Direct neural interface improved by model development

Abstract

The present description proposes a determination of a direct neural interface signal by detecting an evoked potential in an electroencephalogram (EEG) signal of a user in response to a stimulus signal emission intended for the user, comprising: - obtaining electroencephalogram signals (Su(EEG)) in response to successive emissions of respective stimulus signals (ST1, ST2, ST3, ST4), - obtaining magnetic resonance signals (Su(MRI)) in response to successive emissions of said respective stimulus signals (ST1, ST2, ST3, ST4), - implementing artificial intelligence to distinguish between them the electroencephalogram signals (Su(EEG)) obtained, by learning from observations of differences between the magnetic resonance signals (Su(MRI)) obtained in response to the same respective stimulus signals (ST1, ST2, ST3, ST4),and - using said artificial intelligence to identify a current stimulus signal from a current electroencephalogram signal, obtained in response to the emission of said current stimulus signal. Abstract figure: Figure 2,

Description

Title of the invention: Direct neural interface improved by developing a model Technical field

[0001] The present disclosure relates to the field of detection of potentials evoked by direct neural interfaces (or BCI hereinafter, for “Brain Computer Interface”). Prior art

[0002] As an example below, the evoked potentials can be auditory, or alternatively, visual, tactile, etc.

[0003] An evoked potential is a signal that appears in electroencephalogram signals (EEG hereinafter) when a user is subjected to a stimulus. For example, in the SSAEP approach (for "Steady-State Auditory Evoked Potential"), the stimulus is auditory and consists of a sound having a carrier frequency and to which a sinusoidal amplitude modulation is applied. Equipment such as a headset with EEG signal sensors makes it possible to measure the EEG signals and to find the modulation frequency in these signals.

[0004] An example of a stimulus used for an SSAEP application may be a sinusoidal function at a frequency audible to all (e.g., 20Hz to 20kHz), and modulated by a second sine wave at another frequency, e.g., between 1 and 200 Hz. The high-frequency sine wave constitutes the carrier and allows a user to simply hear the “beeps” of the low-frequency modulating signal. It is the modulating signal that constitutes the information of interest.

[0005] Indeed, it is possible to find in the auditory areas of the brain the stimulation frequencies between approximately 1 and 200 Hz. However, at this frequency range of 1 to 200 Hz, these are low and can be difficult to reproduce and / or to hear. This is why a carrier frequency is used as a vector, allowing the user to hear these frequencies more easily. A stimulus (low-frequency modulating sine of type 1 to 200 Hz) can then generate an evoked, auditory and measurable potential in the brain at the same low frequency of 1 to 200 Hz (plus any harmonics). This frequency is measurable in the EEG signals of the person subjected to the stimulus. We can therefore find in the user's EEG signals the frequency of the modulation that was carried out at low frequency (here 200 Hz).

[0006] From then on, it is possible to construct an interface by directly searching for the interface signal in the user's brain waves. The principle of such an interface called "BCI" (for "Brain Computer Interface") is recalled in [Fig. 1]. A UT user wears a headset with electroencephalogram signal sensors (or "EEG signals" hereinafter), this headset being able to be part of a general BCI interface. Different stimuli (sound, visual, or others), having respective modulation frequencies, are delivered to the user, who concentrates on one of the stimuli. The same modulation frequency as that of the stimulus on which the user has concentrated can then be measured in his EEG signals. In the example of a stimulus delivered by a light source, the latter can flash at the aforementioned modulation frequency. In the case of a sound signal, the signal can include a sinusoidal modulation or can consist of successive beeps at the aforementioned modulation frequency. An analysis of the user's EEG signals reveals the presence of a frequency corresponding to the aforementioned modulation frequency.It is therefore possible to design a BCI interface, according to which several stimuli from respective sources HP1, HP2, HP3, having different modulation frequencies and representing for example different respective instructions for controlling a machine (for example "turn left", "turn right", "brake"), are presented simultaneously to a user and the latter concentrates on one of them so that his EEG signals reveal the frequency of one of the stimuli HPi and the function associated with this stimulus is then executed by the machine.

[0007] In the example of [Fig.l], the stimuli are delivered by respective loudspeakers HP1, HP2, HP3, etc., and the user focuses on one of these sound sources. Thus, an embodiment is described below in which the interface signal (stimulus applied to the user UT) is, for example, an audio signal for detecting a physiological signal of the SSAEP type (for “Steady-State Auditory Evoked Potential”). The stimulus is therefore auditory and consists of a sound at a carrier frequency to which a sinusoidal or square-wave modulation is applied, at a modulation frequency lower than the carrier. The stimuli are generally constructed with a carrier frequency, and modulated in amplitude, power, or energy by a signal having a second, lower frequency. If the stimulation signal is effective, the EEG signal measured by the headset of the BCI interface has a frequency corresponding to the modulation frequency.

[0008] In the case of a visual stimulus, a light source may have a particular color, thus corresponding to a wavelength LO associated with a carrier frequency fp by a relationship of the type LO = c / fp, where c is the speed of light. Some users may be more sensitive to certain colors than others to retain their attention. Thus, the choice of color (and therefore of the carrier frequency fp) may be important, particularly at certain times of the day for the same user, to effectively collect EEG signals revealing a potential evoked usable (i.e. having a frequency corresponding to the modulation or "flashing" frequency of the light source). For example, a wavelength of 0.5pm can be used for blue as a visual stimuli color in some subjects, or a wavelength of 0.55pm for green to which some users are more sensitive, or 0.65pm for red, a color to which other users are even more sensitive, etc.

[0009] Appropriately choosing the stimulus signals can indeed make it possible to collect EEG signals in which the modulation frequency can be found with a degree of certainty greater than a threshold.

[0010] However, in general, the EEG signals obtained are very noisy and the exact identification of a modulation frequency, in particular among several possible modulation frequencies, can be difficult. Summary

[0011] The present description improves the situation.

[0012] A method is provided for determining a direct neural interface signal by detecting an evoked potential in an electroencephalogram signal of a user in response to a stimulus signal emission intended for the user, comprising: - obtain electroencephalogram signals in reaction to successive emissions of respective stimulus signals, - obtaining magnetic resonance signals in response to successive emissions of said respective stimulus signals, - implementing artificial intelligence to distinguish between the electroencephalogram signals obtained, by learning from the magnetic resonance signals obtained in reaction to the same respective stimulus signals, and - using said artificial intelligence to identify a current stimulus signal from a current electroencephalogram signal, obtained in reaction to the emission of said current stimulus signal.

[0013] This learning on magnetic resonance signals (or "MRI" hereinafter) may typically include an observation of differences between the different MRI signals obtained for different stimuli, whereas such differences do not appear significant in the electroencephalogram signals (or "EEG" hereinafter) obtained for the same stimuli. Thus, this learning from the MRI signals makes it possible to reinforce the discrimination of the stimuli which may have caused measured EEG signals, in the subsequent phase of use of artificial intelligence.

[0014] The aforementioned current stimulus signal may be one of said respective stimulus signals (used for training artificial intelligence). However, this embodiment is not essential and may remain optional to the extent that Artificial intelligence, having been built by learning on a large number of different stimuli, can prove sufficiently robust during its operational phase to identify a common stimulus which does not necessarily appear among the stimuli which were used for its learning.

[0015] In one embodiment, the implementation of artificial intelligence may comprise the development of a first model capable of identifying a current stimulus signal from a current electroencephalogram signal, obtained in reaction to the emission of this current stimulus signal.

[0016] The implementation of this first model then makes it possible, in a common situation where a stimulus is presented to a user, to determine the stimulus which was presented to the user by measuring his EEG signals, even though these EEG signals could be highly noisy, and this thanks to learning on the MRI signals.

[0017] In such an embodiment, the development of the aforementioned first model may comprise the development of a second model (prior to the development of the first model), this second model being capable of identifying a given stimulus signal from among said respective stimulus signals, from a magnetic resonance signal obtained in reaction to the emission of one of said respective stimulus signals.

[0018] In such an embodiment, the first aforementioned model is obtained from the second model, for example by transfer learning.

[0019] In one embodiment, the aforementioned electroencephalogram signals and magnetic resonance signals are obtained in response to the emissions of said respective stimulus signals intended for at least one user.

[0020] In such an embodiment, the first aforementioned model can for example be developed for this same user.

[0021] In such an embodiment, the electroencephalogram signals and the magnetic resonance signals can be obtained according to one of the following methods: - simultaneously on the said same user (for example by collecting EEG signals through a suitable headset while the user is in an MRI cabin); - separately on the same user (a first sequence by MRI for example, and a following sequence by EEG).

[0022] In one embodiment, the electroencephalogram signals and the magnetic resonance signals are obtained in response to the emissions of said respective stimulus signals intended for a plurality of categories of users, the implementation of the artificial intelligence comprising the development of a plurality of respective first models and one of these first models is chosen for a current user in order to identify a current stimulus signal from a current electroencephalogram signal, obtained on the current user.

[0023] In such an embodiment, only a few categories of individuals (classified by example by age, or other criteria) can be used to build, for example, a model by category (and thus be subject to MRI and EEG tests in the cabin, or separately). Then, it is possible to choose one of these models or a weighting of these models to obtain a first model best suited to a common user.

[0024] Thus, in such an embodiment, the first chosen model, mentioned above, can for example be selected from a plurality of first pre-recorded models as best suited to the current user.

[0025] In such an embodiment, the method may further comprise an adaptation of the chosen model to the current user, by learning based on an analysis of electroencephalogram signals obtained in reaction to stimuli applied to the current user.

[0026] The method may further comprise an adaptation of the chosen model to the current user, by learning based on an analysis of magnetic resonance signals obtained in reaction to stimuli applied to the current user.

[0027] Certainly in this embodiment, the current user makes a passage in the MRI cabin to obtain some MRI signals to reinforce the determination of his first personal model. However, this is only a relatively short MRI sequence compared to the MRI measurements allowing the development of a first model specific to a category of users.

[0028] The present description also relates to a computer program comprising instructions for implementing the above method, when they are executed by a processor. According to another aspect, there is provided a non-transitory recording medium, readable by a computer, on which such a program is recorded.

[0029] The present description also relates to a neural signal detection model generator comprising a processing circuit for implementing the above method.

[0030] In particular, this generator can create the aforementioned second model (based on MRI measurements), then the first model (based on EEG measurements), for example by transfer learning of the second model, and more generally by learning artificial intelligence. Nevertheless, it can still implement the inference of the first model (exploitation of artificial intelligence) in order to possibly refine the first model, if necessary.

[0031] In one embodiment, this model generator may comprise, for example, a server configured to construct and store a plurality of models capable of identifying a current stimulus signal from a current electroencephalogram signal, obtained in response to the emission of this current stimulus signal.

[0032] Thus, this server can store a base of first models from which a selection or a weighting can be carried out to develop a model suitable for a user. given reader.

[0033] The present description also relates to a device for controlling a direct neural interface by using the artificial intelligence implemented in the above method.

[0034] Such a driving device may comprise a processing circuit configured to drive the direct neural interface (BCI), and typically interpret the collected EEG signals in order to identify (or attempt to identify) a stimulus that may have been presented to a current user.

[0035] In particular, this control device may, in one embodiment, be further configured to test a plurality of models capable of identifying a current stimulus signal from a current electroencephalogram signal, obtained in reaction to the emission of this current stimulus signal on a current user, and to select one of these tested models as best suited to the current user (or weight these tested models to construct a model best suited to the current user). Brief description of the drawings

[0036] Other characteristics, details and advantages will appear on reading the detailed description below, and on analyzing the attached drawings, in which: Fig.l

[0037] [Fig.l] schematically illustrates the principle of a BCI interface according to one embodiment. Fig. 2

[0038] [Fig.2] illustrates the combined use of MRI and EEG signals to construct a model of stimulus identification from commonly acquired EEG signals. Fig. 3

[0039] [Fig.3] illustrates an example of steps that a method of the above type may include, according to one embodiment. Fig. 4

[0040] [Fig.4] illustrates an example of an embodiment of a computer device for controlling a BCI interface and implementing an inference phase of a model of the aforementioned type, according to one embodiment. Fig. 5

[0041] [Fig.5] illustrates an exemplary embodiment of a computer device for implementing implements a learning phase to construct a model of the aforementioned type, according to one embodiment. Fig. 6

[0042] [Fig.6] illustrates an example of a detail of the realization of step S10 of [Fig.3], according to a possible embodiment to refine the model intended for a common user of a BCI interface. Description of the embodiments

[0043] For the purposes of an interface such as described above with reference to [Fig. 1], current brain-machine interaction devices (so-called "BCI" for "Brain Computer Interface") mainly use the technique of analyzing electroencephalogram signals or "EEG" hereinafter. Indeed, these EEG signals can be collected by sensors distributed on a headset such as shown in [Fig. 1], and typically connected to an analysis circuit which can include a computer or a tablet (not shown). Such a computer or tablet can, with the sensor headset, constitute a simple, portable and more responsive interface device than other devices for analyzing brain reactions which require heavy equipment such as magnetic resonance imaging devices ("MRI", or "functional MRI").

[0044] An interface based on EEG signal sensors as illustrated in [Fig.l] indeed has the advantage of being portable and easy to implement. However, this technique suffers from a less good spatial resolution than the aforementioned MRI, flRM techniques, or even more invasive techniques using sensors such as EEG electrodes directly implanted in the brain, which are not (today) suitable for a general public target.

[0045] The signal collected by an interface of the type illustrated in [Fig.l] is also often noisy, i.e. having an unfavorable signal-to-noise ratio. The analysis of the collected EEG signals requires the implementation of signal processing methods, in particular assisted by artificial intelligence (in particular by learning, or “Machine Learning”). It may then be planned to carry out learning on identified (or “labeled”, “tagged”) EEG signals and then infer on EEG signals to be classified. The quality of a Machine Learning model depends on the quality of the EEG signals which are by nature very noisy and as mentioned above, suffer from limited spatial resolution.

[0046] The solution proposed by the present description is based on machine learning carried out both on EEG signals and on MRI capture.

[0047] The learning phase can use two types of signals: EEG signals and MRI or functional MRI signals.

[0048] Indeed, it is recalled that the MRI or functional MRI technique makes it possible to measure a parameter linked to blood flow in certain areas of the brain. A technique based on EEG signals makes it possible to measure evoked potentials and therefore electrical activity which requires blood flow of the aforementioned type. Thus, the types of signals measured by MRI and EEG are indeed linked to the same brain activity of the user.

[0049] The MRI signals are then used initially to learn a first model. In a second step, this first model is adapted, by transfer learning, in order to use only the EEG signals as input, but keeping the same purpose as the first model (based on MRI). Such an implementation makes it possible to obtain a new model using EEG signals as input but having retained the learning based on the MRI signals and therefore the advantages linked to the MRI data in the weights of the model.It is recalled that Transfer Learning (also translated into French as "Learning by transfer") aims to transfer knowledge from one or more source tasks to one or more target tasks, and allows a computer entity to recognize and apply knowledge and skills, learned from previous tasks (for example here analysis of MRI signals), for new tasks in areas sharing similarities (typically here the analysis of EEG signals, for the same stimuli).

[0050] During the learning phase on the EEG and MRI signals, it is possible to teach a neural network to make a difference (even barely perceptible) between the EEG signals in reaction to two stimuli by relying on the measurements of MRI signals (which are less noisy than the usual EEG signals). Thus, in the exploitation (or "inference") phase, a difference, even barely perceptible, between the EEG signals for two different stimuli allows the neural network to distinguish the EEG signals obtained in reaction to these two stimuli, the neural network having been effectively trained to make this difference thanks to the support of the MRI technique.

[0051] This situation has been illustrated very schematically in [Fig. 2]. In part A of [Fig. 2] typically, two stimuli STI and ST2 cause the generation of respective EEG signals from which a useful parameter Su(EEG) is derived (for example a frequency in the evoked potentials, or other). The values ​​of this parameter can be sufficiently different despite the noise present in the collected EEG signals, to be able to distinguish the two stimuli STI and ST2 which have been generated because for example the respective error bars linked to the determinations of these values ​​are well separated. On the other hand, the values ​​of this parameter are not sufficiently distinct because of the noise present in the collected EEG signals, to be able to distinguish the two stimuli ST3 and ST4 which have been generated because in this example the respective error bars linked to the determinations of these values ​​partially overlap.Thus, EEG detection (which can be used by a conventional direct neural interface BCI) is not sufficient here on its own to distinguish the two stimuli ST3 and ST4 which were generated, in the illustrated example.

[0052] It is then proposed to rely here on a useful parameter Su(MRI) which can be taken from the measurements of MRI or functional MRI signals, to help an artificial intelligence to correctly distinguish the stimuli (notably ST3, ST4) which have been generated, even though the difference which can be taken from the EEG signals is practically imperceptible due to the noise present in the EEG signals. In the example illustrated, the error bars for the detection of this parameter taken from the MRI signals of part B of [Fig.2] are much narrower than in part A because the MRI signals are much less noisy than the EEG signals and better defined spatially.

[0053] An artificial intelligence can then learn to typically discriminate the useful parameters Su(EEG) measured from the EEG signals in reaction to the STI to ST4 stimuli, thanks to the differences observed between the useful parameter values ​​Su(IRM) measured from the MRI signals in reaction to the same STI to ST4 stimuli. Thus, with reference to part C of [Fig.2], in the exploitation phase, the artificial intelligence can, thanks to this learning, distinguish the useful parameter values ​​Su(EEG)+IA taken from the EEG signals and identify the stimulus (for example among STI, ST2, ST3 and ST4) which caused a given value of this useful parameter.

[0054] In the illustrated example, in the exploitation phase, artificial intelligence makes it possible to recognize the stimulus among the STI to ST4 learning stimuli which causes an EEG signal of measured useful value. Nevertheless, more generally, it is possible to develop a model which learns the correlation between the EEG and MRI signals so that it is not necessary to choose the stimulus to be applied during the exploitation phase among the stimuli which were used for learning.

[0055] It is then possible to develop a new detection model capable of interpreting and discriminating EEG signals, even if they are highly noisy and / or relatively poorly defined spatially. Such a model, improved because it is more robust, makes it possible to carry out the current recognition task based on EEG signals only, even if the latter are noisy.

[0056] To obtain such an improved model, it is possible to use an equivalent learning protocol for both MRI and EEG techniques consisting of emitting identical stimulus signals for the same user or for the same group of users (for example a series of sounds with beeps at given frequencies, or images, or a request to imagine an image, or others). In response to these emissions, EEG and MRI signals are obtained and recorded with the two detection methods (EEG and MRI) implemented for example separately (for example MRI, then EEG) according to a first embodiment. Each user is thus the subject of MRI measurements in the cabin, then of EEG measurements by an EEG headset, from the same stimuli. This first embodiment offers the advantage of avoiding interference effects linked to the specificities of the two techniques (for example, the use of an EEG headset of the type shown in [Fig.l], in an MRI cabin). Nevertheless, the disturbances generated by the magnetic field linked to the use of the MRI technique on the EEG signal sensors can today be controlled and compensated by techniques known per se.

[0057] Alternatively, in a second embodiment, simultaneous recording of the signals from both MRI and EEG techniques (in an MRI cabin and with an EEG headset) can be provided. Such an implementation makes it possible to obtain better synchronization, with identical test conditions. Such an implementation requires placing a user equipped with an EEG headset (with a number and positioning of electrodes in relation to the area of ​​the brain a priori concerned by the task to be learned, or a headset with electrodes distributed uniformly on the skull to learn a set of tasks of different natures: audio, visual, touch, or others), and this in a closed MRI cabin.

[0058] The inference phase is then carried out using only the EEG signals.

[0059] These different steps have been illustrated by way of example in [Fig. 3]. A first step SI consists of placing one or more users of a first group UT1 in an MRI cabin and applying successive stimuli ST to them, to collect MRI signals which are processed and analyzed in step S2 (for example to deduce successive values ​​of useful parameters Su(MRI)). During a following step S3, a model MOD(MRI) is constructed by learning, capable of identifying stimuli which have been applied to the users UT1, as a function of the MRI signals which can be measured.

[0060] It may be carried out simultaneously with step S1 or subsequently with this step S1, step S4 which consists of now measuring the EEG electroencephalogram signals for a user or a group of users UT2, in reaction to the same ST stimuli which are applied to this or these users UT2. The group of users UT2 may be identical or different from the group of users UT1, since the morphological signals measured by MRI or by EEG electroencephalogram are caused by the same ST stimuli. The collected EEG signals are analyzed and processed in step S5 to obtain, by learning, a MOD(EEG) model capable of identifying a stimulus which has been applied, simply from EEG signal measurements, by Transfer Learning in step S6.In step S7, data from this last MOD(EEG) model can then be stored in a memory MEM, for example of a server SER to which a computing device of a BCI interface can subsequently connect for an inference phase as described below.

[0061] In an inference phase in step S8, to operate a BCI interface from STi stimuli, a user UT3 (who is not necessarily among the users readers UT1 and / or UT2) receives a current stimulus STi and a corresponding EEGi signal is captured by the BCI headset. In step S9, the data of this EEGi signal can be transmitted to the aforementioned server SER (via a wide area network, or "the cloud"). Alternatively, they can be processed locally by a computing device of the BCI interface. In step S10, the data of this current EEGi signal are used by the MOD(EEG) model (via the cloud or locally) to determine, in step S11, which stimulus STi caused this current EEGi signal, and thus operate the BCI interface in step S12.

[0062] An embodiment is now described with reference to [Fig. 6] in which, in the aforementioned step S10, it is possible to choose a model best suited to a current user. Thus, during a first step S100, the aforementioned server SER can store a plurality of models M0D1, M0D2, ..., MODN obtained for respective categories of users in steps S1 to S7 of [Fig. 3]. These models can be tested successively on the current user in step 101, to search for the model MODj capable of interpreting the EEG signals of the current user as best corresponding to the ST stimuli applied to the current user. Once this model MODj has been thus selected in step S102, tests can be applied in step S103 to determine whether this model MODj must be further improved by additional learning specific to the current user.For example, at this step S103, an inference of this MODj model can be based on the set of ST stimuli applied at step S4 of [Fig.3], but here on the current user to determine whether the MODj model manages to identify, via the analysis of the EEG signals from the current user, a part above a threshold of all the stimuli applicable at step S4.

[0063] If this is the case (OK arrow at the output of test S103), then the MODj model is satisfactory and retained as such in step S102. On the other hand, if test S103 is negative (KO arrow) for example because a number of stimuli greater than the aforementioned threshold could not be identified by the sole analysis of the EEG signals of this current user), then additional learning is carried out in step S104 to improve the MODj model. This additional learning can be based on the EEG signals to be recognized, following the emission of particular stimuli, applied to this current user. If the learning of step S104 is not sufficient to improve the MODj model and MRI signals are still necessary for the development of a robust model specific to this current user, it may be provided, in step S105 (optional), to record: - a series of MRI signals on the current user (in the MRI cabin) in reaction to stimuli applied to the current user as in step SI of [Fig.3], - build a model based on these MRI signals as in step S3 of [Fig.3], and - apply transfer learning based on EEG signals as in step S6 of [Fig.3], to refine the MODj model, specific to the current user.

[0064] It should be noted that the passage in the MRI cabin of the current user in this new step S105 can be shorter than in step S1 of [Fig.3] because only a limited series of stimuli to be correlated with the observed MRI signals can be applied to adjust the MODj model specific to the current user.

[0065] [Fig.4] schematically illustrates a DIS-BCI device for controlling a BCI interface, comprising: - an INT1 input interface for receiving EEGi current signals, - a memory MEM storing data of the MOD(EEG) model constructed by implementing the above method, and instruction data of a first computer program for using the data of the MOD(EEG) model to identify a stimulus STi which caused a current signal EEGi, and possibly to select this MOD(EEG) model from a plurality of predefined models and possibly adjust the selected model to specificities of a user of the BCI interface, - a processor PROC capable of accessing the memory MEM (which can be local in the DIS-BCI device or remote in the case of a remote server SER) to read the data of the model and execute the instructions of the first computer program to identify the stimulus STi which caused the current signal EEGi, during an inference phase of the model MOD(EEG), and - an INT2 output interface delivering a signal corresponding to the identified stimulus STi in order to control the BCI interface according to the identified stimulus.

[0066] Reference is now made to [Fig. 5] to describe a device configured to implement at least the learning steps presented above. In the illustrated example, this is a DIS-SER server device as presented previously with reference to [Fig. 3]. However, as a variant, it could be the same DIS-BCI device as that illustrated in [Fig. 4]. In the illustrated example, this DIS-SER device comprises: - an input interface IN capable of receiving EEG signals and MRI signals, - a memory MEM' configured to store instructions of a second computer program for the construction of at least one inference model MOD(EEG) based on EEG signals only, and further store the data of this model, - a processor PROC' capable of accessing this memory MEM' to read the instructions of the second computer program and construct at least the aforementioned model MOD(EEG), and - an OUT output interface capable of transferring data from one or more MOD(EEG) models on request (as illustrated in [Fig.6] in step S100), or generating a BCI interface signal when an EEG signal is applied to the IN input of the device, by using the MOD(EEG) model, in particular when it is the same device as that illustrated in [Fig.4].

[0067] The subject of the present description can find a multiplicity of applications and can be integrated into widely distributed headsets such as virtual or augmented reality headsets, audio headsets, or even protective helmets for driving machines (motorcycles, airplanes, etc.), in order to add a modality of interactions, check and / or monitor the mental state of users, or other. Improving the performance of recognizing EEG signals is therefore important, particularly for these use cases.

Claims

Claims

1. Method for determining a direct neural interface (BCI) signal by detecting an evoked potential in an electroencephalogram (EEG) signal of a user (UT) in response to a stimulus signal emission intended for the user, comprising: - obtaining electroencephalogram (EEG) signals in response to successive emissions of respective stimulus signals (ST), - obtaining magnetic resonance (MRI) signals in response to successive emissions of said respective stimulus signals (ST), - implementing artificial intelligence to distinguish between them the electroencephalogram (EEG) signals obtained, by learning from the magnetic resonance (MRI) signals obtained in response to the same respective stimulus signals (ST), and - using said artificial intelligence to identify a current stimulus signal (STi) from a current electroencephalogram (EEGi) signal,obtained in response to the emission of said current stimulus signal (STi).,

2. Method according to claim 1, in which the implementation of artificial intelligence comprises the development of a first model (MOD(EEG)) capable of identifying a current stimulus signal (STi) from a current electroencephalogram signal (EEGi), obtained in reaction to the emission of said current stimulus signal (STi).

3. Method according to claim 2, in which the development of the first model (MOD(EEG)) comprises the development of a second model (MOD(IRM)) capable of identifying a given stimulus signal (ST) among said respective stimulus signals (ST), from a magnetic resonance signal (MRI) obtained in reaction to the emission of one of said respective stimulus signals (ST).

4. Method according to claim 3, wherein the first model (MOD(EEG)) is obtained from the second model (MOD(IRM)) by transfer learning.

5. Method according to one of the preceding claims, in which the electroencephalogram (EEG) signals and the magnetic resonance (MRI) signals are obtained in response to the emissions of said respective stimulus signals (ST) intended for at least one user.

6. The method of claim 5, wherein the electroencephalogram signals- phalograms (EEG) and magnetic resonance imaging (MRI) signals are obtained according to one of the following methods: - simultaneously on the same user; - separately on the same user.

7. Method according to one of claims 2 to 6, in which the electroencephalogram (EEG) signals and the magnetic resonance (MRI) signals are obtained in response to the emissions of said respective stimulus signals (ST) intended for a plurality of categories of users, the implementation of the artificial intelligence comprising the development of a plurality of respective first models (MOD(EEG)), and one of said first models is chosen for a current user in order to identify a current stimulus signal (STi) from a current electroencephalogram (EEGi) signal, obtained on the current user.

8. The method of claim 7, wherein the first selected model (MODj) is selected from a plurality of pre-stored first models (S 100) as best suited to the current user.

9. Method according to one of claims 7 and 8, further comprising an adaptation of the chosen model to the current user, by learning based on an analysis (S 104) of electroencephalogram (EEG) signals obtained in reaction to stimuli applied to the current user.

10. Method according to one of claims 7 to 9, further comprising an adaptation of the chosen model to the current user, by learning based on an analysis (S 105) of magnetic resonance (MRI) signals obtained in response to stimuli applied to the current user.

11. Computer program comprising instructions for implementing the method according to one of the preceding claims, when executed by a processor.

12. Neural signal detection model generator comprising a processing circuit for implementing the method according to one of claims 1 to 10.

13. Model generator according to claim 12, comprising a server (DIS-SER) configured to construct and store a plurality of models capable of identifying a current stimulus signal (STi) from a current electroencephalogram signal (EEGi), obtained in reaction to the emission of said current stimulus signal (Sti).

14. Direct neural interface (DIS-BCI) control device using artificial intelligence implemented in the method according to one of claims 1 to 10.

15. A driving device according to claim 14, further configured to test a plurality of models capable of identifying a current stimulus signal (STi) from a current electroencephalogram (EEGi) signal, obtained in response to the emission of said current stimulus signal (STi) on a current user, and to select one of said tested models as best suited to the current user.

Images