Device for analysis of a signal, in particular a physiological signal such as an ECG signal

ES 2 300 554 T3

DESCRIPTION

Device for analyzing a signal, specifically, a physiological signal such as an ECG signal.

The invention relates generally to the analysis of physiological data, for example data obtained by a self-contained ambulatory appliance or an implanted active healthcare device.

Active medical devices are defined in directive 93/42 / CE of June 14, 1993 of the Council of European Communities and, in relation to active implantable medical devices, these are defined in Council directive 90/385 / CE June 20, 1990.

The invention will be described in particular for the case of the analysis of cardiac activity signals, in particular, of signals obtained by a recorder called "Holter", that is, an apparatus that allows to carry out, continuously and for a prolonged period of time , the recording of signals obtained by implanted electrodes (electrograms) or external electrodes (electrocardiogram), signals that we will call from now on only "ECG signals".

This application, to which the invention is applied in a particularly advantageous way, is not however limiting, and the invention can also be applied to the analysis of other physiological parameters such as electroencephalogram, respiratory rhythm, blood pressure, etc., or non-physiological such such as radar signals, ultrasonic signals, etc.

If we take the analysis of a Holter recording as a case, it requires a rather complex examination work. Indeed, the continuous recording of the ECG signals of a patient for 24 or 48 hours represents around 100,000 PQRST complexes that need to be analyzed, as well as their variability, in order to look for any pathological event such as rhythm disorder, cardiac anoxia , entrainment abnormality of a cardiac stimulator, etc. This analysis, which is carried out automatically by algorithms carried out by a computer or by an outpatient device, provides synthetic intermediate results from which the doctor can later establish a diagnosis.

These algorithms mainly need, due to the large volume of data that must be processed, relatively important calculation means, so their optimization in terms of efficiency in relation to the required calculation sources is an important factor in the field of analysis of physiological signals. , especially Holter signals.

Another difficulty lies in the error rate in signal analysis, which can lead to serious consequences, especially the risk of misdiagnosis. Indeed, in the specific case of an ECG recorded in an outpatient setting, the signal is not regular and presents numerous artifacts. More specifically, the ECG signal is constituted in most cases by the signal of cardiac origin, almost periodic (PQRST complexes), accompanied by signals generated by the muscles, by mechanical disturbances of the electrode-skin interface, or by electrical or even electromagnetic disturbances received by the cables connecting the electrodes to the recorder.

Classical algorithms have quite a bit of difficulty in eliminating all of these parasites and, although they typically allow an error rate of 0.1% to be achieved, this represents about 100 errors in a 24-hour record (100,000 PQRST complexes), which constitutes Even too high a level of error, since these errors can be concomitant with complexes that present important singularities from the diagnostic point of view.

Finally, also for an ECG signal, it is important to be able to observe the variability of the QRS complex, which can be very significant for the diagnosis; the analysis algorithm used must therefore be able to reveal and discriminate a certain number of microvariations.

The automatic analysis of an ECG signal involves three different stages, which are: 1 °) the previous filtering of the signal, in order to eliminate a certain number of parasites in the frequency domain and provide a better quality signal; 2 °) the decomposition and identification of the characteristic waves of the signal; and 3 °) a synthesis of the temporal evolution of the parameters that describe these different characteristic waves.

These results allow the doctor to establish his diagnosis and it is evident that they must be reliable as well as relevant to facilitate this diagnosis.

Reliability is based in part on the robustness and good adaptation of the decomposition and identification performed in the second stage.

More specifically, the ECG signal is presented as illustrated in Figure 1, which is a trace that represents the evolution, over time, of the electrical activity of the heart, with a succession of waves of positive or negative amplitude, from one side and the other of the line called "isoelectric" characteristic of cardiac rest.

During a normal heartbeat (the one illustrated in Figure 1), these positive or negative waves are identified as resulting from well-defined physiological processes, allowing each wave to be assigned a label

ES 2 300 554 Standardized T3, typically P, Q, R, S or T. Physiologically, the P wave is created by the depolarization of the atrium, the QRS waves by the depolarization of the ventricle, and the T wave by the repolarization of the ventricle. ventricle.

It is from the shape and temporal position of these various waves, as well as their variability, that the doctor will be able to recognize a pathology.

Various methods of decomposition and identification of the characteristic waves of the signal have been proposed.

Thus, frequency analysis allows the signal to be described in Fourier space. However, such a decomposition is not fully adapted to the analysis of an ECG signal since this signal is not strictly periodic. Its rich spectral content varies in time and, furthermore, the sinusoidal functions of the decomposition do not allow obtaining the phase of the signal, necessary to identify the waves; in effect, this frequency decomposition lacks temporal information.

To perform a time-frequency decomposition, it is proposed to use the one transformed into orthogonal wavelets, thus identifying the waves in the time-frequency content of the wavelets that model the signal. However, the lack of resolution in these types of methods quickly becomes a limitation for accurate analysis.

This difficulty can be alleviated by a decomposition into non-orthogonal wavelets (or base radial functions, or Gaussian functions, etc.): the ECG signal is decomposed into a sum of Gaussians that are either fixed or adaptable in size. This method has been used often but has a characteristic drawback because the waves to be modeled (that is, the P, Q, R, S and T waves) are not really Gaussian, so their modeling requires the use of a large number of parameters to be of sufficient quality and therefore requires considerable computing power to be carried out in a reasonable time. Furthermore, if the result is correct for the QRS wave, the P wave and the T wave , which are difficult to assimilate to Gaussian waves , are modeled quite poorly unless a large number of parameters are used to obtain a sufficient quantity that thus leads to complexification. excessive analysis algorithm.

Finally, for these various reasons, a simple linear decomposition is commonly used in which the amplitude fluctuations are replaced by line segments as soon as the derivative of the signal becomes significant. The result is a modeled signal made up of a series of line segments that are then very simple to process: for example, a monophasic wave is made up of a series of two segments with opposite directions, and a biphasic wave of three segments with opposite directions. .

This last technique is very effective but reaches its limits when applied to particular cases such as, for example, low-frequency waves with noises that lead to over-decomposition into multiple segments, difficult to analyze a posteriori.

One of the objectives of the invention is to propose a device for analyzing a signal, specifically a physiological signal previously obtained by an active healthcare device, which can, specifically in the case of an ECG signal, alleviate the drawbacks of the techniques used up to now.

The objective of the invention is to propose a device such that, while having an efficiency at least equal to that of the best techniques developed so far, it can work with restricted computing resources (processor power and memory resources), for example, those available with an office microcomputer or a traditional laptop, so that the doctor can have the results of the automatic analysis in a very short period of time, typically a few minutes at most after the transfer of the stored data by the Holter recorder.

For this, the invention proposes an analysis device of the type mentioned above, that is to say a device for analyzing a signal whose variations define a one-dimensional function, said signal having been previously filtered, sampled and numbered. Such a device comprises means for memorizing the numbered signal and analyzing the memorized signal, with extracting means capable of decomposing the signal into a plurality of N elementary waves, and classifying means capable of recognizing at least one characteristic parameter of each one of the N elementary waves and of assigning a standardized label chosen from a plurality of predetermined labels as a function of the said characteristic parameters thus recognized.

According to the invention, the extractor means are means capable of decomposing the signal into N parameterized hump functions in which each hump function is a continuous function defined in three successive intervals by, respectively, a first monotonous parameterized function, an affine function and a second monotonic parameterized function, said monotonic parameterized functions being one increasing and the other decreasing. According to various advantageous subsidiary characteristics:

- the dimension of said one-dimensional function is a time dimension;

- said signal is an electrocardiographic signal that forms a wave of the PQRST type;

ES 2 300 554 T3

- said affine function is a function of zero slope and each one of said parameterized functions is a semi-Gaussian, increasing and decreasing respectively; Said characteristic parameters can then be the five parameters constituted by the standard half-intervals of each of the two semi-Gaussians, the length of the interval for defining the affine part, the ordinate position of this interval and the amplitude at the peak of the semi-Gaussians;

- the extraction means comprise: selection means capable of searching, among a plurality of hump-types gathered in a library of predetermined hump-types and, for each of the N elementary waves, the respective hump-type most pertinent relative to the signal to be decomposed, and means of adaptation of the parameters of each of the N type humps thus determined by the suitable selector means to minimize the separation between the signal and the composition of the parameterized N type humps;

- in the latter case, advantageously, the selection means search for said most pertinent hump-type by orthogonalization and / or the adaptation means adapt said parameters by non-linear optimization under pressure;

- the classifying means operate through the implementation of hidden Markov chains.

In the particular case indicated above, in which the signal is an electrocardiographic signal that forms a wave of the PQRST type:

the means for analyzing the memorized signal further comprise subtraction means capable of subtracting from the memorized signal at least one of the N elementary waves determined by the extracting means and bearing a given label assigned by the classifying means;

there are five elementary waves, said predetermined labels can therefore be those of the P, Q, R, S and T waves of the electrocardiographic signal;

- the electrocardiographic signal is a signal obtained by PCA analysis and projection of the main components on a significant axis, specifically a dynamically recalculated maximum amplitude axis;

- the device comprises means for determining the variability over time of at least a specific factor of at least one of the N elementary waves determined by the extractor means, specifically the amplitude of the T wave, the amplitude of the P wave or the direction of a significant axis determined by PCA analysis;

- the device comprises means for determining a temporal correlation of a specific factor between at least two of the N elementary waves determined by the extractor means, specifically the time interval between the QRS wave and the T wave or the PR interval.

We will now describe in more detail an example of putting the invention into practice, with reference to the attached figures.

Figure 1, cited above, illustrates a PQRST complex of an ECG signal.

Figure 2 illustrates the constituent elements of a hump function.

Figure 3 illustrates the constituent elements and the corresponding parameters of a particular hump function, advantageously used for the implementation of the invention.

Figures 4a, 4b and 4c show, respectively, the result of the decomposition carried out according to the invention, the composition of the waves that give the modeled signal and the corresponding original signal.

The initial idea of ​​the invention consists of using a fast and reliable algorithm that can be put into practice with limited computing resources, carrying out the modeling of the physiological signal for decomposition into humps, the "hump" being a mathematical notion already known by itself but that had not been proposed, until now, in the framework of the analysis of a physiological signal.

A hump is, as illustrated in Figure 2, a parameterized monotonic function defined in three successive intervals by a first parameterized increasing function G1, an affine function D, and a second parameterized decreasing function G2.

According to the invention, the objective of the analysis device is to compose a set of humps so that the graph of the sum of these humps is as similar as possible to a known one-dimensional signal.

The optimization of the parameters of each hump is carried out by any appropriate mathematical method that allows obtaining a graph as close as possible to the original signal to be modeled.

Figure 4 illustrates like this, in the particular case of a beat of an ECG signal:

ES 2 300 554 T3

1 °) the set of humps, five in this case, elaborated in the way that will be described below,

2 °) the graph resulting from the composition of these five humps, and

3 °) the original signal which, as can be seen, is very close to the modeling obtained in (b).

Advantageously, we use a particular type of hump, illustrated in figure 3, derived from the general definition given above with reference to figure 2.

This particular type of hump is defined by a reduced number of parameters (five in this case) and, in practice, the modeling obtained turns out to be extremely reliable and close to the original signal in the case of the analysis of an ECG signal, despite reduced number of hump definition parameters.

To do this, we use a horizontal segment (that is, with zero slope) as the affine part D of the hump and, in order to be able to execute the stage of adaptation of the parameters by classical multidimensional optimization algorithms, the monotonic functions G1 and G2 are here semigaussian of amplitude equal to A.

Under these conditions, the hump is a continuous function defined by five parameters and differentiable with respect to each of these parameters, which are:

- μ: temporal position, for example, the ordinate position of the middle of the segment D;

- σ1: typical half-interval of the first Gaussian G1;

- σΙ .: length of segment D;

- σ2: standard half-interval of the second Gaussian G2;

^Y

- A: amplitude of the function.

The decomposition of the temporal signal into humps is carried out iteratively in two stages:

1 °) Selection of the most pertinent hump by a selection algorithm applied to a set of humps previously constructed and preserved in a library of hump-types in the computer memory; This selection step can advantageously use an orthogonalization method known per se (see for example MJ Korenberg, SA Billings, YP Liu et al., Orthogonal parameter estimation for non-linear stochastic systems, International Journal of Control, 48, 193-210, 1988).

2 °) Adaptation of the parameters of the hump that has been selected in the previous stage, that is to say, search of the five parameters μ, σ1, σ,., Σ2 and A indicated above, so that the particular hump finally obtained approaches as much as possible to the original signal to be modeled; This adaptation step can advantageously use a method of non-linear optimization under pressure known per se (see for example M Minoux, Programmation Mathématique (1983), Dunod, 1983).

Thus, the modeling in N humps of a signal will be done in N times two stages. The maximum number N of humps is either defined in advance, or dynamically adapted depending on the desired modeling precision.

In the common case of an ECG signal, a beat modeled in five humps turns out to be satisfactory in practice: the five hump limit corresponds to the simple case in which one hump represents a characteristic P, Q, R, S or T wave of the ECG recording, as illustrated in figure 4a (figure 4b represents the modeled wave obtained by composition of the five humps of figure 4a and figure 4c the original signal analyzed by the device of the invention).

The hump-modeled signal is then analyzed to find the characteristic waves of cardiac activity.

Each hump is then assigned a label (P, Q, R, S, T or other) according to its shape and its location with respect to the other humps. For this labeling, we can advantageously use a method based on hidden Markov chains (CMO or HMM, Hidden Markov Models), a method known per se and described, for example, by LR Rabiner, A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition, Proceedings of the IEEE, 77 (2), 257-286 (1989).

This labeling of the humps makes it possible to recognize the different components of a typical PQRST complex and thus easily detect atypical waves, which are precisely those that are of the greatest interest for the diagnosis of rhythm disorders.

ES 2 300 554 T3

Having identified each wave of the ECG, it is possible to measure each characteristic parameter of a given wave and above all - to study it dynamically. The curves obtained can be analyzed one by one or in relation to others since the modeling provides stabilized shapes that allow efficient correlations.

In this way, we can analyze the variability of the T wave in a pertinent way by measuring its amplitude variations and / or its time lag with respect to the QRS wave. We can also analyze the parameters of the hump represented by the T wave, evaluate the variability of the PR interval, of the amplitude of the P wave, etc.

Wave recognition also makes it possible to subtract a known wave from a signal; This operation is very useful for unmasking a small amplitude wave (for example, the P wave) that appears synchronously to a large amplitude wave (Q, R, S, or T).

This automatic wave subtraction in the signal allows underlying signals to appear, for example very early P waves, which appear just after the preceding QRS and which were very often hidden in the analysis techniques used up to now.

Labeling the P waves provides results that will later allow the physician to greatly improve the diagnosis.

The example of implementation of the invention that we have just described can be open to many variations.

Thus, rather than directly modeling the signal obtained in each of the ECG pathways after filtering, it is advantageous to previously carry out, dynamically (that is, for each beat), an analysis in principal components (ACP or PCA, Principal Component Analysis) .

This technique, known in itself (see for example IT Jolliffe, Principal Component Analysis, Springer, 1986) consists in using the signals obtained simultaneously in the X, Y and Z tracks and recorded in several tracks, searching in a three-dimensional space for an axis of maximum amplitude (PCA1 axis) and expressing the temporal variation of the beat by its projection on this main axis, the position of which is recalculated at each beat.

This pretreatment makes it possible, in particular, to model the ECG on only one track containing the maximum amount of information, rather than on all the original tracks.

The analysis of the main components of the PQRST complex (or the QRS complex and / or the P wave) also makes it possible to permanently obtain the projection plane of the cardiac electrical signals, a plane that is mainly linked to the position of the heart in space. An analysis of the movement of this plane or, more simply, of the axis of the main component (axis PCA1), allows to obtain an image of the movement of the heart that can be used for various purposes such as:

- the detection of respiration (which has the property of moving the heart at each cycle) and the discrimination between thoracic or abdominal respiration; Indeed, the displacement of the heart, which intervenes in different axes or curves, translates into a modulation of one or more angles of the main components that, analyzed with the help of appropriate mathematical tools, allows the extraction of one or more respiration curves , abdominal and thoracic for example;

- the detection of the position of the body, the heaviness caused by the anatomical changes that are transmitted to the heart and the ECG signal;

- compensation for the deformations of the ECG signal introduced by changes in position that displace the heart: knowing the electrical position, a dynamic matrix calculation applied to the XYZ leads or to the standard leads makes it possible to compensate for the influence of these changes and recreate a signal stabilized. This stability is extremely useful for various analyzes, in particular, comparisons between two recorded ECGs;

- the creation of an original ECG by projection of the spatial signal according to the three components X, Y and Z. In relation to an XYZ ECG, this original ECG has two advantages, namely: (i) a very high stability, since it is not subject to variations in the position of the heart and (ii) a signal level that is maximum in the first path corresponding to the projection on the main axis. It can be completed by information on the angles of the projection axes that would represent the positional changes of the heart, that is, the effect of changes in body position and breathing.

These technical differences resulting from the PCA analysis are known per se, but the implementation of the invention makes it possible to improve its efficiency to a great extent, thanks in particular to the labeling of the humps. Hump ​​labeling makes it possible to improve the efficiency of this PCA-type analysis by performing it if necessary on a particular wave.

The invention is specifically applicable to external ambulatory recorders such as the SYNEFLASH and SYNEVIEW models from ELA Médical, Montrouge, France.

ES 2 300 554 T3

It is equally applicable to implanted devices such as implanted devices with microprocessors, also marketed by ELA Médical, Montrouge, France, to which it is possible to transmit by telemetry programs that will be kept in memory and will be executed to put into practice the functions of the invention described. previously.

The adaptation of these apparatuses to the practice of the invention is within the scope of the person skilled in these techniques and will not be described in more detail.