• 专利附录
  • 专利说明
  • 权利要求
  • 类似技术
  • 同族专利
  • 引用文献
  • 法律状态
  • 优先权
    • 专利摘要:
    The present invention provides a method of continuous motion decoding for a direct neural interface. The decoding method estimates a movement variable from an observation variable obtained by a time-frequency transformation of the neural signals. The observation variable is modeled by an HMM model (140) whose hidden states include at least one active state and one quiescent state. The movement variable is estimated by a Markovian mixture of experts (120), each expert being associated with a state of the model. For a sequence of observation vectors, the probability of the model being in a given state is estimated and a weighting coefficient (150) is deduced for the prediction generated by the expert associated with this state. The movement variable is then estimated by combining (130) the estimates of the different experts using these weights.
    公开号:FR3053495A1
    申请号:FR1656246
    申请日:2016-06-30
    公开日:2018-01-05
    发明作者:Marie-Caroline SCHAEFFER；Tetiana Aksenova
    申请人:Commissariat a lEnergie Atomique CEA；Commissariat a lEnergie Atomique et aux Energies Alternatives CEA；
    IPC主号:
    专利说明:

    Holder (s): COMMISSIONER OF ATOMIC ENERGY AND ALTERNATIVE ENERGIES Public establishment.
    Extension request (s)
    Agent (s): BREVALEX.
    104) DIRECT NEURONAL INTERFACE WITH EXPERT DECODING.
    CONTINUOUS USING A MARKOVIAN MIXTURE
    FR 3 053 495 - A1 (£ /) The invention relates to a method for continuous decoding of movement for a direct neural interface. The decoding method estimates a movement variable from an observation variable obtained by a time-frequency transformation of the neural signals. The observation variable is modeled by an HMM model (140) whose hidden states include at least one active state and one rest state. The movement variable is estimated by a Markovian mixture of experts (120), each expert being associated with a state of the model. For a sequence of observation vectors, we estimate the probability that the model is in a given state and we deduce a weighting coefficient (150) for the prediction generated by the expert associated with this state. The movement variable is then estimated by combining (130) the estimates of the different experts using these weights.

    i
    CONTINUOUSLY DECODING DIRECT NEURONAL INTERFACE USING A MARKOVIAN MIX OF EXPERTS
    DESCRIPTION
    TECHNICAL AREA
    The object of the present invention relates to the field of direct neural interfaces also called BCI (Brain Computer Interface) or BMI (Brain Machine Interface). It is particularly applicable to the direct neural control of a machine, such as an exoskeleton or a computer.
    PRIOR STATE OF THE ART
    Direct neural interfaces use the electro-physiological signals emitted by the cerebral cortex to develop a control signal. These neural interfaces have been the subject of much research, in particular with the aim of restoring motor function to a paraplegic or quadriplegic subject using a prosthesis or a motorized orthosis.
    Neural interfaces can be invasive or non-invasive. Invasive neural interfaces use intracortical electrodes (that is to say implanted in the cortex) or cortical electrodes (arranged on the surface of the cortex) collecting in the latter case electrocorticographic signals (ECoG). Non-invasive neural interfaces use electrodes placed on the scalp to collect electroencephalographic (EEG) signals. Other types of sensors have also been envisaged, such as magnetic sensors measuring the magnetic fields induced by the electrical activity of neurons in the brain. We then speak of magnetoencephalographic signals (MEG).
    Direct neural interfaces advantageously use ECoG type signals, which have the advantage of a good compromise between biocompatibility (matrix of electrodes implanted on the surface of the cortex) and quality of the signals collected.
    The ECoG signals thus measured must be processed in order to estimate the trajectory of the movement desired by the subject and to deduce therefrom the control signals from the computer or the machine. For example, when it comes to controlling an exoskeleton, the BCI interface estimates the trajectory of the desired movement from the electro-physiological signals measured and deduces the control signals allowing the exoskeleton to reproduce the trajectory in question. . Similarly, when it comes to controlling a computer, the BCI interface estimates, for example, the desired trajectory of a pointer or a cursor from the electro-physiological signals and deduces therefrom the cursor control signals. / pointer.
    The trajectory estimation, and more precisely that of the kinematic parameters (position, speed, acceleration) is still called neuronal decoding in the literature. Neural decoding makes it possible in particular to control a movement (of a prosthesis or a cursor) from ECoG signals.
    When the ECoG signals are acquired continuously, one of the main difficulties in decoding resides in the asynchronous nature of the command, in other words in the discrimination of the phases during which the subject actually controls a movement (active periods) of the phases in which there is n '' not on order (rest periods).
    To circumvent this difficulty, direct neural interfaces, called synchronous, give the possibility of controlling a movement only during well-defined time windows (for example intervals of time successively periodically), signaled to the subject by an external index. The subject can then control the movement only during these time windows, which is prohibitive in most practical applications.
    More recently, direct neural interfaces with continuous decoding have been proposed in the literature. The article by M. Velliste et ol. entitled "Motor cortical correlates of arm resting in the context of a reaching task and implications for prosthetic control" published in The Journal of Neuroscience, April 23, 2014, 34 (17), pp. 6011-6022, describes in particular a direct neural interface in which the periods of rest (idle State) are detected by LDA (Linear Discriminant Analysis) from the frequency of emission of the action potentials (neuron firing rate). The kinematic parameters are estimated during the active periods by means of a state transition model, the states being predicted by means of a so-called Laplace-Gauss filter. However, the results were obtained for microelectrode arrays and could not be reproduced for conventional cortical electrodes. In addition, the switching between the rest periods and the active periods result in discontinuities in the decoding of the movement and therefore jolts along the trajectory.
    Another approach was proposed in the article by L. Srinivasam et ol. entitled “General-purpose filter design for neural prosthetic devices” published in J. Neurophysiol. Vol 98, pp. 2456-2475, 2007. This article describes a direct neural interface with continuous decoding in which the sequence of active and resting states is generated by a 1st order Markovian model. A switched Kalman filter or SKF (Switching Kalman Filter) whose observation matrices and transition matrices depend on the hidden switching variable (active state / rest state) allows the kinematic parameters of the movement to be estimated. This type of direct neural interface has been used to control a wheelchair from EEG signals on the simulation database but has not been tested for ECoG signals. In addition, the state detection (active state / rest state) is sometimes erroneous (high rate of false positives and false negatives).
    Finally, the aforementioned direct neural interfaces have been developed to decode the movement of a single member of a subject. They are therefore not suitable for controlling an exoskeleton with several members, in particular for a quadriplegic, paraplegic or hemiplegic subject, which appreciably limits their field of application.
    The aim of the present invention is therefore to propose a method of continuous decoding of movement for direct neural interface, which has a low rate of false detection of active state and which allows a decoding without discontinuity of the movement from the ECoG signals d 'a subject. Another object of the present invention is to allow a decoding of the movement of several members from the ECoG signals of a subject.
    STATEMENT OF THE INVENTION
    The present invention is defined by a continuous motion decoding method for a direct neural interface acquiring neural signals by means of a plurality of electrodes, characterized in that, a control variable, y (z) is estimated. , representing kinematic variables of the movement, by means of a mixture of experts, each expert being associated with one of the hidden states of an HMM model modeling an observation variable represented by an observation vector, x (r) , the HMIVI model comprising at least one active state and one rest state, said method comprising:
    - a calibration phase to estimate the parameters of the HMM model and of a linear prediction parameter of each expert;
    - a preprocessing of neural signals, said processing including a time-frequency transformation on a sliding observation window, to obtain a sequence of observation vectors x [l: t]
    - an estimate (230) of the mixing coefficients (, K) of the different experts from the probability of said sequence of observation vectors (P (x [l: z])) and the probability of the states respectively associated with these experts taking into account said sequence (P (z k (t) = l, x M));
    - an estimate by each expert, of said movement variable from the current observation vector (x (r)) and the linear prediction parameter of said expert;
    - a combination of the estimates of the different experts using the mixing coefficients to provide an estimate of the control variable.
    According to an alternative embodiment, the hidden states of the HMM model include, for each element of a plurality of elements respectively controlled by a component of the movement variable, an active state during which said element is moved and a state of rest for which said element is at rest.
    The hidden states of the HMM model may also include, for each of said elements, a state of preparation for movement immediately following a state of rest and immediately preceding an active state of said element.
    Advantageously, during the calibration phase, the parameters of the HMM models and the linear parameters of the experts are obtained from a sequence of observation vectors and a sequence of measurements of the motion variable, according to a method of supervised learning.
    The linear parameters of the experts are obtained by means of a partial least squares regression, or even by means of a generalized partial least squares regression.
    Alternatively, during the calibration phase, the parameters of the HMM models and the linear parameters of the experts are obtained from a sequence of observation vectors and a sequence of measurements of the motion variable, according to a method of semi-supervised learning.
    In this case, the experts' linear parameters can be obtained by means of a expectation-maximization algorithm.
    The time-frequency transformation is advantageously a decomposition into wavelets.
    The mixing coefficient g k (t) associated with the expert (ζ is obtained by Ρ (ζ λ , χ [ΐ: ί]) e, (i) = -———- where P xl: / is the probability observation of the sequence of
    P (x [l: i]) v L observation vectors and P (z i ., X [l: /]) is the probability of the state associated with the expert taking into account the observed sequence, the probabilities p (x [l: /]) and P ^ z k , x [l: /]) being obtained by means of a forword algorithm.
    Advantageously, the estimate ÿ k (i) of the motion variable by the expert is obtained by ÿ k (f) = β Λ χ (ί) where β ζ is a matrix of size NxM where M is the dimension of the vectors of observation and N the dimension of the control variable. Neural signals are typically electro-cortical signals (signals
    ECoG).
    BRIEF DESCRIPTION OF THE DRAWINGS
    Other characteristics and advantages of the invention will appear on reading a preferred embodiment of the invention, with reference to the attached figures among which:
    Fig. 1 schematically represents the principle of a continuous decoding of movement by Markovian mixture of experts used in the present invention;
    Fig. 2 shows in flowchart form the method of continuous decoding of movement for direct neural interface, according to an embodiment of the present invention;
    Fig. 3 represents examples of continuous decoding of movement according to different methods of the prior art and according to the method of FIG. 2.
    DETAILED PRESENTATION OF PARTICULAR EMBODIMENTS
    We will consider in the following a direct neural interface with continuous decoding of movement in the sense defined above. Decoding is continuous in the sense that it is carried out over time without being synchronized with an external stimulus. The interface acquires neural signals, typically ECoG signals, from a matrix of electrodes previously implanted on the cortical surface of a subject. The direct neural interface is used to decode movement from neural signals so as to control the movement of an external element. By the term "external element" is meant, without limitation, all or part of a prosthesis, an orthosis, an exoskeleton, a wheelchair, an effector. We also hear, without limitation, a visual symbol such as a cursor or an icon whose movement is usually controlled by an HMI (Human Machine Interface).
    The neural signals acquired by the interface undergo a preprocessing (detailed below) to provide an observation variable of dimension M dependent on time, denoted x (r) where / e N is a discretized time. This variable represents an observation of the characteristics of the measured neural signals, from which it is possible to estimate the kinematic parameters. By kinematic parameters is meant in particular the position, the speed and / or the acceleration along the trajectory.
    The idea behind the invention is to decode the movement using a Markov mixture of experts. The estimation by Markovian mix of experts or MME (Markov Mixture of Experts) is based on a hidden Markovian model or HMM (Hidden Markov Model), an estimator (called expert) being associated with each hidden state of the model. A presentation of an estimation method by Markovian mixture of experts can be found in the article by M. Meila et al. entitled "Learning fine motion by Markov mixtures of experts" published in Advances in Neural Information Processing Systems 8 (1567), pp. 1003-1009.
    Fig. 1 schematically represents the principle of a continuous decoding of
    0 movement by Markovian mixture of experts.
    The Markovian mixture of experts involves, on the one hand, a hidden state automaton (or HMM model), 140, which can take K possible states and, on the other hand, a plurality K of estimators, 120, each estimator being associated with a hidden state.
    The estimates of the different experts are combined in a module of
    5 combination (gating network), 130, to give an estimate, ÿ (r) of the variable to be explained, y (r), here the kinematic parameters of the movement, from the observation x (r) representing the characteristics of the neural signals at time t, picked up by means of the electrodes 105. In the following, the input variable x (r) is assumed to be of dimension M and the response y (r) is assumed to be of dimension N.
    The mixture of experts assumes that the starting space, i.e. the space X in which the input variable takes its values, can be partitioned into a plurality
    K
    K of regions, that is to say X. To each region 2Ç corresponds a model of * = i prediction of the variable y (r) by the variable x (r) The prediction or expert model is defined by a local function f k of the region 2Ç from the departure space X to the arrival space Y.
    In a mixture of classical experts (ME), we model the dependence of the variable y (r) as a function of x (r) on the whole of the starting space, in other words the variable to be explained is supposed to follow the model:
    γ (0 = Σ ^) Λ ( χ (ϋ) + η <>) (i) * = 1 where z k (t) = 1 when the variable y (r) is generated by the model f k (and z k (t) = 0 otherwise) and where η (ί) is a noise vector of size N assumed here Gaussian.
    The estimate of the variable to be explained can be broken down into the following form:
    y (i) = E (y (f) | x (i)) = y £ (y (i), z, (f) = l | x (f)) = X /> (z t (i) = l | x (f)) E (y (f) | x (i), z, (i) = 1) * = 1 * = 1
    0 (2) In other words:
    ÿ (/) = £ (y (/) | x (/)) = ^ g k (t) ÿ k (t) k = l (3) where y k (t) = p (y (/) | x (i), z k (t) = l) is the estimate made by the expert (ξ and g k (t) = p (z k (t) = l | x (/)), are the coefficients of the The coefficient g k (t) expresses the conditional probability that the variable to be explained has been generated by the model f k (t) knowing that x (i) has been observed. The weighting coefficients (or mixing coefficients) verify the property :
    Σ ^) = 1 (4) * = i
    The mixing coefficients g k (t) are estimated by the module 150 from the sequence of observation vectors as explained below. These mixing coefficients are supplied to the combination module 130.
    The models f k of the experts are advantageously chosen to be linear, in other words / λ (χ (ί)) = β λ χ (ί) where P ^ is a matrix of size NxM. In this case, the estimate by mixture of experts is simply written:
    γ (ο = Σ ^^ χ ^ (5) * = 1
    We now assume that the sequence of observations x [l: r] = {x (l), x (2), ..., x (r)} is modeled by an underlying Markov process, of first order at transmission continues. At each instant t, the hidden state automaton can take K distinct states corresponding respectively to the cases z k (t) = 1, k -Ι,.,., Κ. The state of the automaton at time t depends on its state at the previous time, i-1, (Markovian model of order 1) as well as on the probability of transition between these states. The sequence of hidden states is therefore entirely characterized by the initial probabilities of occupation of the different hidden states and of the matrix of ίο transition probabilities between states. The probability distribution of the observation variable x (r) at time t depends only on the state z (t) of the HMM automaton at this time. The conditional probability Ρ (χ (ί) | ζ (θ) of the observation variable is called emission probability (emission probobility). The probabilities that the automaton is initially in the K different states, the transition probabilities between states and emission probabilities are called HMM model parameters.
    In a Markovian mixture of experts, it is assumed that the coefficients of the mixture depend on the sequence of observations x [l: t], in other words:
    A (i) = /> (z t (0 = l | x [l :(]) (6)
    The coefficients of the mixture can be determined using the rule of
    Bayes:
    Τ '(r) - l | x [l: /]) - (7) where the probabilities p (x [l: i]) and PU (o = i, x [l: i]) can be obtained so recursive by the forword algorithm, from the parameters of the HMM model, in a manner known per se.
    The parameters of the HMM model and the parameters β ζ of the different experts can be estimated during a calibration phase, using a supervised or semi-supervised estimation method.
    A supervised estimation method can be implemented if a complete set of training data is available, namely a sequence of observations (sequence of characteristics of neural signals), x [l: i ], the sequence of kinematic parameters y [l: i], as well as the sequence of the corresponding states
    The HMM and the K experts are then trained independently of each other. For example, the matrix of transition probabilities between states of the HMM automaton can be estimated by counting the frequencies of transitions between states in the learning sequence z [l: /]. Likewise, the emission probabilities of the HMM can be estimated from the sequences z [l: i] and x [l: i]. The expert can be trained on the basis of the relevant subsequences {xj and {yj extracted respectively from x [l: /] and y [l: /] where {x t } = {x T ex [l: ί] | ζ λ (r) = l} and {y t } = | y r gx [i: î] | Z (t (7) = i}. The matrix β ζ can be estimated by linear regression or by linear PLS regression on these subsequences. Alternatively, the matrix β ζ can be obtained by multivariate PLS regression or NPLS (N-way PLS) from these sub-sequences. A detailed description of the NPLS regression method can be found in R. Bro's thesis entitled "Multi-way analysis in the food industry: models, algorithms and applications", 1998. The method Multivariate PLS was applied to the calculation of the predictive model of trajectories in a BCI interface as described in the article by A. Eliseyev and T. Aksenova entitled "Recursive N-Way Partial Least Squares for Brain-Computer Interface" published in PLOS One, July 2013, vol. 8, n ° 7, e69962.
    A semi-supervised estimation method can be implemented when the sequences x [l: î], y [l: f] and z [l: r] are only partially known. In this case, the parameters of the HMM automaton and the experts β ζ matrices can be estimated using an expectation-maximization algorithm.
    Whatever the learning method implemented during the calibration phase, it is important to note that the continuous decoding method is based on a particular definition of the states of the HMM automaton. Advantageously, these states will include a state of rest, an active state and, optionally, a state of preparation for the movement. In the rest state, the subject does not control the external element (for example prosthesis, cursor). Conversely, in the active state, the subject controls the movement of this element. The preparatory state is a state immediately preceding the active state and immediately following the quiescent state, during which the characteristics of the neural signals are neither those of the quiescent state nor those of the active state.
    When the direct neural interface must control the movement of several external elements (for example several members of an exoskeleton), the states of the HMM automaton include the aforementioned states (rest, active and optionally preparatory) for each of these elements. In other words, if the direct neural interface must control the movement of L elements, the number of states will be K = 2 L or 3 L (if the preparatory states are considered).
    The weighting of the predictions of the K experts associated with the different states makes it possible to avoid jolts of movement. Furthermore, it has been shown that decoding by Markovian mixture of experts makes it possible to significantly reduce the rate of false positives (motion command generated while the subject is actually in the rest state) and the rate of false negatives. (movement command absent while the subject actually wants to make a movement).
    Fig. 2 represents in the form of a flowchart the method of a continuous decoding of movement implemented in a direct neural interface according to an embodiment of the present invention.
    In step 210, the HMM model and the experts are calibrated. More precisely, the parameters of the HMM automaton and of the experts β ζ prediction matrices are estimated.
    The neural signals acquired by the interface during the calibration phase are pretreated to obtain the observation variable x (i).
    This preprocessing can include the subtraction of an average taken from all of the measured signals.
    A time-frequency analysis of the signals thus obtained is then carried out. This analysis is carried out on a sliding observation window of width ΔΓ for all the electrodes, the sliding window of δΤ between two successive positions. The time-frequency analysis is advantageously a decomposition in wavelets.
    Thus, at each instant t - ηδΤ is associated a vector x (r) of size M - SPQ where S = ΔΤ / δΤ, P is the number of electrodes and Q is the number of wavelets (or frequency bands) used in decomposition.
    The kinematic parameters of the movement associated with neural activity are acquired in parallel (for example the components of the speed of a point or the coordinates of a point). The position can be determined by means of a motion tracking system and the components of the speed are deduced therefrom. Conversely, acceleration and / or speed can be measured by a sensor (accelerometers) and the position can be deduced from this. In the case of an imagined movement (by a healthy subject) the kinematic parameters are those of the movement instructions supplied to the subject. Finally, in the case of a movement presented to a subject (mirror neuron theory), in particular to a paralyzed subject, the kinematic parameters are those of the movement presented.
    During this learning phase, it is also known how to determine (for example by means of a speed thresholding), the rest states and the active states. Preparatory states can be determined by intervals of predetermined lengths before the start of active states.
    Thus, we have a set of observation data making it possible to estimate the parameters of the HMM model and the experts' prediction matrices using a supervised or semi-supervised learning method as explained above. The mixing coefficients are also estimated from expression (7).
    Finally, the state of the automaton is initialized, for example to a rest state.
    In step 220, the neural signals from the P electrodes are continuously acquired and a preprocessing of these signals is carried out, identical to that described for the calibration phase. We obtain an observation variable (vector) x (r) at a plurality of successive instants.
    In step 230, the mixing coefficients g k (t), k = l, ..., K are estimated iteratively from the past sequence of observation vectors x [l: r] by means of the forword algorithm. The use of the forword algorithm is particularly advantageous here insofar as it makes it possible to calculate the mixing coefficients without lag time, unlike the Viterbi algorithm.
    In step 240, each of the experts T k , k = l, ..., K, performs a linear prediction of the command, ÿ k (f), assuming that the automaton is in the state z k to the instant t, let y k (t).
    In step 250, the predictions of the K experts are made by weighting them with the mixing coefficients, g k (t). This gives an estimate of the control variable, ÿ (r). This command can be used to move an external element such as a prosthesis part or a cursor on a screen for example. In the case where the control variable comprises components relating to distinct elements, each of the elements is controlled by the corresponding component.
    Fig. 3 represents examples of continuous decoding of movement according to different methods of the prior art and according to the method of FIG. 2.
    Different methods have been used to decode the movement of a primate's wrist from ECoG signals during a task of grasping food. ECoG signals are acquired by means of a matrix (of 32 or 64 electrodes) and sampled at a frequency of 1 kHz. The characteristics of the neural signals were obtained by a complex continuous wavelet transform or CCWT (Complex Continuous Wavelet Transform) applied to each observation window. The frequency analysis was carried out from 1 to 250 kHz, the sampling in the frequency domain being carried out by a mother wavelet and 38 daughter wavelets. The observation window was of duration equal to ls and the sliding step was 100ms.
    The kinematic parameter of the movement was the three-dimensional position of the wrist.
    The states retained were the wrist at rest (¾) and the wrist in motion (^ i).
    For each of the methods, continuous decoding was carried out (calibration phase) over 70% of the duration of the session and provided an estimate of the movement over the remaining duration.
    Continuous motion decoding was done using three methods.
    The first method, MML (Markovian Mixture of Linear Experts), represented in line A, is that described in relation to FIG. 2.
    The second method, W-TH, represented in line B, consisted of a partial least squares regression assuming a linear relationship between the kinematic parameters and the observation variable. The states (rest, active) were estimated a posteriori by comparing the speed with a threshold.
    The third method, SKF, represented in line C, is a switched Kalman filtering, mentioned in the introductory part.
    It is noted that the estimation of the movement according to the MML method is significantly better than that according to the W-TH method or the SKF method, in the sense that it is more consistent with the movement observed. In addition, the movement estimated by the MML method presents only relatively few jolts. Finally, the false positive and false negative rates were 11% and 6.9% respectively for the WTH method, 21.9% and 26.3% for the SKF method, 5.5% and 5, 7% for the MML method. We can thus see that the MML method makes it possible to carry out a continuous decoding of the movement with a low level of error of prediction of the active / rest states.
    权利要求:
    Claims (12)
    [1" id="c-fr-0001]
    1. Continuous motion decoding method for a direct neural interface acquiring neural signals by means of a plurality of electrodes, characterized in that one estimates a control variable, y (i), representing kinematic variables motion, by means of a mixture of experts, each expert being associated with one of the hidden states of an HMM model modeling an observation variable represented by an observation vector, x (zj, the HMM model comprising at minus an active state and a rest state, said method comprising:
    - a calibration phase (210) to estimate the parameters of the HMM model and of a linear prediction parameter of each expert;
    - a preprocessing (220) of the neural signals, said processing including a time-frequency transformation on a sliding observation window, to obtain a sequence of observation vectors x [l: t]
    - an estimate (230) of the mixing coefficients (, K) of the different experts from the probability of said sequence of observation vectors (p (x [l: z])) and the probability of the states respectively associated with these experts taking into account said sequence (p (z i _ (r) = l, x [1: <])>;
    - an estimate (240) by each expert, of said motion variable from the current observation vector (x (zj) and the linear prediction parameter of said expert;
    - a combination (250) of the estimates of the different experts using the mixing coefficients to provide an estimate of the control variable.
    [2" id="c-fr-0002]
    2. A continuous motion decoding method for direct neural interface according to claim 1, characterized in that the hidden states of the HMM model include, for each element of a plurality of elements respectively controlled by a component of the motion variable, an active state during which said element is moved and a rest state during which said element is at rest.
    [3" id="c-fr-0003]
    3. A method of continuous decoding of movement for a direct neural interface according to claim 2, characterized in that the hidden states of the HMM model further comprise, for each of said elements, a state of preparation for movement immediately succeeding a state of rest and immediately preceding an active state of said element.
    [4" id="c-fr-0004]
    4. Continuous motion decoding method for direct neural interface according to one of claims 1 to 3, characterized in that, during the calibration phase, the parameters of the HMM models and the linear parameters of the experts are obtained from a sequence of observation vectors and a sequence of measurements of the motion variable, according to a supervised learning method.
    [5" id="c-fr-0005]
    5. Continuous motion decoding method for direct interface according to claim 4, characterized in that the linear parameters of the experts are obtained by means of a partial least squares regression.
    [6" id="c-fr-0006]
    6. Continuous motion decoding method for direct interface according to claim 4, characterized in that the linear parameters of the experts are obtained by means of a generalized partial least squares regression.
    [7" id="c-fr-0007]
    7. Continuous motion decoding method for direct neural interface according to one of claims 1 to 3 characterized in that, during the calibration phase, the parameters of the HMM models and the linear parameters of the experts are obtained from a sequence of observation vectors and a sequence of measurements of the motion variable, according to a semi-supervised learning method.
    [8" id="c-fr-0008]
    8. Continuous motion decoding method for direct neural interface according to claim 7, characterized in that the linear parameters of the experts are obtained by means of a hope-maximization algorithm.
    [9" id="c-fr-0009]
    9. Continuous motion decoding method for direct neural interface according to one of the preceding claims, characterized in that the time-frequency transformation is a wavelet decomposition.
    [10" id="c-fr-0010]
    10. Continuous motion decoding method for direct neural interface according to one of the preceding claims, characterized in that the coefficient zx / γ Ρ (ζ λ , χ [ΐ: ζ1) of mixture g k t) associated with l expert is obtained by g k (t) = -where p (x [l: z]) is the probability of observation of the sequence of observation vectors and P (z i ., x [l: z]) is the probability of the state associated with the expert taking into account the observed sequence, the probabilities p (x [l: z]) and P (z i ., x [l: z]) being obtained by means of an / orword algorithm.
    [11" id="c-fr-0011]
    11. Continuous motion decoding method for direct neural interface according to one of the preceding claims, characterized in that the estimate y k (z) of the motion variable by the expert is obtained by y k (z) = β Λ χ (Ζ) where β ζ is a matrix of size NxM where M is the dimension of the observation vectors and N the dimension of the control variable.
    [12" id="c-fr-0012]
    12. Method for continuous decoding of movement for direct neural interface according to one of the preceding claims, characterized in that the neural signals are electro-cortical signals.
    S.60465
    1/3
    类似技术:
    公开号 | 公开日 | 专利标题
    FR3053495A1|2018-01-05|NEURONAL DIRECT INTERFACE WITH CONTINUOUS DECODING USING A MARKOVIAN MIXTURE OF EXPERTS
    Botcharova et al.2014|Markers of criticality in phase synchronization
    EP2321769B1|2016-08-31|Method for recognising shapes and system implementing said method
    EP2407912A2|2012-01-18|Method and system for classifying neuronal signals and method for selecting electrodes for direct neuronal control
    Das et al.2014|A statistical method for automatic detection of seizure and epilepsy in the dual tree complex wavelet transform domain
    An et al.2016|Hand motion identification of grasp-and-lift task from electroencephalography recordings using recurrent neural networks
    Zou et al.2019|Removing muscle artifacts from EEG data via underdetermined joint blind source separation: A simulation study
    Zhao et al.2017|On the improvement of classifying EEG recordings using neural networks
    Hettiarachchi et al.2015|Multivariate adaptive autoregressive modeling and Kalman filtering for motor imagery BCI
    Casas et al.2021|Adversarial signal denoising with encoder-decoder networks
    US20160358085A1|2016-12-08|System and method for multimodal human state recognition
    KB et al.2019|BCI cinematics–A pre-release analyser for movies using H2O deep learning platform
    EP3563218B1|2021-01-06|Iterative process for calibrating a direct neural interface
    EP3789852A1|2021-03-10|Method for iterative calibration of a direct neural interface using a markov mixture of experts with multivariate regression
    Mayor-Torres et al.2021|Interpretable SincNet-based Deep Learning for Emotion Recognition from EEG brain activity
    WO2013139979A1|2013-09-26|Method for detecting at least one anomaly in an observed signal, computer program product and corresponding device
    Katthi et al.2020|Deep canonical correlation analysis for decoding the auditory brain
    Liu et al.2008|Quantitative complexity analysis in multi-channel intracranial EEG recordings form epilepsy brains
    EP3190480B1|2018-10-03|Method for calibrating a direct neural interface by penalised multi-channel regression
    Gianfelici et al.2011|An effective classification framework for brain–computer interfacing based on a combinatoric setting
    Kher et al.2013|EEG signal enhancement and estimation using adaptive filtering
    Notley et al.2003|Efficient estimation of a time-varying dimension parameter and its application to EEG analysis
    Bhandari et al.2019|Multimodal data fusion and prediction of emotional dimensions using deep neural network
    EP2921867A1|2015-09-23|Method for extracting electrical power signals from a mixed signal supplying a plurality of separate electrical appliances
    Skinner et al.2007|Distributed classifier migration in XCS for classification of electroencephalographic signals
    同族专利:
    公开号 | 公开日
    FR3053495B1|2018-08-10|
    US10832122B2|2020-11-10|
    US20180005105A1|2018-01-04|
    引用文献:
    公开号 | 申请日 | 公开日 | 申请人 | 专利标题
    US20130165812A1|2010-05-17|2013-06-27|Commissariat A L'energie Atomique Et Aux Energies Alternatives|Direct Neural Interface System and Method of Calibrating It|FR3100352A1|2019-09-04|2021-03-05|Commissariat A L'energie Atomique Et Aux Energies Alternatives|METHOD OF ITERATIVE CALIBRATION OF A DIRECT NEURONAL INTERFACE USING A MARKOVIAN MIX OF MULTIVARIATE REGRESSION EXPERTS|FR2981561B1|2011-10-21|2015-03-20|Commissariat Energie Atomique|METHOD FOR DETECTING MOTION SENSOR ACTIVITY, CORRESPONDING DEVICE AND COMPUTER PROGRAM|WO2015063535A1|2013-10-31|2015-05-07|Commissariat A L'energie Atomique Et Aux Energies Alternatives|Direct neural interface system and method|
    CN109063941A|2018-02-08|2018-12-21|国家电网公司|A kind of method of the straight power purchase transaction risk assessment of Generation Side|
    法律状态:
    2017-06-30| PLFP| Fee payment|Year of fee payment: 2 |
    2018-01-05| PLSC| Publication of the preliminary search report|Effective date: 20180105 |
    2018-06-27| PLFP| Fee payment|Year of fee payment: 3 |
    2020-06-30| PLFP| Fee payment|Year of fee payment: 5 |
    2021-06-30| PLFP| Fee payment|Year of fee payment: 6 |
    优先权:
    申请号 | 申请日 | 专利标题
    FR1656246|2016-06-30|
    FR1656246A|FR3053495B1|2016-06-30|2016-06-30|NEURONAL DIRECT INTERFACE WITH CONTINUOUS DECODING USING A MARKOVIAN MIXTURE OF EXPERTS|FR1656246A| FR3053495B1|2016-06-30|2016-06-30|NEURONAL DIRECT INTERFACE WITH CONTINUOUS DECODING USING A MARKOVIAN MIXTURE OF EXPERTS|
    US15/615,080| US10832122B2|2016-06-30|2017-06-06|Continuous decoding direct neural interface which uses a markov mixture of experts|
    [返回顶部]