WO2001080042A1

WO2001080042A1 - Method and device for cardiac atrial fibrillation by wavelet process

Info

Publication number: WO2001080042A1
Application number: PCT/FR2001/001170
Authority: WO
Inventors: Anestis Antoniadis; Jean-Claude Barthelemy; Frédéric Costes; David Duverney; Jean-Michel Gaspoz; Jean-René Lacour; Vincent Pichot; Frédéric Roche
Original assignee: Novacor
Priority date: 2000-04-14
Filing date: 2001-04-17
Publication date: 2001-10-25
Also published as: AU5486001A; FR2807851A1

Abstract

The invention concerns a method for processing an electrocardiogram RR interval signal, which consists in: detecting zones with high variability of cardiac frequency in the RR signal; classifying said zones into sinus rate zones (RS) and atrial fibrillation zones (FA).

Description

Method and device for detecting cardiac atrial fibrillation by the wavelet method.

TECHNICAL FIELD AND PRIOR ART The invention relates to the field of analysis and interpretation of variations in rhythm occurring in a signal or a physical phenomenon which is initially regular, then which becomes irregular, and which can then become regular again.

It applies in particular to the analysis of variations in heart rate. It relates in particular to a method and a device for detecting a heart rhythm disorder called atrial fibrillation (AF).

Atrial fibrillation (AF) is the most common rhythm disturbance encountered in cardio therapy.

In an electrocardiogram, and as illustrated in FIG. 1, the interval of time separating two maxima 2, 4 of QRS complexes from an electrocardiogram is called RR interval.

Cardiac supraventricular arrhythmias (AF) can lead to severe functional problems and be associated with high cardiovascular mortality. The number of arrhythmias increases with age. The Cardiovascular Health Study Collaboration Research Group showed, in an article by B.M. Psaty et al. entitled "Incidence of and risks factor for atrial fibrillation in older adults" published in Circulation,. 199 ', vol. 96, pages 2455-2461, at least one major electrocardiogram (ECG) abnormality in 29% of a group of 5150 subjects aged 65 years or more. In this group, the incidence of atrial fibrillation (AF) increased, between the groups aged 65-74 and 75-84, from 1.76 per 100 people per year to 4.27, respectively, for men, and from 1.01 to 2.16, respectively, for women.

Other authors (S. Rajala et al., "Prevalence of ECG findings in very old people", Eur Heart J, 1984, 5, 168-174) have described an AF prevalence of around 83% in a population aged 85 or over and have also shown that the associated clinical risk increases with age.

In the United States, one in four ischemic strokes, which means a total of 75,000 accidents per year, is associated with AF, as shown in the article by AJ. Camm et al. "Epidemiology and mechanism of atrial fibrillation and atrial flutter", aw. J. Cardiol. 1996, 78, Pages 3-11. The presence of AF is a strong predictor of the incidence and recurrence of accidents, as well as the increase in mortality after an accident, both in the short and long term. Compared to the groups studied, the absolute risk of accident attributed to AF varied from 3% to 67% with a relative risk ranging from 1.0 to 6.9% (KH Flegel et al., Risk of stroke in non rheumatic atrial fibrillation, Lancet, 1987, 1, pages 526-529). The absolute risk of clinically silent cerebral infarction was in the same proportions, while the relative risk was estimated at 2.5 (see on this subject: H. Yamanouchi et al. "Embolism brain infarction in nonrheumatic atrial fibrillation: a clinico- pathologic study in the elderly, Neurology, 1997, 48, pages 1593-1597).

It is currently believed that the type of AF, permanent or paroxysmal, does not modify the risk of occurrence of an accident.

All these data make the diagnosis of AF a major health priority (D.P. Zipes: "An overview of arhythmies and antiarrhythamic approaches", J. Cardiovasc Electro Physiol, 1999, 10, pages 267-271).

Several approaches have been considered. Detection of AF was achieved using intracardiac recordings at the atrial level, as well as at the level of the conduction system; however, these techniques targeted the initiation of implantable defibrillators. Other trials have been attempted on average ear activity. Esophageal signals could give the best detection power, but are difficult to use on a standard basis. Ambulatory ECG recordings have been shown to be the most suitable for detecting such arrhythmias, in terms of ease and their ability to detect frequent paroxysmal forms. With the surface ECG, of the type obtained by ambulatory recordings ("Holter"), the ideal situation would be to be able to recognize the P waves (identified in Figure 1), since they are the key to diagnostics and classification of atrial arrhythmias; however, the weak signal obtained in most cases makes this option difficult. In this context, several attempts have been made to develop effective methods of recognition of AF in Holter recordings. The simple Lorentz graph makes it possible to recognize specific forms of AF. Other methods were applied to improve the reading of the ECG, such as that of Murgatroyd et al. (FD Murgatroyd FD et al.: "Identification of atrial fibrillation épisodes in ambulatory electrocardiographic recordings: validation of a method for obtaining labeled RR interval files", Pace, 1995; 18: 1315-1320). Neural networks have also been applied to the ECG, analyzing RR intervals, as well as baseline fluctuations, as described in the articles by JF Yang et al. : "Artificial neural network for the diagnosis of atrial fibrillation", med & biol eng & comput 1994; 32: 615-619; "Deterministic logic versus software-based artificial neural networks in the diagnosis of atrial fibrillation", J. Electrocardiol. 1993; 26 (suppl): 90-94; or in the article by Cubanski D. et al. : "A neural network system for detection of atrial fibrillation in ambulatory electrocardiograms", J. Cardiovasc. Electrophysiol. 1994; 5: 602-608. However, in Holter systems, or event recorders commonly used over times of the day, the identification of AF should always be improved by increasing both the sensitivity and the specificity of detection algorithms. Indeed, the sensitivities mentioned in the articles of TF Yang et al. and D. Cubanski et al. cited above range from 82.4% to 96.6% and the specifics from 92 to 92.3%.

STATEMENT OF THE INVENTION

The subject of the invention is a method for processing an RR interval signal from an electrocardiogram, this signal being prerecorded on a data storage medium or in storage means, comprising:

- detection of areas of high variability in heart rate in the RR interval signal,

- classification of these zones into zones of sinus rhythm (RS) and atrial fibrillation (AF). The detection of zones of high variability in heart rate can be carried out by projection or transformation of the RR signal in a base or a family of discrete wavelets, and calculation of the coefficients d. (K) giving the projection of the RR signal on each discreet wavelet function.

For example, we perform a transformation into dyadic wavelets. In order to select the coefficients greater than a minimum value, a step of thresholding the coefficients d _j (k) for several levels j of decomposition is carried out.

In addition, a step of filtering the thresholded coefficients can be provided to obtain homogeneous areas of high energy. Possibly, the variations of the RR signal allow a recalibration of the beginning and the end of the FA episodes. The zones of high heart rate variability can be classified into RS and FA zones by calculating the fractal exponent of each of these zones and comparison with a threshold value Ho.

Preferably 0.5 <H ₀ ≤ 0.9 or 0.7 <Ho ≤ 0.85. It is in these ranges that the sensitivity and specificity are the best.

The invention also relates to a device for analyzing digital signals of RR intervals of electrocardiograms, comprising:

- means for, or specially programmed for, detecting zones of high variability in heart rate, - means for, or specially programmed for, classifying these zones into sinus rhythm zones and zones of atrial fibrillation.

The means for detecting zones of high variability in heart rate include means for projecting or transforming the RR signal into a base or a family of discrete wavelets, and calculating coefficients d. (K) giving the projection of the RR signal on each discrete wavelet function.

According to another aspect, the means for classifying the zones of high variability into zones of sinus rhythm and zones of atrial fibrillation comprise means for calculating the fractal exponent of each of these zones and for comparing it with a threshold value Ho L ' The invention also relates to a system for determining the heart rate, comprising one or more recorders for recording electrocardiograms of patients, and a device according to the invention and as described above.

BRIEF DESCRIPTION OF THE FIGURES

The characteristics and advantages of the invention will appear better in the light of the description which follows. This description relates to the exemplary embodiments, given by way of explanation and without limitation, with reference to the appended drawings in which:

FIG. 1 represents part of an electrocardiogram, with two QRS signals,

FIG. 2 represents a Holter recording,

FIGS. 3 and 4 represent steps of a method according to the invention, FIG. 5 represents steps for obtaining an RR signal from an ECG signal, FIG. 6 represents stages of detection of the zones of high variability of an RR signal,

- Figure 7 represents wavelet decomposition levels of RR intervals of an atrial fibrillation episode, - Figure 8 represents the power spectral density of a signal of the type

RS

FIG. 9 represents the power spectral density of an FA type signal,

FIG. 10 represents a patient fitted with a "Holter" type measurement system,

FIG. 11 represents a "Holter" type recorder,

FIGS. 12 and 13 represent means for analyzing RR signals according to the invention,

DETAILED DESCRIPTION OF EMBODIMENTS OF THE INVENTION

The invention will first be explained in conjunction with the flow charts of Figures 3 to 6.

Beforehand, it is assumed that an RR signal has been obtained (steps 3-1 and 4-1 of FIGS. 3 and 4), for example from an electrocardiogram (ECG) supplied using a Holter device. standard.

An RR record thus obtained is given in Figure 2. In this figure, are represented more than 70,000 consecutive RR intervals obtained from an ECG performed over 24 hours. On the ordinate is indicated the amplitude of the RR and on the abscissa the index of the series.

The invention also applies to the treatment of ECGs obtained over shorter periods. In a first step 3-2 (FIG. 3), the periods of interest are identified, by detection of the zones of high variability in the heart rate. Indeed, atrial fibrillation (AF) increases this variability.

Precise localization of events with a high variability in heart rate can be carried out using a discrete wavelet transform (TOD) of the RR signal beforehand (Figure 4, step 4-2). The technique of transformation into continuous wavelets is described in the article by S. Mallat "A theory for multiresolution signal decomposition: the wavelet representation", IEEE transactions on pattern analysis and machine intelligence, 1989, Volume 11, pages 674-693. The discrete wavelet transform (TOD) is described in the same article by S. Mallat. It is an analysis by a family of functions ψ _jk created by the dilations and translations of a mother wavelet ψ, with ψ _jk (x) = 2 ^{j / 2} ψ (2 ^J xk), and j, ke Z (set of relative integers). The projection or transformation of each function f (signal RR) in the base of ψ _jk is noted: W _f (j, k) = <f, ψ _{j, k} >, where <.,.> Represents the scalar product in L ² (R). For a signal composed of 2 ^J values, this transformation leads to a series of 2 ^{* "1} coefficients noted dj (k), giving the contribution to the projection of the signal at position 2 ^J k on the scale 2 ^J. squared of the coefficients obtained provides the power of the transformed signal for each time and scale index Different time zones are characterized by a power calculated on the highest scales of their TOD (low j).

The discrete wavelet transform of a discrete signal is calculated by an algorithm based on quadrature mirror filters. Such filters are given in the work by A. Cohen, Wavelets and digital signal processing, Edition Masson, Paris, 1992.

From the digitized RR signal, a first decomposition is carried out, which gives a first level of detail (by high-pass filtering) then a first level of approximation (by low-pass filtering).

A second decomposition, by high-pass and low-pass filtering of the first level of approximation, respectively provides a second level of detail then a second level of approximation, which is in turn treated by a third decomposition, .etc.

The quadrature mirror filters make it possible to carry out the various successive high-pass and low-pass filterings. FIG. 7 represents several levels of the wavelet decomposition of the RR intervals of an AF episode. The abscissa corresponds to the index k and the ordinate to the index j. In other words, each line j gives the set of coefficients dj (k). At the center of this figure, a high variability due to an episode of AF is identifiable. The periods of high variability having been identified, they are separated into periods of RS and AF (steps 3-3 and 4-3 of Figures 3 and 4). Indeed, a great variability of similar amplitude can also be found in normal subjects, for example at night, in some young adults.

To this end, the periods of great variability are separated by fractal analysis. Calculation, by Fourier transform, of the spectral power density

(DSP), shows that a normal sinus rhythm (RS) has a general trend in 1 / f ^{* 3} , therefore presents a linear form of negative slope when it is displayed in log-log, as illustrated in figure 8. On this figure, peak 10 corresponds to a ventilation frequency. The FA DSP, on the other hand, presents two different slopes as seen in Figure 9: one in a strip of

1 high frequencies (10 ^" - 10 ^" Hz) and the second in a lower frequency band (ÎO ^ - IO ^" H _z ).

The slope in the high frequency band for AF is greater (β ≈ 0) than the slope obtained in the RS (β ≈ 1). To model the power law in 1 / f ³ , we apply the concept of fractional Brownian motion (mBf), a self-similar and non-stationary process. This concept is described, for example, in B. Mandelbrot et al. "

Fractional brownian motions, fractional noises and application ", SIAM Review,

1968, volume 10, pages 422-437. The self-similarity property is characterized by the exponent of Hurst H which is directly related to β by the relation: β = 2H + 1. The calculation of H is given for example in the article by M. Akay et al . "Fractal analysis of HRV signais: a comparative study" published in Methods of information in Medicine, 1997, volume 36, pages 271-273. The increment of the fractional Brownian motion process can also be used to estimate the Hurst exponent H, via the TOD, by: var [d _2j (n)] = 2 ^(2H - ^{1) ϋ "1)} _σ ² ( 2-2 ^{2H "1} ) where the d ₂ j are the coefficients of the details on the scale j. We then obtain the following equation: log ₂ [var [d _2j (n)]] = (2H - l) j + f (H, σ) where σ is the standard deviation of the signal, and f is a function not known.

The high-frequency part of an atrial fibrillation signal corresponds to a white noise, for which H "0.5. On the contrary, a normal signal moves away from white noise, the coefficient of H being greater than 0.5. We therefore set a threshold Ho, between 0.5 and 1, for example between 0.5 and

0.9 or between 0.7 and 0.85, for example Ho = 0.8. Very good values of sensitivity and specificity of the method were obtained for 0.7 <Ho <0.85. In particular, a sensitivity of 99.2% was obtained for permanent AF and 95.82% for paroxysmal AF, and a specificity of 99.96% for normal RS and 91.55% for paroxysmal AF. Figure 5 shows preliminary steps to obtain an RR signal from a digitized electrocardiogram (ECG) signal. The first two steps (5-1 and 5-2) are carried out conventionally in a standard "Holter" type device.

Typing classifies each complex as normal, ventricular, or supraventricular.

Normal and supraventricular QRS complexes show waves

P, in front of the maximum peak, while the abnormal (ventricular) do not have any. The QRS corresponding to these are in fact beats spontaneously born in the ventricles of the heart. They are generally wider than normal. To detect them, we can determine their width.

Thus is created a table or a database giving the type of QRS for each QRS, identified by its index in the continuation of the registered QRS.

Normal and supraventricular complexes are selected from this table (step 5-3). An RR signal can then be obtained (step 5-4) by calculating the difference between the peaks of successive QRSs (see Figure 1).

A transform into discrete wavelets particularly well suited in the context of the present invention is the transform into dyadic wavelets.

Indeed, there is generally a problem of reversibility of the wavelet transform. If the beginning of an RR signal is correctly identified, its transformation into discrete wavelets is reversible: we can, from each transform, find the initial signal by an inversion operation.

If, on the other hand, the origin of the RR signal is not correctly selected, the inverse transform operation brings about a distortion and does not make it possible to find the initial RR signal. This therefore results in a lack of precision in the identification of the origin or the beginning of atrial fibrillation, which can be very annoying, in particular for AF events of short duration.

The use of the dyadic transform, invariant by translation, solves the problem. Information on these dyadic transforms is given in the work of A. Antoniadis and G. Oppenheim, Springer Verlag, 1995, and in particular in the following articles contained in this work: "The Stationnary Wavelet Tranform And Some Satistical Applications" GPNason et al., page 281-299 and "Translation Invariant De-Noising", RR Coifman et al., pages 125-150. FIG. 6 represents steps for carrying out the detection or the identification of the zones of high variability of the RR signal (step 4-3).

The coefficients obtained by wavelet analysis are compared with a predetermined threshold, in order to retain only the coefficients of higher energy. Thus, this thresholding can be carried out on the decomposition levels j = 2, 4, and 8 (step 6-1).

Then (step 6-2), these coefficients are filtered, in order to obtain homogeneous areas of high energy.

To do this, we proceed to:

- a search for the start of each episode or each high energy zone,

- an accounting of the coefficients in the episode or in the zone,

- a determination of the end of the episode or of the zone.

During the search stage at the start of the episode, the number of coefficients which, in a certain window, are greater than the predetermined threshold is counted. When this number is greater than a certain predetermined number Ns, an episode begins.

In the accounting phase, we count the number of successive events that are below the threshold. If this number is greater than a predetermined number, an end of episode is detected. This step 6-2 allows a fairly precise determination of the events.

In some cases, however, this is not sufficient and a higher degree of precision is required.

For this reason, the start and end of the episodes are readjusted (step 6-3). This readjusting step makes it possible not to miss or let pass unique events and, moreover, to avoid misidentifying the start of an episode in the event of a longer episode.

Unique events can occur, which are just as important as successive events. For this reason, it may be just as important for the practitioner or the cardiologist to have information on simple events as well as on episodes counting several events. In addition, a bad positioning of the beginning of the episode, which can appear in particular on a display screen, can disturb this practitioner or this doctor and make him doubt the reliability of the device implementing the method of analysis according to the invention. To solve this problem, after having identified the homogeneous areas of high energy of the signal RR transformed by wavelets (step 6-2), the areas of high variability of the digital signal RR (step 6-3) are determined.

According to an example, the digital signal RR is taken up again, and the variations of this signal are detected which are greater than a certain predetermined percentage, for example 20% (a threshold between 10% and 40% or between 15% and 25 is predetermined %).

When there is such a variation from one point to another of the signal RR, it can precisely be said that there is the beginning of an event at this point.

We also determine the end of episodes or events. Thus, for each episode, one can obtain an interval NI - N2, which identifies the rank of the RR signals comprised between the start and end of the episodes.

This interval is then transferred to the different levels of decomposition (Figure 7) and only the coefficients in this interval are retained as corresponding to the episode. It is thus possible that coefficients, considered after step 6-2 as being part of the beginning or the end of a homogeneous zone of high energy, are excluded from this zone due to a slight modification, due at the resetting, at the beginning or at the end of an episode, or else that coefficients located at the transition of the homogeneous high energy zones and low energy zones are reclassified from one zone to another.

FIG. 10 represents a patient 20 provided with a device 22 of the "Holter" type, with its electrodes 24, 26, 28, 30, 32 arranged on the patient's body. The Holter records the ECG data for a certain period, for example about 24 hours or over a shorter period (for example a few minutes), or over longer periods (for example several days or several weeks),, and detects and classifies Qrs complexes.

As illustrated in FIG. 11, the data are for example recorded on a memory card 34 of the Flash memory type, for example of about 20 Mbytes. Other devices allow you to make recordings on cassettes inserted in the Holter 22. Figures 9 and 10 schematically show the essential electronic components of a Holter device.

FIG. 12 shows a device 40 for processing signals recorded on a memory medium 34, for example a memory card. The support 34 is introduced into a reader 36, for example a PCMCIA reader, and the data is then subjected to processing according to the steps described above.

The device 40 is therefore programmed to carry out these processing steps.

Such a device comprises a PC microcomputer 42 to which the data read by the reader 36 are transmitted via a link 41.

More specifically, the microcomputer 42 comprises (FIG. 13) a microprocessor 52, a set of RAM memories 54 (for the storage of data), a ROM memory 56 (for the storage of program instructions).

Optionally, a data acquisition card 60 puts the data supplied by the reader into the format required by the microcomputer. These various elements are connected to a bus 62.

Peripheral devices (screen or display device 64, mouse 66, keyboard 68) allow interactive dialogue with a user. In particular, the display means (screen) 64 make it possible to provide a user with a visual indication relating to the recorded ECG, at the calculated RR intervals, to the coefficients calculated by wavelet transform (as in FIG. 7). By selecting from menus and clicking on buttons or icons, an operator, for example a cardiologist, can view these different curves, in whole or in portions, with enlargement. On these curves can appear the intervals determined as being episodes of atrial fibrillation, and in particular the beginning and the end of each episode.

In the microcomputer 42 are loaded the data or the instructions for implementing data processing according to the invention.

This data or instructions for processing the data can be transferred to a memory area of the microcomputer 42 from a floppy disk or from any other medium which can be read by a microcomputer or a computer (for example: hard disk, ROM read only memory, DRAM dynamic random access memory or any other type of RAM memory, compact optical disk, magnetic or optical storage element). The display of episodes classified as episodes of atrial fibrillation does not in itself constitute a diagnosis. Only a doctor cardiologist can, at the sight of the initial electrocardiogram (which has also been digitized, stored in the support 34 and transferred to the memory of the microcomputer 42) confirm the existence of a pathology. In particular, the physician should still observe baseline 3 of the EKG (see Figure 1), to confirm whether an episode, apparently classified as AF, is indicative of atrial fibrillation .

The treatment method presented above is therefore an aid to diagnosis, but not the diagnosis itself, which requires an additional stage of medical interpretation.

EXAMPLE

This example concerns a population of 50 subjects: 19 suffering from chronic atrial fibrillation (FAC), 15 from paroxysmal fibrillation (FAP), and 16 patients controlled with a normal sinus rhythm (RS).

All subjects were recorded for 24 hours using a digital Holter ECG system (DuoHolter, Novacor, Paris, France). The recordings were directly digitized at a frequency of 100 Hz on the recorder; they were then reviewed and edited manually, if necessary, to obtain a list of consecutive RR intervals.

The signal processing was carried out with MATLAB® 5.0 software and the statistical analysis of the results with Statview ® on a Power Macintosh. ® The results are presented in the mean ± standard deviation form. The results are as follows: a. Controlled group ECG recordings with normal rhythm The standard examination of the 24-hour ECG recordings of the 16 SAR subjects showed the absence of any cardiac arrhythmia.

In one subject, two AF episodes were falsely detected (duration of 154 and 193 beats respectively; average heart rate of

81.5 + 14.8 and 76.0 + 9.5 beats per minute (bpm) respectively). There was no significant difference in the time indices between the subject with the false detections and the other 15.

Thus, the specificity of AF detection based on QRS complexes reaches 99.96% (1,083,537 Qrs true negatives versus 447 false positives). As all the false positives were obtained in a single patient, the specificity was 93.7% if the patient is considered as the unit of analysis (16 true negatives versus 1 false positive). b. Chronic AF group

Standard examination of the 24-hour ECG recordings showed permanent AF in the 19 subjects. The recordings were of a duration of

18.8 + 4.6 hours. In the 19 patients, the QRS number was 87,953 ± 27,491 and the heart rate was 75.4 + 1.0 bpm (range 23.1 to 289.7 beats per minute).

RS was falsely identified in 7 patients over 40 episodes (heart rate 82.5 + 9.5, duration 374 + 360 beats); one patient had 24 of 40 episodes while the others appeared in 6 other patients. These false negative episodes were due to atrial flutter and atrial tachycardia in 34 and 1 of 40 cases, respectively. In 5 of 7 patients, AF was not recognized at the start of the recordings; however, the recognition time was limited to 178.4 + 70 beats.

Thus, the detection sensitivity of AF reached 99.2% when considering the QRS complexes (2,064,197 QRS true positive against 16018 false negative). With the exception of the undetected episodes, all the patients were recognized as suffering from AF and the sensitivity obtained by considering the patients was 100%. The largest undetected period reached 2075 beats and the shortest 120. The patient with 24 undetected episodes had 10490 false negative beats; for this particular record, the detection sensitivity of AF based on QRS complexes was 89.9%. vs. Group of intermittent FAs. The 15 subjects with FAP had a total of 36 episodes of FAP with a duration of 2.5 ± 3.4 hours (from 34 s to 15:36). During these episodes, the average heart rate was 86.2 ± 20.7 (16.4 to 163.1) bpm and the number of QRS was 15055 ± 22125 (52 to 94990 QRS).

Only one of the 36 episodes remained undetected (260 QRS, heart rate of 123.2 ± 28.1 bpm). For the episodes detected, there were delays or advances in the recognition of AF at the beginning and at the end of the events:

- start of episodes: in 8 episodes there was a delay of 91 ± 95 beats (from 2 to 255); 23 episodes were detected prematurely (due to supraventricular tachycardias) of 186 ± 621 beats (from 2 to 2994). - end of the episodes: two episodes were premature (one of 3486 beats, the other of 333); 21 had a delay of 60 ± 103 beats (from 10 to 429). During the episodes of FAP, 14 false interruptions were observed, among which 13 in the same patient, with a duration of 1484 ± 151 beats; the last was 2001 beats in another patient; the total number of false negative QRS complexes then reached 28,695 (1.9%). These false negative episodes were due to atrial flutter and atrial tachycardia in 5 and 8 of the 14 cases, respectively.

There were also 37 false detections of 1573 ± 2309 beats (from 147 to 11491).

In the FAP group, the sensitivity based on the QRS complexes reached 95.8% and the specificity 91.5% /

All patients in the group were recognized as having AF, so the patient-based sensitivity was 100%.

The wavelet transform and associated fractal analysis according to the invention provided a sensitivity of 95% for detecting QRS complexes in AF, and a specificity of 99.6%, with a limitation due to the delays in recognition of AF at beginning and end of arrhythmia episodes. Based on the patients, the specificity reached 93% in the normal controlled group and the sensitivity 100% in either the FAC group or the FAP group.

In-ear recordings have often been used to detect AF, as well as analysis of the electrical activity of the atrium at the level of the esophagus, with a sensitivity and specificity ranging from 52.4-96 , 6% and

92.0-92.3%, respectively (see in particular S. Rajala et al., "ECG findings and survival in very old people", Eur Heart J. 1985, volume 6, pages 274-252).

Such recordings were not used since the goal was the non-invasive detection of AF by standard Holter recordings.

To obtain automatic detection, Holter recordings were analyzed by neural networks (T.F. Yang et al. "Artificial neural network for the diagnosis of atrial fibrillation", med. & Bio. Eng. & Comp. 1994, 32, pages

615-619), giving a sensitivity of 92% and a specificity of 92.3% in the recognition of QRS. Using a neural network fed both by the information of RR intervals and a morphological analysis of the baseline of the ECG trace, Cubanski D. et al (J. Cardiovasc. Electrophysiol. 1994, 5, pages

602-608), working on groups of 10 consecutive QRS, obtained a sensitivity of 82.4% and a specificity of 96.6%. Thus, the sensitivity and the specificity obtained by the method according to the invention are better than those of the previous studies. Our method has certain limitations. The use of HRV analysis alone is sensitive to anything that can degrade HRV, including supraventricular extrasystoles, supraventricular tachycardias and other types of arrhythmias. Detections of false positives can be due to the presence of a few supraventricular arrhythmias, isolated or grouped, generating a significant VFC; false positive detections occurred in 0.04% of the QRS in the normal control group and in 4.7% of the QRS in the PAF group. These false positive QRSs can be reclassified correctly using a track recorded by automatic event storage. Conversely, non-detections can occur in other atrial arrhythmias, such as paroxysmal supraventricular tachycardia, or atrial flutter, due to the relative regularity of these rhythms.

In clinical conditions, non-detection is more problematic than detection of false positives, since in this case the trace is not stored for the user, which prevents any subsequent correction. In the cases studied, non-detections were not frequent and represented less than 3% of episodes (1 in 36) in the FAP group and less than 2% of QRS. Improvements to the algorithm could reduce these non-detections, at the expense of specificity, however. The invention is therefore suitable for the automatic detection of AF in large populations using standard Holter recordings. This is particularly important in the elderly, in whom AF, both paroxysmal and chronic, is frequent and presents a major risk of stroke. Implemented in standard Holter devices or in long-term event recorders, the algorithm or method according to the invention makes it possible to better highlight the AF in order to take preventive measures before the appearance of dramatic complications.

Claims

1. Method for processing an RR interval signal from an electrocardiogram, comprising: - the detection of areas of high variability in heart rate in the RR signal,

- the classification of these zones into zones of sinus rhythm (RS) and atrial fibrillation (AF).

2. Method according to claim 1, the detection of the zones of strong variability of heart rate being carried out by projection or transformation of the signal of intervals RR in a base or a family of discrete wavelets, and calculation of the coefficients d. (K) giving the projection of the RR signal on each discrete wavelet function.

3. Method according to claim 2, the projection or transformation of the RR interval signal being carried out in a base or a family of dyadic wavelets.

4. Method according to one of claims 2 or 3, comprising a step of thresholding the coefficients dj (k) on one or more decomposition levels j in order to select the coefficients greater than a minimum value.

5. Method according to claim 4, further comprising a step for obtaining homogeneous areas of high energy.

6. Method according to claim 5, the step for obtaining high energy homogeneous zones comprising a filtering of the thresholded coefficients.

7. Method according to one of claims 4 to 6, comprising, for each homogeneous high energy area, a search for the start of the area, an accounting of the coefficients in this area, and a determination of the end of the area.

8. The method of claim 7, the search for the start of the zone comprising a counting of the coefficients which are greater than the predetermined threshold.

9. Method according to claim 7 or 8, the accounting of the coefficients comprising a counting of the number of successive events which are less than a predetermined threshold.

10. The method of claim 9, an end of episode being detected if the number of events counted is greater than a predetermined number.

11. Method according to one of claims 5 to 10, further comprising a step of registration or identification of the beginnings and ends of the episodes, using variations of the RR signal.

12. The method of claim 11, the coefficients giving the projection of the RR signal being retained as part of a homogeneous area of high energy only if they are between a start and an end of an episode.

13. The method of claim 12, further comprising a display, on display means, of at least part of the electrocardiogram and an identification, on this display, of the zones of atrial fibrillation.

14. Method according to one of claims 1 to 13, the zones of high heart rate variability being classified into zones RS and zone FA by calculation of the fractal exponent of each of these zones and comparison with a threshold value Ho.

15. The method of claim 14, the threshold value Ho being between 0.5 and 0.9.

16. The method of claim 15, the threshold value being between 0.7 and 0.85.

17. Method according to one of claims 1 to 16, the electrocardiogram having been acquired for a period of between a few minutes and a few months.

18. Device for analyzing digital signals from RR intervals of electrocardiograms, comprising: - means for classifying these zones into sinus rhythm zones (RS) and atrial fibrillation zones (AF).

19. Device according to claim 18, the means for detecting zones of high variability in heart rate comprising means for projecting or transforming the signal of RR intervals in a base or a family of discrete wavelets, and calculating coefficients d. (k) giving the projection of the RR signal on each discrete wavelet function.

20. Device according to claim 18 or 19, the means for classifying the zones of high variability into zones of sinus rhythm and into zones of atrial fibrillation comprising means for calculating the fractal exponent of each of these zones and for comparing it with a threshold value Ho.

21. Device according to one of claims 18 to 20, further comprising display means for displaying the zones of atrial fibrillation.

22. Device according to one of claims 18 to 20, further comprising display means for viewing the areas of atrial fibrillation and the electrocardiograms, and for viewing the areas of atrial fibrillation with respect to the electrocardiograms.

23. System for determining the heart rate, comprising one or more recorders for recording electrocardiograms of patients, and a device according to one of claims 18 to 22.