WO2012052642A1

WO2012052642A1 - Method and device for predicting the risk and/or recurrence of a heart attack on the basis of at least one value of a parameter that can be extrapolated from an electrocardiogram wave

Info

Publication number: WO2012052642A1
Application number: PCT/FR2011/052162
Authority: WO
Inventors: Jean Marc Boucher; Salim Graja; Philippe Castellant; Jean Jacques Blanc
Original assignee: Institut Telecom - Telecom Bretagne; Centre Hospitalier Universitaire; Universite De Bretagne Occidentale
Priority date: 2010-09-20
Filing date: 2011-09-20
Publication date: 2012-04-26
Also published as: FR2964849A1; FR2964849B1

Abstract

The present invention relates to a method for predicting the risk and/or recurrence of a heart attack, including the steps of: selecting at least one electrocardiogram (ECG) wave; segmenting said wave; parameterizing said wave selected from the electrocardiogram into a plurality of parameters; selecting parameters and calculating the values thereof for two groups of people that are at risk and not at risk, respectively, wherein the first three steps are used on each electrocardiogram of said persons, and the values of one or more parameters representing one of the groups, respectively, are determined; selecting at least one of said values for a given patient; and using the first three steps on at least one wave of the electrocardiogram thereof, obtaining at least one selected value, referred to as the patient value, and comparing the patient value with all the values obtained for each person from the two groups so as to determine the group to which said patient belongs. The present invention also relates to a device for implementing such a method.

Description

"Method for predicting a cardiac risk and / or recurrence according to at least one value of an extrapolatable parameter of a wave of an electrocardiogram"

DOMAI TECHNICAL NODE AND STATE OF THE ART:

The present invention relates to the field of prediction of a cardiac risk and / or recurrence as a function of at least one extrapolatable parameter of a wave of an electrocardiogram (ECG), representative of the function of the heart.

More particularly, the present invention relates to a method for predicting a patient's cardiac risk by comparing the so-called patient value value of at least one of its directly or indirectly extrapolatable parameters of a wave of his electrocardiogram with the values corresponding to at least said parameter of a first group of non-cardiac risk persons and a second group of cardiac risk persons, the parameter or parameters being chosen because having substantially different values for the two groups, allowing determine the patient's affiliation to one of the two groups.

By directly extrapolable, it is understood respectively one or parameters directly obtained from the electrocardiogram, for example the duration or amplitude of the selected wave or other similar parameters and indirectly extrapolable, it is understood a parameter obtained after mathematical treatment of the selected wave.

In what follows, the present invention will be described in connection with a risk of atrial fibrillation. It should be borne in mind, however, that the present invention is not limited solely to this cardiac risk but includes all the cardiac risks which can be recognized during an electrocardiogram study, this with or without a mathematical treatment. at least one parameter of a selected wave of the electrocardiogram.

It is known that the contraction of the heart is controlled by electrical impulses that are born in the heart itself. Some cardiac muscle cells are said to be autorhythmic, that is, they are capable of spontaneously self-exciting and producing electrical specific frequency. These autorhythmic cells are located in the wall of the right atrium near the entrance of the superior vena cava in a region called the Sino-Auricular Node or sinus node and abbreviated as NSA. The excitement originally created by The cells of the NSA are then able to spread to all the cells.The transmission of the excitation in the cardiac fibers is carried out step by step.

Thus, at each instant, an electrical pulse is generated at this sino-atrial node and propagates to the two atria that will contract under the effect of this electric current. At the same time, the electric current goes to the atrioventricular node. At this node, the current is delayed to allow the contraction of the atria to be completed, then it is transmitted to both ventricles by the two branches of the Purkinje beam. It is then that the current propagates to the two ventricles which will contract in turn. Finally, after the contraction, the cells return to rest and repolarize.

Thanks to this electrical activity, we can control the proper functioning of the heart. For example, a malfunction of the N.S.A. can cause some problems with frequency and rhythm. Too slow (brachycardia) or too fast (tachycardia) may affect cardiac work.

The study of variations in the recording of the electrical activity of cardiac cells is called electrocardiography. The recordable graphic signal is the electrocardiogram (ECG). The electrocardiogram is the recording at reference points (the cardiac leads) by means of electrodes placed on the surface of the body, differences in electrical potentials caused by the propagation of the electrical impulse within the heart. During an observation of the electrocardiogram, it can be recognized mainly a P wave, a QRS complex and a T wave which constitute a heartbeat. The P wave corresponds to the contraction of the atria, the QRS complex corresponds to the contraction of the ventricles, the T wave corresponds to the relaxation of the ventricles. This is shown in Figure 1.

The electrocardiogram has become essential in cardiology. Nevertheless, it gives only limited information on the phenomena of propagation, insofar as it is a weighted average of the activation and is inhomogeneously disturbed by neighboring organs with different conduction properties.

In the case of atrial fibrillation, this fibrillation being chosen to illustrate the present invention, it corresponds to a complete disorganization of the cardiac electrical activity which induces anarchic, irregular and extremely rapid atrial activity. Transmission to the ventricles is also anarchic. This fibrillation is reflected on the electrocardiogram by the disappearance of organized P waves in favor of a continuous activity resembling a sort of irregular sinusoid. The frequency of depolarization is variable from one point to another within the atria, but in any case most often very high. Multiple depolarization fronts thus appear at the level of the atrioventricular node which fulfills its role of "filter" by randomly passing only a few fronts, the frequency of the ventricles becoming irregular, generally between 90 and 140 strokes. by mn depending on the state of permeability of the node. It is called complete arrhythmia by atrial fibrillation.

Other waves of the electrocardiogram may be affected by the presence of atrial fibrillation and therefore may also be considered. For example, the selected wave may be the P wave, the QRS complex, the T wave, the U wave, one or more QT, ST, RT, PT, PQ, PR, PS segments. Thus the word wave must be taken in a broad sense in the present application and concerns any part of the electrocardiogram that can be exploited and not only the P and T waves or the QRS complex.

Currently, the methods used for the detection of atrial fibrillation allow its detection only when it is established, that is to say essentially by the absence of the P wave. This requires the patient to carry a ECG (holter) device for a day or more in the hope that an episode of atrial fibrillation will occur during this time and the analysis algorithm will be able to detect it.

WO / 9841144 relates to an apparatus and method for determining a non-invasive atrial fibrillation prognosis comprising detecting the electrical activity of the heart with external electrodes, such as ECG leads. The system, the apparatus and the method of diagnosis constitute a non-invasive technique for classifying the intensity of atrial fibrillation according to measurement values of at least one predetermined parameter of fibrillation as observed in the frequency domain. The apparatus includes an AN converter, a peak frequency detector, a signal processor, a filter circuit and a memory. The method and apparatus utilize observed correlations between the measured parameter and success or failures with various treatment modalities to obtain useful atrial fibrillation treatment information, including prediction of conversion with anti-inflammatory drugs. -arythmics such as ibutilide and the estimation of a recurrence time after cardioversion.

In this document, there is no reference to the constitution of groups of people respectively healthy and at risk with a sufficiently large number of people to reliably determine whether a patient is at risk or not. The classification of the patient is rather carried out according to the criterion of being able to react to a given treatment or not. In addition, the decision-making process is not disclosed in this document.

The ECG Segmentation and P-Wave Feature Extraction: Application to Patients Prone to Atrial Fibrillation, IEEE, Oct. 25, 2001, pages 298-301 shows an analysis method based on the discriminant analysis method. The segmentation used in this document is based on a hidden Markov tree algorithm that is not optimal. This document also does not provide an update of the composition of groups of healthy or non-risk patients and a group of patients at cardiac risk to have a measure of confidence and credibility of decision-making.

The object of the present invention is to provide a method of signal processing and classification applied to the analysis of patients' electrocardiograms for assessing the risk of experiencing an episode of heart disease.

SUMMARY OF THE INVENTION

The present invention intends to overcome the disadvantages of the prior art by proposing a method for predicting a cardiac risk and / or recurrence comprising the following steps: selecting at least one wave of an electrocardiogram (ECG), which may be particularly representative of said cardiac risk,

segmentation of said wave after removal of baseline fluctuations, removal of background noise and determination of a reference element in the electrocardiogram,

parameterizing said selected wave of an electrocardiogram into several parameters,

- selection of the parameters with calculation of their values for two groups of people, the people of one of the groups being not disposed to a cardiac risk while the people of the other group being at risk, with the application of the first three stages on each electrocardiogram of the persons and determination of the values of one or the parameters respectively representative of one of the groups,

selecting at least one of said values,

for a given patient, applying the first three steps to at least one wave of his electrocardiogram and obtaining at least one selected value called the patient value,

comparing the patient value with the set of values obtained for each person of the two groups for the determination of the group belonging to said patient, said determination of the group of membership of said patient is carried out by a transductive SVM model, this model allowing an updating of the composition of the risk groups and not disposed of a cardiac risk and / or a refresh of at least one selected parameter.

In the context of the present invention, parametrization is understood to mean a parameter selection capable of receiving a constant value for a given case. Such parameterization can be done in several parameters or even in one.

According to advantageous features of the present invention:

the determination of the group of membership of said patient is done by tracing at least one nonlinear separation curve between the values obtained, on the one hand, by the persons of the non-risk group and, on the other hand, by people in the at-risk group.

the determination of the group belonging to said patient is carried out by a transductive SVM model, this model allowing an updating of the composition of risk groups and not at risk of cardiac risk and / or a refresh of at least one selected parameter.

- two separation curves are drawn, one with the patient value included in the risk group and the other with the patient value included in the non-risk group, the patient's group being the group presenting the value patient it contains furthest from the respective separation curve.

- the patient value is included in its home group for future membership determinations of at least one other patient.

- the parameters of the selected wave are discretized before being used by the SVM model.

the selected wave is the P wave, the QRS complex, the T wave, the U wave, one or more QT, ST, RT, PT, PQ, PR, PS segments.

the elimination of the fluctuations of the baseline is done using an algorithm based on a discrete wavelet transform (TOD).

- The baseline is estimated using a multiresolution decomposition of the electrocardiogram on a wavelet basis of Daubechies, Haar or Meyer.

during the segmentation, the electrocardiogram waves are modeled according to a Markov tree hidden from the wavelets, the Markov tree being advantageously combined with a Haar wavelet base.

during the segmentation step, the detection of the beginning and end of a selected wave is made according to the maximum likelihood criterion of the observations to the models of the different waves and segments of the electrocardiogram.

the statistical distribution of the observations of the P wave of the electrocardiogram is based on the statistical distribution of the observations of the T wave, the P wave being then assimilated to a low amplitude T wave.

- the statistical distribution of observations of the electrocardiogram isoelectric line is based on the statistical distribution of ST segment observations.

The invention also relates to an apparatus for predicting cardiac risk and / or recurrence, comprising an electrocardiogram and means electronic treatment of at least one wave of the electrocardiogram with storage means of the parameters of this wave for a group of people at risk and a group of people at no risk and for at least one patient to obtain a diagnosis concerning said patient concerning his belonging to one of the two groups, characterized in that the electronic processing means are in the form of electronic components or software included or received in said apparatus, said means allowing the implementation of a such a method.

BRIEF DESCRIPTION OF THE FIGURES:

The invention will be better understood from the description given below of an embodiment given by way of non-limiting example, with reference to the following appended figures:

FIG. 1 is a curve illustrating the various waves of an electrocardiogram during a heart beat,

FIG. 2 is a diagram illustrating the principle of eliminating fluctuations in the baseline of an electrocardiogram,

FIG. 3 is a diagram illustrating the principle of the filtering system of an electrocardiogram,

FIG. 4 is a diagram illustrating a model for the segmentation of the electrocardiogram,

Figure 5 illustrates a breakdown of the measured values of a parameter for non-risk and at-risk patients,

Figure 6 illustrates a distribution of the measured values of a parameter for non-risk patients and at-risk patients with the introduction of a patient value of a parameter of a patient whose membership is to be determined. the curve according to FIG. 5, this integration of the patient value occurring in the non-risk group,

Figure 7 illustrates a distribution of the measured values of a parameter for non-risk and at-risk patients with the introduction of a patient value of a parameter of a patient whose membership is to be determined. the curve according to FIG. 5, this integration of the patient value occurring in the risk group, Figure 8 illustrates a distribution of the measured values of a parameter for non-risk patients and at-risk patients with the introduction of another patient value of a parameter of a patient whose membership is to be determined. in the curve according to Figure 5.

PREFERRED EMBODIMENT OF THE INVENTION

The method for predicting cardiac risk and / or recurrence according to the present invention may be in the following succession of several steps:

selecting at least one wave of an electrocardiogram (ECG), which may be particularly representative of said cardiac risk,

selecting at least one of said values,

comparing the patient value with the set of values obtained for each person of the two groups for determining the group of membership of said patient.

Advantageously, the comparison of the patient value is done using a transducer system as will be explained later.

These various steps will now be detailed.

The first step consists in selecting at least one wave of the electrocardiogram that may be particularly representative of said heart risk we want to study. For example, for atrial fibrillation, it is known that this results in a modified P wave on the electrocardiogram of a patient. However, other waves of the electrocardiogram may be affected by the presence of atrial fibrillation and therefore may also be considered. For example, the selected wave may be the P wave, the QRS complex, the T wave, the U wave, one or more QT, ST, RT, PT, PQ, PR, PS segments. Thus the word wave must be taken in a broad sense in the present application and concerns any part of the electrocardiogram that can be exploited and not only the P and T waves or the QRS complex.

The second step consists in segmenting said wave after removing baseline fluctuations, eliminating background noise and determining a reference element in the electrocardiogram.

One embodiment of the elimination of baseline fluctuations is given by the scheme of Figure 2. For the elimination of baseline fluctuations an algorithm based on the discrete wavelet transform can be used ( TOD).

Advantageously, a wavelet transform is applied to the raw (original) ECG signal, the wavelet used may be under a Daubechie wavelet base, in particular of order 8. Next, an estimate of the baseline LB is given by the 10-scale ECG approximation. Finally, the baseline LB is subtracted from the ECG signal to obtain the new NECG electrocardiogram signal.

Alternatively or in combination, a wavelet of Haar or Meyer can be used.

This treatment improves the segmentation of the ECG for certain types of waves.

Once the baseline fluctuations have been eliminated, the signal obtained is then filtered to eliminate other types of noise. The filtering system used, given in FIG. 3, is advantageously composed of two stages: the first stage is used to estimate the useful ECG signal and the noise, and the second stage performs MAP (Maximum a posteriori) filtering in the wavelet domain. .

To obtain an estimate of the useful ECG signal, a W1 discrete wavelet transform is applied to the ECG signal, the wavelet used being able to to be that of Haar. Then, hard thresholding is applied to the wavelet coefficients. Finally, a WV ¹ inverse wavelet transform makes it possible to obtain the pilot signal. Thus, it will also be possible to estimate the ECG noise - Pilot signal.

The MAP filter maximizes the posterior probability of obtaining the useful values, given the noisy observations. In this case, the filtering is applied in the domain of the discrete wavelet transform invariant to the translation W2.

As a reference element in the electrocardiogram, it can be taken, for example, the peak R of the QRS complex.

As a result of these three substeps of eliminating baseline fluctuations, eliminating background noise and determining a reference element in the electrocardiogram, the actual segmentation of the electrocardiogram is performed. electrocardiogram.

As shown in Figure 4, for the segmentation of the ECG, it is possible to break down the electrocardiogram into 5 classes: a first class ISO-C1 - before the P wave, a second class P-C2- consisting of 1 P wave, a third class QRS-C3- consisting of the QRS complex, a fourth class ST-C4- between QRS complex and T wave and a fifth class T-C5- consisting of the T wave.

The basic principle is to characterize each wave (class) by its wavelet coefficients using the Markov tree model hidden in the wavelet domain, otherwise known as Wavelets-AMC thus extracting the maximum information of these coefficients. Thus, each class C, we associate a vector 9 _C including settings Wavelet-AMC model that characterize the best.

Once these vectors are calculated, the segmentation of the electrocardiogram will be based on the likelihood of the wavelet coefficients of each class.

One of the phases of segmentation is learning. At the end of this phase, each wave and segment of the electrocardiogram were modeled by a hidden Markov tree with different models:

Isoelectric line: Wavelet model-AMC 1 (M1)

Wave P: Wavelet Model-AMC 2 (M2)

QRS complex: Ondelette-AMC 3 (M3) model ST segment: Wavelet-AMC 4 (M4) model

Wave T: Wavelet Model-AMC 5 (M5)

To segment the ECG of a new patient, P-wave detection is performed not with the M1 and M2 models but with the M4 and M5 models. This gives more robustness for the detection of the beginning and end of the P wave. Thus, the statistical distribution of the observations of the P wave of the electrocardiogram (ECG) can be based on the statistical distribution of the wave observations. T, the P wave being then assimilated to a low amplitude T wave.

The detection of the QRS complex is carried out not with the models M3 and M4 but with the models M1 and M3. This gives more robustness for the detection of beginning and end of the QRS. Thus, the statistical distribution of observations of the electrocardiogram (ECG) isoelectric line can be based on the statistical distribution of ST segment observations. The phases followed to segment the ECG are as follows:

a learning phase which consists in calculating for each class C _l the parameters of the wavelet-AMC model which characterize it best. These parameters will be stored in a 9 _C vector. This phase is carried out thanks to the EM algorithm,

- a phase of segmentation proper. For an electrocardiogram to be segmented, it is calculated its wavelet transform and the likelihood of each coefficient to each class. Then, based on the maximum likelihood criterion, the coefficients of each scale are classified. This makes it possible to determine the beginning and the end of each wave of the electrocardiogram at the different scales,

- a phase of fusion between scales, in which it is exploited the dependence of the coefficients of wavelets between the different scales to refine the classification can also be used. For this, a context is defined for each wavelet coefficient that takes into account the state of the coefficients associated with it at the previous scale, using a merge algorithm between scales.

Advantageously, the Markov tree can be advantageously combined with a Haar wavelet base. The third step consists of the parametrisation of said selected waveform into several parameters. For example, in the case of a detection of atrial fibrillation for which the wave P can be the wave of the electrocardiogram selected, it is necessary to define parameters of this wave.

Several examples of possible parameters will now be given in the context of a detection of atrial fibrillation. These examples are non-limiting.

As a first parameter, for example, an increase in the duration of the P wave may be indicative of the presence of electrical conduction disorders in the atria. Indeed, the duration of the P wave in normal patients is less than 130 ms. This parameter is calculated directly from the segmentation data. To characterize the shape of the P wave two methods can be used.

In the first method, the best polynomial approximation of the P wave in the least squares sense is sought. The coefficients of the polynomial considered will represent the parameters retained.

In the second method, each wave P is likened to a probability density with which it is associated a distribution function. To characterize the shape of this wave, it is compared to a reference form. The comparison is done by calculating a function that links the two distribution functions.

As second parameters, it is possible to use so-called spectral parameters, these parameters notably relating to the frequency.

For example, it may be possible to generate a relative power spectrum with respect to the frequency by Fourier transform.

In terms of frequency, the useful information in the electrocardiogram is between 0 and 40 Hz, in particular most of the information of the P wave is located mainly between 0 and 25 Hz.

In order to explore the frequency content of the P wave, one can use the Morlet wavelet. The time-frequency plane is thus divided into several sub-bands where three types of energies relating to:

- the start energy, from the beginning of the P wave to the beginning + 60 ms,

- the end energy, going from the end the P-60 ms wave to the end of the P wave and

the energy of the wave P, from the beginning of the wave P to its end. As third parameters, it can be used wavelet-entropy parameters. For these, the entropy in a wavelet plane is calculated by the following formula:

with w _{i (fc} the _wavelength coefficients on the scale s at the position k calculated by a bi-orthogonal wavelet filter bank.

As a fourth parameter, a form-duration coupling parameter can be used. For the calculation of this parameter, we start by calculating an intermediate parameter R representing the ratio of the surface of the wave P by its duration squared. Then, the coupling parameter is obtained by the ratio of the average surface area to the average of the R parameter.

Thus, during the parametrization of a P wave, 122 parameters (duration, shape, energy, entropy, ...) have been determined, these parameters being directly extrapolatable from the electrocardiogram or indirectly extrapolable, which can be obtained after mathematical transformation. .

It should be kept in mind that not all of these parameters are informative. Indeed, some of them may be insignificant, correlated or irrelevant for the detection of a cardiac risk, for example atrial fibrillation. Thus, a step of selecting the relevant parameters is necessary. But before this step the parameters can be discretized, which can be done in particular by the CAIM discretization algorithm.

The number of parameters is then reduced by filtering approaches, which makes it possible to go from 122 parameters to 43. Then an enveloping selection approach using a genetic algorithm (AG) and the SVM classification algorithm was used. The genetic algorithm makes it possible to explore the space of the possible solutions and the classification algorithm is used to calculate the classification performance.

By way of nonlimiting indication, a parameter set consisting of twelve parameters comprising the duration, the coefficient of the polynomial ax ⁴ , the coefficient of the polynomial ax ² , the coefficient of the polynomial ax, the form by the method of distribution functions, the entropy by the Wigner Ville transform, the average energy of the end of the P wave between 0.9 to 1 .9 Hz, the average energy of the end of the wave P between 1 .9 to 7.8 Hz, the average energy of the end of the P wave between 7.8 to 15.9 Hz, the variance of the energy of beginning of the P wave, the variance of the energy of the end It should be noted that these parameters can be combined with each other, for example with different coefficients, in order to establish a parameter that would be more representative of the functioning of the core.

The fourth step of the method according to the present invention resides in the selection of the parameters and their values with their calculation for two groups of people, the persons of one of the groups being not disposed to a cardiac risk while the persons of the other group being at risk, with the application of the first three steps on each electrocardiogram of persons and determination of values of one or more parameters respectively representative of one of the groups

Taking the case of atrial fibrillation, in order to start this step, it is possible to form a database comprising patients. This database can be divided into two groups. The first group is the control group consisting of patients without known fibrillation or tachycardia and having normal atrial ultrasound. The second group is the study group consisting of patients with at least one documented atrial fibrillation episode.

This choice is based on the following hypothesis: the electrocardiograms of such patients have the same characteristics as those preceding the fibrillation.

Having such a database and the twelve selected parameters characterizing the P wave of each patient, a mathematical model is sought which makes it possible to discriminate between the risk group and the non-risk group. For this purpose, the machine vector carriers (SVM) were used. The SVM model was fixed by an experimental study: Gaussian kernel function, of standard deviation 160 and infinite bound C. This resulted in an average specificity (SP) of 88% and an average sensitivity (SE) of 82%.

It is advantageous to use a transductive SVM model in order to have a measure of confidence and credibility of decision-making. Such a model transductive allows the reactualization of the values of the parameters taking into account the new measurements made on the electrocardiogram.

Starting from a learning base formed by the observations of x _l , ..., x _l and their corresponding classes j _; e {- 1, 1}, to classify a new observation or possibly several observations, a new model is built to allow decision-making. Thus, for a transductive method there can be as many models as there are points to estimate. This type of model also makes it possible to calculate a measure of confidence and credibility of the decision taken.

A particularly preferred embodiment of the present invention for determining the patient's home group and advantageously its classification in this group for subsequent determination to be made for another patient will be described with reference to Figures 5 to 8.

As previously mentioned, this step of determining and classifying the patient is carried out after the elimination of the noise and the fluctuation of the baseline, the segmentation of the ECG by detecting for example the P, QRS and T waves, the characterization of a wave, for example the P wave, by extracting a set of relevant parameters for the detection of cardiac risk, for example a risk of atrial fibrillation and the discretization of this set of parameters.

So, we now have a database consisting of a group of healthy or non-risky patients and a group of patients at cardiac risk, such as atrial fibrillation. For example, each patient is characterized by a set of parameters of his P wave. For a new patient, one begins by applying the previously indicated steps to its ECG signal. Then, to classify it as "healthy" or "at risk of atrial fibrillation," as shown in Figure 5, where the values of the healthy group are symbolized by circles and where the values of the risk group are symbolized by squares, a transductive SVM model is used.

The patient value of the new patient is symbolized in Figures 5 to 8 by a triangle.

The principle of classification is as follows: It is assumed that the new introduced patient belongs to the group of healthy patients, as shown in Figure 6. Then we use the SVM algorithm to determine a curve of S1 separation between the two groups at no risk or at risk. Finally, a value P1 is calculated representing the degree of belonging of the patient to the healthy group. This value is a function of the position of the patient with respect to the separation curve S1, this curve being the most frequently non-linear.

The same operation is performed, but this time it is assumed that the new patient entered belongs to the group of patients at risk. This is shown in FIG. 7. Thus, the SVM algorithm is used to determine a separation curve S2 between the two non-risk and at-risk groups. Finally, we calculate a value P2 representing the degree of belonging of the new patient to the group at risk. This value is a function of the position of the new patient with respect to the separation curve.

The decision is then made: If the value P1 of FIG. 6 is greater than the value P2 of FIG. 7, then the patient is healthy and therefore included in the non-risk group. Otherwise the patient is at risk of heart failure.

It can be defined also a confidence value, called Trust, of the decision like:

Confidence = [1 - min (P1, P2)]. 100%.

min being the minimum value of P1 or P2

For example, in the case of Figures 6 and 7, the patient is healthy with a confidence of 70%.

We can also define a credibility value (trust confidence), called Credibility, with:

Credibility = max (P1, P2).

max being the maximum value of P1 or P2

This value serves to differentiate between the case of Figures 6 and 7 and the case of Figures 7 and 8. In both cases, the patient is classified as not at risk with a confidence of 70%. But we note that the patient belongs more to the healthy group in Figure 6 than in Figure 8. So if we calculate the Credibility, we find 75% in the first case and 41% in the second case.

Since the model is transductive, the patient value is included in its home group for future membership determinations of at least one other patient. Thus, the transductive SVM model allows an updating of the composition of the groups at risk and not disposed to a cardiac risk and / or a refresh of at least one selected value. The present invention also relates to an apparatus for predicting a cardiac risk and / or recurrence, comprising an electrocardiogram and electronic means for processing at least one wave of the electrocardiogram with means for memorizing the parameters of this wave for a group of people at risk and a group of non-risk persons as well as for a patient to obtain a diagnosis concerning said patient concerning his belonging to one of the two groups, characterized in that the electronic means of treatment are in the form of electronic components or software included or received in said apparatus, said means allowing the implementation of such a method.

As previously mentioned, the software can be integrated in the device or introduced into it, for example by being in a medium such as a CD or a USB key. It can also be in the form of an electronic circuit embedded in said apparatus.

Claims

1 - A method for predicting cardiac risk and / or recurrence comprising the following steps:

selecting at least one wave (P, T, QRS) of an electrocardiogram, which may be particularly representative of said cardiac risk,

segmentation of said wave (P, T, QRS) after elimination of baseline fluctuations, elimination of background noise and determination of a reference element in the electrocardiogram, parameterization of said selected wave of an electrocardiogram into several parameters,

selecting at least one of said values,

2. The method of claim 1, for which the determination of the group belonging to said patient is done by the tracing of at least one nonlinear separation curve (S1, S2) between the values obtained, on the one hand, by the persons in the non-risk group and, on the other hand, by the persons in the at-risk group.

3- Method according to the preceding claim, for which two separation curves (S1, S2) are plotted, one (S2) with the patient value included in the risk group and the other (S1) with the patient value included in the non-risk group, the patient's group being the group having the patient value it contains furthest from the respective separation curve (S1 or S2).

4- Method according to the preceding claim, wherein the patient value is included in its home group for future membership determinations of at least one other patient.

5. Method according to any one of the preceding claims, wherein the parameters of the selected wave are discretized before being used by the SVM model.

6. Process according to any one of the preceding claims, for which the selected wave is the P wave, the QRS complex, the T wave, the U wave, one or more QT, ST, RT, PT segments. PQ, PR, PS.

7. A method according to any one of the preceding claims, wherein the elimination of baseline fluctuations is done using an algorithm based on a discrete wavelet transform. 8- Method according to the preceding claim, wherein the baseline is estimated using a multiresolution decomposition of the electrocardiogram on a basis of wavelets Daubechies, Haar or Meyer.

9- Method according to any one of the preceding claims, for which, during the segmentation step, it is proceeded to the electrocardiogram waves are modeled according to a hidden Markov tree of wavelets, the Markov tree being advantageously combined with a Haar wavelet base.

10- Method according to the preceding claim, wherein, during the segmentation step, the detection of the beginning and end of a selected wave is made according to the maximum likelihood criterion of the observations to the models of the different waves and segments of the electrocardiogram.

1 1 - Method according to claim 9 or 10, for which the statistical distribution of the observations of the P wave of the electrocardiogram is based on the statistical distribution of the observations of the T wave, the P wave being then assimilated to a T wave of small amplitude.

12- Method according to claim 9 or 10, wherein the statistical distribution of the observations of the isoelectric line of the electrocardiogram is based on the statistical distribution of observations of the ST segment.

13- Apparatus for predicting a cardiac risk and / or recurrence, comprising an electrocardiogram and electronic means for processing at least one wave (P, T, QRS) of the electrocardiogram with means for memorizing the parameters of this wave (P, T, QRS) for a group of people at risk and a group of non-risk persons as well as for at least one patient in order to obtain a diagnosis concerning said patient concerning his belonging to one of the two groups, characterized in that the electronic processing means are in the form of electronic components or software included or received in said apparatus, said means allowing the implementation of the method according to any one of claims 1 to 12.