阅读说明：本技术 一种基于脑网络动态连接特征的运动意图预测方法 (Movement intention prediction method based on brain network dynamic connection characteristics ) 是由 曾洪 刘兴 黄孝妍 沈俊杰 王新志 宋爱国 于 2020-03-24 设计创作，主要内容包括：本发明公开了一种基于脑网络动态连接特征的运动意图预测方法,步骤：构建时变动态贝叶斯网络模型,将时域的EEG信号矩阵转变为时域的电极间的有向加权连接矩阵；将得到的有向加权连接矩阵按定义的运动意图状态和静息状态对每个电极对进行配对t检验,筛选得到在两种状态下差异性显著的电极对,并按筛选得到的电极对和时间重构为特征矩阵；将得到的特征矩阵重构为特征向量,并将该特征向量输入训练好的分类器中,输出运动意图的预测结果。本发明弥补了传统脑电运动意图预测方法忽视大脑皮层不同区域之间的动态联系的缺陷。(The invention discloses a movement intention prediction method based on brain network dynamic connection characteristics, which comprises the following steps: constructing a time-varying dynamic Bayesian network model, and converting an EEG signal matrix of a time domain into an inter-electrode directed weighting connection matrix of the time domain; carrying out pairing t test on each electrode pair according to a defined movement intention state and a defined rest state on the obtained directed weighted connection matrix, screening to obtain electrode pairs with remarkable difference in the two states, and reconstructing the electrode pairs obtained by screening and time into a feature matrix; and reconstructing the obtained feature matrix into a feature vector, inputting the feature vector into a trained classifier, and outputting a prediction result of the movement intention. The invention makes up the defect that the traditional electroencephalogram movement intention prediction method neglects the dynamic connection between different areas of the cerebral cortex.)

1. A movement intention prediction method based on brain network dynamic connection characteristics is characterized by comprising the following steps:

(1) constructing a time-varying dynamic Bayesian network model, wherein a coefficient matrix of the model is a directional weighted connection matrix to be solved, decomposing the directional weighted connection matrix according to two orthogonal coordinate axis directions of time and electrodes, converting the decomposed directional weighted connection matrix into L1 regularized weighted least square regression problem, and solving the problem, so that an EEG signal matrix of a time domain is converted into a directional weighted connection matrix between electrodes of the time domain, wherein each element in the directional weighted connection matrix corresponds to a connection coefficient between the two electrodes at the current moment;

(2) performing pairing t test on each electrode pair by the directional weighting connection matrix obtained in the step (1) according to a defined movement intention state and a defined rest state, screening to obtain electrode pairs with obvious difference in the two states, and reconstructing the electrode pairs obtained by screening and time into a feature matrix;

(3) and (3) reconstructing the feature matrix obtained in the step (2) into a feature vector, inputting the feature vector into a trained classifier, and outputting a prediction result of the movement intention.

2. The method for predicting motor intention based on brain network dynamic connection characteristics according to claim 1, wherein in step (1), the following time-varying dynamic bayesian network model is constructed based on a first order markov theory:

X^t＝A^tX^t-1+,～N(0,σ²I)

in the above formula, the first and second carbon atoms are, the value of the electrode i at the time t is 1,2, …, n, and n is the number of electrodes; a. the^tA directed weighted connection matrix at time t; n (0, σ)²I) Expressed as the variance is σ²I is an identity matrix.

3. The method for predicting the movement intention based on the brain network dynamic connection characteristics as claimed in claim 2, wherein the inter-electrode connection coefficients are weighted by adopting a Gaussian radial basis kernel function, and the sparse of the directional weighted connection matrix is performed by L1 regularization:

in the above equation, T represents the length of the EEG time series;is shown at t^*The ith row of the time-wise weighted connection matrix,is composed ofAn estimated value of (d); λ is a regularization parameter;to estimate t^*Calculating a weight value according to time t when the weighted connection matrix is directed at the moment; t is observed value corresponding time, t^*Time is corresponding to the estimated value;

decomposing the directional weighted connection matrix according to two orthogonal axes of time and electrodes, converting the directional weighted connection matrix into L1 regularized weighted least squares regression problem, and iteratively solving the problem by adopting a coordinate descent method until convergence.

4. The method according to claim 3, wherein the Gaussian radial basis function isThen the weight valueIs calculated as follows:

in the above formula, h is the nuclear bandwidth.

5. The method for predicting the movement intention based on the dynamic connection characteristics of the brain network according to claim 1, wherein before the step (1), the EEG signals are subjected to frequency division, the low-frequency EEG signals are directly used as the input of the time-varying dynamic Bayesian network model, and the signal envelope of the high-frequency EEG signals after Hilbert transformation is used as the input of the time-varying dynamic Bayesian network model.

6. The method as claimed in claim 5, wherein the EEG signal is frequency-divided by using a fourth-order zero-phase-shift Butterworth filter.

7. The method for predicting motor intention based on dynamic connection characteristics of brain network as claimed in claim 5, wherein EEG signals of different frequency bands are spatially filtered before being input into the time-varying dynamic Bayesian network model.

8. The method for predicting motor intention based on brain network dynamic link characteristics as claimed in claim 7, wherein the EEG signals of different frequency bands are spatially filtered by superposition of a co-average reference CAR and a current source density CSD.

9. The method for predicting motor intention based on brain network dynamic connection characteristics according to claim 1, wherein in the step (2), the motor intention state is defined to be within 1s before the motor starting time, and the rest state is defined to be within 1s before the motor indication.

10. The method for predicting motor intention based on brain network dynamic link characteristics as claimed in claim 1, wherein in the step (3), the classifier adopts a random forest classifier and adopts five-fold cross validation to evaluate the classification effect.

Technical Field

The invention belongs to the field of EEG signal analysis, and particularly relates to a movement intention prediction method.

Background

Generally, the limb activities of a person are manipulated by brain will, so it is important to predict the motor intention of a patient through a brain-computer interface in order to assist and restore the motor communication ability of a patient with impaired motor abilities, such as disabilities and paralysis. For the prediction of the motor intention, most of the published studies at present are based on the time domain features or frequency domain features of motor-related cortical potentials (motor-related cortical potentials). Although temporal or frequency domain features of motor-related cortical potentials have proven successful in predicting motor intent, these methods ignore the dynamic association of time-varying brain electrical activity in different regions of the brain. In order to take into account the dynamic link between different regions of the brain, connectivity analysis methods must be employed. Methods for analyzing the connectivity of brain waves are roughly classified into two categories, namely functional connectivity (functional connectivity) and effective connectivity (effective connectivity). Whether neural activity is coherent or correlated for a pair of regions of the brain can be measured using a functional connectivity approach, but correlation or coherence does not imply a causal relationship, i.e., functional connectivity does not reflect the directionality of the interaction between two regions of the brain. In contrast, an operative connection is a directional and causal relationship that reflects the interaction between two regions of the brain.

Disclosure of Invention

In order to solve the technical problems of the background art, the invention provides a method for predicting a movement intention based on a brain network dynamic connection characteristic.

In order to achieve the technical purpose, the technical scheme of the invention is as follows:

a movement intention prediction method based on brain network dynamic connection characteristics comprises the following steps:

(1) constructing a time-varying dynamic Bayesian network model, wherein a coefficient matrix of the model is a directional weighted connection matrix to be solved, decomposing the directional weighted connection matrix according to two orthogonal coordinate axis directions of time and electrodes, converting the decomposed directional weighted connection matrix into L1 regularized weighted least square regression problem, and solving the problem, so that an EEG signal matrix of a time domain is converted into a directional weighted connection matrix between electrodes of the time domain, wherein each element in the directional weighted connection matrix corresponds to a connection coefficient between the two electrodes at the current moment;

(2) performing pairing t test on each electrode pair by the directional weighting connection matrix obtained in the step (1) according to a defined movement intention state and a defined rest state, screening to obtain electrode pairs with obvious difference in the two states, and reconstructing the electrode pairs obtained by screening and time into a feature matrix;

(3) and (3) reconstructing the feature matrix obtained in the step (2) into a feature vector, inputting the feature vector into a trained classifier, and outputting a prediction result of the movement intention.

Further, in step (1), constructing a time-varying dynamic bayesian network model based on a first order markov theory as follows:

X^t＝A^tX^t-1+,～N(0,σ²I)

in the above formula, the first and second carbon atoms are, the value of the electrode i at the time t is 1,2, …, n, and n is the number of electrodes; a. the^tA directed weighted connection matrix at time t; n (0, σ)²I) Expressed as variance σ²I is an identity matrix.

Further, the inter-electrode connection coefficients are weighted by using a gaussian radial basis kernel function, and the sparse of the directional weighted connection matrix is performed by L1 regularization:

in the above equation, T represents the length of the EEG time series;is shown at t^*The ith row of the time-wise weighted connection matrix,is composed ofAn estimated value of (d); λ is a regularization parameter;to estimate t^*Calculating a weight value according to time t when the weighted connection matrix is directed at the moment; t is observed value corresponding time, t^*Time is corresponding to the estimated value;

decomposing the directional weighted connection matrix according to two orthogonal axes of time and electrodes, converting the directional weighted connection matrix into L1 regularized weighted least squares regression problem, and iteratively solving the problem by adopting a coordinate descent method until convergence.

Further, the Gaussian radial basis kernel function isThen the weight valueIs calculated as follows:

in the above formula, h is the nuclear bandwidth.

Further, before the step (1), frequency division is performed on the EEG signal, the low-frequency EEG signal is directly used as the input of the time-varying dynamic Bayesian network model, and the signal envelope of the high-frequency EEG signal after Hilbert transformation is used as the input of the time-varying dynamic Bayesian network model.

Further, the EEG signal is banded using fourth order zero phase shift butterworth filtering.

Further, the EEG signals of different frequency bands are spatially filtered before being input into the time varying dynamic bayesian network model.

Further, the EEG signals of different frequency bands are spatially filtered by means of superposition of the co-averaged reference CAR and the current source density CSD.

Further, in the step (2), the exercise intention state is defined to be within 1s before the exercise start time, and the rest state is defined to be within 1s before the exercise instruction.

Further, in the step (3), the classifier adopts a random forest classifier and adopts five-fold cross validation to evaluate the classification effect.

Adopt the beneficial effect that above-mentioned technical scheme brought:

the invention utilizes brain network connection characteristics to predict movement intention, and in order to make up for the defect that the traditional brain electrical movement intention prediction method neglects dynamic connection between different areas of the cerebral cortex, the invention solves a directed weighting connection matrix according to a time-varying dynamic Bayesian network model, and the directed weighting connection matrix reflects connection weights between different electrodes, namely the directed weighting connection matrix can reflect the dynamic changes of the strength and direction of the mutual connection between different brain areas and reflect the causal connection between different brain areas.

Drawings

FIG. 1 is a flow chart of the method of the present invention.

Detailed Description

The technical scheme of the invention is explained in detail in the following with the accompanying drawings.

The invention designs a movement intention prediction method based on brain network dynamic connection characteristics, as shown in figure 1, the steps are as follows:

step 1, constructing a time-varying dynamic Bayesian network model (TV-DBN), decomposing a directional weighted connection matrix in the model according to two orthogonal coordinate axis directions of time and electrodes, converting the decomposed directional weighted connection matrix into L1 regularized weighted least square regression problem, and solving the problem, so as to convert an EEG signal matrix of a time domain into a directional weighted connection matrix between electrodes of the time domain, wherein each element in the directional weighted connection matrix corresponds to a connection coefficient between the two electrodes at the current moment;

step 2: performing pairing t test on each electrode pair according to the defined movement intention state and rest state of the directional weighting connection matrix obtained in the step 1, screening to obtain electrode pairs with remarkable difference in the two states, and reconstructing the electrode pairs obtained by screening and time into a feature matrix;

and step 3: and (3) reconstructing the feature matrix obtained in the step (2) into a feature vector, inputting the feature vector into a trained classifier, and outputting a prediction result of the movement intention.

As shown in FIG. 1, in order to analyze the good and bad effects of the EEG signal of different frequency bands for motion intention prediction, the preprocessed EEG signal is subjected to quadric zero-phase shift Butterworth filtering, in this embodiment, divided into five frequency bands of [0.1,1] Hz, [1,3] Hz, [4,7] Hz, [8,13] Hz, [14,30] Hz, wherein the amplitude is directly input into the model calculation in the low frequency band of [0.1,7] Hz, and the Hilbert transformed signal envelope is used as the input of the model in the high frequency band of [8,30] Hz.

As shown in fig. 1, the acquired EEG signals are preprocessed and then spatially filtered to reduce the influence of brain volume conductor effects, and in this embodiment, the spatial filtering method selects a method of superposition of a Common Average Reference (CAR) and a Current Source Density (CSD).

In this embodiment, preferably, the step 1 can be implemented by the following preferred scheme:

according to a simplified first order Markov model, if the activity of the j electrode at time t is affected by the activity of the i electrode at time t-1, the connection of electrodes i and j is considered to be enhanced at time t. The simplified first-order Markov principle is combined with the dynamic Bayes network to form a time-varying dynamic Bayes network model, namely a linear regression model as follows:

X^t＝A^tX^t-1+,～N(0,σ²I)

in the above formula, the first and second carbon atoms are, the value of the electrode i at the time t is 1,2, …, n, and n is the number of electrodes; a. the^tAt time t isConnecting the matrix to the weights; n (0, σ)²I) Expressed as variance σ²I is an identity matrix.

In order to reduce noise and ensure smoothness of time-varying dynamics, the present embodiment weights the connection coefficients of neighboring points by using a gaussian radial basis kernel function, and L1 regularization performs sparseness of a directional weighted connection matrix:

in the above equation, T represents the length of the EEG time series;is shown at t^*The ith row of the time-wise weighted connection matrix,is composed ofAn estimated value of (d); λ is a regularization parameter;to estimate t^*Calculating a weight value according to time t when the weighted connection matrix is directed at the moment; t is observed value corresponding time, t^*The time is an estimate.

Gaussian radial basis kernel functionWeight calculation formula:

in the above formula, h is the nuclear bandwidth.

For solving the directed weighted connection matrix, the directed weighted connection matrix is decomposed according to two orthogonal axes of time and electrodes.

Firstly, ordering:

then, the solution can be optimized to a standard L1 regularized weighted least squares regression problem, namely:

by coordinate descent method, fixingThe remaining n-1 coordinates except the jth coordinate are unchanged, namely:

for the jth coordinateCalculating a partial derivative:

where sign denotes the sign function, i.e.:

order:

therefore:

gradually updating according to the above process that j is 1, …, nUp toConvergence, i ═ 1, …, n. I.e. the EEG signal matrix in the time domain can be transformed into a directed weighted connection matrix in the time domain.

In this embodiment, preferably, the step 2 can be implemented by the following preferred scheme:

defining the motion intention state as 1s before the motion starting time, namely [ -1,0] s, and 0 being the motion starting time; the resting state is defined as 1s before the motion indication. And then carrying out pairing t test on the connection coefficients of all electrode pairs in the rest state and the movement intention state (t is less than 0.01), screening to obtain the electrode pairs with remarkable difference in the two states, and reconstructing into a matrix according to the screened electrode pairs and time.

In this embodiment, preferably, the step 3 can be implemented by the following preferred scheme:

the classifier adopts a random forest classifier, and the random forest classifier is suitable for classifying high-dimensional sparse features without considering dimension reduction of feature vectors. And evaluating the classification effect by adopting five-fold cross validation.