阅读说明：本技术 一种基于张量分解模型的多元脑电数据识别分析方法 (Multi-element electroencephalogram data identification and analysis method based on tensor decomposition model ) 是由 安钰 冯毅隆 陈善恩 张玺 于 2021-11-19 设计创作，主要内容包括：本发明公布了一种基于张量分解模型的多元脑电数据识别分析方法,利用采集到的多元脑电信号,建立基于张量分解模型的多元脑电数据识别分析模型,实现对脑电数据的分析与识别；包括：利用CP张量分解方法,构建基于多元脑电数据的张量分解模型,作为多元脑电数据识别分析模型；建立用于建模的脑电信号数据的分类类别信息约束项,即标签矩阵；利用多元脑电数据识别分析模型进行个体脑电信号数据的分类识别。采用本发明提供的技术方案,有助于实现对脑电数据分类类别的识别分析,提升数据识别的准确性,为远程个体数据的分析识别提供启发式的辅助,能够为个体提供切身的便利与帮助。(The invention discloses a multivariate electroencephalogram data identification and analysis method based on a tensor decomposition model, which is characterized in that a multivariate electroencephalogram data identification and analysis model based on the tensor decomposition model is established by utilizing collected multivariate electroencephalogram signals, so that electroencephalogram data are analyzed and identified; the method comprises the following steps: constructing a tensor decomposition model based on the multivariate electroencephalogram data by using a CP tensor decomposition method, and using the tensor decomposition model as a multivariate electroencephalogram data identification analysis model; establishing a classification category information constraint item, namely a label matrix, of the electroencephalogram signal data for modeling; and carrying out classification and identification on the individual electroencephalogram signal data by utilizing the multivariate electroencephalogram data identification analysis model. By adopting the technical scheme provided by the invention, the electroencephalogram data classification type identification and analysis can be realized, the data identification accuracy is improved, heuristic assistance is provided for the analysis and identification of remote individual data, and the individual can be provided with the body-cutting convenience and help.)

1. A multivariate electroencephalogram data identification and analysis method based on a tensor decomposition model is characterized in that a multivariate electroencephalogram data identification and analysis model based on the tensor decomposition model is established by utilizing collected multivariate electroencephalogram signals, and identification and analysis of electroencephalogram data are realized; the method comprises the following steps:

1) constructing a tensor decomposition model based on multivariate electroencephalogram data by using a CP tensor decomposition method, and using the tensor decomposition model as a multivariate electroencephalogram data identification analysis model; the method comprises the following steps:

11) defining an input tensor for electroencephalogram data；

Representing each EEG signal segment as an adjacent matrix through mutual information，Representing the number of electrodes of the electroencephalogram data collector; defining an input tensor by stacking a contiguous matrix of all of the segments of the electroencephalographic signalI.e. by(ii) a Wherein the content of the first and second substances,value of 1 to；Is the total number of all the electroencephalogram signal segments,；representing the number of individuals used for modeling,representing the number of individuals for identification;is as followsThe number of electroencephalogram signal segments for an individual,；

12) inputting tensor by CP tensor decomposition methodIs decomposed intoA group vector, represented by formula 1:

formula 1

Wherein the content of the first and second substances,is the input tensor(ii) an estimate of (d);is a unit tensor, i.e.WhereinRepresenting a dirac function;is an electroencephalogram electrode information matrix;an electroencephalogram fragment information matrix;is the first of tensorA modal product;

2) establishing classification category information constraint items, namely label matrixes, of electroencephalogram signal data for modeling；

Defining an objective function for minimizing the distance between two electroencephalogram segments with similar physiological states, and expressing the objective function as formula 2:

formula 2

Wherein the content of the first and second substances,is composed ofThe 3 rd modality expansion matrix of (a);is the total number of all electroencephalogram signal segments used for modeling; label matrix，Representing the number of classifications of the electroencephalogram data;is based on a label matrixOf the core matrix ofRepresents the firstA fragment and the firstSimilarity between individual segments;is thatThe Frobenius norm of (a);

3) carrying out classification and identification on individual electroencephalogram signal data;

defining electroencephalogram fragment information matrixThe corresponding mapping matrix isThe identification process of the electroencephalogram data is expressed as a formula 3 by using ridge regression:

formula 3

Wherein the content of the first and second substances,representing that the existing individual electroencephalogram signal segment is used only in the training process;is a parameter of the control constraint;representation matrixThe Frobenius norm of (a);

and (3) carrying out classification and identification on the individual electroencephalogram signal data by solving the optimization solution of the formula 4:

formula 4

Wherein the content of the first and second substances,is the parameter of all control items;

the optimized solution obtained by solving is recorded asWhereinSolving the obtained electroencephalogram electrode information matrix;is to solve the obtained electroencephalogram segment information matrix,is a solved electroencephalogram fragment information matrixA corresponding mapping matrix;

according toAnd obtaining a label matrix of the electroencephalogram data to be identified, namely identifying the classification information of the electroencephalogram data, thereby realizing the analysis and prediction of the multivariate electroencephalogram data based on the tensor decomposition model.

2. The method for multivariate electroencephalogram data identification and analysis based on tensor decomposition model as set forth in claim 1, wherein step 12) is to input tensorIs decomposed intoGroup vectors, each group comprising three vectors,individual vectors of the same type are represented as a corresponding matrix.

3. The tensor decomposition model-based multivariate electroencephalogram data identification and analysis method as recited in claim 1, wherein a label matrixIn (1), each row contains only 1 "element anda "0" element.

4. The method for multivariate electroencephalogram data identification and analysis based on tensor decomposition model as recited in claim 1, wherein a kernel function is applied to solve and analyze equation 4.

5. The method for multivariate electroencephalogram data identification and analysis based on tensor decomposition model as recited in claim 1, which is characterized in that a multi-block multiplier alternating direction method is adopted to solve formula 4.

6. The method as claimed in claim 1, wherein the classification categories of the electroencephalogram data represented in the tag matrix include pre-onset period and inter-onset period, and are represented by one-hot coding.

7. The method for multivariate electroencephalogram data identification and analysis based on tensor decomposition model as recited in claim 1, wherein definition is performedWhereinIs the tag matrix of the new individual; obtained according to solutionAnd obtaining the label matrix of the electroencephalogram data to be identified, thereby obtaining the category information of the electroencephalogram data to be identified.

8. Tensor decomposition model based multi as claimed in claim 1The method for identifying and analyzing the meta-electroencephalogram data is characterized by controlling parametersAre all positive values.

Technical Field

The invention provides a multivariate electroencephalogram data identification and analysis technology, in particular relates to a multivariate electroencephalogram data identification and analysis method based on a tensor decomposition model, and belongs to the technical field of industrial engineering data analysis.

Background

Sudden abnormal firing of neurons in the brain can lead to transient cerebral dysfunction. With the development of communication and computing technologies, the related technologies are widely applied to acquisition, identification and analysis of electroencephalogram data of brain neurons. The electroencephalogram data acquisition cost is low, the flexibility is high, the time resolution is high, and the electroencephalogram data acquisition method has the characteristics of non-invasiveness, usability, portability, safety and the like. The electroencephalogram data contains a large amount of information with potential value, and plays an important role in the explanation of remote medical treatment and pathological mechanisms and the like. Therefore, automatic identification and analysis of multivariate electroencephalographic data is an important step to reduce the dependence on professionals.

The electroencephalogram data is usually specific to a specific object, and is reflected in that the height of the electroencephalogram data is influenced by various individual differences. For example, the physiological records of different patients with epilepsy can vary widely. In recent years, academic and industrial people have developed many methods for the identification and analysis of brain electrical data. The prior art focuses more on the subject-specific modeling of the electroencephalogram data of a specific individual population, and a subject-independent modeling method based on the electroencephalogram data of a non-specific individual population in an actual scene is lacked.

At present, the identification and analysis of multivariate electroencephalogram data still face the following challenges: firstly, electroencephalogram data has properties of non-linearity, non-stationarity, non-gaussian property and the like which are difficult to capture. Second, the characteristics of the electroencephalogram data are complex and difficult to distinguish. For example, the characteristics of the electroencephalogram data of the pre-seizure stage and the inter-seizure stage are complex and difficult to distinguish, and difficulty is brought to the identification of the electroencephalogram data of different stages. Thirdly, the individual difference of the electroencephalogram data is obvious, so that the modeling result based on the existing electroencephalogram data is not high in applicability and difficult to effectively apply.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides the multivariate electroencephalogram data identification and analysis method based on the tensor decomposition model, which fully considers the specificity among different individual electroencephalogram data, establishes the multivariate electroencephalogram data identification and analysis model based on the tensor decomposition model by utilizing the collected multivariate electroencephalogram data, and realizes the identification and analysis of new electroencephalogram data.

The technical scheme provided by the invention is as follows:

a multivariate electroencephalogram data identification and analysis method based on a tensor decomposition model is characterized in that a multivariate electroencephalogram data identification and analysis model based on tensor decomposition is established by utilizing collected individual multivariate electroencephalogram signals, and electroencephalogram data are analyzed and identified. The method comprises the following steps:

1) constructing a tensor decomposition model based on multivariate electroencephalogram data by using a CP tensor decomposition method, and using the tensor decomposition model as a multivariate electroencephalogram data identification analysis model; the method comprises the following steps:

11) defining an input tensor for electroencephalogram data；

Representing each EEG signal segment as an adjacent matrix through mutual information，Representing the number of electrodes of the electroencephalogram data collector; defining an input tensor by stacking a contiguous matrix of all of the segments of the electroencephalographic signalI.e. by(ii) a Wherein the content of the first and second substances,is the total number of all the electroencephalogram signal segments,；representing the number of individuals used for modeling,representing the number of individuals for identification;is as followsThe number of electroencephalogram signal segments for an individual,；

12) inputting tensor by CP tensor decomposition methodIs decomposed intoA group vector, represented by formula 1:

(formula 1)

Wherein the content of the first and second substances,is the input tensor(ii) an estimate of (d);is a unit tensor, i.e.WhereinRepresenting dirac functions, i.e. only if satisfiedUnder the condition ofIs 1, and the rest cases are 0;is an electroencephalogram electrode information matrix;an electroencephalogram fragment information matrix;is the first of tensorThe modal product.

2) Establishing classification category information constraint items, namely label matrixes, of electroencephalogram signal data for modeling；

If the two electroencephalogram fragment data are similar in physiological state, the two electroencephalogram fragment data are similar to each other in physiological data structure. We define the objective function that minimizes the distance between two brain electrical segments with similar physiological states as:

(formula 2)

WhereinIs the input tensor estimate obtained by equation 1The 3 rd modality expansion matrix of (a);is the total number of all electroencephalogram signal segments used for modeling;is based on a label matrixOf the core matrix ofRepresents the firstA fragment and the firstSimilarity between individual segments.

3) Establishing an individual electroencephalogram signal data classification identification method;

the method provided by the invention is designed for further analyzing and identifying the electroencephalogram data state of a non-specific individual. Defining electroencephalogram fragment information matrixThe corresponding mapping matrix isThe identification and analysis process of the electroencephalogram data can be expressed as formula 3 by using ridge regression.

(formula 3)

WhereinOnly the electroencephalogram signal segments of the existing individuals are ensured to be used in the training process;is an electroencephalogram fragment information matrix;is an electroencephalogram fragment information matrixA corresponding mapping matrix;is a label matrix for the modeled electroencephalogram signal data;is a parameter that controls the constraint term and takes a positive value.Representation matrixFrobenius norm of (1).

Therefore, the method proposed by the present invention can be expressed as solving the following optimization problem:

(formula 4)

WhereinAll the parameters controlling each item take positive values.

In the present invention, we can apply various kinds of kernel functions for analysis. The optimization problem in the formula 4 is solved by adopting a multi-block multiplier alternating direction method, and the solved optimization solution is recorded asWhereinSolving the obtained electroencephalogram electrode information matrix;is to solve the obtained electroencephalogram segment information matrix,is a solved electroencephalogram fragment information matrixA corresponding mapping matrix. According toThe information of the electroencephalogram data to be identified can be obtained, so that the category of the electroencephalogram data of the segment can be analyzed and judged. Through the steps, multivariate electroencephalogram data identification and analysis based on a tensor decomposition model are realized.

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

the invention provides a multivariate electroencephalogram data identification and analysis method based on a tensor decomposition model, which is characterized in that a tensor decomposition model framework based on multivariate electroencephalogram data is constructed by utilizing collected multivariate electroencephalogram signals, a classification category information constraint item of electroencephalogram signal data for modeling is established, an individual electroencephalogram signal data classification identification method is established, and identification of electroencephalogram data to be identified is further completed. By adopting the technical scheme provided by the invention, the identification of classification categories of the electroencephalogram data is facilitated, the identification accuracy is improved, the body-cutting convenience and help can be provided for individuals, and heuristic auxiliary guidance is provided for remote medical treatment and precise medical treatment.

Drawings

FIG. 1 is a block flow diagram of the method of the present invention.

Detailed Description

The invention will be further described by way of examples, without in any way limiting the scope of the invention, with reference to the accompanying drawings.

The invention provides a multivariate electroencephalogram data identification and analysis method based on a tensor decomposition model, and a process of the method is shown in figure 1 and comprises the following steps: the collected multivariate electroencephalogram signals are utilized to construct a tensor decomposition model frame based on multivariate electroencephalogram data, a classification category information constraint item of the electroencephalogram signal data for modeling is established, an individual electroencephalogram signal data classification identification method is established, and identification of the electroencephalogram data to be identified is further completed. By adopting the technical scheme provided by the invention, the identification of classification categories of the electroencephalogram data is facilitated, the body-cutting convenience and help can be provided for individuals, and heuristic auxiliary guidance is provided for remote medical treatment and precise medical treatment.

In the present invention, settingRepresenting the number of classes of individual electroencephalogram data, i.e. having electroencephalogram dataA category;representing the number of electrodes of the electroencephalogram data collector;representing the number of individuals used for modeling (data in the modeling database),representing the number of individuals (new individuals not included in the database) used for identification.Is as followsThe number of individual electroencephalogram signal segments, where each electroencephalogram signal segment belongs to and only belongs to one category;as the total number of all electroencephalogram signal segments, i.e.；Is the number of segments of the individual electroencephalographic signal data (in the modeling database) used for modeling, i.e.；The number of segments of the electroencephalogram signal data (data of new individuals not contained in the database) for identification, i.e.. Each EEG signal segment can be represented as a contiguous matrix by mutual information (mutualformation). The input tensor of the present invention can be defined by stacking a contiguous matrix of all the electroencephalogram signal segmentsThat is to say that,. Defining a tag matrix of electroencephalogram signal data for modeling asThe label matrixContains only 1 "element per row anda "0" element. The multivariate electroencephalogram data identification and analysis method based on the tensor decomposition model provided by the invention utilizes the collected individual multivariate electroencephalogram signals to establish the multivariate electroencephalogram data identification and analysis model based on the tensor decomposition, thereby realizing the analysis and identification of the electroencephalogram data.

The following examples illustrate actual collected electroencephalographic data from epileptic patients, including 14 epileptic patients (8 males, age: 25-71 years; 6 females, age: 20-58 years). The data contained a record of the patient's brain activity obtained by taking scalp brain electrical data using a 512 Hz sampling rate, with the electrodes arranged according to the international 10-20 system. The clinical and electrophysiological data for each patient have been carefully revised by clinical experts. Seizure electroencephalography records include inter-seizure, pre-seizure, and seizure periods. The method realizes the multivariate electroencephalogram data identification and analysis based on the tensor decomposition model, provides assistance for identifying the concerned epilepsy stages, and pays more attention to the identification and analysis of the early stage of the seizure so as to further take intervention measures in time. Specifically, a 5 minute period of 1 minute to 6 minutes prior to the seizure may be used as the pre-seizure phase signal for each seizure in each subject. In the same manner, a 5 minute time period of 50 to 55 minutes prior to the seizure was taken as the interpulse signal for each seizure in each subject. The specific implementation mode for identifying the new electroencephalogram data to be identified based on the collected electroencephalogram data of the epileptic patient by utilizing the method of the invention is as follows:

5 patients (PN 01, PN03, PN09, PN13, PN 14) were scored as existing patients, and 1 patient (PN 10) as new. Consider thatIndicates the number of categories of the subject,representing the number of electrodes collecting brain electrical data.Indicating the number of patients that are present,indicating the number of new patients.Is as followsThe number of segments of the electroencephalogram signal of each patient, where each segment belongs to and belongs only to one category;the total number of segments of the electroencephalogram signal for all patients, i.e.；Is the total number of electroencephalogram signal segments of all existing patients, i.e.；The total number of segments of the electroencephalogram signal for all new patients, i.e.. Each EEG signal segment can be represented as a contiguous matrix by mutual information (mutualformation). The input tensor of the present invention can be defined by stacking a contiguous matrix of all the electroencephalogram signal segmentsThat is to say that,. Defining a tag matrix of electroencephalogram signal data for modeling asEach row containing only 1 "element andand a "0" element representing classification category information of the electroencephalogram signal data.

1) Constructing a tensor decomposition model based on the multivariate electroencephalogram data;

by using CP (CANDECOMP/PARAFAC) tensor decomposition method, the electroencephalogram recognition model is established by using multivariate electroencephalogram data, and the input tensor can be decomposed by the CP decomposition methodIs decomposed intoGroup vectors, each group comprising three vectors,the vectors of the same type can be represented by a corresponding matrix, and the CP decomposition is expressed as formula 1:

(formula 1)

Wherein the content of the first and second substances,is the input tensor(ii) an estimate of (d);is a unit tensor, i.e.；Is an electroencephalogram electrode information matrix;an electroencephalogram fragment information matrix;is the first of tensorThe modal product.

2) Establishing classification category information constraint items, namely label matrixes, of electroencephalogram signal data for modeling；

If the two electroencephalogram fragment data are similar in physiological state, the two electroencephalogram fragment data are similar to each other in physiological data structure. We define the objective function that minimizes the distance between two brain electrical segments with similar physiological states as:

(formula 2)

WhereinIs the input tensor estimate obtained by equation 1The 3 rd modality expansion matrix of (a);is the total number of all electroencephalogram signal segments used for modeling;is based on a label matrixOf the core matrix ofRepresents the firstA fragment and the firstSimilarity between individual segments.

3) Establishing an individual electroencephalogram signal data classification identification method;

the method provided by the invention is designed for further analyzing and identifying the electroencephalogram data state of a non-specific individual. Defining electroencephalogram fragment information matrixThe corresponding mapping matrix isThe identification and analysis process of the electroencephalogram data can be expressed as formula 3 by using ridge regression.

(formula 3)

WhereinOnly the electroencephalogram signal segments of the existing individuals are ensured to be used in the training process;is an electroencephalogram fragment information matrix;is an electroencephalogram fragment information matrixA corresponding mapping matrix;is a label matrix for the modeled electroencephalogram signal data;is a parameter that controls the constraint term and takes a positive value.Representation matrixFrobenius norm of (1).

Therefore, the method proposed by the present invention can be expressed as solving the following optimization problem:

(formula 4)

WhereinAll the parameters controlling each item take positive values.

In the present invention, we can apply various kinds of kernel functions for analysis. The optimization problem in the formula 4 is solved by adopting a multi-block multiplier alternating direction method, and the solved optimization solution is recorded asWhereinSolving the obtained electroencephalogram electrode information matrix;is to solve the obtained electroencephalogram segment information matrix,is a solved electroencephalogram fragment information matrixA corresponding mapping matrix. Then according toInformation of the electroencephalogram data to be recognized can be obtained, whereinThe label matrix is a label matrix of a new individual, the label matrix represents classification category information (the category comprises pre-attack period and inter-attack period) of the electroencephalogram signal data, and the classification category information is represented by One-Hot Encoding (One-Hot Encoding). Through the steps, multivariate electroencephalogram data identification and analysis based on a tensor decomposition model are realized.

The results of comparing the present invention with two advanced deep learning methods in the current academia, MSSA-TN and CCNN, and some classical statistical learning methods such as SVM, LR, KNN, RF and LDA are shown in Table 1. As can be seen from Table 1, the accuracy of the method provided by the invention is superior to that of two deep learning methods and a classical statistical learning method provided by the current academia, which illustrates the superiority of the invention.

TABLE 1 comparison of the accuracy of identification of the method of the invention with existing models

Meanwhile, the method provided by the invention is subjected to an ablation experiment, and the result is shown in table 2. As can be seen from table 2, the constraint term proposed by the present invention (i.e., equation 2) is significant. Compared with the method without constraint items, the accuracy of analyzing and identifying the electroencephalogram data categories by adopting the method provided by the invention is improved by 11.62%.

TABLE 2 model ablation test results

It is noted that the disclosed embodiments are intended to aid in further understanding of the invention, but those skilled in the art will appreciate that: various alternatives and modifications are possible without departing from the invention and scope of the appended claims. Therefore, the invention should not be limited to the embodiments disclosed, but the scope of the invention is defined by the appended claims.