WO2011004078A1

WO2011004078A1 - Data-processing method enabling the filtering of artefacts, in particular for electroencephalography

Info

Publication number: WO2011004078A1
Application number: PCT/FR2009/051340
Authority: WO
Inventors: Laurent Peyrodie; Samuel Boudet; Philippe Gallois
Original assignee: Hautes Etudes D'ingenieur; Institut Catholique De Lille
Priority date: 2009-07-07
Filing date: 2009-07-07
Publication date: 2011-01-13
Also published as: EP2451346A1; FR2947933B1; FR2947933A1

Abstract

The present invention relates to a method for processing at least one multichannel signal (X) broken down, over a plurality of time-frequency windows (P_i ^j), into a plurality of multichannel signals (X_pj_i), said multichannel signal (X) resulting from a linear combination of at least one main signal (sp) obtained by measuring data from a main data source (S) and at least one secondary signal (ss) obtained by measuring data from a secondary data source (B) corresponding to at least one artefact, said method comprising at least one learning step (SO) including a comparative multi-component analysis (AMC) mainly comprising the comparison of an index of said signal (X), preferably representing the power thereof, over two separate time-frequency windows (P_i ¹ and P_i ²), so as to separate the components (w) of the main signal (sp) from the components of at least one portion of the secondary signal (ss).

Description

METHOD OF PROCESSING DATA FOR FILTERING ARTIFACTS, ESPECIALLY ADAPTED TO ELECTROENCEPHALOGRAPHY

Technical field of the invention

The subject of the present invention relates to the field of filtering artifacts on multichannel signals; the subject of the present invention being particularly adapted to the analysis of multichannel signals from the measurement of cerebral electrical activity by electroencephalography.

More generally, the object of the present invention finds an advantageous application in the medical field, by promoting for the practitioner the reading and the analysis of the electro-physiological activity of one or more parts of the human or animal body, and especially the electrical activity of the human brain.

Artefact means throughout the present description all types of alteration or disturbance originating from external phenomena that prevent the precise analysis of the signals that one seeks to study and / or determine.

State of the art

The purpose of neurology is to study the anatomy, physiology and pathologies of the nervous system, in general the human nervous system.

Among the various pathologies of the nervous system, epilepsy is a chronic neurological condition characterized by excessive discharges of brain neurons, the symptoms of which are manifested by intermittent chronic cerebral hyperactivity, manifested for example in the form of convulsions. eclampsia, loss of consciousness, or hallucinations.

To diagnose such a pathology that affects more than 40 million people worldwide, neurologists must measure the electrical activity of the brain during epileptic seizures in order to advocate appropriate treatment.

Electroencephalography is a method of measuring the electrical activity of the brain, the principle of which is to amplify the electrical signals collected on the surface of a patient's scalp via a set of placed electrodes. on the surface of the scalp. There are other technologies that are also interesting for assessing the activity and physiology of the brain.

These include technologies related to medical imaging such as the Magnetic Resonance Pulse method, better known as MRI.

However, electroencephalography due to its low cost, its non-invasive nature, its small size and its high temporal resolution is often used for the diagnosis of neurological diseases such as epilepsy and / or preliminary examinations.

In addition, electroencephalography, by its technical characteristics, is the only known non-invasive technology that can reveal the pathological abnormalities due to epilepsy.

We note, however, that a standard electroencephalograph has many problems that must be resolved to allow the practitioner a good analysis of the measured electrical activity.

Indeed, when using an electroencephalograph, the number of electrodes that can be placed on the surface of the skull is limited; this implies a low spatial resolution for the study of the electrical activity of the patient's brain.

Moreover, as mentioned above, the electrodes are placed on the surface of the skull at the level of the scalp; this involves artifacts and the loss of certain signals (deep sources, high frequencies, signals specific to very localized areas, etc.).

These artifacts, which contaminate electroencephalographic signals, make medical diagnosis often very difficult for the practitioner or neurologist who must read, analyze and interpret these contaminated signals.

To improve the quality of these signals, it is possible to use subcutaneous electrodes or intracranial electrodes on the cortical surface (electrocorticogram) or in depth (electrogram).

Nevertheless, these invasive methods involve heavy surgery; they are practiced only in exceptional cases.

It is therefore necessary to improve the existing electroencephalographs by improving the quality of the measured signals, and in particular by reducing the maximum the different alterations and disturbances of the main signals due to artifacts while retaining these main signals.

Artifacts include ocular artifacts that may be due to blinking of the eyes or eye movements, for example, a patient who, while in hospital, watches a tennis game and follows the ball. on the screen without moving your head.

Blink artifacts are attributed to conductance alterations from contact between the eyelid and the cornea.

The artefacts due to the movements of the eyes are, in turn, caused by the eyeball which acts as a dipole whose positive pole is the cornea and the negative pole is the retina; the movements of the eyeball causing the positive pole to move towards or away from the electrodes placed on the forehead and the temporal areas of the patient.

This kind of artefacts of ocular origin is mainly in a frequency band between 0.5 and 4 Hertz, and has an amplitude generally between 20 and 150μV.

Artifacts of muscular origin are also found which may be due, for example, to jaw clenching or frowning, or artifacts of electrocardiographic origin. These so-called muscular artifacts cause myogenic potentials likely to disturb the reading of cerebral electrical activity signals (amplitude ranging from 0 to 200μV and frequency> 13 Hertz).

There are also artifacts of mechanical origin that can be caused when an electrode and / or the wire of an electrode move, especially because of the throbbing of a vein, breathing or during movements of the face, for example, when the patient speaks (variable amplitude and frequency between

0.5 and 4 Hertz).

There is also an artifact called galvanic sweat which has the same characteristics as these artefacts of mechanical origin.

Another type of artifact is caused by the power supply of the electroencephalograph. In France, this artifact corresponds to interference at 50 Hertz. Thus, it is easily understood that it is necessary, in order to improve the clinical analysis of the signals measured by the electroencephalograph, to carry out a suppression, or at least a reduction, of these different types of artifacts while retaining the signals. main features that are characteristic of brain activity and more particularly the characteristic signals of neurological pathology.

In the particular case of epilepsy, among the various characteristic signals, we will seek to maintain the normal rhythm including the alpha rhythm that occurs during the closing of the eyes, and pathological rhythms such as paroxysms, spikes and slow waves, which can all be indicators of epilepsy.

However, we note that the artifacts often have an amplitude greater than that of brain activity. It can therefore be very difficult to distinguish between a simple artifact and a cerebral signal representing a part of brain activity.

In known manner, electroencephalograph systems implement methods using frequency filtering to reduce these artifacts.

More specifically, these systems include a high-pass filter for at least partially removing artifacts of mechanical or ocular origin, and a low-pass filter for at least partially removing artifacts of muscular origin.

These filters have the advantage of letting pass a large part of the cerebral activity signals which are generally in the bandwidth.

In addition, these frequency filters can be implemented easily in real time on the electroencephalograph so that after frequency filtering the filtered signals representative of the brain activity can be made in real time on screen.

However, these types of high-pass filters and low-pass filters do not pass signals representative of the desired pathology. Indeed, some cerebral activity signals (most often those representative of the pathology) belong to the same frequency band as certain artifacts.

Let us take the example of two signals SpI and Sp2 which respectively correspond to an artifact and a principal signal, and whose spectra are illustrated in Figure 1 corresponding to a graph representing ordinate amplitude and abscissa frequency.

In this example, it is impossible to distinguish by frequency filtering the spectrum of the main signal Sp2 from that of the artefact Sp1; these being included in the same frequency bands.

For the various reasons mentioned above, frequency filtering can not be used continuously in a clinical environment to allow the practitioner to detect accurately and certain symptomatic signals of epilepsy or all other types of neurological pathology .

Alternatively, it has already been envisaged to use spatial filtering using the distribution of the signals on the different electrodes.

Among the various types of known spatial filtering, mention may be made of the Independent Component Analysis proposed by T. Jung et al in the publication "Removing electroencephalographic artifacts by Blind Source Separation", published in 2000 in the journal "Psychophysiology".

This kind of treatment involves a manual selection of the components to determine for each of them whether they represent an artefact or brain activity.

The Chinese patent application CN 1 883 384 discloses a treatment using the Independent Component Analysis for filtering.

We can also mention the international patent application WO

2006/072150 proposing an electroencephalograph containing a processor implementing a blind source separation algorithm

Canonical Correlation Analysis; this algorithm allows to reduce artifacts of muscular origin only.

This kind of statistical processing has the constraint of limiting the number of components to the number of electrodes.

Thus, when the signals have a lot of artifacts, the components that result from this processing mainly represent the artifacts, and the cerebral signal is strongly deteriorated, even unreadable.

In addition, automated spatial filtering methods generally allow filtering for only one type of artifact (ocular or muscular). Only Levan, in the publication "A System for automatic artifact removal in ictal scalp EEG based on independent component analysis and Bayesian classification", published in 2006 in the journal "Clinical neurophysiology by Elsevier", managed to automate a method using the Analysis Independent Components to reduce all types of artifacts, and to maintain largely normal and pathological brain activities.

However, the method proposed by Levan still has the drawbacks and limitations of Independent Component Analysis, which is that when the signal (s) have too many artifacts, this analysis allows only a partial reduction of the artifacts and results in a significant decrease in cerebral signals.

The results and performances thus obtained are not satisfactory and sufficient for a current clinical application.

Another disadvantage of the spatial filters compared to the frequency filters remains that the results and performances remain inconstant. Indeed, for the same type of wave, depending on the rest of the traced route, the source may or may not be correctly characterized and preserved.

It is therefore difficult to use spatial filters in the clinic knowing that their results and their performances are random.

Obiet and summary of the invention

The object of the invention is therefore to provide a solution to the aforementioned problems among other problems.

One of the technical problems solved by the present invention consists in particular in the reduction, or even the automatic removal, of all the artefacts, while retaining most of the cerebral signals, particularly those which characterize the pathology; the present invention being particularly suitable for long-term recording filtering.

For this purpose, the subject of the present invention relates to a method of processing at least one multichannel signal decomposed over a plurality of time-frequency windows into a plurality of multichannel signals.

The decomposition of the multichannel signal, over a plurality of time-frequency windows, into a plurality of multichannel signals corresponds in the the present invention at a prior stage of extracting from the multichannel signal a plurality of time-frequency windows to obtain a plurality of multichannel signals.

The multichannel signal results from a linear combination of at least one primary signal from the measurement of data from a primary data source and at least one secondary signal from the measurement of data from a secondary data source. corresponding to at least one artifact.

The method according to the present invention comprises a learning step comprising a comparative m ulticomotive analysis consisting in particular of a comparison of an index, preferably representative of the power, of said signal on two distinct time-frequency windows so as to separate the components of the main signal of the components of at least a part of the secondary signal, preferably on each of the channels, and for the two time-frequency windows.

The method according to the present invention further comprises a step of overall construction of a global filter for eliminating or reducing the secondary signal.

This global construction step of a global filter comprises a first step of constructing at least a first filter making it possible to preserve in said multichannel signal, and on a third time-frequency window, the main signal and to eliminate at least one part of the secondary signal.

More precisely, this first construction step consists in particular of a linear regression of the components of the main signal or of said components of said at least part of the secondary signal, possibly on said two time-frequency windows.

Preferably, the comparison in a first variant is performed, for at least one frequency window, on a window of time at rest and on an artefacted time window. In a second variant, the comparison is made, for the same time window, on two windows of different frequencies. Preferably, the idle time window and the artifact time window are obtained manually or automatically from a pre-established patient database.

According to a preferred variant embodiment, the multicomponent comparative analysis of the multichannel signals comprises the steps of:

calculating the covariance matrices of the multichannel signals on the two time-frequency windows;

- diagonalize simultaneously the two matrices of covariances; and

obtain the vector representing the power ratio of each component between the two time-frequency windows.

Preferably, the comparative multicomponent analysis further comprises a step of eliminating the components that have the largest eigenvalues, for example either by predetermining a threshold value not to be exceeded, or by predetermining a number of components to be eliminated.

In a variant embodiment in which the first construction step makes it possible to determine a first and a second filter, the overall filter is obtained by a convolution step of said first and second filters.

Preferably, the method according to the present invention comprises a filtering step consisting in the application of one of the aforementioned filters to the components of the multichannel signal for a plurality of time-frequency windows, in order to obtain signals filtered for each time-frequency window.

The method according to the present invention further comprises a step of reconstructing at least one signal over the entire time period by subtracting the multichannel signals for each time-frequency window, and adding the filtered signals for each window time-frequency.

Advantageously, the overall construction step according to the present invention comprises a stabilization step consisting in particular of an orthogonal projection of the components of one of said filters previously obtained on the third time-frequency window so as to increase the difference between the distribution components of the secondary signal and components of the main signal. Preferably, the multichannel signal is a signal that contains electrophysiological information.

More particularly, the multichannel signal is the signal of an electroencephalograph; this signal resulting from a linear combination of at least one main signal derived from measurement derived from brain activity, including normal brain activity and pathological brain activity, and at least one secondary signal from the measurement coming from at least one artefact of the mechanical, ocular, muscular artifact type.

The use of the method according to the present invention thus makes it possible to improve the reading of the channels of an electroencephalograph, and to improve the detection of neurological pathology of the epilepsy type.

Correlatively, the object of the present invention relates to a computer program comprising instructions adapted to the execution of the steps of the processing method as described above, and when said program is executed by a computer.

The object of the present invention also relates to a computer-readable recording medium on which is recorded a computer program comprising instructions for executing the steps of the processing method as described above.

The object of the present invention further relates to a medical device comprising means adapted for implementing the treatment method as described above; said medical device may consist of an electroencephalograph or an electrocardiograph.

Thus, besides electroencephalography, the present invention finds other advantageous applications such as applications for electrocardiography with, for example, the removal of artifacts for a better analysis of the electrical activity of the heart and a better diagnosis of disorders. cardiacs such as arrhythmia.

Brief description of the drawings

Other features and advantages of the present invention will emerge from the description below, with reference to FIGS. 2 to 5, which illustrate an example of embodiment that is devoid of any limiting character and on which: - Figure 2 shows a device according to the invention, in a particular embodiment.

- Figure 3 shows schematically the positioning in top view of the device electrodes on the surface of the skull of a patient.

FIG. 4 represents, in flowchart form, the main steps of the processing method according to a particular embodiment of the invention.

FIGS. 5a to 5d show examples of results obtained clinically using the method according to the present invention.

Detailed description of an embodiment

An electroencephalograph-type medical device and a treatment method according to a particular embodiment of the invention will now be described with reference to FIGS. 2 to 5.

The invention is described herein in the particular context of the signal processing of an electroencephalograph to improve and facilitate the analysis and interpretation by a practitioner of the brain activity of a patient who suffers from a neurological pathology such as epilepsy; this electroencephalograph can be a simple signal reading device and / or a simple signal processing device.

The electroencephaloalogram EEG

In known manner and as illustrated in FIGS. 2 and 3, the electroencephalograph, referred to herein as EEG, comprises a plurality of measurement sensors E adapted to be positioned on the surface of the skull of a patient in order to measure the electrical activity of the patient. his brain.

In the embodiment described here, the measurement sensors E consist in particular of electrodes.

Preferably, and as illustrated in FIG. 3, a 21-electrode "10/20" system is used in the embodiment of the present invention, for precise positioning of each of the electrodes to the patient's skull surface. kind of getting a precise measure of the cerebral activity of the brain.

In this well-known system, the letters of each of the electrodes respectively indicate the brain lobe on which each of the electrodes is positioned: FP for Polar Front, F for Frontal, P for Parietal, C for Central, T for Temporal, and O for Occipital.

Then, an even number is added behind this letter for each of the electrodes positioned on the right side, and an odd number is added behind this letter for each of the electrodes positioned on the left side.

A z suffix is added behind each of the letters for the electrodes positioned on the center line.

The electroencephalograph EEG further comprises a module Ml adapted to acquire the multichannel signals representative of the measurement of the electrical activity of the brain.

These multichannel signals are materialized by an X matrix of dimensions (m, n) where the m lines represent the EEG electroencephalograph tracks and the n columns represent the temporal samples.

In general, it can be considered that the one or more multichannel signals X of an EEG electroencephalograph result from a linear combination between at least one main signal sp resulting from the measurement of data coming from a main data source S corresponding here to brain activity (normal brain activity and possibly pathological abnormalities), and at least one secondary signal ss derived from the measurement of data from a secondary data source B corresponding to at least one artifact.

More particularly, it is possible to assume that these multichannel signals X result from a simple sum between the signals of these different types of main sources S and secondary B.

The electroencephalograph EEG may also include an M3 display module for displaying a graphical representation of this brain activity.

Thus, the signal or signals that will be processed according to the method of the present invention, then displayed via the module M3 on the EEG electroencephalograph, may be analyzed, for example visually by the practitioner or automatically, so as to allow the pathological detection of epilepsy, early or early detection of epileptic seizures, characterization of sleep phases, bi-spectral analysis (anesthesia), or even pain detection for unconscious patients.

This M3 display module also allows the "biofeedback" which can especially help the patient to relax.

Obviously, the display can be performed on a device other than the EEG electroencephalograph, a computer connected to the electroencephalograph, for example.

This visual analysis can detect and characterize neurological disorders in the context of epilepsy, or in the context of dementia, multiple sclerosis, or any other type of neurological pathology.

Such EEG electroencephalograph can also detect brain death in comatose patients.

Advantageously, the electroencephalograph EEG according to the present invention further comprises a digital processing module M2 which cooperates with a microprocessor μP, or possibly a microcontroller, comprising a computer program PG.

This computer program PG includes instructions adapted for the execution of the steps of a processing method for obtaining a filter capable of filtering the multichannel signals X so as to retain the main signals sp representative of the brain activity S while partially or totally eliminating secondary signals characteristic of B artifacts.

The object of the present invention is therefore to implant on a conventional EEG electroencephalograph such a microprocessor μP.

Preferably, the microprocessor μP is further arranged to control the modules M1 and M3.

In other embodiments, part of the treatments can also be performed analogically.

Process according to the present invention

The various steps of the treatment method according to the present invention will now be described in the following with reference directly to FIG. 4. The general principle! of the present invention is based on a duality between time and frequency; this duality makes it possible to fifter the different types of secondary signals ss corresponding to the artifacts B.

Thus, in a characteristic way, thanks to this duality between the time and the frequency, it is possible to proceed to the different stages of the process by replacing the time by the frequency and vice versa, and to obtain a filter making it possible to filter most of the artifacts while retaining the main signals characteristic of the main activity.

More specifically, the invention consists in breaking down the multichannel signals X into one or more multichannel signals.

on one or more time-frequency windows and then build a set of spatial filters

(matrices mxm).

The construction of such filters

requires in particular a step SO consisting in the use of a method called multicomponent comparative analysis AMC and a global construction step S1, S2, S3 consisting in particular of a linear regression; these different steps will be described in more detail below.

Each filter

obtained is then applied during a filtering step S4 on the time-frequency window

multichannel signal

; the application of such a filter on a multichannel signal

to perform the operation

The method then comprises a SS reconstruction step; such a reconstruction can be considered as a filtering of a time-frequency window Α consisting in extracting the multichannel signal

from this window

I filter spatially according to the formula

, then reintroduce it into the original signal so as to obtain a signal X 'equal to:

In general, the reconstruction step S5 can be considered as a function of recomposition Fr defined such that; this

function being applied to recompose at least partially the main sp signal (s) without artifact

Time / frequency windowing

As mentioned above, the method implanted on the microprocessor μP EEG electroencephalograph allows treatment of the multichannel signal or signals X over a period of time T; this time period T being decomposed into a plurality of time-frequency windows

By time-frequency window of a signal is meant throughout the present description a linear function of a signal to another signal; this signal does not necessarily have the same duration, nor the same sampling frequency, nor the same base (for example frequency base).

Preferably, a time-frequency window corresponds to the resultant of the extraction of a signal on a window (or period) of time, and then of a frequency filtering.

According to the present invention, the application of a time-frequency window for a multichannel signal then consists in applying the same time-frequency window for each of the channels.

In this case, we therefore note

the multichannel signal resulting from the application of the time-frequency window

on the multichannel signal X.

A decomposition of a signal X into time-frequency windows thus consists in extracting a plurality of time-frequency windows from this signal.

It may be noted that these time-frequency windows are defined such that there is a function for at least partially reconstructing the original signal from only time-frequency window signals. There is therefore a function of recomposition Fr which makes it possible to define the final signal X from the decomposition into multichannel signals

Preferably, the time-frequency window decomposition according to the present invention consists of breaking down the signal into different disjoint time windows and then filtering it in different frequency windows disjoined by zero-phase filters (for example FIR type filters). symmetrical or HR, applied in one direction then in the other).

Preferably, the redial function is a sum. Definition of a spatial filter

A spatial filter F according to ia present may be represented by an idempotent matrix (m, m); that is to say a matrix whose product gives the following result - ^ = F,

On the other hand, a idempotent matrix is characterized by the fact that it is diagonalizable, and that these eigenvalues can only be 1 and 0.

Therefore, such a matrix can be obtained by looking for its two proper subspaces on the right

respectively corresponding to the own values 1 and 0, or by its two subspaces proper on the left

But typically, the right and left specific subspaces are such that

designating the unique orthogonal space at E,

The diagonalization can thus be written in the form F ^≈ SIDW with M = W * matrix of passage:

where (α = 0 or 1) is a dimension matrix

whose column vectors form a basis for

a subspace vector dimension

, and is a matrix dimension matrix whose column vectors

form a basis for

_t a vector subspace of dimension

It should be noted that any database of

and

gives the same filtering matrix, likewise for any database of

More specifically, in the embodiment described here,

represents the vector subspace of mixing (or distribution) of artifacts; and,

represents the vector subspace of mixing (or distribution) of the main sources.

represents the subspace vector of separation of artifacts; and,

represents, for its part, the vector subspace of separation of the main sources.

On lexical yaws, it is decisive to distinguish the distribution of sources which corresponds to the distribution of the sources on the different pathways. and the source separation which corresponds to the manner in which these sources are obtained from the different channels.

The method according to the present invention therefore consists of searching, for each time-frequency window, the bases of the vector subspaces.

; this is to define the filter h.

More specifically, the idea is to define

by a SQ learning step, and then

by a first step of construction S1, so as to construct a filter designating

Dimension identity matrix (mxm), Step SO: Determination of

The step SO according to the present invention consists in searching for a set of components representing the main signal (s) stemming from the measurement of the cerebral activity S (or the ss secondary signal or signals representing the artefacts).

B); these components forming a base

of a vector subspace

By component, i! is meant throughout the present description a linear combination which is represented by a vector w of dimension m; the source being a signal which is the result of the linear combination w of a multichannel signal:

For this purpose, a comparative multicomponent analysis method AMC is applied to the signals of two time-frequency windows.

previously defined.

In general, such an AMC method between two multichannel signals thus consists in finding the components whose index difference between

two multichannel signals is maximal (or minimal); this index being

representative of the power for example. Comparative Muitic Component Analysis

In general, the AMC multicomponent comparative analysis consists in defining two multi-channel signals on two time-frequency windows.

, and to look for the components of these.

At this effeζ one determines a subscript of the signal that we compute for and

; this index can be the power, the variance, the covariance matrix, the autocorrelation, the expectation, any order moment, the kurtosls, or any other type of index representative of a signal.

The components must have a maximum or minimum index difference between the two multi-channel signals, depending on the intended purpose and

the chosen gap.

By difference, here means a difference, a ratio, or another function between the two indices.

According to a first variant embodiment, the AMC multicomponent comparative analysis of two multi-channel signals corresponds to the method

so-called "CSP" for "Common Spatial Pattern" whose purpose is to find components maximizing (or minimizing) a variance ratio on the components between multi-channel signals

This first embodiment variant consists in particular of:

- calculate the covariance matrices of the multi-channel signals

and on two time-frequency windows

- diagonize simultaneously the two matrices of covariances

where P is the transition matrix and D is the diagonal matrix of the eigenvalues sorted in ascending order; and

obtain the vector v representing the power ratio of each component between its signals, this vector v corresponding to the diagonal

matrix D; and finally - get your components from

by the vectors of P which have the largest (or smaller) eigenvalues, for example either by predetermining a threshold value on v not to be exceeded, outputs by predetermining a number of components to be eliminated.

According to a second variant embodiment, the AMC comparative multi-component analysis of the two multi-channel signals corresponds to a

a method called "ACP" for Principal Component Analysis which consists in particular in calculating the difference of the covariance matrices in the following manner;

According to a third variant embodiment, the AMC comparative multi-component analysis of the two multi-channel signals consists in particular in

perform the method called "QA" for Independent Component Analysis, then compare the index computes on the components between

Step S1: Determination of

Once the base

determined by this multicompassing comparative analysis AMQ a new window of time-frequency

is considered, and a first construction step Sl consisting in particular of a linear regression of the sources is applied to the different channels for this time-frequency window P * in order to reconstruct a portion of the signal.

More specifically, linear regression consists of calculating:

is the covariance matrix of the multi-channel signal on the window

.

The matrix allowing this linear regression contains column vectors which are a base of

can then be considered as the vector subspace that is -orthogonal to

(or

, the vector-orthogonal subspace to

.

Choice of time-frequency windows

The time-frequency duality mentioned above is decisive in the choice of time-frequency windows; this choice being specific to the electrophysiological signals in the embodiment relating to the electroencephalograph I ¹ EEG Two different embodiments then make it possible to filter different artifacts: the time mode and the frequency mode.

In these two modes, your time-frequency windows

are the same ;

possibly taking into account edge effects.

Time mode

In this mode, the frequency windows of the time-frequency windows

are the same ; however, their time windows are different.

In other words, for at least one determined frequency window (for example Δ (0-4Hz), θ (4-8Hz), α (8-13Hz) or β (> 13Hz)), we will consider a window on a time window corresponding to a rest period

(here called window of rest time); this window having a minimum of secondary signals ss, but during which the main signals sp are active.

For each of the frequency windows above, we will then consider a time-frequency window.

on a time window corresponding to an artefacted period (here called artefacted time window) during which the secondary signals ss are active.

The present mode is based on the assumption that the distribution of artifacts (or mixture of secondary signals) is the same on windows

and, and on all your windows

this same frequency window; this even if the artifacts are inactive.

In order to define this so-called artefact window

, the treatment method according to a first embodiment variant provides an operating protocol that the patient must perform.

More precisely, before each recording, the latter must perform a series of movements (eg for TEEG: blinking, jaw movements, forehead stretching, smiles, etc.) to simulate artifacts, preferably for two minutes, so that you can determine the time window artificially for all i.

In this first variant, the patient will also have to simulate a rest period in order to determine the window of rest time.

.

According to a second variant embodiment, it is possible to avoid this protocol and to select, manually or automatically, artefacted windows and / or at rest directly on the recording or from a set of recordings, and use a pre-established artifact database.

Initiating the method with a learning step to determine the spatial distribution of the artifacts for which the distribution is constant over time is characteristic of the present invention.

Thus, once the idle time window and the artefacted time window determined by one of these two variants, the processing method comprises a learning step SO comprising a comparative multicomponent analysis AMC which compares for each of the windows frequency, particularly the frequency windows Δ (0-4Hz), θ (4-8Hz), α (8-13Hz) and β (> 13Hz), an index, representative of the power for example, of the multichannel signal for the two windows temporal; these two time windows respectively corresponding to a period of time at rest and an artefacted time period.

More specifically, this learning step SO consists in particular of:

calculating an index, preferably representative of the power, for the components on the multichannel signals respectively on the

time-frequency window

for example the calculation of covariance matrices; then

to maximize or minimize a difference between the indices, preferably the ratio between these indices.

Using this multicomponent comparative analysis AMC, it is possible to determine the components of multichannel signals X which have a deviation (power) approximately constant between the two windows of time-frequency

and In other words, the components corresponding to the separation of the main sources are those whose index increases the least between the two time windows studied.

It should be noted that determining the spatial distribution of the secondary source (s) is equivalent to determining the separation of the main source (s).

Preferably, for all time windows

corresponding to the same frequency window, the windows

are identical and therefore the bases of vector subspaces

are the same.

Once this learning step SO is performed on each time window, it is possible to reconstruct the entire signal on the time-frequency window.

For this purpose, the method according to the present invention comprises a first construction step Sl consisting in particular of a linear regression with

corresponding to

possibly with a slightly larger time window to avoid edge effects.

Thus, it is possible to determine the distribution of the main sources according to the current time.

More particularly, this first construction step S1 makes it possible to determine a first filter making it possible to filter the artifacts whose distribution

is constant in time; namely mainly eye artifacts and certain muscular artifacts.

However, it is found that the mechanical artifacts are not filtered; these mechanical artefacts are characterized by their random distribution on the electrodes. Frequency mode

As mentioned above, the mechanical artifacts are non-constant in time; however, these are characterized by a very low frequency compared to that of the main sources.

Thus, in order to obtain a spatial filter to eliminate them, it is possible to determine the components of which most of the signal is located in the very low frequencies. This mode can also be used to filter out muscular artifacts that have the characteristic of carrying much of their power in high frequencies.

In this frequent mode, the time windows of the time-frequency windows

are identical (except edge effects for;

Again, the frequency windows are different.

The construction of the filter according to ia here therefore comprises a learning step SO of applying a multicomponent comparative analysis AMC comparing an index of multichannel signals on two windows

different frequencies for the same window of time; preferably between 0 and 2Hz and between 4 and 8Hz for the filtering of mechanical artifacts.

More precisely, the components of mechanical artifacts can be considered as those whose power in the 0-2 Hz band is greater than the power in the 4-8 Hz band.

Previously, the extraction of a frequency window of at least one multi-channel signal requires the application of high-pass, low-pass, and / or band-pass or other type of frequency filter frequency filters, and in particular those resulting from an optimization for the purpose of separating as much as possible the main signals from the secondary signals.

The method further comprises a step Sl consisting in particular of reconstructing the signal by linear regression on the time-frequency window P * (corresponding to

) possibly with a slightly larger time window to avoid edge effects. .

We note that the number of components of the bases can

vary according to the components to be filtered on the frequent window.

This first construction step S 1 thus makes it possible to determine a second filter

allowing a spatial characterisatioπ of the sources to eliminate which correspond to mechanical and / or muscular artefacts.

Combination of filters

Then, in the embodiment described here, it is possible to combine the characteristics of the first and second filters

previously obtained from so as to obtain a global filter capable of simultaneously filtering different types of artifacts: mechanical, muscular, ocular, etc.

In a first variant embodiment, obtaining such a filter F ₁ "consists in particular in combining the first filter of the time mode and the

second filter

frequency mode, previously described.

To do this, we consider

two windows to learn the distribution of a first type of artifacts, and

two windows to learn the distribution of a second type of artifact.

We also consider the windows used to learn the

distribution of the main signals. Preferably, and corresponds

to but with more edges.

In the embodiment described here, for a time-frequency window a first filter is calculated

by applying the steps of comparative multicomponent analysis SO and linear regression Sl on

; then we calculate a second filter

by applying comparative multicomponent analysis steps

SO and linear regression Sl on

The global filter F * is then calculated by a convolution step S3 of the first and second filters according to the formula

In this first variant, the combining step includes the learning step SQ and the linear regression step S1.

In a second variant embodiment, the construction of the first filter F comprises a combination step of searching from

time-frequency windows

(two windows to learn the repartions of a first type of artifact) and (two windows allowing

to learn the distribution of a second type of artifact).

This variant embodiment consists in particular of:

- perform comparative multicomponent analysis of X on

in order to get to determine the components with no artifact of a

first type but possibly artifacts of a second type;

- define a new multichannel signal corresponding to the sources of

these components; - perform the AMC multicomponent comparative analysis of

to obtain in order to extract the components without artifact from the

second type; then

- to define

In this second variant, the combination step incurs the learning step SO. Step S1 is then applied normally to

in order to build

Of course, all other types of method allowing the combination of the first and second filters to obtain a general filter P ₁ * will be possible.

be used in the present invention.

In general, we note that, in a typical way, it is possible to calculate the inverse filter in order to reverse the hypotheses by replacing this filter Fx by I-Fx; this is equivalent to replacing

, and

A filter is thus obtained which retains the signals whose distribution is constant over time and which eliminates the signals whose distribution varies.

Stabilization

The global filter

thus obtained is not optimal.

Indeed, the actual distribution of artifacts on a current window can

be slightly different from the distribution of artifacts on learning windows

If this is the case, the method may seek to reconstruct the secondary signal (s) (artifacts) instead of reconstructing the sp (brain source) main signal (s); the secondary signals may be much larger than the main signals.

This instability is materialized by a gap between the vector subspace

space of distribution of artifacts, and the vector subspace

space of distribution of brain sources.

Preferably, the measurement of the difference between

can be considered as a minimal angle.

To avoid this instability, the overall construction step S1, S2, 53 according to a first embodiment advantageously comprises a step additional stabilization S2 consisting of orthogonalisatonon, possibly progressive, the filtering matrix by projection of components on the space of principal components.

This stabilization step 52 is performed on the principle of the principal component analysis method called ACP.

More precisely, this stabilization step S2 consists of increasing the difference (preferably the minimum angle) between

the space of distribution of the artifacts and

the space of distribution of the brain sources, until the latter is above a threshold.

To do this, this stabilization step S2 consists of an orthogonal projection of the vector subspace

on the main components of

so to reduce the gap between the true spatial distribution of secondary sources and its estimate for most of the signal.

So, in order to get the filter matrix after having determined by

step SO, we perform the analysis of the principal components ACP on the window

of X.

More precisely, we calculate the covariance matrix of this window,

then we diagonalize it is the diagonal material of the values

clean sorted in descending order and P _t the passage matrix.

We denote the matrix formed by the first k columns of the matrix

of passage

To make this projection, it is necessary to define a base of

named and project it orthogonally

Once this step has been performed, a base of the orthogonal space is again computed and the linear regression step Sl is carried out with this new one.

base in order to construct the filter matrix of the first filter according to the formula:

Advantageously, stabilization step S2 is repeated by decreasing k until a difference (preferably minimum angi) between the two is reached. subspaces vectorieis is greater than a threshold, for example an angle

greater than 16 °,

This stabilization is particularly useful for filtering ocular artifacts, since low frequency signals make it difficult to estimate the covariance matrix.

In a second substantially equivalent embodiment, in order to obtain the filter matrix F _ir and after having determined, the analysis is carried out

in principal components ACP on the window Pf of X.

For this purpose, we calculate the covariance matrix of this window ₁ then

we diagonalize it where _t is the diagonal matrix of the eigenvalues

sorted in descending order and P _t the passage matrix.

Then, a step of projecting orthogonally on

the k first components of the principal components analysis ACP (Preferentially k is initially set at m and gradually decreases).

What is next?

the matrix formed by the first k columns of P ₁ and the matrix formed by the last nk.

To make this projection, it is also necessary to define a base of named and put it in another base formed by the matrix of

passage P,.

The last nk components are kept as they are, and the k

first

form a new workspace in which linear regression is performed.

So we calculate corresponding to a base of in the base

the first k components of PCA Principal Component Paralysis,

A base of the orthogonal of; the regression

Linear is then performed on the muitican signals (matrix of

covariance and from the components to get the

matrix *

Finally, the filter matrix in the original database is given by

(the components are kept as they are and the linear regression step Sl is applied to the components

Of course, any other type of stabilization method may be used in the present invention.

As far as EEG filtering is concerned, this stabilization is particularly useful for filtering ocular artifacts. Indeed, the low frequency signals make it difficult to estimate the covariance matrix.

Results

Figures 5a and 5c show different multichannel signal recordings for a patient during an epileptic seizure.

Reading these recordings and diagnosing epilepsy is clearly difficult for a neurologist in both cases; these recordings being particularly artefacted.

The use of the filter obtained by the process according to the present invention makes it possible to obtain filtered signals which have good results; these results being displayed on the M3 display means.

On these filtered signals, most of the artifacts are removed while the rhythm Δ is kept up to the position of the focus. We can also discern the spread of the crisis.

The results obtained with such a method are particularly advantageous:

There is an 80% decrease in the amplitude of ocular and muscular artifacts; while the amplitudes of the cerebral signals are correctly conserved (95% of the rhythms α, 80% of the rhythms θ, 70% of the rhythms Δ, 90% of the points, and 75% of the spikes are preserved).

The results obtained thus allow a reliable use in a clinical environment and facilitate for the neurologist the reading of the EEG electroencephalograph tracks.

Claims

A method of processing at least one decomposed multichannel signal (X) over a plurality of time-frequency windows into a plurality of signals

multichannel said multichannel signal (X) resulting from a linear combination

between at least one main signal (sp) from the measurement of data from a main data source (S) and at least one secondary signal (ss) from the measurement of data from a secondary data source ( B) corresponding to at least one artifact, said method being characterized in that it comprises:

a learning step (SO) comprising a comparative multicomponent analysis (AMC) consisting in particular in a comparison of an index, preferably representative of the power, of said signal (X) on two distinct time-frequency windows;

so as to separate the components of the main signal (sp) from the components of at least a part of the secondary signal (ss) on each of the channels and for said two distinct time-frequency windows

and;

a step of global construction (S1, S2, S3) of a global filter

for eliminating or reducing the secondary signal (ss), comprising a first step of construction (Sl) of at least a first filter allowing

retain in said multichannel signal (X), and on a third time-frequency window

the main signal (sp) and eliminating at least a portion of the secondary signal (ss), said first building step (Sl) consisting in particular of a linear regression of said components of said main signal (sp) or said components of said at least one part of the secondary signal (ss).

2. Treatment method according to claim 1, characterized in that the comparison is carried out, for at least one frequency window, on a window of time at rest and on an artefacted time window, or is performed for the same window of time, on two windows of different frequencies, so as to get the first filter

The treatment method according to claim 2, characterized in that the idle time window and the artefact time window are obtained manually or automatically from a pre-established patient database.

4. Processing method according to any one of the preceding revendeccattons, characterized in that the multicomponent comparative analysis (AMC) of the muiticaπaux signals (X) comprises the steps of:

- calculate the covariance matrices of the multichannel signals and

on both time-frequency windows

- diagonaHser simultaneously the two matrices of covariances:

obtain the vector (v) representing the power ratio of each component between the two time-frequency windows

this vector (v) corresponding to the diagonal of the matrix D.

5. Treatment process according to claim 4, characterized in that the comparative multi-component analysis (AMC) further comprises a step of eliminating from the diagonal matrix D the components which have the largest eigenvalues, for example either by predetermining a threshold value not to be exceeded, ie by predetermining a number of components to be eliminated.

6. Treatment method according to any one of claims 1 to 5, the first step of construction (Sl) being a step of constructing a first and a second filter

characterized in that the global filter

is obtained by a convolution step (S3) of said first and second filters

7. Treatment method according to any one of the preceding claims, characterized in that! further comprises a filtering step (S4) by applying one of said first and second filters, or said global filter

_f on multichannel signals

to multichannel signal (X) for a plurality of windows time-frequency

to obtain filtered signals

for each time-frequency window

8. The treatment method according to claim 1, characterized in that! further comprises a step of reconstructing (S5) at least one signal (XO over the entire period of time (T) by subtracting the multi-channel signals

for each time-frequency window

multi-channel signal (X), and adding the filtered signals

for each time-frequency window

; in order to obtain:

9. A method of treatment according to any one of the preceding claims, characterized in that the overall construction step (S1, S2, S3) further comprises a stabilization step (52) consisting in particular of an orthogonal projection of the components d one of said first and second filters on ia

third time-frequency window

so as to increase the difference between the distribution of the components of the secondary signal (ss) and the components of the main signal (sp).

10. Treatment method according to any one of the preceding claims, characterized in that the multi-channel signal (X) is a signal which contains electro-physiological information,

11. A method of treatment according to any one of the preceding claims, characterized in that the multi-channel signal (X) is the signal of an electroencephaiograph (EEG) which results from a linear combination of at least one main signal (sp). ) derived from the measurement of brain activity, including normal brain activity and pathological brain activity, and at least one secondary signal (ss) from the measurement from at least one artifact of the type mechanical, ocular, muscular artifacts.

12. The method as claimed in claim 1, for improving the reading of the electroencephalograph (EEG) pathways, and in particular for detecting neurological pathology of the epilepsy type.

13. Computer program (PG) comprising instructions adapted for performing the steps of the method of treatment according to any one of claims 1 to 12 when said program is executed by a computer.

A computer-readable recording medium on which a computer program (PG) is recorded including instructions for performing the steps of the processing method according to any one of claims 1 to 12.

15. Medical device comprising means adapted for implementing the treatment method according to any one of claims 1 to 12.