about 5 and 100 times greater than the noise energy, respectively, for the J point and QRSoff. However, the optimal values are different by changing the acquisition hardware, and then the signals characteristics. It is, therefore, necessary to test and analyze the results of a substantial number of acquired ECGs, until obtaining a threshold level that allows optimal localization of the points in all signals.
Figure
Example of J point and QRSoff location on VM with and without VLP.
Once the J point has been determined, the measurement of temporal parameters is almost immediate. The standard parameters for the time analysis [ QRSd: QRS complex duration. RMS40: root mean square voltage of the terminal 40 msec of the QRS complex. LAS40: amount of time that the QRS complex remains below 40
Definition of QRSd (top), RMS40 (middle), and LAS40 (bottom).
The measurement of QRSd requires the knowledge of the QRS complex onset. However, the VLPs are present only at the end of the complex. For this reason we have redefined the parameter QRSd as the temporal distance between the R wave peak and J point, influenced equally by the VLP (Figure
The time-frequency representations (TFRs) are a very effective tool for VLPs detection, due to the localization of VLPs in both domains. Unlike the time analysis, a standard for this approach has not yet been defined. Laciar and Orosco proposed the use of normalized energies as parameters for VLP detection [
Representation of the VM signal (top) and its WVD (bottom) in the time-frequency plane and in 3D.
The normalized frequency in Figure
The first value of energy is calculated in an interval around the J point. The extremes of the window that gave the best results, based on the characteristics of VLP signal, are as follows:
The energy in this area, called
We introduced a second energy value, called
This value is very important in case of wrong localization of the J point. However, the calculated energies require a normalization. For this purpose, they are divided by the energy of the QRS complex (
The normalized energy indexes are, therefore, as in (
Electrophysiological abnormalities may change on a beat-to-beat basis, resulting in a failure of signal-averaged recordings to identify changes related to arrhythmogenesis. The VLP detection during the ECG acquisition is therefore very important. An adequate noise reduction is required in order to obtain a good effectiveness. The impossibility of using the SAECG to increase the signal-noise ratio is the main problem of real-time analysis.
The real-time application of the algorithm requires the modification of the input phase. In particular, a peak detector should be introduced in order to identify the individual beats exploiting the high amplitude of the QRS complex. Each heartbeat is then sent to the processing phase. We propose a scheme as in Figure
Input stage steps for real-time processing.
The processing phase is the same used in the postacquisition analysis, without the SAECG calculation step.
To test the method we used 60 real electrocardiographic signals with artificial VLP (aVLP). The VLP component has been simulated with a sum of sine signals with a definite amplitude and frequency, added to the first milliseconds of the ST segment [
The constant
The results obtained can be displayed by histograms. Figure
QRSd (top), RMS40 (middle), and LAS40 (bottom) parameters related to healthy ECG (blue) and ECG with aVLP (red).
The parameters are distributed according to a Gaussian law. Our optimal threshold values, as well as the corresponding effectiveness, are reported in Table
Effectiveness of temporal parameters.
Temporal parameters | ||
---|---|---|
Parameter | Threshold | Effectiveness |
QRSd | 95 | 96.7% |
RMS40 | 63.5 | 91.7% |
LAS40 | 44.2 | 85% |
For time-frequency parameters we chose to use bivariable analysis, to improve the effectiveness [
Bivariable time-frequency parameters separation between healthy ECG (blue) and ECG with aVLP (red).
Using the straight line ENEND = EN/10 we obtained a complete separation between the two groups. This separation is not unique but heavily depends on acquisition hardware and signals characteristics.
To test the effectiveness of the real-time processing we have generated signals with a known number of aVLPs. The effectiveness represents the ratio between the number of VLPs detected and the real number of VLPs in the signal. For the temporal analysis the presence of VLPs was confirmed if at least two of the three parameters (QRSd, LAS40, and RMS40) were outside the established range. Instead, for the time-frequency analysis, VLPs were confirmed if the point in the plane (EN, ENEND) exceeds the separation line ENEND = EN/10. Figure
Values of effectiveness for each analyzed signal, with both time-frequency analysis (graphic at the top) and temporal analysis (graphic in the lower). VLP amplitude of 10
It is clear that 3
Average effectiveness obtained with VLPs of different amplitude (3, 4, 6, 8, and 10
We developed a new method for ventricular late potentials (VLP) detection in high-resolution electrocardiographic (HRECG) signals, suitable for both postacquisition analysis and real-time applications. The method was evaluated with a group of 60 healthy ECG, provided by “PTB Diagnostic ECGs Database,” with the addition of realistic artificial VLP (aVLP).
The innovative features of the work, in addition to taking advantages of the best methods available in the literature, concerning a phase of preprocessing with the simultaneous use of two denoising techniques (the SAECG and the denoising wavelet) and a detection phase based on the bivariable separation of time-frequency parameters. Very important is the independence of the parameter ENEND from the J point location, which may be easily subject to wrong localization.
A comparison between the time and the time-frequency analysis demonstrated the real improvement achievable with the time-frequency parameters. The results obtained show how, with the time-frequency representation, it is possible to have an effectiveness of 100%, using the signal averaged electrocardiography (SAECG), and an effectiveness higher than 90% for the beat-to-beat analysis, considering VLP of amplitude greater than 6
The algorithm for automatic VLP detection can be implemented by software, for example in the Holter ECG, or by firmware, directly associated with an acquisition hardware. Future developments may include a statistical study of a large number of ECGs in order to standardize the thresholds of parameters and optimize the time-frequency analysis.