CN111012337A

CN111012337A - Brain network and regularized discriminant analysis-based electroencephalogram analysis method

Info

Publication number: CN111012337A
Application number: CN201911282373.1A
Authority: CN
Inventors: 付荣荣; 王涵; 王世伟
Original assignee: Yanshan University
Current assignee: Yanshan University
Priority date: 2019-12-13
Filing date: 2019-12-13
Publication date: 2020-04-17
Anticipated expiration: 2039-12-13
Also published as: CN111012337B

Abstract

The invention provides an electroencephalogram analysis method based on a brain network and regularized discriminant analysis, which uses the brain network method to extract characteristics of electroencephalogram signals, constructs the brain network according to an obtained adjacency matrix, makes the activity state of the brain visualized by drawing degree information of each node in a brain topographic map, can more intuitively observe the neural mechanism of brain activity, and finally obtains a classification result through regularized discriminant analysis, thereby having higher classification recognition rate.

Description

Brain network and regularized discriminant analysis-based electroencephalogram analysis method

Technical Field

The invention relates to an electroencephalogram signal analysis technology, in particular to an electroencephalogram analysis method based on a brain network and regularized discriminant analysis.

Background

The human brain is a very complex system in nature, and all the cognitive activities of the human need to be completed by the brain, so that the research on the brain is very meaningful. The brain electrical signal is a synchronous physiological signal for recording brain neuron through scalp electrode, it has higher time resolution, and the signal contains a large amount of physiological information, and can be used as a sensitive index for brain function evaluation. In the aspect of engineering application, the method is often used for realizing a brain-computer interface, and achieves certain control purposes by extracting and classifying signals according to the dissimilarity of human electroencephalograms with different senses, motions and cognitive activities.

The brain network method is an analysis method which can visually and vividly show dynamic interaction conditions between brain areas in the brain. When limbs move or brain motor imagery is performed, a brain function topological network is constructed, each brain area is taken as a node, the relation among the brain areas is taken as the relation among the nodes, the activity of each brain area is reflected, and the reliability is provided for motor imagery classification. In the traditional functional brain network analysis process, a network constructed according to an original signal is complex, a plurality of connections exist among nodes, in order to better reflect network differences of different motor imagery categories, the network needs to be thinned through threshold setting, however, threshold selection can directly affect the network, too high threshold can affect too few network connections and further affect network connectivity, too low threshold can increase false connections in the network and make the differences not obvious,

compared with other discriminant analysis methods such as LDA (linear discriminant analysis) and DLDA (linear discriminant analysis), the regularized discriminant analysis method is a covariance matrix estimation method, and overcomes the influence caused by high-dimensional data or small-sample singular data through two introduced parameters.

Disclosure of Invention

The technical problem to be solved by the invention is to provide a method for dynamically adjusting a threshold value, wherein the threshold value is dynamically adjusted according to the average degree and the network density so as to furthest retain the information of the network, and then the network parameters are extracted as classification characteristics according to the optimized network.

In order to solve the technical problems, the invention provides an electroencephalogram analysis method based on a brain network and regularized discriminant analysis, which can show the activity state of the brain in the aspect of neural mechanisms and can provide better classification recognition rate. The method comprises the following specific steps:

step 1: and (3) acquiring experimental data: simulating a task scene in real life, collecting electroencephalogram signals of a subject and preprocessing the electroencephalogram signals;

step 2: calculating phase lag coefficients of every two time sequences in the electroencephalogram signal data, and constructing an adjacent matrix, wherein the adjacent matrix is an original brain network;

and step 3: performing dynamic threshold optimization on the original brain network constructed in the step 2;

and 4, step 4: carrying out binarization on the adjacent matrix according to the optimized threshold value obtained in the step 3, and constructing a sparse functional brain network;

and 5: calculating the degree of the brain network constructed in the step 4, and taking the degree as a characteristic; and

step 6: and (5) performing regularized discriminant analysis on the features in the step (5), and performing parameter optimization according to the evaluation indexes to finally obtain the respective recognition rates.

Preferably, in the step 1, a virtual bowl and ball system with dynamic energy constraint is used as an experimental paradigm to simulate an experimental task in real life; and preprocessing the acquired electroencephalogram signals by using a digital band-pass notch filter.

Preferably, in step 3, the dynamic threshold optimization is performed on the data matrix obtained in step 2 according to two indexes, namely, the set network average degree and the set network connection density.

Preferably, in the step 3, an initial value of the threshold is 1; when the adjacency matrix simultaneously satisfies that the brain network average degree is more than 2lnN, N is the number of the brain network nodes, and the brain network connection density is less than 50%, the threshold value is reduced by 0.001, and the step 3 is repeated; otherwise, executing step 4.

Preferably, in step 6, the features extracted from the raw data are decoded by using regularized discriminant analysis, and parameter optimization is performed according to the evaluation index.

Compared with the prior art, the invention has the following beneficial effects:

the invention uses a new brain network method to complete the feature extraction of the electroencephalogram signal, constructs the brain network according to the obtained adjacency matrix in the extraction process, enables the activity state of the brain to be visualized by drawing the degree information of each node in the brain topographic map, can more intuitively observe the neural mechanism of brain activity, and finally obtains the classification result through regularized discriminant analysis, thereby having higher classification recognition rate.

Drawings

FIG. 1 is a diagram of the overall system architecture of an embodiment of the present invention;

FIG. 2 is an electrode distribution diagram of the electroencephalogram signal acquisition process;

FIG. 3 is a schematic diagram of a bowl experiment according to an embodiment of the present invention;

FIG. 4 is a flow chart of an algorithm for functional brain network construction in an embodiment of the present invention;

FIG. 5a is a graphical representation of AUC versus all possible parameters of subject 1 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5b is a graphical representation of AUC versus all possible parameters of subject 2 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5c is a graphical representation of AUC versus all possible parameters of subject 3 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5d is a graphical representation of AUC versus all possible parameters of subject 4 and the AUC values versus the best parameter in an embodiment of the present invention;

FIG. 5e is a graphical representation of AUC versus all possible parameters of subject 5 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5f is a graphical representation of AUC versus all possible parameters of subject 6 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5g is a graphical representation of AUC versus all possible parameters of subject 7 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5h is a graphical representation of AUC versus all possible parameters of subject 8 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5i is a graphical representation of AUC versus all possible parameters of subject 9 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5j is a graphical AUC plot for all possible parameters of subject 10 and AUC values for the best parameter in an embodiment of the present invention;

FIG. 5k is a graphical AUC plot for all possible parameters of subject 11 and AUC values for the best parameter in an embodiment of the present invention;

FIG. 5l is a graphical representation of AUC versus all possible parameters of subject 12 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5m is a graphical representation of AUC versus all possible parameters of subject 13 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5n is a graphical representation of AUC versus all possible parameters of subject 14 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5o is a graphical representation of AUC versus all possible parameters of subject 15 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5p is a graphical representation of AUC versus all possible parameters of subject 16 and AUC values versus the best parameter in an embodiment of the present invention;

FIG. 5q is a graphical AUC plot for all possible parameters of subject 17 and AUC values for the best parameter in an example of the invention;

FIG. 5r is a graphical representation of AUC versus all possible parameters of subject 18 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5s is a graphical representation of AUC versus all possible parameters of subject 19 and AUC values versus the best parameter in an embodiment of the invention;

FIG. 5t is a graphical representation of AUC versus all possible parameters of subject 20 and AUC values versus the best parameter in an embodiment of the present invention;

FIG. 6a is the left and right brain region values obtained after brain network analysis of all subjects using the left hand according to the embodiment of the present invention;

FIG. 6b is a degree distribution topographic map plotted after the degree averaging for all subjects using the left hand in accordance with an embodiment of the present invention;

FIG. 7a is the left and right brain region values obtained after brain network analysis for all subjects using the right hand according to the embodiment of the present invention; and

FIG. 7b is a topographic plot of the degree distribution plotted after the right-handed degree averaging for all subjects according to embodiments of the present invention.

Detailed Description

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

The overall system structure diagram of the embodiment of the invention is shown in fig. 1, and comprises three components, namely keyboard control with visual feedback, brain decision and acquisition of high-dimensional electroencephalogram data; the first part is mainly used for completing the acquisition of electroencephalogram data, the second part is mainly used for completing feature extraction and feature classification, and the third part is mainly used for completing the synthesis of data to obtain high-dimensional electroencephalogram signals.

The electroencephalogram analysis method based on the brain network and the phase fluctuation, which is used by the embodiment of the invention, can better finish decoding information contained in the electroencephalogram signal and provide better classification recognition rate, and comprises the following steps:

step 1: acquiring experimental data, acquiring multi-channel motor imagery electroencephalogram data, and preprocessing:

the subject is provided with an electrode cap, the electrode distribution diagram is shown in fig. 2, 14 channels of scalp electrodes are used according to a standard 10-20 system, the motor imagery data of the left hand and the right hand are recorded, the subject sits in front of a computer and carries out corresponding two types of motor imagery according to the experimental paradigm of fig. 3, the subject carries out systematic trial before the experiment to be familiar with the operation flow, in the formal experiment, a bowl filled with a small ball appears on a screen, the subject completes the control of the virtual bowl through control keyboards "←" and "→", and the whole bowl system is moved from A to B under the condition that the small ball is in the bowl.

The acquired electroencephalogram signals are preprocessed through a digital notch band-pass filter, and an 8-15 Hz band-pass filter and a 50Hz notch filter are arranged. 8-15 Hz is the electroencephalogram signal frequency related to motor imagery, and 50Hz is the power frequency interference signal frequency.

And carrying out data format processing on the preprocessed data, changing the two-dimensional electroencephalogram signals into a four-dimensional tensor form, sequentially intercepting the original two-dimensional data for 1s, and finishing the data to form the data, namely the data in the four-dimensional tensor form of the sample point number multiplied by the channel number multiplied by the experiment times multiplied by the class structure.

Step 2: constructing a brain network, as shown in fig. 4, comprising the following specific steps:

step 21, in the original electroencephalogram data, each channel signal is a time sequence, and the analog signal ψ (t) corresponding to the time sequence s (t) can be subjected to Hilbert transform by the time sequence s (t) and the time sequence s (t)

And (3) calculating to obtain:

in the formula: t represents time, p.v. is the cauchy principal value, a (t) represents the amplitude of the signal,

indicating the phase of the signal.

Instantaneous phase of analog signal psi (t) at time t

Comprises the following steps:

step 22, calculating the instantaneous phase difference of the electroencephalogram signals between the channel i and the channel j as follows:

the phase lag factor (PLI) of the brain electrical signal between channel i and channel j may be calculated by:

in the formula: sign is a sign function;

step 23, constructing an adjacency matrix by calculating phase lag coefficients among all channels, wherein the constructed adjacency matrix is a two-dimensional matrix with the channel number multiplied by the channel number, and each element of the matrix represents the phase lag coefficient among different channels;

and step 3: and (3) performing dynamic threshold optimization on the brain network constructed in the step (2): for the original brain network obtained in step 23, there exists a connection between each node; in order to maximally reserve information contained in the network and make the network sparse, the method optimizes the network according to two indexes of set network average degree and network connection density; before optimization, initializing a threshold value to be 1, and when the network simultaneously satisfies that the network average degree is more than 2lnN, wherein N is the number of network nodes and the network connection density is less than 50%, reducing the threshold value by 0.001 and repeating the step 3; otherwise, carrying out binarization on the adjacent matrix; and constructing a corresponding brain network according to the two-dimensional matrix, wherein connecting lines among different channel nodes in the constructed brain network represent the element sizes of the binary matrix.

And 4, step 4: obtaining a final sparse network;

and 5: calculating the degree D of the network constructed in the step 4_iAnd as a feature:

wherein N is the number of nodes, a_ijRepresenting the degree of connectivity between nodes;

step 6: after the step 5, the degree of each node is extracted from each brain network and is used as a feature, and as each subject performs multiple experiments, a time sequence acquired by each experiment can obtain one brain network, so that the features with the same number as the experiment times can be obtained, and all the obtained features are subjected to regularized discriminant analysis; the regularization discriminant analysis is a covariance matrix estimation method, which can overcome the singularity of various covariances of high-dimensional small sample data, and the embodiment of the invention applies the method to the classification of two different motion characteristics in the motor imagery and can be used for distinguishing two adjustable parameters in the whole method;

the parameter lambda estimation is introduced to estimate the sample covariance for each class as:

in this connection, it is possible to use,

where v represents the amount of observation on the training sample, c (v) represents the observation class,X _vrepresenting the measured values, k representing the category,

the average of the measured values under this category is indicated.

Represents the sum of the weights of each observation under each and every category,

the parameter γ is introduced to control the degree of shrinkage of the estimate of the covariance matrix under each class to the hybrid covariance matrix as:

here, λ and γ represent two parameters in discriminant analysis, P represents the dimension of the argument, and I is an identity matrix in dimension P.

In order to better select two parameters with the best prediction effect, the embodiment of the present invention uses the area under the operation curve of the subject as an evaluation index to complete dynamic selection of the two parameters, and the final results are shown in table 1, fig. 5a, fig. 5b, fig. 5c, fig. 5d, fig. 5e, fig. 5f, fig. 5g, fig. 5h, fig. 5i, fig. 5j, fig. 5k, fig. 5l, fig. 5m, fig. 5n, fig. 5o, fig. 5p, fig. 5q, fig. 5r, fig. 5s, and fig. 5 t:

TABLE 1 values of parameters obtained by dynamic selection

In order to verify the effectiveness of the invention, the embodiment of the invention is used on a data set to complete verification, the experimental process is as follows, firstly, preprocessing of data is completed, network construction is carried out on the preprocessed electroencephalogram data, in order to enable the activity state of a brain to become visual and enable the neural mechanism of brain activity to be observed more intuitively, degree distribution of each subject in the data set is expressed in the form of a histogram, as shown in fig. 6a and 7a, degree information of all subjects is averaged and then drawn on a brain topographic map, as shown in fig. 6b and 7b, then, optimization of the network is completed, the network is changed into a sparse network, the degree is taken as a feature, finally, extracted features are classified by utilizing regularized discriminant analysis, and the final experimental result is shown in table 2:

TABLE 2 regularization discriminant analysis classification accuracy on this data

According to the table 2, the method obtains good classification accuracy in classification of the electroencephalogram signals.

In summary, the embodiment of the invention designs an electroencephalogram analysis method based on a brain network and regularized discriminant analysis, and the method can be used for electroencephalogram intention analysis in a system for neural rehabilitation.

The above-mentioned embodiments are merely illustrative of the preferred embodiments of the present invention, and do not limit the scope of the present invention, and various modifications and improvements made to the technical solution of the present invention by those skilled in the art without departing from the spirit of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. An electroencephalogram analysis method based on brain network and regularized discriminant analysis is characterized by comprising the following steps:

2. The brain network and regularized discriminant analysis-based electroencephalogram analysis method according to claim 1, wherein in the step 1, a virtual bowl system with dynamic energy constraints is used as an experimental paradigm to simulate an experimental task in real life; and preprocessing the acquired electroencephalogram signals by using a digital band-pass notch filter.

3. The electroencephalogram analysis method based on brain network and regularized discriminant analysis according to claim 1, wherein in the step 3, the data matrix obtained in the step 2 is subjected to dynamic threshold optimization according to two indexes of set network average degree and network connection density.

4. The brain network and regularized discriminant analysis-based electroencephalogram analysis method of claim 3, wherein in said step 3, an initial value of said threshold is 1; when the adjacency matrix simultaneously satisfies that the brain network average degree is more than 2lnN, N is the number of the brain network nodes, and the brain network connection density is less than 50%, the threshold value is reduced by 0.001, and the step 3 is repeated; otherwise, executing step 4.

5. The electroencephalogram analysis method based on the brain network and the regularized discriminant analysis according to claim 1, wherein in the step 6, the features extracted from the raw data are decoded by using the regularized discriminant analysis, and parameter optimization is completed according to the evaluation index.