阅读说明：本技术 一种经颅直流电刺激和功能性电刺激的运动功能康复方法 (motor function rehabilitation method by transcranial direct current stimulation and functional electrical stimulation ) 是由 陆晟 罗志增 孟明 佘青山 孙曜 席旭刚 于 2019-09-27 设计创作，主要内容包括：本发明提出了一种采用经颅直流电刺激和功能性电刺激的运动功能闭环康复方法。首先对大脑皮层区域进行tDCS刺激,改善皮层活性并且促进神经重塑,提高MI-BCI的准确性。在此基础上采集多通道运动想象的脑电信号,构建通道重要性测度,获得通道特征权值矩阵,为各通道特征向量加权,构造特征加权logistic分类机。最后利用logistic分类机进行运动想象识别,根据识别结果进行FES刺激完成上肢动作,促进本体感觉上行反馈至中枢,与tDCS刺激引发重塑和运动控制的神经冲动下行一起,构成双刺激干预,塑造了“控制下行-感觉上行”的运动功能闭环康复回路,促使患者自然和全面的康复。(The invention provides motor function closed-loop rehabilitation methods adopting transcranial direct current stimulation and functional electrical stimulation, which comprises the steps of firstly carrying out tDCS stimulation on a cerebral cortical area, improving cortical activity, promoting nerve remodeling and improving the accuracy of MI-BCI, collecting multi-channel motor imagery electroencephalogram signals on the basis, constructing channel importance measure, obtaining a channel feature weight matrix, weighting each channel feature vector, constructing a feature weighting logistic classifier, finally carrying out motor imagery identification by using the logistic classifier, carrying out FES stimulation according to an identification result to finish upper limb actions, promoting proprioceptive uplink feedback to a central nervous system, forming double stimulation intervention with nerve impulse downlink of remodeling and motor control caused by tDCS stimulation, forming a motor function closed-loop rehabilitation circuit of 'control downlink-sensory uplink', and promoting natural and comprehensive rehabilitation of a patient.)

The motor function rehabilitation method of 1, kinds of transcranial direct current stimulation and functional electrical stimulation is characterized in that motor cortex stimulation of transcranial direct current and functional electrical stimulation according to an electroencephalogram recognition result achieve motor function closed loop rehabilitation through double stimulation intervention, and the method comprises the following steps:

step 1, performing transcranial direct current stimulation (tDCS) on a patient, improving the cortical liveness, enabling nerve concussion to extend downwards to a spinal cord bundle, promoting brain function network recombination and improving motor imagery event-related desynchronizing ERD indexes;

step 2, collecting a multi-channel electroencephalogram signal EEG of the patient after the step 1, and performing feature extraction on the collected electroencephalogram signal by adopting an electroencephalogram signal feature extraction method based on Hilbert-Huang transform to obtain initial feature vectors of all channels and construct an initial feature matrix;

step 3, constructing an electroencephalogram channel importance measure, and grading the importance of each acquisition channel by adopting a multi-threshold strategy so as to obtain a characteristic weight matrix of each channel;

step 4, weighting the initial feature matrix in the step 2 by using the feature weight matrix obtained in the step 3, realizing sparsity of a nuclear logistic classifier in a feature vector selection mode, constructing the feature weighted logistic classifier, and improving the accuracy of motor imagery electroencephalogram pattern recognition;

step 5, controlling the functional electrical stimulation FES upper limb rehabilitation robot according to the identification result obtained in the step 4, and performing functional electrical stimulation on the patient to complete upper limb movement; meanwhile, the FES can feed the proprioception back to the motor sensory cortex in an ascending way, and the FES and the tDCS form double stimulation intervention to fulfill the aim of reshaping a motor nerve transmission descending control and ascending sensory feedback inner closed loop system, thereby realizing natural and comprehensive motor rehabilitation of a patient.

2. The motor function rehabilitation method based on transcranial direct current stimulation and functional electrical stimulation according to claim 1, characterized in that the method for extracting the electroencephalogram signal features based on Hilbert-Huang transform in step 2 comprises the following specific processes:

2-1, performing Hilbert-Huang transformation on the electroencephalogram signals of each channel obtained by experiments, decomposing the signals into series intrinsic mode functions IMF by empirical mode decomposition, and performing Hilbert transformation on all IMFs to obtain Hilbert spectrums H (v, t) with good resolution;

2-2, the existing research proves that the event-related synchronization/desynchronization phenomenon is mainly embodied in the mu rhythm of 8-13 Hz and the beta rhythm of 14-30 Hz of the human brain movement sensory cortex; dividing 8-30 Hz frequency bands into z sub-bands at equal intervals, and calculating the instantaneous energy function of each sub-band;

2-3, extracting the autoregressive AR model coefficient of each sub-band instantaneous energy function, cascading into the initial characteristic vector of the acquired channel signal, wherein if the number of the acquired channels is Q, the constructed initial characteristic matrix M is as follows:

3. the motor function rehabilitation method by transcranial direct current stimulation and functional electrical stimulation according to claim 1, wherein the channel feature weight matrix in step 3 is constructed by the following specific process:

3-1, determining the strength of the correlation between all the pairwise combinations of each channel, firstly establishing an autoregressive model for the collected EEG of each channel, then converting the model into a frequency domain, obtaining the power spectrum of each channel through a multivariate spectral decomposition theorem, and further constructing a quantitative measure function of the strength of the correlation between the two channels based on the multivariate power spectrum:

wherein S_abIs the EEG signal on channels a and bQ, Q is the number of channels collected;

taking the data of each experiment to obtain C² _ab(f) Mean value of

3-2, constructing a channel characteristic weight matrix; based on the multivariate power spectrum correlation coefficient between every two channels, the importance measure function of each channel can be obtained:

according to the importance measure function, a feature weight vector of the channel a can be constructed:

wherein

then, a multi-threshold strategy is adopted to grade the importance of each acquisition channel, and a characteristic weight matrix of each channel is constructed on the basis:

Technical Field

The invention belongs to the field of pattern recognition, and relates to an motor function closed-loop rehabilitation method for realizing upper limb rehabilitation movement by functional electrical stimulation of an upper limb rehabilitation robot after cortex electrical stimulation, in particular to motor function closed-loop rehabilitation methods adopting transcranial direct current stimulation and functional electrical stimulation.

Background

The stroke is common cerebrovascular disease worldwide, the patient has lack or disorder of motor cortex control and sensory function due to tissue damage caused by cerebral hemorrhage or ischemia, and motor dysfunction such as abnormal muscle tension appears, the development of modern neurophysiology and the cross fusion of electronic information, robots and artificial intelligence technology bring a new means for the motor rehabilitation of the stroke patient, so the motor rehabilitation based on the brain muscle fusion also becomes the hot spot research subject in the field.

Functional Electrical Stimulation (FES) technology is used as neuromuscular stimulation means and is often used for clinical rehabilitation treatment of motor functions of stroke patients, FES can help patients to rebuild damaged peripheral nervous systems and recover or improve muscle action functions, and more importantly, FES intervention measures can input sensory impulses to cerebral cortex through sensory (ascending) conduction pathways to promote remodeling of the motor cortex.

Because the motor imagery process can directly excite the motor cortex of the brain, the MI-BCI which adopts motor imagery electroencephalogram (EEG) as an information source is combined with the FES, so that a conduction path of motor impulse of a stroke patient can be reconstructed in vitro, and the activation degree of the motor cortex is improved. MI-BCI-FES therapy takes advantage of the subjective motor intent of the patient to increase their involvement in FES rehabilitation training with greater efficiency compared to FES treatment alone.

The identification accuracy and reliability of MI-BCI are important for improving the active participation degree and the intervention effect of a patient in FES treatment, but as the motor imagery ability of a cerebral apoplexy patient is influenced by nerve injury, the motor imagery task completion quality is not high, so that MI-BCI-FES has a large false triggering risk, and finally, the rehabilitation training cannot achieve the expected effect, the invention provides a method for combining tDCS and an MI-BCI-FES system to be used as new motor function closed loop rehabilitation, , the tDCS directly stimulates related brain areas, improves the motor cortex activity, enables nerve oscillation to extend downwards to a spinal nerve network, promotes nerve remodeling and long-term effects thereof, even forms new functional connection and induces a spinal tract passage to remodel a nerve impulse downlink path, and , the improvement of motor cortex activity caused by the tDCS is relied on, contributes to the improvement of the quality of motor brain electrical signal quality and the enhancement of the characteristic of the motor brain electrical signal of the patient, thereby improving the accuracy and reliability of the BCI, and the correct motor cortex stimulation of the movement desire can be realized by giving FES stimulation corresponding to the patient, so that the brain electrical signal quality can be comprehensively controlled by a subjective motor cortex remodeling nerve, and the feedback of the rehabilitation nerve remodeling nerve sensory stimulation in a brain nerve channel, and the feedback possibility of the rehabilitation target of the rehabilitation can be greatly increased by a brain rehabilitation nerve remodeling nerve.

Disclosure of Invention

The invention provides motor function closed-loop rehabilitation methods adopting transcranial direct current stimulation and functional electrical stimulation, which comprises the steps of firstly carrying out tDCS stimulation on a cerebral cortical area, improving cortical activity, promoting nerve remodeling, and improving the accuracy and reliability of MI-BCI, then collecting motor imagery electroencephalogram signals on the basis of the above steps, constructing channel importance measure, obtaining a channel feature weight matrix, weighting each channel feature vector, and constructing a feature weighted logistic classifier, wherein each channel feature vector can be obtained by an electroencephalogram signal feature extraction method based on Hilbert-Huang transform, finally carrying out motor imagery identification by using the logistic classifier, carrying out FES stimulation according to an identification result, promoting proprioceptive uplink feedback to a central center while realizing upper limb movement, forming double stimulation intervention with nerve impulse downlink for inducing remodeling and motion control by the tDCS stimulation, forming a double stimulation intervention, and shaping a 'downlink-sensory uplink controlled motor function closed-loop rehabilitation circuit', and promoting natural and comprehensive rehabilitation of a patient.

In order to achieve the above object, the method of the present invention mainly comprises the following steps:

step 1, transcranial direct current stimulation tDCS is carried out on a patient, the cortical liveness is improved, nerve concussion is enabled to extend downwards to a spinal cord bundle, brain function network recombination is promoted, and motor imagery event-related desynchronization ERD indexes are improved.

Step 2, acquiring a motor imagery electroencephalogram (EEG) of the patient after the step 1, performing feature extraction on the acquired EEG by adopting an EEG feature extraction method based on Hilbert-Huang transform to obtain initial feature vectors of all channels, and constructing an initial feature matrix M, wherein the method specifically comprises the following steps:

2-1, decomposing the electroencephalogram signal into series intrinsic mode functions IMF through empirical mode decomposition, setting each channel signal obtained through experiments as a fixed-length -dimensional time sequence x, and after the empirical mode decomposition is carried out on the x, expressing the x as:

in the formula c_i(t) denotes the ith IMF, r_n(t) represents the final residual signal, n being the number of IMFs.

2-2. Hilbert transform on all IMFs, yielding Hilbert spectra H (v, t) with good resolution:

where v denotes frequency, t denotes time, a_i(t) is the instantaneous amplitude of the signal,is an instantaneous frequency, wherein

Is the instantaneous phase.

2-3. the existing research has proved that the event-related synchronization/desynchronization phenomenon is mainly embodied in the mu rhythm of 8-13 Hz and the beta rhythm of 14-30 Hz of human brain motor sensory cortex, the frequency band of 8-30 Hz is divided into z sub-bands at equal intervals, the width of the sub-band is recorded as h, and the instantaneous energy function IE (t) of each sub-band is calculated as:

wherein g is p h, p is 0.

2-4, extracting the autoregressive AR model coefficient of each instantaneous energy function IE (t), cascading into the initial characteristic vector of the acquired channel signal, wherein the number of the acquired channels is Q, and the constructed initial characteristic matrix M is as follows:

step 3, constructing an electroencephalogram channel importance measure, and grading the importance of each acquisition channel by adopting a multi-threshold strategy so as to obtain a characteristic weight matrix K of each channel, wherein the specific steps are as follows:

and 3-1, determining the strength of the correlation between all the channel combinations. Using the matrix X ═ { X for each channel EEG collected_q(l) Denotes, L1., L, Q1., Q, L is the number of sampling points. X ═ X_q(l) The P-order multivariate autoregressive model of the method is as follows:

where X (l) represents the EEG signal for each channel, A (k) is the k-th order autoregressive coefficient, and U (l) is white noise.

And after the frequency domain is switched to, obtaining:

Y(f)＝H(f)U(f)

where H (f) is a transform array,

and the diagonal elements of H (f) represent the power spectrum of each channel, U (f) is the spectrum of U (l).

Can be obtained according to the multivariate spectrum decomposition theorem

In the formula, the superscript H represents the Hermite matrix transformation.

Setting a and b as any two acquisition channels, wherein a and b are 1 and 2, and Q, constructing a quantitative measure function of the correlation strength based on the multivariate power spectrum between the two channels:

wherein S_abIs the cross-power spectrum of the brain electrical signals on channels a and b.

The obtained data of each experiment are compared to obtain C² _ab(f) Mean value of

The correlation coefficient is defined as a multivariate power spectrum correlation coefficient and is used as an index for measuring the correlation strength of the channel a and the channel b.

And 3-2, constructing a channel characteristic weight matrix. Based on the multivariate power spectrum correlation coefficient between every two channels, the importance measure function of each channel can be obtained:

using channels

And (2) grading the importance of each acquisition channel by adopting a multi-threshold strategy, endowing a larger feature weight to the important channel, and constructing a feature weight matrix of each channel on the basis:

wherein k is_aRespectively substituting the average frequencies of the z sub-bands divided in step 2

The construction is as follows:

wherein

Represents the average frequency of the z-th sub-band divided in step 2.

Step 4, weighting the eigenvectors of each channel in the step 2 by using the weight matrix K of the features obtained in the step 3, realizing sparsity of a kernel logistic classifier in a mode of selecting the eigenvectors, constructing the feature weighted logistic classifier, and improving the accuracy of the motor imagery mode recognition, wherein the specific steps are as follows:

4-1, defining a characteristic weighting kernel function. Known as M_αAnd M_βIs any two characteristic matrix samples in a training sample set, and the characteristic weight matrix is K ═ K₁,k₂,...,k_Q]And Q is the number of acquisition channels, and the obtained characteristic weighted Gaussian radial basis kernel function is as follows:

where γ is a kernel function parameter and M is an input feature matrix.

And 4-2, constructing a feature weighting logistic classifier. The prediction function is:

the classification rule is as follows:

in the formula

For the model parameters, N is the number of samples, τ is the number of classes, and Γ is the number of classes.

And 4-3, performing parameter fitting of the logistic classification model by adopting maximum likelihood estimation. For supervised classification tasks of the Γ -type, a training sample set is given

Wherein M is_μIs a feature matrix; t is t_μIs an output vector of dimension Γ. When inputting M_μT is t if it belongs to the category τ, τ ∈ {1, 2.,. Γ }, then_μτ1, otherwise, t_μτ0. The best classification model parameters F can be obtained as:

and 5, controlling the functional electrical stimulation FES upper limb rehabilitation robot according to the identification result obtained in the step 4, and carrying out FES on the patient. The FES can feed the proprioception back to the motor sensory cortex in an ascending way, and the FES and the tDCS form double stimulation intervention to achieve the aim of remolding a motor nerve impulse descending control and ascending sensory feedback inner closed loop system, thereby realizing natural and comprehensive motor rehabilitation of a patient.

Meanwhile, FES can feed proprioception upwards to the kinesthetic cortex and form double-stimulation intervention with tDCS.

The rehabilitation training method provided by the invention builds a motion function closed loop rehabilitation circuit for controlling descending and sensory ascending, and double stimulation intervention is formed by utilizing FES and tDCS, so that the possibility of opening a motor nerve transmission inner closed loop can be greatly increased, and the aim of natural and comprehensive motion rehabilitation of a patient is fulfilled.

Drawings

FIG. 1 is a functional block diagram of an implementation of the present invention;

FIG. 2 is a flow chart of an electroencephalogram feature extraction method based on Hilbert-Huang transform;

Detailed Description

The embodiments of the present invention will be described in detail below with reference to the accompanying drawings:

as shown in fig. 1, the present embodiment includes the following steps:

step , building a rehabilitation experiment platform, and making a rehabilitation training scheme, wherein the specific process is as follows:

the motor function rehabilitation of the upper limb is taken as an object, and an electric stimulation intervention brain-machine fusion motor function closed-loop rehabilitation experiment platform is set up. The multi-channel FES and tDCS devices are adopted to respectively form electrical stimulation intervention channels at two ends of cerebral cortex and limb muscles, and a multi-channel EEG signal acquisition device is utilized to form an EEG signal acquisition channel so as to acquire EEG signals in the MI-BCI-FES rehabilitation training process.

According to the clinical embodiment of upper limb rehabilitation training, the relatively complete specific upper limb actions of the medical examination and evaluation method are selected as target actions of electrical stimulation intervention and motor imagery, including the basic actions of elbow flexion, elbow extension, wrist flexion, wrist extension, fist making and fist making, the experimental subjects are 40 patients with upper limb dysfunction after cerebral apoplexy, which meet the selection standard.

The invention adopts an electroencephalogram characteristic extraction method based on Hilbert-Huang transformation, firstly obtains series Intrinsic Mode Functions (IMFs) through empirical mode decomposition, performs Hilbert transformation on each IMF to obtain a Hilbert spectrum, then obtains an instantaneous energy function of each IMF, extracts an Autoregressive (AR) model coefficient of the instantaneous energy function, and obtains an initial characteristic vector, wherein a characteristic extraction flow chart is shown in figure 2, and the specific process is as follows:

decomposing an electroencephalogram signal into IMFS through empirical mode decomposition, setting each channel signal obtained through experiments as a fixed-length -dimensional time sequence x, and after performing empirical mode decomposition on x, expressing x as:

in the formula c_i(t) denotes the ith IMF, r_n(t) represents the final residual signal, n being the number of IMFs.

2-2. Hilbert transform on all IMFs, yielding Hilbert spectra H (v, t) with good resolution:

where v denotes frequency, t denotes time, a_i(t) is the instantaneous amplitude of the signal,

is an instantaneous frequency, wherein

Is the instantaneous phase.

2-3. the existing research has proved that the event-related synchronization/desynchronization phenomenon is mainly embodied in the mu rhythm of 8-13 Hz and the beta rhythm of 14-30 Hz of human brain motor sensory cortex, the frequency band of 8-30 Hz is divided into z sub-bands at equal intervals, the width of the sub-band is recorded as h, and the instantaneous energy function IE (t) of each sub-band is calculated as:

wherein g is p h, p is 0.

2-4, extracting the autoregressive AR model coefficient of each instantaneous energy function IE (t), cascading into the initial characteristic vector of the acquired channel signal, wherein the number of the acquired channels is Q, and the constructed initial characteristic matrix M is as follows:

step 3, constructing an electroencephalogram channel importance measure, and grading the importance of each acquisition channel by adopting a multi-threshold strategy so as to obtain a characteristic weight matrix K of each channel, wherein the specific steps are as follows:

and 3-1, determining the strength of the correlation between all the channel combinations. Using the matrix X ═ { X for each channel EEG collected_q(l) Denotes, L1., L, Q1., Q, L is the number of sampling points. X ═ X_q(l) The P-order multivariate autoregressive model of the method is as follows:

where X (l) represents the EEG signal for each channel, A (k) is the k-th order autoregressive coefficient, and U (l) is white noise.

And after the frequency domain is switched to, obtaining:

Y(f)＝H(f)U(f)

where H (f) is a transform array,

and the diagonal elements of H (f) represent the power spectrum of each channel, U (f) is the spectrum of U (l).

Can be obtained according to the multivariate spectrum decomposition theorem

In the formula, the superscript H represents the Hermite matrix transformation.

Setting a and b as any two acquisition channels, wherein a and b are 1 and 2, and Q, constructing a quantitative measure function of the correlation strength based on the multivariate power spectrum between the two channels:

wherein S_abIs the cross-power spectrum of the brain electrical signals on channels a and b.

The obtained data of each experiment are compared to obtain C² _ab(f) Mean value of

The correlation coefficient is defined as a multivariate power spectrum correlation coefficient and is used as an index for measuring the correlation strength of the channel a and the channel b.

And 3-2, constructing a channel characteristic weight matrix. Based on the multivariate power spectrum correlation coefficient between every two channels, the importance measure function of each channel can be obtained:

using channels

And (2) grading the importance of each acquisition channel by adopting a multi-threshold strategy, endowing a larger feature weight to the important channel, and constructing a feature weight matrix of each channel on the basis:

wherein k is_aRespectively substituting the average frequencies of the z sub-bands divided in step 2

The construction is as follows:

wherein

Represents the average frequency of the z-th sub-band divided in step 2.

Step 4, weighting the eigenvectors of each channel in the step 2 by using the weight matrix K of the features obtained in the step 3, realizing sparsity of a kernel logistic classifier in a mode of selecting the eigenvectors, constructing the feature weighted logistic classifier, and improving the accuracy of the motor imagery mode recognition, wherein the specific steps are as follows:

4-1, defining a characteristic weighting kernel function. Known as M_αAnd M_βIs a trainingAny two characteristic matrix samples in the training sample set are provided, and the characteristic weight matrix is K ═ K₁,k₂,...,k_Q]And Q is the number of acquisition channels, and the obtained characteristic weighted Gaussian radial basis kernel function is as follows:

where γ is a kernel function parameter and M is an input feature matrix.

And 4-2, constructing a feature weighting logistic classifier. The prediction function is:

the classification rule is as follows:

in the formula

For the model parameters, N is the number of samples, τ is the number of classes, and Γ is the number of classes.

And 4-3, performing parameter fitting of the logistic classification model by adopting maximum likelihood estimation. For supervised classification tasks of the Γ -type, a training sample set is given

Wherein M is_μIs a feature matrix; t is t_μIs an output vector of dimension Γ. When inputting M_μT is t if it belongs to the category τ, τ ∈ {1, 2.,. Γ }, then_μτ1, otherwise, t_μτ0. The best classification model parameters F can be obtained as:

and 5, controlling the FES upper limb rehabilitation robot according to the identification result of the motor brain electrical signals obtained in the step 4, giving the patient functional electrical stimulation corresponding to the subjective intention of the patient to realize limb actions, and if the correct limb actions are finished after the FES stimulation, feeding the proprioception back to the motor sensory cortex in an ascending way, and further enhancing the neural plasticity of the brain functional area.