Multivariate electroencephalogram and electromyogram coupling analysis method and system based on structural sparse regularization

Abstract

The application discloses a multivariate brain-muscle coupling analysis method and system based on structural sparse regularization. The method comprises the following steps: synchronously collecting brain-muscle signals; respectively performing data preprocessing on the collected brain-muscle signals, and dividing the preprocessed data into a plurality of different sample data segments; performing frequency domain analysis on the brain-muscle signals in the sample data segments by using a fast Fourier algorithm, so as to calculate the self-power spectrum matrix of the brain-muscle signals and the cross spectrum matrix therebetween; based on the self-power spectrum matrix and the cross spectrum matrix, and by using a structural sparse typical coherence algorithm, extracting the cortical muscle coupling characteristics between the multi-lead brain-muscle signals, so as to reconstruct a brain topography related to the cortical muscle coupling characteristics. The application can robustly identify the cortical muscle coupling characteristics of healthy people and stroke patients and stably recover the brain topography related to the cortical muscle coupling under the conditions of low sample quantity and high noise level.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of brain-muscle electro signal coupling analysis, in particular to a multivariate brain-muscle electro coupling analysis method and system based on structural sparse regularization. BACKGROUND

[0002] Rhythmic synchronization of neurons plays a crucial role in various brain functions such as motor control, cognition, and sensory processing. Effective estimation of these neuronal interactions is an important area of research in brain structure and functional remodeling in neurodegenerative diseases, and multi-modal fusion analysis methods combining various neurophysiological and imaging techniques have become advanced means of measuring neuronal connections. In recent years, the development of neurophysiological acquisition technology has made it possible to synchronize the acquisition of electroencephalogram (EEG) and electromyogram (EMG), and functional corticomuscular coupling (FCMC) has been widely used to analyze the information exchange between the brain and muscles, thereby helping to understand the pathological mechanism of post-stroke movement disorders.

[0003] Many multi-modal fusion analysis methods for FCMC mainly focus on single variable coupling features between single sensors, often ignoring the spatiotemporal characteristics of central-peripheral neural information transmission. Several recent studies have proposed multivariate analysis methods based on canonical correlation analysis, but due to the high-dimensional characteristics and noise levels of brain-muscle electro data, these multivariate methods based on spatial optimization often face overfitting problems in feature extraction.

[0004] In Multiscale Canonical Coherence for Functional Corticomuscular Coupling Analysis, a multiscale canonical coherence algorithm was proposed to reduce the frequency band aliasing effect and avoid overfitting problems through a multiscale framework and multivariate method, thereby improving the performance of coupling feature extraction. However, the above research still has some shortcomings: first, principal component analysis (PCA) is needed as a preprocessing step to achieve feature selection for each modality, but the optimal hyperparameters of PCA are difficult to determine. In addition, when dealing with high-dimensional, low signal-to-noise ratio brain-muscle electro data, this algorithm is still prone to overfitting problems. SUMMARY

[0005] The present application aims to at least partially solve one of the technical problems in the related art.

[0006] To this end, the application provides a multivariate brain-muscle electrical coupling analysis method based on structural sparse regularization, which overcomes the overfitting problem of the existing FCMC feature extraction method.

[0007] Another object of the application is to provide a multivariate brain-muscle electrical coupling analysis system based on structural sparse regularization.

[0008] To achieve the above object, the application provides a multivariate brain-muscle electrical coupling analysis method based on structural sparse regularization, which comprises:

[0009] Synchronously collecting brain and muscle signals;

[0010] Respectively performing data preprocessing on the collected brain and muscle signals, and dividing the preprocessed data into a plurality of different sample data segments;

[0011] Performing frequency domain analysis on the brain and muscle signals in the sample data segments by using a fast Fourier algorithm to calculate the self-power spectrum matrix of the brain and muscle signals and the cross-spectrum matrix therebetween;

[0012] Based on the self-power spectrum matrix and the cross-spectrum matrix, and by using a structural sparse canonical coherence algorithm, extracting the cortical muscle coupling features between the multi-lead brain and muscle signals to reconstruct a brain topography related to the cortical muscle coupling features, comprising:

[0013] Defining the cross-spectrum matrix as:

[0014]

[0015] Wherein, X∈C n×p represents a complex data matrix of the electroencephalogram signal, Y∈C n×q represents a complex data matrix of the electromyogram signal, n represents the number of data segments, p represents the number of leads of the electroencephalogram, q represents the number of leads of the electromyogram, and H represents the Hermitian transpose; the optimization problem of the structural sparse canonical coherence algorithm is as follows:

[0016]

[0017] Wherein, α∈R p×1 and β∈R q×1 are real value projection vectors of the electroencephalogram signal and the electromyogram signal respectively, φ is an auxiliary variable for solving the optimization problem, I is a unit matrix, λ1, λ2, γ1 and γ2 are regularization hyperparameters of the optimization problem, L α and L β are Laplacian matrices of the electroencephalogram signal and the electromyogram signal respectively, and D-C is the difference between a diagonal matrix D and an adjacency matrix C.

[0018] To solve the optimization problem (2), an alternating optimization algorithm is adopted as shown below: initializing α and β as unit vectors, and φ as π / 4; fixing the optimization parameters β and φ, the original optimization problem is simplified as:

[0019]

[0020] The approximate point operator f is defined as the following form:

[0021]

[0022] For the unconstrained convex optimization problem (3), the gradient descent method is re-formulated as a proximal regularization form:

[0023]

[0024] Wherein, is the 1-norm of the projection vector α, T H is the real part of Yexp(-iφ), P α is the 1-norm of the projection vector α, T α is the iterative step length of the projection vector α; the LASSO problem (5) is solved by using a fast iterative threshold shrinkage algorithm (FISTA);

[0025] When fixing the optimization parameters α and φ, the original optimization problem is simplified as:

[0026]

[0027] Meanwhile, the iterative formula of the optimization parameter β is:

[0028]

[0029] Wherein, β is the 1-norm of the projection vector β, T β is the iterative step length of the projection vector β;

[0030] When fixing the optimization parameters α and β, the original optimization problem is simplified as:

[0031]

[0032] The optimal solution of the optimization parameter φ is solved by using a Levenberg-Marquardt algorithm, and after several iterations, the optimal solution of the original optimization problem (2) is obtained.

[0033] The multivariate brain-muscle coupling analysis method based on structural sparse regularization can further have the following additional technical features:

[0034] In an embodiment of the present application, the brain-muscle signals are synchronously collected, comprising:

[0035] The electroencephalogram acquisition system and the wireless electromyogram acquisition system are used for synchronously acquiring electroencephalogram signals and electromyogram signals, wherein the sampling frequencies are both 1000 Hz.

[0036] In one embodiment of the present application, the acquired brain-muscle signals are respectively subjected to data preprocessing, including:

[0037] The electroencephalogram signals are subjected to common average reference, and are subjected to zero-phase filtering by using a FIR filter to obtain the electroencephalogram signals after primary processing, and are subjected to 50 Hz power frequency noise removal by using the standardized preprocessing framework of EEGLAB, and are subjected to artifact interference removal by using an independent component analysis method, so as to obtain the electroencephalogram signals after preprocessing.

[0038] The electromyogram signals are subjected to processing by using a FIR filter of the same frequency band, so as to obtain the electromyogram signals after preprocessing.

[0039] In one embodiment of the present application, the brain-muscle signals in the sample data segment are subjected to frequency domain analysis by using a fast Fourier algorithm, so as to calculate the self-power spectrum matrix of each brain-muscle signal and the cross spectrum matrix therebetween, including:

[0040] The discrete Fourier spectrum X of the electroencephalogram signal x is calculated by using a fast Fourier algorithm, and the discrete Fourier spectrum Y of the electromyogram signal y is calculated by using a fast Fourier algorithm. n×p n×q n×p

[0041] The electroencephalogram self-power spectrum matrix S is obtained, the electromyogram self-power spectrum matrix S is obtained, and the electroencephalogram-electromyogram cross spectrum matrix S is obtained. XX H YY H XY H

[0042] To achieve the above purpose, another aspect of the present application provides a multivariate brain-muscle electrocoupling analysis system based on structural sparse regularization, including:

[0043] A brain-muscle signal acquisition module is used for synchronously acquiring brain-muscle signals.

[0044] A data preprocessing module is used for respectively subjecting the acquired brain-muscle signals to data preprocessing, and dividing the data after preprocessing into a plurality of different sample data segments.

[0045] A frequency domain analysis module is used for subjecting the brain-muscle signals in the sample data segment to frequency domain analysis by using a fast Fourier algorithm, so as to calculate the self-power spectrum matrix of each brain-muscle signal and the cross spectrum matrix therebetween.

[0046] ​​​​​​​​​A coherence analysis module is configured to extract cortical muscle coupling features between multi-channel brain and muscle signals based on a self-power spectrum matrix and a cross-spectrum matrix, and by a structured sparsified canonical coherence algorithm, so as to reconstruct a brain topography related to the cortical muscle coupling features; the coherence analysis module comprises:

[0047] The cross-spectrum matrix is defined as:

[0048]

[0049] Wherein, X∈C n×p represents a complex data matrix of the electroencephalogram signal, Y∈C n×q represents a complex data matrix of the electromyogram signal, n represents a data segment number, p represents a number of leads of the electroencephalogram, q represents a number of leads of the electromyogram, and H represents a Hermitian transpose; an optimization problem of the structured sparsified canonical coherence algorithm is as follows:

[0050]

[0051] Wherein, α∈R p×1 and β∈R q×1 are real-value projection vectors of the electroencephalogram signal and the electromyogram signal respectively, φ is an auxiliary variable for solving the optimization problem, I is a unit matrix, λ1, λ2, γ1 and γ2 are regularization hyperparameters of the optimization problem, L α and L β are Laplacian matrices of the electroencephalogram signal and the electromyogram signal respectively, and D-C is a difference between a diagonal matrix D and an adjacency matrix C.

[0052] The optimization problem (2) is solved by using an alternating optimization algorithm as shown below: the α and β are initialized as unit vectors, and the φ is taken as π / 4; the optimization parameters β and φ are fixed, and the original optimization problem is simplified as:

[0053]

[0054] The approximation point operator f is defined as the following form:

[0055]

[0056] For the unconstrained convex optimization problem (3), the gradient descent method is re-formulated as a proximal regularization form:

[0057]

[0058] Wherein, is a real part of X H Yexp(-iφ), P α is a 1-norm of the projection vector α, T α is an iterative step size of the projection vector α; the LASSO problem (5) is solved by using a fast iterative threshold shrinkage algorithm (FISTA).

[0059] When fixing the optimization parameters alpha and phi, the original optimization problem is simplified as:

[0060]

[0061] Meanwhile, the iterative formula of the optimization parameter beta is:

[0062]

[0063] Wherein, P β is the 1-norm of the projection vector beta, T β is the iterative step length of the projection vector beta;

[0064] When fixing the optimization parameters alpha and beta, the original optimization problem is simplified as:

[0065]

[0066] The optimal solution of the optimization parameter phi is solved by using the Levenberg-Marquardt algorithm, and after several iterations, the optimal solution of the original optimization problem (2) is obtained.

[0067] The multivariate brain-muscle coupling analysis method and system based on structural sparse regularization can robustly identify the cortical muscle coupling characteristics of healthy people and stroke patients and robustly recover the brain topography related to the cortical muscle coupling under low sample size and high noise level, and can be used as a potential biomarker for post-stroke motor function recovery.

[0068] Additional aspects and advantages of the application will be set forth in part in the description which follows, and in part will become apparent to those skilled in the art upon examination of the following and / or can be learned by practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0069] The above and / or additional aspects and advantages of the application will become apparent and be readily understood from the following description, taken in conjunction with the accompanying drawings, in which:

[0070] Figure 1 is a flowchart of the multivariate brain-muscle coupling analysis method based on structural sparse regularization according to the embodiments of the application;

[0071] Figure 2 is an algorithm framework diagram of structural sparse canonical coherence according to the embodiments of the application;

[0072] Figure 3 is a feature extraction performance comparison diagram of the structural sparse canonical coherence algorithm in a simulation data set according to the embodiments of the application;

[0073] Figure 4is a feature extraction performance comparison chart of a typical coherence algorithm based on structural sparsification according to an embodiment of the present application in a healthy person data set;

[0074] Figure 5 is a structural diagram of a multivariate brain electromyographic coupling analysis system based on structural sparse regularization according to an embodiment of the present application. DETAILED DESCRIPTION

[0075] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0076] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0077] The multivariate brain electromyographic coupling analysis method and system based on structural sparse regularization according to the embodiments of the present application will be described below with reference to the accompanying drawings.

[0078] Figure 1 is a flowchart of a multivariate brain electromyographic coupling analysis method based on structural sparse regularization according to an embodiment of the present application, as shown in the figure, the method comprises: Figure 1

[0079] S1, synchronously collecting brain and muscle signals.

[0080] Specifically, in the brain and muscle electrical signal synchronous collection part of the present application, a 64-channel Neuracle.W64 electroencephalogram acquisition system and a Neuracle.WM wireless electromyographic acquisition system can be used to synchronously collect electroencephalogram signals and electromyographic signals, and the sampling frequency is 1000Hz. The electromyographic device collects the activities of seven muscles, including the first dorsal interosseous muscle, the finger flexor muscle, the finger extensor muscle, the radial wrist flexor muscle, the ulnar wrist flexor muscle, the ulnar wrist extensor muscle, and the brachioradialis muscle. The electrodes are placed in the muscle belly. During the experiment, the subject is required to perform a grip force control task with the right hand. Specifically, the subject is required to control the force to be as close as possible to the preset target force, and the peak value of the preset target force is 20% of the maximum voluntary contraction. After 3s of each test, the subject performs a 6s static grip force control task, and finally performs a 3s rest.

[0081] It can be understood that by synchronously collecting electroencephalogram and electromyographic signals, the time consistency between the two is ensured, providing a reliable basis for subsequent coupling analysis.

[0082] ​It can be understood that the high sampling frequency and multi-channel setting ensure high resolution of data, improve signal quality and reduce noise influence.

[0083] It can be understood that the design of the grip control task ensures consistency and repeatability of experimental conditions, making the results between different subjects comparable. The setting of the preset target force helps to standardize the experimental task, ensuring that each subject completes the task under similar conditions, thereby improving the effectiveness of data analysis.

[0084] In summary, the present application realizes high-quality, synchronous brain and muscle electrical signal acquisition by using advanced brain and muscle electrical acquisition systems and combining with carefully designed experimental tasks. This technology not only improves data quality and experimental effect, but also provides a solid foundation for in-depth study of brain-muscle coupling relationship and development of related applications.

[0085] S2, the collected brain and muscle signals are respectively preprocessed, and the preprocessed data is divided into a plurality of different sample data segments.

[0086] Specifically, in the preprocessing process of the brain and muscle electrical signals, the electroencephalogram signal is first taken as a common average reference, and a 4-60Hz FIR filter is used for zero-phase filtering. Subsequently, the 50Hz power frequency noise is removed by using the standardized preprocessing framework (early stage preprocessing, PREP) of EEGLAB, and the interference of electrooculogram, electrocardiogram and other artifacts is removed by using independent component analysis technology for the electroencephalogram signal, so as to obtain the preprocessed electroencephalogram signal.

[0087] Meanwhile, the electromyogram signal is also subjected to FIR filtering of the same frequency band to obtain the preprocessed electromyogram signal.

[0088] The embodiment of the present application ensures the purity and consistency of the electroencephalogram signal and the electromyogram signal in a wide frequency band through common average reference and frequency band filtering, providing high-quality basic data for subsequent processing. Each preprocessing step aims to gradually optimize signal quality, from removing unnecessary frequency components to eliminating power frequency noise of a specific frequency, and then removing various artifacts, to finally obtain electroencephalogram signals with high purity. High-quality brain and muscle electrical signals provide a solid foundation for advanced analysis such as frequency domain analysis, time-frequency analysis and cortical muscle coupling feature extraction, which helps to more deeply study the relationship between brain cortex activity and muscle activity. It can be used for evaluation and optimization of rehabilitation training programs to help patients recover motor function, and also provides valuable resources for the development of brain-computer interface (BCI) systems, promoting the progress of human-computer interaction technology.

[0089] S3, using fast Fourier algorithm to perform frequency domain analysis on the brain and muscle signals in the sample data segment to calculate the self-power spectrum matrix of the brain and muscle signals and the cross-spectrum matrix between them.

[0090] Specifically, the fast Fourier transform algorithm is used to calculate the EEG signal x∈R n×p and EMG signal y∈R n×q The discrete Fourier spectrum X∈C n×p Y, and further obtain the EEG autopower spectrum matrix S XX =X H X, myoelectric autopower spectrum matrix S YY =Y H Y, brain and myoelectric cross spectrum matrix S XY =X H Y.

[0091] The FFT of the embodiment of the present invention converts the time domain signal into a frequency domain signal, reveals the component distribution of the signal at different frequencies, and improves the understanding of the time variation characteristics of the signal. The autopower spectrum matrix intuitively displays the energy distribution of each frequency component, reflects the intensity change of the signal at different frequencies, and helps to identify the key frequency bands. The cross-spectral matrix reveals the synchronization and coherence between EEG signals and EMG signals, and enhances the understanding of neural regulatory mechanisms, especially in terms of motor control. High-quality frequency domain analysis provides a solid foundation for subsequent frequency domain analysis, time-frequency analysis, and cortical muscle coupling feature extraction, and enhances the depth and breadth of research.

[0092] S4, based on the auto-power spectrum matrix and cross-spectrum matrix, and through the structured sparsification canonical coherence algorithm, the corticomuscular coupling features between multi-lead brain-muscle signals are extracted to reconstruct the brain topography related to the corticomuscular coupling features.

[0093] To avoid overfitting of the optimization model, regularization techniques usually impose appropriate restrictions on the optimization variables to narrow the solution space of the optimal solution, thereby reducing the possibility of obtaining an overfitting solution. First, the cross-spectral block matrix obtained by the above calculation can be defined as:

[0094]

[0095] Where X∈C n×p The complex data matrix representing the EEG signal, Y∈C n×q represents the complex data matrix of the EMG signal, n represents the number of data segments, p represents the number of EEG leads, and q represents the EMG lead data. H represents the Hermitian transpose. Therefore, the optimization problem of the typical coherence algorithm for structured sparsification is as follows:

[0096]

[0097]

[0098] Among them, α∈R p×1and β ∈ R q×1 are the real-valued projection vectors of electroencephalogram and electromyogram signals, φ is the auxiliary variable for solving the optimization problem, I is the identity matrix, λ1, λ2, γ1, and γ2 are the regularization hyperparameters of the optimization problem, L α and L β are the Laplacian matrices of electroencephalogram and electromyogram signals, D is a diagonal matrix and C is the adjacency matrix.

[0099] To solve the optimization problem (2), we adopt the following alternating optimization algorithm: first, initialize α and β as unit vectors, and φ as π / 4; then, fix the optimization parameters β and φ, the original optimization problem can be simplified as:

[0100]

[0101] Here, we define the proximal point operator f as the following form:

[0102]

[0103] For the unconstrained convex optimization problem (3), the gradient descent method can be reformulated as the proximal regularized form

[0104]

[0105] where, is X H the real part of Yexp(-iφ), P α is the 1-norm of the projection vector α, T α is the iterative step size of the projection vector α. Therefore, the LASSO problem (5) can be solved by using the fast iterative thresholding algorithm (FISTA).

[0106] Similarly, when fixing the optimization parameters α and φ, the original optimization problem can be simplified as:

[0107]

[0108] Meanwhile, the iterative formula of the optimization parameter β is:

[0109]

[0110] where, P β is the 1-norm of the projection vector β, T β is the iterative step size of the projection vector β.

[0111] Further, when fixing the optimization parameters α and β, the original optimization problem can be simplified as:

[0112]

[0113] It is easy to prove that the optimization problem (8) can be simplified into a trigonometric function of the optimization parameter φ, so the Levenberg-Marquardt algorithm can be used to find the optimal solution for the optimization parameter φ. After several iterations, the optimal solution to the original optimization problem (2) can be obtained.

[0114] like Figure 2 As shown in the figure, this is the algorithm framework diagram of the typical coherence of structured sparsification. First, the EEG and EMG signals are obtained by synchronously collecting and preprocessing the EEG and EMG signals. Then, with maximizing the coherence as the optimization goal, the latent space representation of the EEG and EMG signals is calculated by solving the optimization problem, which is specifically reflected in the real projection vector α∈R p×1 and β∈R q×1 , and finally obtain the brain and myoelectric fusion characteristics and the corresponding brain topology map.

[0115] like Figure 3 As shown, Figure 3 (A) Figure 3 (B) Figure 3 (C) Figure 3 Figures (D) in the figure compare the feature extraction performance of the structured sparsification canonical coherence algorithm proposed in the present invention in a simulation dataset. As can be seen, in the simulation dataset, under different sample sizes and noise levels, the significant canonical coherence calculated by the multivariate EMG coupling analysis method based on structured sparsity regularization proposed in the present invention outperforms previous multivariate methods.

[0116] like Figure 4 As shown, Figure 4 (A) Figure 4 (B) in the figure is a comparison diagram of the feature extraction performance of the structured sparsified typical coherence algorithm in the present invention in a healthy human dataset. It can be observed that in the healthy human dataset, the significant typical coherence calculated by the multivariate brain-myoelectric coupling analysis method based on structured sparsity regularization proposed in the present invention is significantly better than the previous multivariate method. At the same time, the method proposed in the present invention can more accurately extract the target brain areas related to corticospinal motor control by introducing sparsity and structured constraints, thereby generating a more centralized and interpretable brain topology map.

[0117] In summary, the beneficial effects of the embodiments of the present invention are:

[0118] 1) By introducing a structural sparsification regularization term, a bio-inspired optimization problem is constructed, combining feature selection with the model fitting process. This effectively avoids the overfitting problem that is prone to occur in traditional multivariate methods and significantly improves the coupled feature detection and pattern recovery capabilities.

[0119] 2) The application solves the optimization problem by proposing an alternating optimization algorithm, taking into account the robustness and real-time performance of the cortical muscle coupling feature extraction, and assisting in the design of a real-time neural feedback system.

[0120] According to the multivariate brain-muscle coupling analysis method based on structural sparse regularization, the cortical muscle coupling features of healthy people and stroke patients can be robustly identified and the brain topography related to the cortical muscle coupling can be robustly recovered under low sample size and high noise level, so that the central and peripheral nervous circuit function of the subject can be accurately monitored, and the method can be used as a potential biomarker for post-stroke motor function recovery.

[0121] To achieve the above-embodiment, as Figure 5 shown, the embodiment further provides a multivariate brain-muscle coupling analysis system 10 based on structural sparse regularization, comprising:

[0122] The brain-muscle signal acquisition module 100 is configured to synchronously acquire brain-muscle signals.

[0123] The data preprocessing module 200 is configured to perform data preprocessing on the acquired brain-muscle signals respectively, and divide the preprocessed data into a plurality of different sample data segments.

[0124] The frequency domain analysis module 300 is configured to perform frequency domain analysis on the brain-muscle signals in the sample data segments by using a fast Fourier algorithm, to calculate the self-power spectrum matrix of each brain-muscle signal and the cross-spectrum matrix therebetween.

[0125] The coherence analysis module 400 is configured to extract the cortical muscle coupling features between the multi-lead brain-muscle signals based on the self-power spectrum matrix and the cross-spectrum matrix, and by using a structural sparse canonical coherence algorithm, to reconstruct the brain topography related to the cortical muscle coupling features.

[0126] The cross-spectrum matrix is defined as:

[0127]

[0128] Wherein, X C n×p represents a complex data matrix of the electroencephalogram signal, Y C n×q represents a complex data matrix of the electromyogram signal, n represents the number of data segments, p represents the number of leads of the electroencephalogram, q represents the number of leads of the electromyogram, and H represents the Hermitian transpose. The optimization problem of the structural sparse canonical coherence algorithm is as follows:

[0129]

[0130] Wherein, α R p×1 and β R q×1are the real-valued projection vectors of electroencephalogram and electromyogram signals, φ is the auxiliary variable for solving the optimization problem, I is the identity matrix, λ1, λ2, γ1 and γ2 are the regularization hyperparameters of the optimization problem, L α and L β are the Laplacian matrices of electroencephalogram and electromyogram signals, D is a diagonal matrix and C is an adjacency matrix;

[0131] Solving the optimization problem (2), an alternating optimization algorithm is adopted as shown below: initializing α and β as unit vectors, and φ as π / 4; fixing the optimization parameters β and φ, the original optimization problem is simplified as:

[0132]

[0133] Defining the proximal point operator f as the following form:

[0134]

[0135] For the unconstrained convex optimization problem (3), the gradient descent method is reformulated as a proximal regularized form:

[0136]

[0137] where, is X H the real part of Yexp(-iφ), P α is the 1-norm of the projection vector α, T α is the iterative step size of the projection vector α; the LASSO problem (5) is solved by using the fast iterative thresholding algorithm (FISTA);

[0138] When fixing the optimization parameters α and φ, the original optimization problem is simplified as:

[0139]

[0140] Meanwhile, the iterative formula of the optimization parameter β is:

[0141]

[0142] where, β is the 1-norm of the projection vector β, T β is the iterative step size of the projection vector β;

[0143] When fixing the optimization parameters α and β, the original optimization problem is simplified as:

[0144]

[0145] The Levenberg-Marquardt algorithm is used to find the optimal solution of the optimization parameter φ. After several iterations, the optimal solution of the original optimization problem (2) is obtained.

[0146] Furthermore, the brain muscle signal acquisition module 100 is also used to:

[0147] The EEG acquisition system and the wireless EMG acquisition system are used to synchronously acquire EEG signals and EMG signals; the sampling frequency is 1000 Hz.

[0148] Furthermore, the data preprocessing module 200 is further configured to:

[0149] The EEG signal is subjected to common mean reference and zero-phase filtering using an FIR filter to obtain the initially processed EEG signal. The 50 Hz power frequency noise of the initially processed EEG signal is removed using the EEGLAB standardized preprocessing framework. The artifact interference of the EEG signal after power frequency noise removal is removed using the independent component analysis method to obtain the preprocessed EEG signal.

[0150] The electromyographic signal is processed using an FIR filter of the same frequency band to obtain a preprocessed electromyographic signal.

[0151] Furthermore, the frequency domain analysis module 300 is further configured to:

[0152] The fast Fourier transform algorithm is used to calculate the EEG signal x∈R n×p and EMG signal y∈R n×q The discrete Fourier spectrum X∈C n×p Y;

[0153] Obtain the EEG autopower spectrum matrix S XX =X H X, myoelectric autopower spectrum matrix S YY =Y H Y, brain and myoelectric cross spectrum matrix S XY =X H Y.

[0154] According to the multivariate brain-myoelectric coupling analysis system based on structured sparse regularization according to an embodiment of the present invention, the corticomuscular coupling characteristics of healthy people and stroke patients can be robustly identified under low sample size and high noise level, and the brain topography related to corticomuscular coupling can be robustly restored, thereby realizing accurate monitoring of the central peripheral neural circuit function of the subjects, and can serve as a potential biomarker for motor function recovery after stroke.

[0155] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the description of the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. In addition, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples, without contradiction.

[0156] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise specifically limited.

Claims

1. A method for multivariate brain-muscle electro-coupling analysis based on structural sparse regularization, characterized in that, The method comprises the following steps: Synchronously collecting brain and muscle signals; Respectively performing data preprocessing on the collected brain and muscle signals, and dividing the preprocessed data into a plurality of different sample data segments; Performing frequency domain analysis on the brain and muscle signals in the sample data segments by using a fast Fourier algorithm to calculate self-power spectrum matrices of the brain and muscle signals and cross-spectrum matrices therebetween; Based on the self-power spectrum matrices and the cross-spectrum matrices, and by using a structured sparsifying canonical coherence algorithm, cortical muscle coupling characteristics between the multi-lead brain and muscle signals are extracted to reconstruct a brain topography related to the cortical muscle coupling characteristics; the method comprises the following steps: The cross-spectrum matrix is defined as: Where X∈C n×p The complex data matrix representing the EEG signal, Y∈C n×q represents the complex data matrix of the electromyographic signal, n represents the number of data segments, p represents the number of EEG leads, q represents the electromyographic lead data, and H represents the Hermitian transpose. The optimization problem of the typical coherence algorithm for structured sparsification is as follows: where a e R p×1 and β e R q×1 are the real-valued projection vectors of electroencephalogram and electromyogram signals, respectively, φ is an auxiliary variable for solving the optimization problem, I is an identity matrix, λ1, λ2, γ1, and γ2 are regularization hyperparameters of the optimization problem, L α and L β are Laplacian matrices of electroencephalogram and electromyogram signals, respectively, and D - C is the difference between a diagonal matrix D and an adjacency matrix C. Solving the optimization problem (2) by using an alternating optimization algorithm as shown in the following formula: initializing α and β as unit vectors, and taking φ as π / 4; fixing the optimization parameters β and φ, and simplifying the original optimization problem into: The approximate point operator f is defined in the following form: For the unconstrained convex optimization problem (3), a gradient descent method is re-formulated into a proximal regularization form: where is X H the real part of Yexp(-iφ), P α is the 1-norm of the projection vector α, T α is the iteration step size of the projection vector α; solve the LASSO problem (5) using the fast iterative thresholding algorithm (FISTA); When the optimization parameters α and φ are fixed, the original optimization problem is simplified into: At the same time, the iterative formula of the optimization parameter β is: where P β is the 1-norm of the projection vector β, T β is the iteration step size for the projection vector β; When the optimization parameters α and β are fixed, the original optimization problem is simplified into: The optimal solution of the optimization parameter φ is solved by using a Levenberg-Marquardt algorithm, and after a plurality of iterations, the optimal solution of the original optimization problem (2) is obtained.

2. The method of claim 1, wherein, Respectively performing data preprocessing on the collected brain and muscle signals, comprising: Performing common average reference on the electroencephalogram signals, and performing zero-phase filtering on the electroencephalogram signals by using a FIR filter to obtain the brain signals after initial processing, and performing 50Hz power frequency noise removal on the brain signals after initial processing by using an EEGLAB standardized preprocessing framework, and performing artifact interference removal on the brain signals after power frequency noise removal by using an independent component analysis method to obtain the preprocessed electroencephalogram signals; Processing the electromyogram signals by using a FIR filter of the same frequency band to obtain the preprocessed electromyogram signals.

3. The method of claim 2, wherein, Performing frequency domain analysis on the brain and muscle signals in the sample data segments by using a fast Fourier algorithm to calculate self-power spectrum matrices of the brain and muscle signals and cross-spectrum matrices therebetween, comprising: The discrete Fourier spectrum X e C n×p and y e R n×q of the electroencephalogram signal x e R n×p Y; The electroencephalogram self-power spectrum matrix S XX = X H X, the electromyogram self-power spectrum matrix S YY = Y H Y, the electroencephalogram-electromyogram cross spectrum matrix S XY = X H Y.

4. A multivariate electroencephalography and electrocorticography coupling analysis system based on structural sparse regularization, characterized in that, The method comprises the following steps: A brain and muscle signal collecting module is configured to synchronously collect brain and muscle signals; A data preprocessing module is configured to respectively perform data preprocessing on the collected brain and muscle signals, and divide the preprocessed data into a plurality of different sample data segments; A frequency domain analysis module is configured to perform frequency domain analysis on the brain and muscle signals in the sample data segments by using a fast Fourier algorithm to calculate self-power spectrum matrices of the brain and muscle signals and cross-spectrum matrices therebetween; A coherence analysis module is configured to, based on the self-power spectrum matrices and the cross-spectrum matrices, and by using a structured sparsifying canonical coherence algorithm, extract cortical muscle coupling characteristics between the multi-lead brain and muscle signals to reconstruct a brain topography related to the cortical muscle coupling characteristics; the method comprises the following steps: The cross-spectrum matrix is defined as: where X ∈ C n×p Y ∈ C n×q Y ∈ C where n represents the number of data segments, p represents the number of EEG leads, q represents the number of EMG leads, and H represents the Hermitian transpose; the optimization problem of the structure sparse canonical correlation algorithm is as follows: where α ∈ R p×1 and β ∈ R q×1 are the real-valued projection vectors of the electroencephalogram and electromyogram signals, respectively, φ is an auxiliary variable for solving the optimization problem, I is an identity matrix, λ1, λ2, γ1, and γ2 are regularization hyperparameters of the optimization problem, L α and L β are the Laplacian matrices of the electroencephalogram and electromyogram signals, respectively, and D - C is the difference between a diagonal matrix D and an adjacency matrix C. Solving the optimization problem (2) by using an alternating optimization algorithm as shown in the following formula: initializing α and β as unit vectors, and taking φ as π / 4; fixing the optimization parameters β and φ, and simplifying the original optimization problem into: The approximate point operator f is defined in the following form: For the unconstrained convex optimization problem (3), the gradient descent method is reformulated as a proximal regularized form: wherein is X H the real part of Yexp(-iφ), P α is the 1-norm of the projection vector a, T α is the iteration step size of the projection vector a; solve the LASSO problem (5) using the fast iterative thresholding algorithm (FISTA); When fixing the optimization parameters α and φ, the original optimization problem is simplified as: Meanwhile, the iterative formula of the optimization parameter β is: where P β is the 1-norm of the projection vector β, T β is the iteration step size of the projection vector β; When fixing the optimization parameters α and β, the original optimization problem is simplified as: The optimal solution of the optimization parameter φ is solved by using the Levenberg-Marquardt algorithm, and after several iterations, the optimal solution of the original optimization problem (2) is obtained.

5. The system of claim 4, wherein, The data preprocessing module is further configured to: The electroencephalogram signals are co-averaged and referenced, and the electroencephalogram signals are zero-phase filtered by using a FIR filter to obtain the electroencephalogram signals after initial processing, and the electroencephalogram signals after initial processing are subjected to 50Hz power frequency noise removal by using an EEG LAB standardized preprocessing framework, and the electroencephalogram signals after power frequency noise removal are subjected to artifact interference removal by using an independent component analysis method, to obtain the electroencephalogram signals after preprocessing. The electromyogram signals are processed by using a FIR filter of the same frequency band, to obtain the electromyogram signals after preprocessing.

6. The system of claim 5, wherein, The frequency domain analysis module is further configured to: The discrete Fourier spectrum X e C n×p and y e R n×q of the electroencephalogram signal x e R n×p Y; The electroencephalogram self-power spectrum matrix S XX = X H X, the electromyogram self-power spectrum matrix S YY = Y H Y, the electroencephalogram-electromyogram cross spectrum matrix S XY = X H Y.

Images