Patent Application
20070066907
The present invention relates to a method and device for analyzing a periodic or semiperiodic signal and more particularly but not exclusively to analyzing a signal having the form of a typical electrocardiograph signal, again more particularly but not exclusively to obtaining an improved signal to noise ratio from such a signal.
The electrocardiograph (ECG) signal describes the electrical activity of the cardiac muscle as it generates the various stages of the heart wave. Each cycle in the ECG signal may be subdivided into segments corresponding to stages of the heart wave, such as the P wave, the QRS complex, the T wave, the ST segment etc. Thus, the P wave of the ECG signal is due to depolarization of the atria, the QRS complex to depolarization of the ventricles, and the T wave to repolarization of the ventricles. Detection of an altered ECG signal is an important non-invasive tool in the diagnosis of cardiac abnormalities. Analysis of the ECG signal usually focuses on the ST segment due to its low-noise and its well-known correlation with cardiac abnormalities such as coronary artery disease (CAD). As physical stress is known to introduce features into the ECG signal indicative of CAD not present in signals obtained at rest it is common to obtain an ECG signal from a subject during a stress test comprising phases of rest, exercise and recovery from exercise.
In order to obtain sufficient information to provide useful diagnostic information from ECG measurements it is thus current practice to obtain data under the above stress test. It is desirable, moreover, to maximize the amount of information that can be obtained on line, for example when a patient arrives in intensive care.
The significant frequency range of the ECG signal is traditionally considered to be in the range of 0-100 Hz. However, valuable information is also known to exist in ECG components having frequencies above this range. For example, low amplitude components above 100 Hz at the end of the QRS complex, known as late potentials, correlate with delayed activation of the heart wave; decreased energy levels of the high frequency components in the mid-QRS portion of the signal are associated with ischemia. However, these high frequency components have very low amplitude and are usually masked by the noise at this frequency. Thus, a simple band-pass filter will not be enough for the detection of such phenomena, and more elaborate signal processing techniques should be used to enhance the signal-to-noise ratio.
U.S. Pat. No. 4,732,158, to Sadeh discloses a process for averaging ECG cycles over a time period to produce an average or representative cycle for the time period having decreased noise in the QRS segment. By the process of this patent, an ECG signal is obtained and band passed filtered. At each stage of the process, a new wave is averaged with the wave that was obtained in the previous stage if its maximum cross-correlation coefficient with a given reference wave exceeds a predetermined value. The averaging process is terminated when a predetermined number (e.g., about 100, or more) of successive ECG signal waves have been averaged. This averaging procedure, however, incorporates several significant drawbacks:
The improvement of the signal to noise ratio at the output of the process is very limited. More precisely, the noise level of the averaged signal at the output of the process can, at best, reach half the original noise level.
The averaging process tends to attenuate transient changes that may be of interest for the analysis of the signal.
The method is insensitive to the variability of the noise level in the signal.
A high level of alignment is a prerequisite for any effective averaging process. Thus, filtering the signal, before performing the alignment, as suggested in the above method, makes it practically inapplicable to signals, such as the HF component of ECG, where the signal to noise ratio in the frequencies of interest is very high.
Thus the methods known in the art are not effective in the analysis of signals with a varying level of relatively high noise, especially when transient changes in the signals are of importance. A typical example of such a signal is the HF ECG component obtained during the above-mentioned exercise test, which has been shown to be of importance in early detection of ischemia—it has been shown by Beker et al. (Proceedings on computers in Cardiology, IEEE Computer Society, 33-35,1992) and Beker et al. (Pacing and Clinical electrophysiology 12:2040, 1996) that decrease in the energy level of the HF component of the QRS interval during the course of the exercise test may be indicative of ischemia.
There is therefore a need for a method to enhance the signal-to-noise ratio in periodic or semi-periodic signals, such as ECG signals—a method that will eliminate or substantially reduce the disadvantages of the prior art methods.
Embodiments of the present invention provide a process, referred to herein as “adaptive averaging”, for effecting an improvement in the signal-to-noise ratio of periodic or semi-periodic signals. Preferred embodiments comprise the following features:
According to a first aspect of the present invention there is thus provided a device for reducing noise in signals having successive substantially repetitive portions, comprising:
an iterative averager operative to superimpose and average said substantially repetitive portions to produce a running average thereof,
and an iteration ender comprising a noise analyzer for determining a noise level in said running average and ending operation of said iterative averager when said noise level reaches a predetermined level.
Preferably, said iterative averager is operative to take said successive portions in successive iterative steps.
Preferably, the device further comprises an aligner for aligning at least some of said substantially repetitive portions one with another, wherein said signal comprises first frequency and second frequency components and said aligner comprises a first frequency correlated band pass filter and a second frequency correlated band pass filter to extract respective first and second frequency components, thereby to use said first frequency components to locate an alignment point in successive portions and to use said alignment point to align said second frequency components.
Preferably, said iteration ender is further operative to end said operation of said iterative averager when said repetitive portions are exhausted.
Preferably, said iteration ender is further operative to end said operation of said iterative averager when said running average reaches a preset maximum of included repetitive portions.
Preferably, the device further comprises a repetitive portion selector for selecting repetitive portions for passing to said iterative averager, the repetitive portion selector comprising
a reference portion store for storing a reference portion,
a cross correlator for computing a cross correlation between a current repetitive portion and said reference portion, and
a comparator for comparing a result of said cross-correlation with a predetermined threshold to produce a comparison output,
and wherein said selector is operable to pass said current repetitive portion to said iterative averager in accordance with said comparison output.
Preferably, the device further comprises a reference portion determination unit associated with said repetitive portion selector, operable to determine as a reference portion any one of a group comprising a first repetitive portion of a current length of said signal, a final result of a running average of a previous set of iterations and a prior determined typical wave.
Preferably, said reference portion determination unit is operable to dynamically change between members of said group over the course of a set of iterations.
Preferably, said reference portion determination unit further comprises a reference portion updater for dynamically updating said reference portion during the course of a set of iterations.
Preferably, the device further comprises a reference portion determiner, said reference portion determiner comprising,
a first store for storing a first set of repetitive portions from said signal,
a second store for storing a second set of repetitive portions from said signal,
a cross correlator for cross-correlating repetitive portions from said second set in turn with repetitive portions from said first set to produce a plurality of cross-correlation results for respective repetitive portions in said second set,
and a reference selector for selecting one of said repetitive portions in said second set as a reference portion in accordance with its respective cross-correlation results.
Preferably, said reference selector comprises a threshold level comparator for comparing each cross-correlation result with a threshold and which is operable to select as said reference portion a repetitive portion having a highest number of respective cross-correlation results exceeding said threshold.
Preferably, the device further comprises a summation unit for summing cross-correlation results of respective repetitive portions and which reference selector is operable to select as a reference portion a repetitive portion having the highest sum of respective cross-correlation results.
Preferably, said reference selector comprises:
a threshold level comparator for comparing each cross-correlation result with a threshold, and
a summation unit for summing cross-correlation results of respective repetitive portions exceeding said threshold,
and which reference selector is operable to select as a reference portion a repetitive portion having a highest sum of respective cross-correlation results.
Preferably, the device further comprises a signal extractor for extracting said repetitive portion.
Preferably, the device further comprises an RMS computation unit for calculation of the energy level of segments of wave obtained by averaging a series of said repetitive portions.
Preferably, the device further comprises an RMS value analysis unit for detecting a falloff in said RMS energy value over succeeding averages.
Preferably, the device further comprises a cross-correlation unit for computing the cross correlation coefficient of an average of a series of said repetitive portions and a reference wave.
Preferably, the device further comprises a cross-correlation value analysis unit for detecting a falloff in said cross-correlation value over succeeding averages.
Preferably, said aligner further comprises a cross-correlator for cross-correlating a current input with said running average at a plurality of successive alignments and for aligning said signal on the basis of an alignment giving a maximum cross-correlation.
Preferably, said aligner further comprises:
an interpolator for interpolating between said cross-correlations at said successive alignments to determine a higher accuracy sub-sample alignment, and a wave shifter for shifting said current input in accordance with said determined sub-sample alignment.
According to a second aspect of the present invention, there is provided a device for reducing noise in signals having successive substantially repetitive portions, comprising:
an aligner for aligning at least some of said substantially repetitive portions one with another,
a repetitive portion selector for selecting repetitive portions for passing to said iterative averager on the basis of a comparison with a reference portion, and
an iterative averager operative to superimpose and average said aligned, selected portions to produce a running average thereof.
Preferably, the device further comprises a repetitive portion selector for selecting repetitive portions for passing to said iterative averager, the repetitive portion selector comprising
a reference portion store for storing a reference portion,
a cross correlator for computing the cross correlation between a current repetitive portion and said reference portion, and
a comparator for comparing a result of said cross-correlation with a predetermined threshold to produce a comparison output,
and wherein said selector is operable to pass said current repetitive portion to said iterative averager in accordance with said comparison output.
Preferably, the device further comprises a reference portion determination unit associated with said repetitive portion selector, operable to determine as a reference portion any one of a group comprising a first repetitive portion of a current length of said signal, a final result of a running average of a previous set of iterations and a prior determined typical wave.
Preferably, said reference portion determination unit is operable to dynamically change between members of said group over a course of a set of iterations.
Preferably, said reference portion determination unit further comprises a reference portion updater for dynamically updating said reference portion during a course of a set of iterations.
Preferably, the device further comprises a reference portion determiner, said reference portion determiner comprising,
a first store for storing a first set of repetitive portions from said signal,
a second store for storing a second set of repetitive portions from said signal,
a cross-correlator for cross-correlating repetitive portions from said second set in turn with repetitive portions from said first set to produce a plurality of cross-correlation results for respective repetitive portions in said second set,
and a reference selector for selecting one of said repetitive portions in said second set as a reference portion in accordance with its respective cross-correlation results.
Preferably, said reference selector comprises a threshold level comparator for comparing each cross-correlation result with a threshold and which is operable to select as said reference portion a repetitive portion having a highest number of respective cross-correlation results exceeding said threshold.
Preferably, the device further comprises a summation unit for summing cross-correlation results of respective repetitive portions and which reference selector is operable to select as a reference portion a repetitive portion having a highest sum of respective cross-correlation results.
Preferably, said reference selector comprises:
a threshold level comparator for comparing each cross-correlation result with a threshold, and
a summation unit for summing cross-correlation results of respective repetitive portions exceeding said threshold,
and which reference selector is operable to select as a reference portion a repetitive portion having a highest sum of respective cross-correlation results.
According to a third aspect of the present invention there is provided a waveform frequency component alignment device for aligning first frequency components of waveforms having first frequency and second frequency components, the device comprising:
band pass filters for extracting respective first and second frequency components of said waveform,
a first frequency component aligner for determining a first frequency alignment point of a current waveform with another waveform based on respective first frequency components,
and a second frequency aligner for aligning said second frequency components of said respective waveforms based on said first frequency alignment point.
Preferably, said other waveform is a running average of preceding waveforms.
Preferably, said first frequency component aligner further comprises a cross-correlator for cross-correlating a current waveform with said other waveform at a plurality of successive alignments and for determining said first frequency alignment point on the basis of a one of said successive alignments giving a maximum cross-correlation.
Preferably, said first frequency component aligner further comprises an interpolator for interpolating between said cross-correlations at said successive alignments to determine a sub-sample accuracy alignment point between said successive alignments.
According to a fourth aspect of the present invention there is provided a device for analyzing high frequency components of ECG signals, comprising
a data extractor for extracting said high frequency components and
a data analyzer for determining, from a change over time in at least a part of said high frequency component, whether said ECG signal contains an indication of the presence of ischemia.
Preferably, said at least a part of said ECG signal is a QRS complex.
Preferably, said change over time is a fall in the energy level of succeeding QRS complexes.
Preferably, said change over time is a fall in a cross-correlation value of succeeding QRS complexes.
Preferably, said data extractor comprises a waveform averager for performing iterative averaging over successive ones of said high frequency components to obtain a reduced noise version of said components.
Preferably, the device further comprises a selector for passing to said waveform averager only those ones of said successive components which exceeds a threshold cross-correlation with a reference component.
According to a fifth aspect of the present invention there is provided a method for reducing noise in signals having successive substantially repetitive portions, comprising:
superimposing one by one weightwise in iterative steps weighted instances of at least some of said successive substantially repetitive portions,
forming a running average of said portions,
determining a noise level in said running average, and
ending said iterative steps when said noise level reaches a predetermined level, thereby to produce an average of said substantially repetitive portions having reduced noise.
The method preferably further comprises a step of aligning at least some of said substantially repetitive portions one with another,
Preferably, said signal comprises first frequency and second frequency components and said step of aligning comprises substeps of
extracting said respective first and second frequency components,
using said first frequency components to locate an alignment point in each of successive portions, and
using said alignment point to align said second frequency components of each of said successive portions.
Preferably, said step of ending said iterative steps further comprises ending when said repetitive portions are exhausted.
Preferably, said step of ending said iterative steps further comprises ending when said running average reaches a preset maximum of included repetitive portions.
The method preferably further comprises the step of selecting repetitive portions for passing to said iterative averager, the step of repetitive portion selecting comprising substeps of:
storing a reference portion,
Computing the cross correlation between a current repetitive portion and said reference portion,
comparing a result of said cross-correlation with a predetermined threshold to produce a comparison output, and
passing said current repetitive portion for iterative averaging in accordance with said comparison output.
The method preferably further comprises the step of determining as a reference portion any one of a group comprising a first repetitive portion of a current length of said signal, a final result of a running average of a previous set of iterations and a prior determined typical wave.
Preferably, said step of selecting comprises the further substep of dynamically changing between members of said group over the course of a set of iterations.
Preferably, said step of selecting further comprises dynamically updating said reference portion during the course of a set of iterations.
The method preferably further comprises a step of determining a reference portion by:
storing a first set of repetitive portions from said signal,
storing a second set of repetitive portions from said signal,
cross-correlating repetitive portions from said second set in turn with repetitive portions from said first set to produce a plurality of cross-correlation results for respective repetitive portions in said second set, and
selecting one of said repetitive portions in said second set as a reference portion in accordance with its respective cross-correlation results.
The method preferably further comprises:
comparing each cross-correlation result with a threshold, and
selecting as said reference portion a repetitive portion having a highest number of respective cross-correlation results exceeding said threshold.
Preferably, said step of determining a reference further comprises:
summing cross-correlation results of respective repetitive portions and
selecting as a reference portion a repetitive portion having the highest sum of respective cross-correlation results.
Preferably, said step of determining a reference portion further comprises:
comparing each cross-correlation result with a threshold,
summing cross-correlation results of respective repetitive portions exceeding said threshold, and
selecting as a reference portion a repetitive portion having a highest sum of respective cross-correlation results.
The method preferably further comprises the step of extracting QRS complexes from an ECG signal to provide said repetitive portion.
The method preferably further comprises extracting an RMS energy value from an average of a series of said repetitive portions.
The method preferably further comprises a step of analyzing said RMS energy to detect for the presence of a falloff in said RMS energy value over succeeding averages.
The method preferably further comprises extracting a cross-correlation value from an average of a series of said repetitive portions.
The method preferably further comprises the step of analyzing succeeding ones of said cross correlation value to detect the presence of a falloff in said cross-correlation value over succeeding averages.
Preferably, said alignment step further comprises cross-correlating a current input with said running average at a plurality of successive alignments, and
aligning said signal on the basis of an alignment giving a maximum cross-correlation.
The step of alignment preferably further comprises further comprising interpolating between said cross-correlations at successive alignments to determine a high accuracy alignment between said successive alignments.
According to a sixth aspect of the present invention there is provided a method for reducing noise in signals having successive substantially repetitive portions, comprising:
aligning at least some of said substantially repetitive portions one with another,
selecting repetitive portions for passing to said iterative averager on the basis of a comparison with a reference portion, and
superimposing and averaging said aligned, selected portions to produce a running average thereof.
The method preferably further comprises a step of selecting from said repetitive portions for passing to said iterative averager, the step of selecting comprising substeps of:
storing a reference portion,
carrying out a cross correlation between a current repetitive portion and said reference portion,
comparing a result of said cross-correlation with a predetermined threshold to produce a comparison output, and
passing said current repetitive portion to said iterative averager in accordance with said comparison output.
The method preferably further comprises the step of selecting as a reference portion any one of a group comprising a first repetitive portion of a current length of said signal, a final result of a running average of a previous set of iterations and a prior determined typical wave.
Preferably, said step of selecting a reference portion includes a substep of dynamically change between members of said group over the course of a set of iterations.
Preferably, said step of selecting a reference portion includes dynamically updating said reference portion during the course of a set of iterations.
The method preferably further comprises a step of determining a reference point, said step comprising,
storing a first set of repetitive portions from said signal,
storing a second set of repetitive portions from said signal,
cross-correlating repetitive portions from said second set in turn with repetitive portions from said first set to produce a plurality of cross-correlation results for respective repetitive portions in said second set, and
selecting one of said repetitive portions in said second set as a reference portion in accordance with its respective cross-correlation results.
The method preferably further comprises the further steps of:
comparing each cross-correlation result with a threshold, and
selecting as said reference portion a repetitive portion having a highest number of respective cross-correlation results exceeding said threshold.
The method preferably further comprises the further steps of
summing cross-correlation results of respective repetitive portions, and
selecting as a reference portion a repetitive portion having the highest sum of respective cross-correlation results.
The method preferably comprises the further steps of:
comparing each cross-correlation result with a threshold,
summing cross-correlation results of respective repetitive portions exceeding said threshold, and
selecting as a reference portion a repetitive portion having a highest sum of respective cross-correlation results.
According to a seventh aspect of the present invention there is provided a method of aligning waveforms having first and second frequency components, said second frequency components being more subject to noise than said first frequency components, the method comprising:
extracting respective first and second frequency components of said waveform,
determining an alignment point of a current waveform with another waveform based on respective first frequency components, and
aligning said second frequency components of said respective waveforms based on said alignment point.
Preferably, said other waveform is a running average of preceding waveforms.
The method preferably further comprises the further steps of:
cross-correlating a current waveform with said other waveform at a plurality of successive alignments, and
determining said alignment point on the basis of a one of said successive alignments giving a maximum cross-correlation.
The method preferably comprises the further step of interpolating between said cross-correlations at said successive alignments to obtain a sub-sample alignment point.
According to an eighth aspect of the present invention there is provided a method for analyzing high frequency components of ECG signals, comprising the steps of:
extracting said high frequency components and
determining, from a change over time in at least a part of said high frequency component, whether said ECG signal contains an indication of the presence of ischemia.
Preferably, said at least a part of said ECG signal is at least part of a QRS complex.
Preferably, said change over time is a fall in an RMS energy level of succeeding QRS complexes.
Preferably, said change over time is a fall in a cross-correlation value of succeeding QRS complexes.
The method preferably further comprises the further step of performing iterative averaging over successive ones of said high frequency components to obtain a reduced noise version of said components.
The method preferably further comprises a selection step of comparing successive waveforms with a reference component and selecting only those ones of said successive waveforms which exceeds a threshold cross-correlation level with said reference component for said step of iterative averaging.
According to a ninth aspect of the present invention there is provided a method of obtaining an indication of ischemia in a patient using an ECG signal therefrom, the method comprising:
extracting an ECG signal over a duration,
extracting from said ECG signal a series of at least partial QRS complexes over said duration,
extracting high frequency components of said QRS complexes,
analyzing said high frequency components over said duration for at least one of a predetermined quality, and
inferring from said predetermined quality an indication of ischemia.
Preferably, said predetermined quality is a falloff in a cross-correlation level with a reference component.
Preferably, said predetermined quality is a falloff in the energy level of said component.
Preferably, said step of extracting said high frequency components comprises carrying out iterative steps of averaging over preselected ones of successive components to reduce noise
Preferably, said step of extracting an ECG signal is carried out over a duration of a stress test comprising placing the patient in at least one of a group of phases comprising rest, stress, and recovery from stress.
Preferably, said step of extracting an ECG signal is carried out over a duration of an event being any one of a group comprising: acute myocardial ischemia, other forms of heart failure, coronary occlusion, and coronary angioplasty, said duration being any one of a group comprising before, during and after said event.
Preferably, said ECG signal is masked by another ECG signal.
According to a tenth aspect of the presetninvention there is provided a method of producing a noise reduced waveform from a series of substantially similar repeated waveforms having superimposed noise, the method comprising:
selecting waveforms having a highest cross-correlation with a preselected reference waveform, and
carrying out iterative averaging steps using said selected waveforms.
The method preferably further comprises the step of ending said iterative averaging when a signal to noise ratio of a result of said iterative averaging has a level below a predetermined threshold.
For a better understanding of the invention and to show how the same may be carried into effect, reference will now be made, purely by way of example, to the accompanying drawings.
With specific reference now to the drawings in detail, it is stressed that the particulars shown are by way of example and for purposes of illustrative discussion of the preferred embodiments of the present invention only, and are presented in the cause of providing what is believed to be the most useful and readily understood description of the principles and conceptual aspects of the invention. In this regard, no attempt is made to show structural details of the invention in more detail than is necessary for a fundamental understanding of the invention, the description taken with the drawings making apparent to those skilled in the art how the several forms of the invention may be embodied in practice. In the accompanying drawings:
Before explaining at least one embodiment of the invention in detail, it is to be understood that the invention is not limited in its application to the details of construction and the arrangement of the components set forth in the following description or illustrated in the drawings. The invention is applicable to other embodiments and of being practiced or carried out in various ways. Also, it is to be understood that the phraseology and terminology employed herein is for the purpose of description and should not be regarded as limiting.
In this specification, terminology is used in accordance with the following definitions:
To simplify the following description it is assumed that, unless explicitly stated otherwise, the adaptive averaging procedure takes as its input:
a digitized signal,
a set of waves extracted from that signal and aligned using any method known in the art, and
for each wave, pointers indicating its correct beginning and end within the signal. The skilled person will appreciate that, in order to arrange such inputs, a certain amount of pre-processing using known systems will be required.
The output of the procedure is an array of waves, obtained from the original set of waves by applying noise reduction to the set. Preferably, the output for each waveset is an array of “cleaned” waves. I.e. whenever the averaged wave resulting from noise reduction reaches a satisfactory noise level the algorithm does not stop but rather restarts the same procedure, either with the next wave in the same waveset or with a new waveset, as appropriate in the circumstances.
The noise reduction procedure of the preferred embodiments is based on averaging sets of succeeding waves in an iterative process. At each iteration of the adaptive averaging procedure a new wave is incorporated into a set of waves created during the previous stages and an average is calculated over the modified set. As will be described below, an alternative embodiment stores at each iteration only the averaged wave, and then obtains a new averaged wave using a weighted averaging wherein, at the nth iteration the averaged wave has weight (n-1) and the new wave has weight 1.
Unlike the averaging procedure suggested in the above mentioned prior art patent, in which each new wave is averaged with the wave produced in the preceding step, embodiments of the present invention are able, at least conceptually, to reduce the signal to noise ratio to any desired level.
A preferred feature of the present embodiments is that they permit a user to define a target signal to noise ratio. Such a target may be either constant, or may dynamically change according to a function, supplied by the user, taking as arguments such parameters as the local noise level, the local signal amplitude etc. The use of a selected target level is advantageous in that it provides assurance that as soon as the selected target signal to noise ratio has been reached no further processing takes place, thus minimizing attenuating effects caused by surplus noise reduction steps.
Thus, whenever the required signal to noise ratio level is attained, the current averaged wave is stored in an array of results, and the procedure is repeated with the next wave in the queue. If the target level is not reached within a predetermined number of iterations, the procedure stores the result after the predetermined number of iterations and moves on to the next wave or waveset in a new noise reduction procedure as described above.
In order to improve the averaging procedure and reduce attenuation in the signal produced by the averaging, waves with exceedingly high noise level are preferably identified so that they are not incorporated into the average. Therefore, in a preferred embodiment, a reference wave is used in order to rate the waves—for example using the cross correlation coefficient between the candidate input wave and the reference wave—such that only those waves whose rating is above a predetermined level are included in the average.
Reference is now made to
The next wave is judged against a series of criteria as will be described in more detail below and, if it is found to qualify, is added to a running set of waves taken from the current input waveset.
In a step 16, an average is calculated of the running set of waves, as updated by the next wave in step 14. In the case of the first qualifying wave in the set the wave itself constitutes the average. Generally, even though a waveform may vary over time, waves that are close to one another vary very little and thus averaging such closely related waves is an effective method of removing noise whilst retaining as much as possible of the essential structure of the waveform.
The step of averaging is repeated by taking succeeding waves in the waveset, and thus carrying out an iteration of steps 14, and 16 until one of the following events occur:
1) a reference low noise threshold is reached,
2) the input waveset is exhausted of all waves, or
3) a predetermined number of iterations has been reached.
The reference low noise threshold is preferably calculated based on the reference noise level defined in step 12 above.
The continuation or ending of the iterative process is illustrated by process step 18, and such control of the iterative process is preferred because it provides a means for ending the noise reduction process before waveform structure is adversely effected. To this end the reference noise threshold level and the number of iterations are preferably both set to optimize for later analysis of the waveform, which demands minimal noise but maximal surviving waveform structure.
At the end of the iteration series of the current waveform set, in step 20, the resulting average waves are stored for later analysis. The original waves that have already been processed from the current waveset set and the running waveset may now be deleted if no longer required.
If the current waveset is not exhausted, the running set is deleted, a next wave is chosen, and the process continues from step 14. The choice of the next wave may be the first wave in the temporally arranged waveset that has not been processed. Another way of choosing the next wave may be by choosing, for example the second wave in the running set. Such a choice preferably ensures that the new average is very close to the previous one, resulting in a much smoother change in the array of averaged waves. When the current waveset is exhausted, a new waveset is chosen for iterative analysis, and the process returns to step 10. The wavesets selected for analysis may be contiguous segments of the original signal. Alternatively, they may be segments taken at suitable intervals in accordance with an anticipated level of change in the signal.
In an alternative to the above, instead of storing the running waveset and separately calculating an average over all the waves therein at each iteration, it is also possible to calculate a new average at each iteration by taking the old average and weighting it according to the number of waves it is constructed of and then adding the new wave with a weight of 1.
Reference is now made to
In step 32, the wave currently being considered, the “next” wave, of step 14, is preferably cross-correlated with the reference wave and the cross-correlation is itself compared against a predetermined threshold in step 34. Provided the threshold is exceeded, then, in step 36 the result is added to the running set as described above. If not then step 38 is entered and the wave is discarded.
Reference is now made to
Referring back to
The procedure preferably continues until all of the waves in the second set have been processed, step 66, and then one of the waves of the stored waveset is selected as the reference wave. There are a number of possibilities for selecting a reference wave from the stored waveset that may be considered by the skilled person. One preferred possibility is to take the wave having the highest sum of cross-correlations and another preferred possibility is to select the wave having the highest k value as a reference value.
There is thus described, in FIGS. 1 to 4, a system of adaptive averaging which allows an input signal comprising a repetitive waveform to be segmented and for a clean version of the waveform to be produced for later analysis. The system is useful for any input signal having a repetitive portion and unwanted noise and is particularly useful when analysis is dependent on careful preservation of fine structural portions of the signal.
Reference is now made to
The signal is typically an analogue signal containing fine structure, as will be described in further detail below and is generally initially mixed with noise. As will be discussed in more detail below, the noise forms a higher proportion of the overall input at frequency bands of interest, such that other frequency parts of the signal are generally usable directly, but the signal at the desired frequency parts generally requires noise reduction. However, important information carried in parts of the signal in the frequency band of interest is liable to be distorted or lost by conventional noise reduction techniques.
A signal amplification and digitization unit, 72, processes the signal so that it can be supplied for digital signal processing. The QRS complex part of the signal is then detected by a QRS detection unit 74 and then an alignment unit 76 detects individual waves and marks corresponding portions of the waves for subsequent alignment.
The alignment unit is connected in series to noise reduction unit 78, HF filtering unit 80 and fine alignment unit 82 respectively, the functions of which will be discussed below. Finally a data extraction unit 84 extracts averaged waveforms for subsequent data analysis.
In order to understand the embodiment of
Enlarging on what was stated above concerning the need for noise reduction, the HF ECG signal is typically weaker and thus harder to detect in a meaningful manner than the standard ECC, and therefore, cannot be usefully dealt with before significantly improving the signal to noise ratio. However, even during a comparatively long test such as an exercise test (typically 10-20 min. long) the decrease in the HF ECG level may quite often be observed only in relatively short intervals (2-3 min.). Thus a noise reduction method that requires averaging large number of waves can significantly attenuate the effect for which it was employed, which means that the decrease in the HF ECG signal is barely noticeable in the noise reduced signal.
From the above preliminary remarks it is clear that the analysis of HF ECG preferably requires a noise reduction method that will not abolish transient changes, since the transient changes are of great diagnostic significance in that signal. The adaptive averaging method of the embodiments described above in respect of
The block diagram of
It should be noted, moreover, that individual steps of the iterations as outlined above do not generally deal with large quantities of raw data at a time. Therefore, it is envisaged that the skilled person will be able to provide additional memory and other data handling facilities of the appropriate type as deemed necessary in order to provide an on line implementation of the technique.
Acquisition 70, amplification and digitization 72, of the signal are preferably standard operations. Typically, the standard ECG signal is obtained by placing electrodes on an individual's skin. The signal is then amplified and digitized using any standard signal amplifier and an A/D converter with an appropriate sampling rate (i.e. greater than twice the highest frequency in the signal to be analyzed). As a rule most of the important information contained in the HF ECG signal is considered to be in the range of 100-500 Hz, calling for a sampling rate higher than 200 Hz, and in one embodiment 1000 Hz.
The QRS complex detector 74 is preferably implemented with a standard QRS detection algorithm. One preferred algorithm consists of locating all local maxima (or minima, depending on the signal's polarization) of the overall signal that have a value higher (or respectively lower) than a predetermined value corresponding to an expected level of the QRS amplitude. Another preferred algorithm consists of creating a template QRS complex (for example, by manually selecting one of the waves in the raw data), cross correlating it with the digitized signal and obtaining a cross correlation function. Local maxima of the cross-correlation function having a value exceeding a predetermined threshold may correspond to QRS complexes of the ECG signal (provided the threshold used is high enough).
Following the QRS complex detector 74 is a QRS alignment stage 76. In stage 76 a procedure is implemented for marking QRS complexes so that they may be aligned and then aligning them. In order to perform an alignment procedure it is preferable that the complexes are free of any DC components. Preferably, any such DC components are removed at least by the time that QRS component detection has been completed, either by rejecting the DC of the raw signal before performing the detection and extraction of the waves, or by performing QRS detection using the raw signal but extracting the waves themselves from a signal that has already been high-pass-filtered.
It is noted that from a practical point of view, the alignment unit 76 and the noise reduction unit 78 of the system of
In order to simplify the implementation of the alignment unit 76, it is preferable to assume that the waves extracted from the signal around each maximum point of the cross correlation function share substantially the same length. However, such an assumption is very much dependent on the circumstances of individual embodiments, in particular on different data extraction methods, and the procedure used by the alignment unit is preferably modified to suit any extraction method chosen by the user.
As stated above, the HF ECG signal is relatively weak compared with the expected or typical noise level in the signal. Consequently, the cross correlation of a template HF ECG wave with a wave of raw data may be expected to be dominated by noise, resulting in poor alignment results. However, the HF of the QRS complex is considered to correlate effectively with the standard low frequency wave, i.e. alignment of the low frequency signal will automatically give an alignment of the HF signal. As a result, by contrast with the method of U.S. Pat. No. 4,732,158, discussed above, alignment is preferably carried out using the low frequency wave. The low frequency wave generally has significantly better S/N, and thus HF alignment may effectively be carried out on signals with levels of noise that in fact completely mask the whole HF ECG
Before considering the remaining sections of
In the process of
In a stage 92, a first wave is selected. In a stage 94, a reference wave is selected, preferably as has already been described in respect of FIGS. 1 to 4. Then, in a stage 96, the selected wave is shifted against the reference wave and a series of cross-correlations are calculated for different relative positions of the two waves. In a stage 98 a position corresponding to the maximum cross-correlation is then determined as a first approximation of the correct alignment of the selected wave. A more accurate location is then preferably obtained in a stage 100 by interpolation, as will be described in more detail below. In a step 102, the selected complex is aligned with the reference wave in accordance with the interpolation result of step 100, and this is preferably followed by a step 104 of updating of the reference wave. If there are more waves (query step 106) then a new wave is loaded as the selected wave (step 108) and cross-correlation/alignment is repeated.
Particular emphasis in
The reference wave is preferably used to define the fiducial point with which alignment is preferably made. Each selected wave is aligned, using the fiducial point, with the reference wave, and by assuming transitivity, alignment of the whole set of waves is thus achieved. In order for transitivity to apply, the reference wave is preferably selected and updated in such a way as to ensure that alignment responds to changes of the wave with time (such a change may for example be indicative of development of a Bundle Branch Block during the exercise test) but, on the other hand does not respond to local or transient changes (such as Premature Ventricular Contractions, should these not have been taken care of earlier by previous processing of the signal).
As discussed in detail above, there are many methods for the choice of the reference wave that may be used in step 94. Among these methods the following are preferred:
For step 104, the update of the reference wave, a few approaches may be considered, for example:
The update itself may be carried out in numerous ways, including:
Step 100 of the process comprises interpolation of a sub-sample location for the fiducial point. Consider, for example, the analysis of the 150-300 Hz band of the QRS complex. An A/D converter with a sampling rate of 1 KHz assures that no information in the desired band is lost. However the accuracy of an A/D converter of 1 KHz in the time domain is of 1 ms, which does not permit sufficiently accurate alignment (at 250 Hz an error of 1 ms corresponds to a phase shift of 90°).
For each wave, the fiducial point, used as the basis for alignment, is the point giving maximum cross correlation of the selected wave with the reference wave (steps 96 and 98). Thus, obtaining a sub-sample accuracy (‘sub-sample accuracy’ meaning accuracy in excess of the sampling rate of the AID converter) for the fiducial point amounts to interpolating the exact location of the maximum of the cross correlation function CC(t) defined in step 96. This can be effectively achieved as the cross correlation function may be considered a very smooth function. Thus relatively incomplete data may be analytically interpolated by any method known in the art to give an exact location of the maximum beyond the sampling resolution. A simple method for the interpolation of the exact location of the maximum of the cross-correlation function involves approximating it to a polynomial of degree k (depending on the desired accuracy) using the values of CC(t) at, say, 2 k points about the maximum indicated by stage 98. In general the result of interpolation stage 100 is given as a fraction corresponding to an abstract point between two given points of the digitized signal. Thus, in order to correctly perform the alignment the whole wave must be shifted by that fraction (i.e. the correct value of the signal at intermediate points (n+fraction) has to be calculated). This can be easily achieved (e.g. by using an interpolation filter) as the sampling rate of the signal is assumed to be adequate.
The alignment algorithm of
Reference is now made to
In
Then, in decision stages 122, 124 and 126, the process ends the noise reduction stage if the noise level has dropped below the reference level, or if the number of waves averaged has reached a predetermined threshold, or there are no more waves left. Otherwise the process is repeated with the next wave in the set.
Particular emphasis in
The choice of the reference wave in the procedure of
In the procedure of
The reference noise level is needed to calculate the S/N ratio in the raw signal and to assess an expected amount of averaging needed to reach the target S/N. Alternatively, the reference noise level may be used to determine a target S/N that is deemed sufficient for the analysis of the signal. The target S/N is preferably a level that may be reached by averaging over a sufficiently small number of waves such that transient changes, which may be of importance in the subsequent analysis, will not be attenuated.
In either of the above cases the reference noise level is preferably regarded as an input for a function, preferably supplied by the user, that defines when the S/N level of the averaged signal has reached a satisfactory level (step 122) and that also defines a threshold on the maximal size of the set of averaged waves (step 124).
Whichever of the two possibilities is selected for use of the reference noise level as described above, the reference noise level is preferably calculated from a signal to noise analysis of the raw signal in the following way:
Each wave not eliminated after step 116, is now preferably incorporated into the running set, that is to say along with all previous waves of the same set which have not been eliminated. At each iteration, the average wave of the updated set is preferably calculated (step 120).
The noise level of the average wave is now calculated in step 122 as described above. Step 122 bears close attention as it ensures that as soon as the desired S/N ratio of the average wave (as determined by the user) has been reached no further averaging of the present set takes place, thereby reducing attenuating effects due to surplus noise reduction. As long as the required S/N has not been attained the algorithm keeps adding new waves to the averaged set, until a threshold number of waves (preferably determined according to a function supplied by the user) are reached. When such a threshold is reached (step 124) iteration on the current waveset is stopped, and a new iteration process starts with the next waveset (segment etc.) in the queue, as discussed with regard to the continuation of the process of
Returning now to
Filter parameters selected in accordance with either one of the two considerations above may be mutually exclusive. In practice an optimal filter is selected based on the individual traits of the phenomenon to be dealt with.
It should be stressed that different implementations of the algorithm may call for different filters and different frequency bands. These parameters should preferably be selected in accordance with the requirements of the data analysis in the given case being taken into account.
Although, for the purpose of the alignment procedure it has been assumed above that the HF signal of the QRS complex is highly correlated with the low frequency or standard QRS, local phenomena in the low frequency ECG, which have no trace in the HF ECG (e.g. transient “notches” that appear or disappear during exercise) may cause slight deviations in the alignment of the HF signal. Realignment of the filtered waves (say as in step 102 in
Preferably, an additional stage of noise reduction (not appearing in the diagram of
It is noted that, in order to implement the system of
In an alternative embodiment, the fine alignment procedure (using fine alignment unit 81 in
The above procedure avoids relying on the low frequency signal for performing alignment, thus bypassing the inaccuracies referred to above requiring a separate fine alignment procedure.
Preferably, each segment selected for examination contains a QRS complex and only one such complex. Thus, the restriction of the alignment procedure to a single wave at a time makes it possible to perform alignment over waves with relatively high S/N ratio. Furthermore, the cross-correlation function of the LF signals (See
However, as the S/N ratio in the frequency band of interest is typically much lower than in the standard low frequency part of the ECG signal there may be circumstances in which the S/N in the required frequency band will not permit alignment in that band even using the above-described procedure, whilst in the standard ECG, an acceptable alignment could yet be achieved. Thus in order to implement the above-described modified procedure some assumption regarding the S/N in the band of interest is preferably made to give a yardstick to determine whether or not the alignment according to the modified procedure is sufficiently accurate. Generally, in the case of the QRS complex, the part thereof being of interest and having the higher noise level is a high frequency component thereof.
As mentioned above, analysis of the HF QRS preferably serves as a diagnostic tool for early detection of ischemia. Therefore, various data extraction procedures (unit 84 of
In the following, for the sake of simplicity it will be assumed that data extraction procedures described below have access to the following data:
The following specific embodiments of the data extraction unit 84 are taken from experiments which explored the 150-250 Hz frequency band. The embodiments are not, of course, in any way limited to those specific frequencies.
As shown by Beker et al, a decrease in the total energy of the HF QRS signal during exercise gives a strong correlation with an ischemic condition of the heart. In what follows the RMS of an HF QRS signal is taken to represent the energy. A decrease in the total energy of the signal during exercise can be looked for in any of the following ways:
In a paper by Abboud (Progress in Cardiovascular Diseases Vol XXXV. No. 5 March/April 1993) it has been demonstrated that transient ischemia in patients undergoing percutaneous transluminal coronary angioplasty (PTCA) of a critical stenosis in the left anterior descending (LAD) coronary artery can be detected in the HF QRS wave. It has been shown that the inflation of the balloon (and the transient ischemia it induces) corresponds to a sensible decrease on a graph depicting the normalized cross-correlation coefficient of a (constant) template HF QRS and the real signal. Thus, any of the methods of sections 1-8 of the previous paragraph can be applied using the cross-correlation function in place of the RMS function or in addition thereto.
Thus, analysis of the HF ECG may serve as a very sensitive non-invasive diagnosis tool for the detection of ischemia. Furthermore, the methods for HF ECG analysis suggested herein may be designed to supply on-line results without using hardware additional to standard ECG equipment. Thus, the present embodiments may be readily applied in all situation where standard ECG is used to monitor the heart's condition. Thus, different specific embodiments of the said methods include, among others:
1) Screening exercise tests for early detection of ischemia. The methods suggested herein may serve to gauge the patient's HF ECG signal at rest (i.e. before exercise), the evolution of the signal under stress (i.e between rest and peak exercise) and during a recovery period (i.e. from peak exercise till the heart rate returns to its normal level) carry out a comparison and thus extract diagnostically significant data. The information thus obtained can be compared to the data of previous exercise tests undergone by the same patient, in order to follow the evolution of the cardiac condition over time.
2) Monitoring the evolution over time of coronary perfusion in CAD patients undergoing a drug treatment or during a rehabilitation period after cardiac surgical interventions. That is, using data of several consecutive HF ECG tests over a relatively short period of time (several days or weeks) it is possible to monitor the improvement in the coronary perfusion of the patient in order to assess the effectiveness of the treatment.
3) ER and surgery room on-line monitoring of patients during heart failure such as acute myocardial infarction or during PTCA. On-line analysis of the HF ECG in such situations may help the diagnosis in ERs (whether or not the patient suffers from ischemia) and to assess the immediate improvement of the coronary perfusion during a cardiac surgical intervention.
4) Integration of the embodiments into any (not necessarily cardiac) monitoring systems, including but not in any way limited to standard ECG monitoring.
In addition, the various embodiments of the present invention may be applied to the extraction of a low amplitude signal from input containing a high amplitude signal masking the low amplitude signal. That is, as long as noise is not correlated with the low amplitude signal that it masks the embodiments of the present invention are applicable, even if the noise is not evenly distributed in the spectrum domain. A typical embodiment of the method in such a case is fetal ECG monitoring, where the signal of interest is masked by the mother's much stronger signal. In that case the method is applied to detect the mother's ECG, and to create therefrom a set of dynamically chaining templates that may subsequently be subtracted from the original data, thus to leave only the fetal ECG.
In accordance with the above described embodiments there are thus provided embodiments of the present invention which provide in various aspects:
1) The use of the high frequency part of an ECG signal, more specifically the high frequency components of the QRS complex, in early detection of cardio-vascular infarction, tests involving the high frequency part of the signal being typically more sensitive than the standard ECG signal. In general, a decrease of the HF ECG of the QRS complex during stress test serves as an indication of such a condition.
2) Alignment of high frequency parts of ECG waves using the low frequency parts of the waves.
3) Reducing eccessive-noise related artifactes by by selecting waves having highest cross-correlation levels with a reference wave.
4) A noise reduction procedure involving successive averaging of aligned waves, the procedure being modified for minimal distortion of an overall waveform by selecting a predetermined SNR and terminating iteration when the predetermined SNR is reached.
5) Extraction of an RMS of a high frequency component of a QRS complex of an ECG signal to determine whether there is a successive decrease in the level of such an RMS. The presence of such a successive decrease may be used to indicate the probability of the presence of ischemia.
6) Extraction of a cross-correlation function of a high frequency component of a QRS complex of an ECG signal to determine whether there is a successive decrease in the level of such a function. The presence of such a successive decrease may be used to indicate the probability of the presence of ischemia.
Generally, the probability of the presence of any of the conditions referred hereinbefore in the human (or animal) body infers a signal given by a machine to draw the attention of medical personnel to the possibility of the presence of a condition.
It is appreciated that certain features of the invention, which are, for clarity, described in the context of separate embodiments, may also be provided in combination in a single embodiment. Conversely, various features of the invention which are, for brevity, described in the context of a single embodiment, may also be provided separately or in any suitable subcombination.
It will be appreciated by persons skilled in the art that the present invention is not limited to what has been particularly shown and described hereinabove. Rather the scope of the present invention is defined by the appended claims and includes both combinations and subcombinations of the various features described hereinabove as well as variations and modifications thereof which would occur to persons skilled in the art upon reading the foregoing description.
| Number | Date | Country | Kind |
|---|---|---|---|
| 133780 | Dec 1999 | IL | national |
This Application is a Divisional of U.S. patent application Ser. No. 10/168,673, filed on Jun. 25, 2002, which is a U.S. National Phase of PCT Patent Application PCT/IL00/00871, filed on Dec. 28, 2000, which claims priority of Israel Patent Application No. IL133780, filed Dec. 29, 1999.
| Number | Date | Country | |
|---|---|---|---|
| Parent | 10168673 | Jun 2002 | US |
| Child | 11601773 | Nov 2006 | US |
