Review of Data analysis methods for Denoising and Characterizing ECG.

Bioelectrical signals express the electrical functionality of different organs in the human body. The electrocardiogram, also called ECG signal, is an important signal among all bioelectrical signals. It gives valuable information about the functional aspects of the heart. The ECG pattern consists of a recurrent wave sequence of P-, QRS- and T-wave associated with each beat. The automatic detection of ECG waves is important to cardiac disease diagnosis. In a clinical setting, such as intensive care units, it is essential for the automated systems to accurately detect and classify ECG wave components. The correct performance of these systems depends on several important factors, including the quality of the ECG signal, the applied denoising methods and classification rule, the learning and testing dataset used.

Fig. 1 Typical ECG cycle.

Some of the heart disease can be silent killer which needs the sudden attention immediately after the considerable change in the ECG and symptoms of the patient. Cardiologists say that in order to classify the heart disease 2 out of 3 criteria should be satisfied. First, the ECG and the lead information, second, patient symptoms and third Enzymatic test. Since Enzymatic tests are expensive, where one should take a sample from the patient and the testing should be carried out in a chemical laboratory, ECG and the symptoms of the patient are sufficient to decide the disorder. Abnormalities can also be due to exercise, heavy carbohydrates, tobacco, drugs and consuming potassium. Normally doctors prefer to monitor ECG as seen from various leads such as limb, augmented and precordial leads. There should be a verbal communication between the patient and the doctor in order to acquire the symptoms of the patient.

The first deflection, termed as the P-wave is due to the depolarization of the atria. The large QRS complex is due to the depolarization of the ventricles. This is the complex with largest amplitude and is easy to detect. Numerous methods are reported in literature for the detection of QRS complex. The last deflection is T-wave. It corresponds to the ventricular repolarization of the heart. Reliable detection of P- and T-wave is more difficult than QRS-complex detection for several reasons including low amplitudes, low signal-to-noise ratio, amplitude and morphological variability. The P-wave may be even absent from some ECG recordings. Over the last few years, the P- and T-wave detection and delineation problem has been addressed using different approaches. In most of these methods, P- and T-waves are detected relative to the position of QRS-complex by applying appropriate threshold. The main problems of the thresholding techniques are their high noise sensitivity and their low efficiency when dealing with odd morphologies. This results in more false positive and false negative detections. Further, some of the algorithms detects only monophasic P- and T-waves and not suitable for biphasic P- and T-waves. Therefore, more sophisticated techniques are needed to facilitate the development of new detection schemes with higher detection accuracy and suitable for all kind of morphologies of QRS-complex, P- and T-waves.

Most biomedical signals appear as weak signals in environment that is consisting of many signals of various origins. The sources of noise could be physiological, the instrumentation used or the environment of the equipment.

ECG recordings are often contaminated by high frequency noise known as residual power-line (PL) interference despite the very high common mode rejection ratio of the biomedical amplifiers. This is due to differences in the electrode impedances and to stray currents through the cables, resulting in transformation of the common mode voltage into a false differential signal. The residual PL interference (PLI) may corrupt the proper function of automatic ECG analysis, which presumes correct delineation of ECG wave boundaries. Sophisticated but conceptually traditional digital filters are known to suppress to different extent the components intrinsic to ECG around the PL frequency.

Low frequency artifacts and base line drift may be caused in chest lead ECG signals by coughing or breathing with large movements of the chest, or when a arm or limb is moved in limb lead ECG acquisition

A comprehensive suite of programs can be developed for processing intra cardiac signals. This includes:

1.Visual inspection program: Two channels of data can be visually examined using the graphical display unit. Any necessary software gain and DC offset correction can be applied independently to either channel prior to entry to the averaging program. The cursors of the graphics display may be used to automatically measure time intervals between epochs of interest,

2. Ensemble averaging: A technique is used to extract intra cardiac signals from overwhelming noise is that of signal averaging. Thus an ensemble averaging program must:

The second of these requirements, i.e. accurate alignment of samples to be averaged using a time reference point, is of crucial importance to the success of the averaging process. Moreover, as averaging must be time-referenced to different sections of the ECG complex for different applications, e.g. to the P-wave for intra-atrial signal enhancement and to the QRS complex for P-R segment signal enhancement, an accurate method of fiducial point detection is essential. The time reference point for averaging may be established using one of three methods which are available in the averaging program.

These are:(i) voltage threshold (ii) slope detection(iii) pattern matching.

Voltage-threshold triggering is straightforward to implement. The input data waveform is monitored in the program to establish the point at which the magnitude exceeds some preset threshold value. If, for example, the total excursion of the input is ± 5 V, a convenient threshold level is + 3 V. The waveform crosses the threshold for the first time in each complete cardiac cycle when the rising edge of the R wave occurs. The time reference point established by this method is thus on this rising edge. Slope-detection triggering is more complex to implement, and is hence more costly in computer time. It is particularly suitable for establishing a time reference point within the Qwave, should one be present, of the reference ECG. Initially a threshold reference point is found as described above. The program then calculates the slope of the reference ECG from the point at which the threshold is exceeded backwards in time towards the P-R segment and Q wave.

Storage of the complete waveform in the computer facilitates this process. The slope is calculated as the average difference between a number of adjacent samples around each sample point. An average difference of several samples is taken at each point to reduce any possible effect of high-frequency noise within the reference waveform. The method permits detection of a time reference point either:

Pattern-matched trigger-point detection allows a reference template to be compared with the input waveform. The point at which the two most closely match is then defined as the reference. The template may be established interactively by the system user. He is able to examine the input reference waveform on a cycle-by-cycle basis, and can select a representative waveform and the portion of it to be used for matching by moving delimiters on the display screen. When a suitable template has been established in this way it is compared on a sample-by-sample basis with a limited section of each succeeding input waveform. The comparisons achieved by calculation Ek = 2 \Sn - Tn\at each sample point where the summation is performed overall the samples in the section of input waveform chosen Ek is the error between the template and input at the sample point. Sn and Tn are the nth samples of the signal and template, respectively, to be included in the summation. The best match is achieved at the value of k for which Ek is a minimum. The maximum value of Ek which is allowable for a match to be accepted can be set interactively. Thus the 'closeness' of the fit between the template and the signal can be specified.

The methods of fiducial point detection described above all have particular areas of applicability. Threshold detection is fast and simple, and is useful in determining an approximate reference position as a preliminary to the use of one of the other methods. However, it is of limited applicability on its own because of the variability of the reference point detected due to baseline drift in the ECG. This may cause variations of several milliseconds in the reference point position.

Slope detection eliminates the problem of baseline drift, and is especially useful in selecting a reference point for averaging when events in the P-R segment which are locked to the QRS complex, e.g. the His bundle signal, are to be detected . A reference point at the start of the Q wave is most appropriate in these circumstances.

Pattern matching allows any portion of the reference signal to be chosen for trigger-point establishment. It is thus appropriate for averaging when events which are time-related to other sections than the QRS complex are to be investigated. Pattern matching has been successfully applied to a number of waveforms including the P wave and QRS complex.

When the averaging process is complete the result of it is automatically displayed. It can also be stored in a file for subsequent inspection and manipulation by other processes. Gain and DC offset correction can be applied, and the result can be processed using a linear phase digital filter as discussed below. A hard copy of the final result can be obtained.

3.Digital filter design program:

When an ensemble of several realizations of an event is not available synchronized averaging is not possible. We are then forced to consider temporal averaging for noise removal. As temporal statistics are computed using few samples of the signal along the time axis and temporal window is moved to obtain the output at various points of time and the filtering procedure is called moving average filter.

where

are the input and output of the filter respectively. N is the order of the filter and values are filter coefficients.

A simple moving average filter can be a 4 point Hanning filter for filtering the noise given by

Though the description of the filter in time domain, subsequent analysis will be performed in frequency domain using the z transform and the frequency response. It is seen in Hanning filter that components beyond 20% of the sampling frequency are reduced in amplitude more than 3dB, less than half of their levels in the input.

Increased smoothing by having more number of sample points for average, with longer window , at the expense of increased time delay can be achieved as shown in the figure.

The DC component or the baseline drift in ECG may be removed by high pass filter. The derivative operator in time domain removes the part of the input that are constant. The ideal

Operator in time domain is multiplication of the Fourier transform of original signal by jω in the frequency domain. The basic derivative operator is given by I order difference operator.

It follows readily that the second order derivative operator

has the frequency response with quadratic in gain for higher frequencies. The second order derivative operator may be used to obtain higher gain for higher frequencies.

The gain of the filter increases for higher frequencies up to the folding frequency, that is half the sampling frequency. Any high frequency noise present in the signal will be amplified significantly.

The noise amplification in the first order difference operator may be controlled by taking the average of two successive output values.

The transfer function above is also known as three point central difference operator which is a product of the transfer functions of the first order difference operator and a two point Moving Average filter.

The time domain filters have the drawback of less attenuation in the stopband and attenuation is non uniform. Filters may be designed in frequency domain to provide specific lowpass, highpass, bandpass or band reject characteristics. The frequency domain filters may be implemented by taking the Fourier transform of the input signal.

The simplest method to remove the periodic artifacts is to compute the Fourier transform of the signal, delete the undesired components from the spectrum and then compute the inverse Fourier transform. But this method will remove the noise components as well as the signal components at the frequencies of concern. The periodic interference may be removed by notch filters with zeros on the unit circle in the z domain with the specific frequencies to be rejected. If multiple zeros are introduced on unit circle in z domain then such a filter is called “Comb” filter.

The designs of a lowpass, highpass or bandpass digital FIR ,linear phase filter is important in ECG filtering. The design requires the following input parameters: (i) band-edge frequencies (ii) desired pass band and stop band ripple (iii) filter type (high-, low- or bandpass). The result of the design is a set of impulse response coefficients which are stored in a disc file for subsequent use as discussed below. These coefficients can be used to calculate the frequency response (phase and magnitude) and step response of the filter. The design is interactive in as much as the filter parameters (e.g. the number of impulse response coefficients) are calculated and displayed. The number of impulse response coefficients needed can be reduced by 'relaxing' other requirements such as the specified value of stop band attenuation. The frequency response of a Type 1.FIR filter is shown in the figure 2.

Fig2. Frequency Response of FIR.

Spectral analysis program:A spectral analysis program has been developed around an efficient real-data FFT algorithm. It may be used to calculate and display the power

spectrum of either: (i) the isoelectric portion of the T-P segment of an ECG or (ii) any arbitrary signal.

In summary for denoising ECG we may go for time domain filters or frequency domain filters. But as noise will be the integral part of the signal one may need to go either for adaptive filtering or signal averaging techniques to remove the random noise. Baseline drift can be removed from a high pass filter with 8 -14 Hz cutoff frequency, structured noise like power line interference can be removed from 50 Hz notch filter and or “Comb” filter.

Content actions

Footer

This work is licensed by veena hegde under a Creative Commons Attribution License (CC-BY 3.0), and is an Open Educational Resource.

Last edited by veena hegde on Nov 14, 2009 1:11 am US/Central.