CN115346676A - Movement function reconstruction dynamic model construction method based on cortical muscle network - Google Patents
Movement function reconstruction dynamic model construction method based on cortical muscle network Download PDFInfo
- Publication number
- CN115346676A CN115346676A CN202210905321.0A CN202210905321A CN115346676A CN 115346676 A CN115346676 A CN 115346676A CN 202210905321 A CN202210905321 A CN 202210905321A CN 115346676 A CN115346676 A CN 115346676A
- Authority
- CN
- China
- Prior art keywords
- network
- cortical
- eeg
- muscle
- emg
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
- G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
- G16H50/50—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for simulation or modelling of medical disorders
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/369—Electroencephalography [EEG]
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/389—Electromyography [EMG]
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
- A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Health & Medical Sciences (AREA)
- Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Public Health (AREA)
- Medical Informatics (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Pathology (AREA)
- Molecular Biology (AREA)
- Veterinary Medicine (AREA)
- Animal Behavior & Ethology (AREA)
- Surgery (AREA)
- Heart & Thoracic Surgery (AREA)
- Theoretical Computer Science (AREA)
- Psychiatry (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- Epidemiology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Psychology (AREA)
- Databases & Information Systems (AREA)
- General Engineering & Computer Science (AREA)
- Primary Health Care (AREA)
- Fuzzy Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physiology (AREA)
- Signal Processing (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a method for constructing a motor function reconstruction dynamic model based on a cortical muscle network. Based on multi-node interconnection and synchronous triggering technology, high-density electroencephalogram-electromyogram information acquisition is realized by using high-density electroencephalogram-electromyogram information acquisition equipment, and grading pretreatment is carried out on the high-density electroencephalogram information and electromyogram information; constructing a multidimensional generalized complex network model based on hypergraph learning so as to obtain a high-order network index; constructing a data association mapping model, acquiring a cortex-muscle function network feature coding vector based on an automatic encoder, and converting the cortical-muscle function network feature coding vector into physiological state features based on a feature converter and a linear mask unit; and generating a sparse cortical-muscle function network significant structure by a cortical-muscle network significant structure generator, and realizing the construction and visual presentation of a cortical-muscle function network dynamic simulation model. The method starts from the relevance of the brain and the cortical muscle network, researches the evolution law of the cortical-muscle function network, and has important significance for understanding the cortical-muscle function coupling and the cooperative mechanism.
Description
Technical Field
The invention relates to the technical field of biomedical engineering research, in particular to a method for constructing a motor function reconstruction dynamic model based on a cortical muscle network.
Background
In daily life, people may suffer from brain damage caused by some diseases and present different degrees of limb movement dysfunction, which seriously affects the daily life ability of patients to a certain extent and also brings heavy burden to families and society. Therefore, the study of the cortex-muscle function coupling rule in the user motor function reconstruction process is of great significance.
Electroencephalograms (EEG), which are electrophysiological acquisition systems that can record Electroencephalogram activity from multiple electrodes on the scalp, are widely used in brain science research because of their advantages such as non-invasive property and high time resolution. Electromyogram (EMG) is a bioelectricity graph of muscles recorded by an Electromyogram, and has an important meaning for evaluating activities of people in a man-machine system. The study of brain diseases with EEG and EMG has taken a number of important results over the years. However, while it is desirable, there are many problems, the most important of which is that studies based on EEG and EMG are mostly manual intervention to screen for eligible indicators or to consider some relevant variables for discriminatory statistics, and indicators and parameters often involve a large number of subjective special considerations, and lack of objective law thinking. Therefore, establishing a more objective calculation model is very important for promoting the reconstruction process of the motor function of the stroke.
Disclosure of Invention
The invention aims to provide a more objective method for constructing a dynamic model for reconstructing a motor function based on a cortical muscle network.
A method for constructing a motor function reconstruction dynamic model based on a cortical muscle network. The construction method comprises the following steps:
s1, based on a multi-node interconnection and synchronous triggering technology, synchronous acquisition of high-density EEG and EMG information is realized by using 128-channel EEG-EMG synchronous acquisition equipment, and grading pretreatment is carried out on the high-density EEG and EMG information;
and S2, constructing a data association mapping model for the dynamic evolution model based on the preprocessed EEG and EMG data. Firstly, acquiring time-frequency space three-dimensional feature tensors of EEG and EMG; secondly, constructing a multi-dimensional information structure based on the acquired tensor; further, obtaining a high-order network index of the cortex-muscle function network; finally, a data association mapping model is constructed based on the acquired information, which specifically comprises the following steps:
s21, decomposing the EEG and EMG signals based on short-time Fourier transform, wavelet transform and other frame theories to obtain a time-space-frequency three-dimensional feature tensor of the EEG and EMG signals, and further constructing a time-space-frequency three-dimensional feature tensor space;
and S22, acquiring the required node characteristics and edge characteristics according to the three-dimensional characteristic tensor acquired in the S21. Construction of an EEG-EEG homogeneity map G 1 Homogeneity map G between EMG and EMG 2 And EEG-EMG heterogeneity map G 3 Then, a hypergraph G is obtained based on hypergraph learning, and a multi-dimensional information structure is constructed;
s23, calculating network related indexes according to the map in the S22, and further obtaining cortex-muscle function network high-order network indexes, wherein the method specifically comprises the following steps:
s231, firstly, respectively extracting a time sequence of each channel of EEG and EMG signals in each window by using a sliding window technology; secondly, a biased directional coherent analysis method is introduced to construct an information transfer coefficient matrix PDC between EEG and EMG sequences in each window; further carrying out binary conversion on the PDC sparse into SPDC, and carrying out integrity judgment to determine and select the optimal threshold value of the dynamic cortex-muscle function network(ii) a Finally, the network connectivity D is calculated based on the map in S22 t Cluster coefficient CC t Shortest path length SPL t Local efficiency LE t And global efficiency GE t Obtaining dynamic time-varying network index DTV = [ SPDC, D ] by using the indexes t ,CC t ,LE t ,GE t ]；
S232, firstly, acquiring EEG and EMG components in a specific frequency range based on a band-pass filter, introducing Hilbert transform to calculate envelope lines of all sub-band signals, calculating full-connection correlation coefficients among the envelope lines by using a Pearson correlation method, and constructing a weighted adjacency matrix; secondly, calculating the overall strength of coupling between layers and in layers in the multilayer network to finish thresholding processing on the network; further based on the map in S22, each vertex participation coefficient PC is analyzed and calculated i And a multi-layer network participation coefficient PC, and a connectivity D of the multi-layer network and the sub-network f Cluster coefficient CC f Shortest path length SPL f Local efficiency LE f And global efficiency GE f And (4) waiting for indexes, and obtaining a functional network frequency domain index FCMN = [ PC = i ,PC,D f ,CC f ,SPL f ,LE f ,GE f ]；
S233, firstly, defining relative distance criterion to optimize effective EEG and EMG channels for spatial combination; secondly, decomposing the original signal into a plurality of space domain modes to obtain a new time sequence Z, and then calculating the variance v of the new time sequence Z p And apply the variance vector v p As its spatial domain characteristics F = [ v = 1 ,v 2 ,…,v M ] T (ii) a Further based on the map in S22, analyzing and calculating the connectivity D of the functional network s Cluster coefficient CC s Shortest path length SPL s Local efficiency LE s And global efficiency GE s Obtaining the space domain function network index SBN = [ F, D ] by the indexes s ,CC s ,SPL s ,LE s ,GE s ]。
S24, constructing a data association mapping model, realizing low-dimensional expression of high-order information in the S23, and reducing the complexity of data, wherein the data association mapping model specifically comprises the following steps:
s241, acquiring an adjacent matrix A of the constructed hypergraph G from the first layer of the model based on the multi-dimensional information structure constructed in the S22, and using an automatic encoder to obtain a node embedding X by taking the adjacent matrix A as an input, wherein the specific expression is as follows:
X i ＝T(W i *A i +b i )
in the formula, T represents a Tanh function, X i Node embedding, W, representing a node type of i i A weight matrix representing the node type i, b i Representing the deviation of a node type i, wherein the node types are mainly two, namely EEG and EMG; due to the particularity of the different types of nodes, the potential space specific to the different types of nodes needs to be learned, each type of node has its own self-encoder, and for all types of nodes, the loss function is defined as:
where is the index of the inode type, sign is the sign function,
representing the original features, the expression is as follows:
and S242, the second model layer is a fully-connected layer with a nonlinear activation function, the node embedding X obtained in S241 is used as an input, the node embedding X is nonlinearly mapped to a public potential space L, and the joint expression of the node embedding X in the potential space is as follows:
L ij ＝T(W i *X i +W i *X j +b)
s243, the third layer of the model is to map the potential space L in S242 to the probability space to obtain the similarity:
S ij ＝T(W*L ij +b)
and S244, optimizing the model by using random gradient descent, further improving the calculation efficiency of the model, solving the problem of local optimization, and finally obtaining a data association mapping model for constructing the dynamic evolution model.
And S3, constructing a dynamic simulation model and visualizing the dynamic simulation model. Firstly, acquiring a feature tensor according to the data association mapping model constructed in the S2 so as to construct a dynamic simulation model; secondly, realizing dynamic visual expression of the dynamic simulation model, and the specific process is as follows:
s31, constructing a dynamic simulation model, specifically as follows:
s311, acquiring a cortical-muscle network mixed feature embedding tensor et and a model predictive coding C (x) according to the data association mapping model constructed in the S2, and taking the et as the input of a cortical-muscle functional network feature automatic coding machine E to further obtain an FCMN feature coding vector E (et);
s312, converting the model predictive coding C (x) and the feature coding vector E (et) in the S311 into adjustable physiological state features { A } such as motor function damage parts, rehabilitation stages and user ages by adopting a feature expression converter and a linear mask unit;
s313, further generating a sparse cortical muscle function network significant structure G (t) with similar characteristics to the original high-dimensional cortical muscle function network through the state characteristics { A } by a dynamic simulation cortical-muscle network significant structure generator G, and further realizing construction of a cortical-muscle function network dynamic simulation model.
And S32, realizing visual dynamic expression of a cortical-muscle function network according to different physiological state information of the user, and presenting a change process of the user along with time in different physiological states.
In the preferable technical scheme of the construction method of the dynamic simulation model, the dynamic simulation model is constructed based on the cortical-muscle network characteristics of the time-frequency space-equal domain.
In the preferred technical solution of the above-mentioned construction method of the dynamic simulation model, the multidimensional information structure in step S22 is based on the functional connection characteristics of dynamic time variation, rhythm oscillation, topological coupling, etc. of the cortical-muscle functional network, and is combined with the homogeneity among the same kind of EEG or EMG signals and the heterogeneity among different EEG-EMG signals, and further combined with the multidimensional information structure constructed by the hypergraph.
In a preferred embodiment of the above method for constructing a dynamic simulation model, in step S23, the obtained cortical-muscle function network high-order network index is based on a synchronous causal and functional connectivity analysis of the multichannel EEG and EMG signals, and is used to describe dynamic information of the cortical-muscle function network in motor function control.
In the preferred technical solution of the above construction method of the dynamic simulation model, in step S24, the constructed data association mapping model compresses the high-order network structure that takes the cortex-muscle function connection characteristic as input, so that the high-order information of the cortex-muscle function network is retained while the complexity of data is reduced, and the local and global structure information in the network construction process is retained.
In the preferred technical solution of the above-mentioned method for constructing a dynamic simulation model, in step S31, the constructed dynamic simulation model takes the cortical-muscular function network and its related features as input, combines with the physiological state information of the user, and comprehensively considers the human dynamics and the dynamics characteristics of the biological neural network based on the feature self-encoding and feature expression conversion technology, so as to embody the complex evolution law of the cortical-muscular function network in the process of reconstructing the motor function.
In the preferred technical solution of the above-mentioned construction method of the dynamic simulation model, in step S32, the cortical-muscle function network is visualized and dynamically expressed, and the change of the user' S motor function region with time can be presented from the perspective of individualization and groupwise according to the different information of the user in the physiological states of different rehabilitation stages, different lesions, different age stages, and the like.
Compared with the prior art, the invention has the following beneficial effects:
the invention relates to a method for deep learning, hypergraph learning, deep reinforcement learning and the like based on a graph, which is characterized in that the method combines multi-level characteristics of time domain dynamic time variation, frequency domain rhythm oscillation, spatial domain topological coupling and the like of a cortex-muscle function network, and comprehensively considers the dynamics of human body and the dynamics of a biological nerve network aiming at physiological state information of different focuses, different rehabilitation stages, different ages and the like, so that the complex evolution rule of the cortex-muscle function network in the process of rebuilding the motor function is reflected, and the construction of a dynamic simulation model is carried out on the rebuilding of the motor function in the cortex-muscle function network, thereby solving the problem that the traditional research based on EEG and EMG contains a large number of indexes and parameters which are considered subjectively and specially, and simultaneously establishing a more objective construction method of the dynamic simulation model for rebuilding the motor function, and having great significance for deeply researching the cortex-muscle function coupling rule in the process of rebuilding the motor function of a user and disclosing the mechanism of the motor function.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the drawings without creative efforts;
FIG. 1 is a flow chart of the present invention;
FIG. 2 is a graph of results after signal pre-processing (before and after EEG and EMG signal processing);
FIG. 3 is a flow chart of data association mapping model construction;
FIG. 4 is a comparison of a normal map and a hypergraph;
FIG. 5 is a diagram of a sliding window forming process;
FIG. 6 is a flow chart of dynamic simulation model construction;
fig. 7 is a diagram of a self-encoder structure.
Detailed description of the invention
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the accompanying drawings in conjunction with the following detailed description. It should be understood that the description is intended to be exemplary only, and is not intended to limit the scope of the present invention. Moreover, in the following description, descriptions of well-known structures and techniques are omitted so as to not unnecessarily obscure the concepts of the present invention.
The following detailed description of the embodiments is made with reference to the accompanying drawings.
A dynamic simulation model construction method for stroke motor function reconstruction comprises the steps of preprocessing original signals, constructing a cortical-muscle network state mapping model in self-adaptive correlation with physiological information and visualizing a dynamic simulation model.
A method for constructing a motor function reconstruction dynamic model based on a cortical muscle network is disclosed, as shown in figure 1, firstly, based on a high-density EEG-EMG acquisition device, realizing the acquisition of high-density EEG and EMG information, and carrying out grading pretreatment on the high-density EEG and EMG information; secondly, acquiring three-dimensional feature tensors of EEG and EMG, constructing a multi-dimensional information structure based on graph theory and hypergraph learning, acquiring high-order network indexes on the basis, and completing construction of a data association mapping model; then, acquiring a cortical-muscle function network feature coding vector based on an automatic encoder, and further converting the cortical-muscle function network feature coding vector into a physiological state feature based on a feature converter and a linear mask unit; and finally, generating a sparse cortical-muscle function network significant structure by a cortical-muscle network significant structure generator, and further realizing the construction and visual presentation of a cortical-muscle function network dynamic simulation model.
The invention is realized by the following steps:
s1, based on a multi-node interconnection and synchronous triggering technology, synchronous acquisition of high-density EEG and EMG information is realized by using 128-channel EEG-EMG synchronous acquisition equipment, and grading pretreatment is carried out on the high-density EEG and EMG information;
s2, constructing a data association mapping model for the dynamic evolution model based on the preprocessed EEG and EMG data;
and S3, constructing a dynamic simulation model and visualizing the dynamic simulation model.
The present invention will be described in detail below.
S1, based on multi-node interconnection and synchronous triggering technology, synchronous acquisition of high-density EEG and EMG information is achieved by using 128-channel EEG-EMG synchronous acquisition equipment, and grading pretreatment is carried out on the high-density EEG and EMG information.
The method mainly comprises the steps of synchronously acquiring high-density brain electromechanical data and preprocessing the acquired data. The data acquisition is specifically as follows:
firstly, a plurality of matrix type EMG acquisition modules (32-channel Sessaqualtro, OT) which are arranged in a laboratory are synchronously interconnected, a 128-channel high-density EMG network is constructed by utilizing a multi-vertex information networking technology, a multi-information transceiver is used as an acquisition vertex unified manager, the matrix type EMG acquisition modules and the multi-information transceiver jointly form a local area network, and the matrix type EMG acquisition modules can be communicated with an upper computer for transmission. Further, in order to realize synchronous acquisition of multi-channel EEG and EMG multi-source heterogeneous data, a synchronous pulse trigger and a cooperative transmission protocol are developed, a 128-channel EMG acquisition system (actiCHamp Plus, BP) which is already provided in a laboratory is synchronously integrated with the EEG network system, and multi-source synchronous acquisition of high-density EMG and EEG information is realized.
The data preprocessing strategy comprises filtering and removing artifacts such as baseline drift, 50Hz power frequency and harmonic interference, eye electrical signals and action interference in EEG and EMG. Each pretreatment method will be described below.
Filtering: the method is mainly based on high-pass filtering and low-pass filtering, signals outside frequency are filtered out, interference can be effectively removed, and a frequency sampling method-based finite impulse response Filter (FIR) is used for designing the high-pass filter and the low-pass filter aiming at the collected high-density EMG data.
Removing baseline drift: the baseline drift will cause the original signal to have a slow, slight tendency to float up and down, which is observed in the low frequency coefficients of the decomposition, taking advantage of the difference in the spectral and energy distributions of the signal and noise, and subtracting this baseline tendency from the original signal.
Removing power frequency and harmonic interference: because alternating current is coupled to interference caused by EEG and EMG signals through space, when the electrode is higher in impedance due to grease, dirt, too thick horny layer and the like, power frequency interference is more easily introduced, and the interference is removed by using a filter.
Removing artifacts: artifacts are interference signals that affect electrophysiological signals and can cause difficulties in analyzing normal signals. In order to avoid the generation of the artifacts, the human subject is told not to blink or do some actions which may generate the artifacts before the experiment, then some obvious artifact signals are found out through a method of observation or automatic identification, the obvious artifact signals are directly deleted, and finally the data signals and the artifacts are decomposed into different signal component areas through an Independent Component Analysis (ICA) method to be eliminated. The pre-processed signal is shown in figure 2.
And S2, constructing a data association mapping model for the dynamic evolution model based on the preprocessed EEG and EMG data. The model structure is as shown in fig. 3, first, time-space and time-space three-dimensional feature tensors of EEG and EMG are obtained; secondly, constructing a multi-dimensional information structure based on the acquired tensor; further, obtaining a high-order network index of the cortex-muscle function network; and finally, constructing a data association mapping model based on the acquired information, which specifically comprises the following steps:
s21, decomposing the EEG and EMG signals based on short-time Fourier transform, wavelet transform and other frame theories, obtaining the space-time frequency three-dimensional feature tensor of the EEG and EMG signals, and further constructing a space-time frequency three-dimensional feature tensor space.
And S22, acquiring the required node characteristics and edge characteristics according to the three-dimensional characteristic tensor acquired in the S21. Construction of an EEG-EEG homogeneity map G 1 Homogeneity map G between EMG and EMG 2 And EEG-EMG heterogeneity map G 3 The expression is as follows:
G 1 ＝<V 1 ,E 1 >
G 2 ＝<V 2 ,E 2 >
G 3 ＝<V 3 ,E 3 >
wherein G is 1 And G 2 Representing a homogeneous graph containing only one type of node and one type of edge, V 1 And V 2 Representing the EEG and EMG signals, respectively, as a set of nodes, E 1 And E 2 Respectively representing edges constituted by homogeneity between EEG-EEG and EMG-EMG like signals; g 3 Representing a heterogeneous graph comprising nodes of two classes EEG and EMG and different types of edges, V 3 Representing a set of nodes jointly composed of EEG and EMG, E 3 Representing EEG-EMGThe heterogeneity between signals. On the basis, a hypergraph G is constructed based on the functional connection characteristics of dynamic time variation, rhythm oscillation, topological coupling and the like of a cortex-muscle functional network, and the specific expression is as follows:
G＝<V,E,W>
where V is the finite set of vertices of the hypergraph, E is the set of hyper-edges of the hypergraph, and W is the set of weights for the hyper-edges. Compared with the common graph theory, the Hypergraph (Hypergraph) can more accurately describe the relationship between the objects with the multivariate association. The main difference between the hypergraph and the ordinary graph is the number of vertices on the graph edges, in the ordinary graph, one edge connects two vertices, in the hypergraph, the edge is called as a Hyperedge (Hyperedge), one Hyperedge connects multiple vertices, and the graph and hypergraph pair is as shown in fig. 4. Due to the characteristics of the hypergraph, high-dimensional information in data can be acquired so as to construct a multi-dimensional information structure.
S23, calculating network related indexes according to the map in the S22, and further obtaining cortex-muscle function network high-order network indexes, wherein the method specifically comprises the following steps:
s231, firstly, time sequences of EEG and EMG signals of each channel in each window are respectively extracted by using a sliding window technology, the sliding window is an idea based on double pointers, a window is formed between elements pointed by the two pointers, the problem complexity can be reduced by using the sliding window technology, the loop nesting depth is further reduced, and the specific process of window formation is shown in FIG. 5. First, left and right pointers left and right point to the 0 th element, the window is left and right, and here, the window is left closed and right open, so there is no element in the initial window [0,0) interval; step two, starting to circularly traverse the elements of the whole array, judging whether the current right pointer exceeds the length of the whole array, if so, exiting the circulation, otherwise, executing the step three; thirdly, the right pointer starts to move rightwards for a length, and interval data in the window is updated; fourthly, when the data of the window interval meets the requirement, the right pointer right is kept unchanged, and the left pointer left starts to move until the left pointer moves to an interval which does not meet the requirement any more, and the left pointer does not move any more; and returning to execute the second step.
Secondly, a biased directional coherent analysis method is introduced to construct an information transfer coefficient matrix PDC between EEG and EMG sequences in each window, and the biased directional coherent analysis method is a Grave causal frequency domain measurement method based on a multivariate autoregressive processing time sequence model. The calculation formula is as follows:
wherein, C i For N-lead EEG (EMG) signals X (X) 1 ,x 2 ,...,x n ) The time-domain AR model coefficients in the j-th column of the representation. The presentation information is represented by x j Flow direction x k Accounting for all outflow x j Represents the information flow intensity of the lead j flowing to the lead k, and has a value range of [0,1 ]]The closer the value is to 1, the stronger the correlation between the two leads is, and the stronger the information flow intensity is. Finally, an N-dimensional matrix is obtained, and the strength and the direction of information flow between N (N-1) lead pairs are represented.
Further carrying out binary operation on the PDC sparsity to obtain SPDC, and carrying out integrity judgment to determine and select an optimal threshold value of the dynamic cortex-muscle function network; finally, the network connectivity D is calculated based on the map in S22 t Cluster coefficient CC t Shortest path length SPL t Local efficiency LE t And global efficiency GE t Obtaining dynamic time-varying network index DTV = [ SPDC, D ] by using the indexes t ,CC t ,LE t ,GE t ]；
S232, firstly, acquiring EEG and EMG components in a specific frequency band range based on a band-pass filter, introducing Hilbert transform to calculate envelope lines of sub-band signals, and calculating a full-connection correlation system between the envelope lines by using a Pearson correlation methodAnd a weighted adjacency matrix is constructed. The hilbert transform is a common means in signal processing, and is usually used to construct an analytic signal, so that the signal spectrum only contains positive frequency components, thereby reducing the sampling rate of the signal; to represent band pass signals, thereby providing a means for signal modulation in radio communications; combined with other transformations and decompositions, a spectral analysis of the non-stationary signal is performed. The mathematical definition is: provided with a real-valued function x (t), in which the Hilbert transform is denoted
(Or, take as H [ x (t)]) The expression is:
the inverse transform is:
the pearson correlation method is to use a pearson correlation coefficient to measure the correlation between two variables, which is defined as the quotient of the covariance and the standard deviation between the two variables:
wherein X and Y are two different variables, μ X And mu Y Is the mean of X and Y, σ X And σ Y Is the standard deviation of X and Y.
Due to the fact that
μ X ＝E(x)
E[(X―E(X))(Y―E(Y))]＝E(XY)―E(X)E(Y)
The correlation coefficient can also be expressed as:
the pearson correlation coefficient varies from-1 to 1. A coefficient value of 1 means that X and Y can be well described by a straight line equation, all data points fall well on a straight line, and Y increases with increasing X, a coefficient value of-1 means that all data points fall on a straight line, and Y decreases with increasing X, a coefficient value of 0 means that there is no linear relationship between the two variables, a larger absolute value indicates a stronger correlation between the two.
Secondly, calculating the overall strength of coupling between layers and in layers in the multilayer network according to the obtained adjacency matrix, and finishing thresholding processing on the network; further based on the map in S22, each vertex participation coefficient PC is analyzed and calculated i And a multi-layer network participation coefficient PC, and a connectivity D of the multi-layer network and the sub-network f Cluster coefficient CC f Shortest path length SPL f Local efficiency LE f And global efficiency GE f And (4) waiting for indexes, and obtaining a functional network frequency domain index FCMN = [ PC = i ,PC,D f , CC f ,SPL f ,LE f ,GE f ]；
S233, firstly, defining relative distance criterion to optimize effective EEG and EMG channels for spatial combination; secondly, decomposing the original signal into a plurality of spatial domain modes based on a common spatial domain subspace decomposition method, and solving a signal X A And X B Of the covariance matrix R A And R B The method comprises the following steps:
where T denotes the transpose, tra (. + -.) denotes the traces of the matrix, X A And X B Are EEG-EMG signals under different tasks. Performing characteristic decomposition on the sum of the average covariance matrixes:
where Σ is the eigenvalue matrix, U 0 For the corresponding eigenvector matrix, further construct a whitening value transformation matrix P, and apply the mean covariance matrix R A And R B Performing form conversion and feature decomposition:
then the mean covariance matrix R A And R B Can be converted into the following forms:
S A ＝PR A P T ,S B ＝P B P T
then the formula is shown as A And S B And (3) carrying out characteristic value decomposition:
wherein A Sum-sigma B As a matrix of eigenvalues, U A And U B Is the corresponding eigenvector matrix. S A And S B Having the same eigenvector matrix, i.e. U A ＝U B And the eigenvalue matrix satisfies ∑ A + Σ B = I, for the same eigenvector, if S A Has a larger eigenvalue, then S B The corresponding characteristic values are smaller and vice versa.
Further constructing a spatial filter W, as can be seen from principal component analysis A Front m of a Sum of characteristic values ∑ B M after b The feature vector corresponding to each feature value can represent A, B two kinds of spatial features. To distinguish between the two types of signals, the slave U A (or U) B ) Simultaneously extracting characteristic vectors of the front m columns and the rear m columns, combining the characteristic vectors into a matrix U, and constructing a spatial filter W according to the following formula:
Q＝U T P
wherein m is the number of eigenvectors selected from the eigenvector matrix, m is an integer and
int (, denotes the rounding operation. Filtering the signal X to obtain a new time sequence Z
Z＝WX
Let line p of Z be Z p (p =1,2, …,2 m) and its variance v is defined p The following were used:
var (—) represents the variance of the time series and this variance is vector v p As its spatial domain characteristics F = [ v = 1 ,v 2 ,…,v M ] T (ii) a Further based on the map in S22, analyzing and calculating the connectivity D of the functional network s Cluster coefficient CC s Shortest path length SPL s Local efficiency LE s And global efficiency GE s Obtaining the airspace function network index SBN = [ F, D ] according to the indexes s ,CC s ,SPL s ,LE s , GE s ]。
S24, constructing a data association mapping model, realizing low-dimensional expression of high-order information in the S23, and reducing the complexity of data, wherein the data association mapping model specifically comprises the following steps:
s241, acquiring an adjacent matrix A of the first layer of the model from the constructed hypergraph G based on the multi-dimensional information structure constructed in the S22, and using an automatic encoder to obtain a node embedding X by taking the adjacent matrix A as an input, wherein the specific expression is as follows:
A＝HH T ―D v
X i ＝T(W i *A i +b i )
where H represents the correlation matrix in the hypergraph with dimensions | V | × | E | dimension, dv represents the vertex order of the hypergraph, T represents the Tanh function, and X represents the order of the vertex i Node embedding, W, representing a node type of i i Representing classes of nodesWeight matrix of type i, b i Representing the deviation of a node type i, wherein the node types are mainly two, namely EEG and EMG; due to the particularity of the different types of nodes, the potential space specific to the different types of nodes needs to be learned, each type of node has its own self-encoder, and for all types of nodes, the loss function is defined as:
where is the index of the inode type, sign is the sign function,
representing the original features, the expression is as follows:
and S242, a second model layer is a fully-connected layer with a nonlinear activation function, the node embedding X obtained in the S241 is used as an input, and the node embedding X is nonlinearly mapped to a public potential space L, and the joint expression of the node embedding X in the potential space is as follows:
L ij ＝T(W i *X i +W i *X j +b)
s243, the third layer of the model is to map the potential space L in S242 to the probability space to obtain the similarity:
S ij ＝T(W*L ij +b)
s244, optimizing the model by using random gradient descent, further improving the calculation efficiency of the model, and solving the problem of local optimization, wherein the optimization process specifically comprises the following steps:
first, the concept of gradient descent is described, where a function y = f (x), where x and y are real numbers, and the derivative of this function is denoted as f' (x) or
The derivative f' (x) represents f (x) at point xThe slope. In other words, it shows how small changes in the input are scaled to obtain corresponding changes in the output: f (x +. Epsilon.) is approximately equal to f (x) +. Epsilon.f' (x). The derivative is therefore useful to minimize a function that would indicate how x is modified to improve y slightly. X may be moved a small step in the opposite direction of the derivative to reduce f (x). This technique is called gradient descent.
For functions with multidimensional inputs, the concept of partial derivatives needs to be used. Partial derivative
(x) Only x at measurement point x i How f (x) changes when increasing. The gradient (gradient) is the derivative with respect to a vector derivative: the derivative of f is a vector containing all partial derivatives, noted
(x) In that respect The ith element of the gradient is f with respect to x i Partial derivatives of (a). In the multi-dimensional case, the critical point is the point where all elements in the gradient are zero. The point at f' (x) =0 is referred to as a critical point or a stagnation point. A local minimum means that f (x) for this point is smaller than all neighboring points, so it is not possible to reduce f (x) by moving an infinitesimal step. A local maximum means that f (x) for this point is larger than all neighboring points, so it is not possible to increase f (x) by moving an infinitesimal step size. Some critical points are neither minimum nor maximum points. These points are called saddle points.
The directional derivative in the u (unit vector) direction is the slope of the function f in the u direction. In other words, the directional derivative is the derivative of the function f (x + α u) with respect to α (taken at α = 0). Using the chain rule, it can be seen that when a =0,
to minimize f, it is desirable to find the direction in which f drops fastest. Calculating the directional derivative:
where θ is the angle of u with the gradient. Will | hollow u | telecom 2 Substitution of =1 and ignoring the term independent of u can be simplified to
This is minimal when u is opposite to the gradient direction. In other words, a gradient vector points up a hill and a negative gradient vector points down a hill. Moving in the negative gradient direction may decrease f. This is called the steepest descent method or gradient descent.
The model is optimized using a Stochastic Gradient Descent (SGD), the core of which is that the gradient is expected. It is desirable to approximate the estimate using a small sample. Specifically, at each step of the algorithm, a small batch of samples B = { x = is evenly drawn from the training set (1) ,…,x (m′) }. The number m' of small batches is usually a relatively small number, from one to several hundred. Importantly, as the training set size m grows, m' is typically fixed.
The estimate of the gradient can be expressed as
Where L is the loss per sample
L(x,y,θ)＝―logp(y│x；θ)
Samples from small lot B were used. Then, the random gradient descent algorithm uses the following gradient descent estimation:
θ←θ―∈g
where e is the learning rate.
The main advantage of the stochastic gradient descent is that the calculation rate is fast, the key step is to calculate the partial derivative of the parameter θ, and the expression of θ is:
and S3, constructing and visualizing the dynamic simulation model, wherein a model constructing flow chart is shown in FIG. 6. Firstly, acquiring a feature tensor according to the data association mapping model constructed in the S2 so as to construct a dynamic simulation model; secondly, realizing dynamic visual expression of the dynamic simulation model, and the specific process is as follows:
s31, constructing a dynamic simulation model, specifically as follows:
s311, acquiring a cortical-muscle network mixed feature embedding tensor et and a model predictive code C (x) according to the data association mapping model constructed in the S2, and taking the et as the input of a cortical-muscle functional network feature automatic coding machine E to further obtain an FCMN feature coding vector E (et).
An autoencoder (autoencoder) is a type of neural network that is trained to attempt to copy an input to an output. The self-encoder (autoencoder) has a hidden layer h inside, which can generate an encoded (code) representation input. The network can be seen as being composed of two parts: one encoder represented by the function h = f (x) and one decoder r = g (h) generating the reconstruction. The encoder is used for encoding the high-dimensional input X into a low-dimensional hidden variable h so as to force the neural network to learn the characteristics with the most information quantity; the decoder is used for restoring the hidden variable h of the hidden layer to the initial dimension, and the best state is that the output of the decoder can perfectly or approximately recover the original input, namely X R X. The structure of the self-encoder is shown in fig. 7.
The encoding process of the original data from the input layer to the hidden layer is as follows:
wherein the content of the first and second substances,
it is a data encoding process that encodes a high-dimensional input et into a low-dimensional hidden variable h.
The decoding process from the hidden layer to the input layer is:
wherein the content of the first and second substances,
is a data decoding process that decodes the encoded input h into a high-dimensional E (et). The optimization objective function of the self-encoder is:
wherein
Denotes x and
is called the reconstruction error function.
S312, converting the model predictive coding C (x) and the feature coding vector E (et) in the S311 into adjustable physiological state features { A } such as motor function damage parts, rehabilitation stages and user ages by adopting a feature expression converter and a linear mask unit.
Predictive coding is based on the fact that there is a certain correlation between discrete signals, using one or more signals from the front to predict the next signal, and then coding the difference between the actual and predicted values (prediction error). If the prediction is accurate, the error will be small. Under the condition of equal precision requirement, the method can use less bits for coding to achieve the purpose of compressing data. The performance of predictive coding depends on the performance of a predictor, and the best predictor is a predictor which can achieve the best performance of predictive coding under a certain criterion. Some of the criteria commonly used are: the criterion of minimum mean square error value, the criterion of maximum probability of zero (no) error, the criterion of minimum mean distribution entropy of error, and the like. The structure of the optimal predictor is not only related to the criterion but also to the statistical properties of the source. The great advantage of predictive coding is that it is easy to implement and quite efficient for most practical sources.
Since machine learning algorithms all perform linear algebraic calculations on matrices, the features involved in the calculations must be numerical, and encoding processes are required for non-numerical features. There are generally two ways to implement, label coding and unique heat coding. The label coding is to encode the label, convert the original category variable into a numerical variable, solve the problem of classification coding, customize the quantitative variable according to the needs, have no meaning for the numerical value, are only the functions of identification or sequencing, but have poor interpretability. Such as: for the sex variable, converting a male into 1 and a female into 2, thus realizing the label coding; one-hot encoding is the conversion of original variables into multidimensional variables and represents unique variables with 0,1. The problem that a classifier cannot well process classification variables is solved, and meanwhile the function of the features can be expanded, but when the categories are too many, the feature space is enlarged, and dimension disasters are easily caused.
S313, further generating a sparse cortical muscle function network significant structure G (t) similar to the original high-dimensional cortical muscle function network characteristic from the state characteristic { A } through a dynamic simulation cortical-muscle network significant structure generator G, and further realizing construction of a cortical-muscle function network dynamic simulation model.
And S32, realizing visual dynamic expression of a cortical-muscle function network according to different physiological state information of the user, and presenting a change process of the user along with time in different physiological states.
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 of the technical solution of the present invention by those skilled in the art should fall within the protection scope defined by the claims of the present invention without departing from the spirit of the present invention.
Claims (7)
1. A method for constructing a dynamic model for reconstructing motor functions based on a cortical muscle network is characterized by comprising the following steps:
s1, based on a multi-node interconnection and synchronous triggering technology, synchronous acquisition of high-density EEG and EMG information is realized by using 128-channel EEG-EMG synchronous acquisition equipment, and grading pretreatment is carried out on the high-density EEG and EMG information;
s2, constructing a data association mapping model for the dynamic evolution model based on the preprocessed EEG and EMG data; firstly, acquiring time-frequency space three-dimensional feature tensors of EEG and EMG; secondly, constructing a multi-dimensional information structure based on the acquired tensor; further, obtaining a high-order network index of the cortex-muscle function network; and finally, constructing a data association mapping model based on the acquired information, which specifically comprises the following steps:
s21, decomposing the EEG and EMG signals based on short-time Fourier transform, wavelet transform and other frame theories to obtain a time-space-frequency three-dimensional feature tensor of the EEG and EMG signals, and further constructing a time-space-frequency three-dimensional feature tensor space;
s22, acquiring required node features and edge features according to the three-dimensional feature tensor acquired in the S21; construction of an EEG-EEG homogeneity map G 1 Homogeneity map G between EMG and EMG 2 And EEG-EMG heterogeneity map G 3 Then, a hypergraph G is obtained based on hypergraph learning, and a multi-dimensional information structure is constructed;
s23, calculating network related indexes according to the map in the S22, and further obtaining cortex-muscle function network high-order network indexes, wherein the method specifically comprises the following steps:
s231, firstly, respectively extracting a time sequence of each channel of EEG and EMG signals in each window by using a sliding window technology; secondly, a biased directional coherent analysis method is introduced to construct an information transfer coefficient matrix PDC between EEG and EMG sequences in each window; further carrying out binary operation on the PDC sparsity to obtain SPDC, and carrying out integrity judgment to determine and select an optimal threshold value of the dynamic cortex-muscle function network; finally, the network connectivity D is calculated based on the map in S22 t Cluster coefficient CC t Shortest path length SPL t Local efficiency LE t And globalEfficiency GE t Obtaining dynamic time-varying network index DTV = [ SPDC, D ] by using the indexes t ,CC t ,LE t ,GE t ]；
S232, firstly, acquiring EEG and EMG components in a specific frequency range based on a band-pass filter, introducing Hilbert transform to calculate envelope lines of all sub-band signals, calculating full-connection correlation coefficients among the envelope lines by using a Pearson correlation method, and constructing a weighted adjacency matrix; secondly, calculating the overall strength of coupling between layers and in layers in the multilayer network to finish thresholding processing on the network; further based on the map in S22, each vertex participation coefficient PC is analyzed and calculated i And a multi-layer network participation coefficient PC, and a connectivity D of the multi-layer network and the sub-network f Cluster coefficient CC f Shortest path length SPL f Local efficiency LE f And global efficiency GE f And (4) waiting for indexes, and obtaining a functional network frequency domain index FCMN = [ PC = i ,PC,D f ,CC f ,SPL f ,LE f ,GE f ]；
S233, firstly, defining relative distance criterion to optimize effective EEG and EMG channels for spatial combination; secondly, decomposing the original signal into a plurality of space domain modes to obtain a new time sequence Z, and then calculating the variance v of the new time sequence Z p And apply the variance vector v p As its spatial domain feature F = [ v = 1 ,v 2 ,…,v M ] T (ii) a Further based on the map in S22, analyzing and calculating the connectivity D of the functional network s Cluster coefficient CC s Shortest path length SPL s Local efficiency LE s And global efficiency GE s Obtaining the space domain function network index SBN = [ F, D ] by the indexes s ,CC s ,SPL s ,LE s ,GE s ]；
S24, constructing a data association mapping model, realizing low-dimensional expression of high-order information in the S23, and reducing the complexity of data, wherein the data association mapping model specifically comprises the following steps:
s241, acquiring an adjacent matrix A of the first layer of the model from the constructed hypergraph G based on the multi-dimensional information structure constructed in the S22, and using an automatic encoder to obtain a node embedding X by taking the adjacent matrix A as an input, wherein the specific expression is as follows:
X i ＝T(W i *A i +b i )
in the formula, T represents a Tanh function, X i Node embedding, W, representing a node type of i i A weight matrix representing the node type i, b i Representing the deviation of a node type i, wherein the node types are mainly two, namely EEG and EMG; due to the particularity of the different types of nodes, the potential space specific to the different types of nodes needs to be learned, each type of node has its own self-encoder, and for all types of nodes, the loss function is defined as:
where is the index of the inode type, sign is the sign function,
representing the original features, the expression is as follows:
and S242, the second model layer is a fully-connected layer with a nonlinear activation function, the node embedding X obtained in S241 is used as an input, the node embedding X is nonlinearly mapped to a public potential space L, and the joint expression of the node embedding X in the potential space is as follows:
L ij ＝T(W i *X i +W i *X j +b)
s243, the third layer of the model is to map the potential space L in S242 to the probability space to obtain the similarity:
S ij ＝T(W*L ij +b)
and S244, optimizing the model by using random gradient descent, further improving the calculation efficiency of the model, solving the problem of local optimization, and finally obtaining a data association mapping model for constructing the dynamic evolution model.
S3, constructing and visualizing a dynamic simulation model; firstly, acquiring a feature tensor according to the data association mapping model constructed in the S2 so as to construct a dynamic simulation model; secondly, realizing dynamic visual expression of the dynamic simulation model, and the specific process is as follows:
s31, constructing a dynamic simulation model, specifically as follows:
s311, acquiring a cortical-muscle network mixed feature embedding tensor et and a model predictive coding C (x) according to the data association mapping model constructed in the S2, and taking the et as the input of a cortical-muscle functional network feature automatic coding machine E to further obtain an FCMN feature coding vector E (et);
s312, converting the model predictive coding C (x) and the feature coding vector E (et) in the S311 into adjustable physiological state features { A } such as motor function damage parts, rehabilitation stages and user ages by adopting a feature expression converter and a linear mask unit;
s313, further generating a sparse cortical muscle function network significant structure G (t) with similar characteristics to the original high-dimensional cortical muscle function network through the state characteristics { A } by a dynamic simulation cortical-muscle network significant structure generator G, and further realizing the construction of a cortical-muscle function network dynamic simulation model;
and S32, realizing visual dynamic expression of a cortical-muscle function network according to different physiological state information of the user, and presenting a change process of the user along with time in different physiological states.
2. The method for constructing a cortical muscle network-based motor function reconstruction dynamic model as claimed in claim 1, wherein the dynamic simulation model is constructed based on the cortical-muscle network characteristics of the time-frequency space-equal domain.
3. The method as claimed in claim 1, wherein the multidimensional information structure in step S22 is a multidimensional information structure constructed by combining the homogeneity of the same kind of EEG or EMG signals and the heterogeneity of different EEG-EMG signals based on the functional connection characteristics of the cortical-muscle functional network, such as dynamic time variation, rhythm oscillation, and topological coupling.
4. The method as claimed in claim 1, wherein in step S23, the obtained cortical-muscle function network high-order network index is based on synchronous causal and functional connectivity analysis of multi-channel EEG and EMG signals, and is used to describe the dynamic information of cortical-muscle function network in motor function control.
5. The method as claimed in claim 1, wherein in step S24, the constructed data association mapping model compresses the high-order network structure using the cortical-muscle function connection feature as input, so as to reduce the complexity of data and retain the high-order information of the cortical-muscle function network, and also retain the local and global structure information during the network construction process.
6. The method as claimed in claim 1, wherein in step S31, the constructed dynamic simulation model takes the cortical-muscle function network and its related features as input, combines with the physiological status information of the user, and embodies the complex evolution law of the cortical-muscle function network in the motor function reconstruction process by comprehensively considering the characteristics of human dynamics and biological neural network dynamics based on the feature self-coding and feature expression transformation technology.
7. The method of claim 1, wherein in step S32, the cortical-muscle network visualization dynamic representation is able to show the change of the user' S motor function region with time from the perspective of individuation and grouping according to the different information of the user in physiological states such as different rehabilitation stages, different focuses, and different age stages.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210905321.0A CN115346676A (en) | 2022-07-29 | 2022-07-29 | Movement function reconstruction dynamic model construction method based on cortical muscle network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210905321.0A CN115346676A (en) | 2022-07-29 | 2022-07-29 | Movement function reconstruction dynamic model construction method based on cortical muscle network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115346676A true CN115346676A (en) | 2022-11-15 |
Family
ID=83950904
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210905321.0A Pending CN115346676A (en) | 2022-07-29 | 2022-07-29 | Movement function reconstruction dynamic model construction method based on cortical muscle network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115346676A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116153522A (en) * | 2023-03-02 | 2023-05-23 | 之江实验室 | Model training method and device based on cortical-muscle-cortical network diagram |
| CN117153379A (en) * | 2023-10-31 | 2023-12-01 | 深圳市前海蛇口自贸区医院 | Prediction device for thoracic outlet syndrome |
-
2022
- 2022-07-29 CN CN202210905321.0A patent/CN115346676A/en active Pending
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116153522A (en) * | 2023-03-02 | 2023-05-23 | 之江实验室 | Model training method and device based on cortical-muscle-cortical network diagram |
| CN116153522B (en) * | 2023-03-02 | 2023-09-19 | 之江实验室 | Model training method and device based on cortical-muscle-cortical network diagram |
| CN117153379A (en) * | 2023-10-31 | 2023-12-01 | 深圳市前海蛇口自贸区医院 | Prediction device for thoracic outlet syndrome |
| CN117153379B (en) * | 2023-10-31 | 2024-02-20 | 深圳市前海蛇口自贸区医院 | Prediction device for thoracic outlet syndrome |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Abdelhameed et al. | A deep learning approach for automatic seizure detection in children with epilepsy | |
| Zhou et al. | Classification of motor imagery EEG using wavelet envelope analysis and LSTM networks | |
| Mane et al. | Hand motion recognition from single channel surface EMG using wavelet & artificial neural network | |
| Phinyomark et al. | Application of wavelet analysis in EMG feature extraction for pattern classification | |
| CN110969108B (en) | Limb action recognition method based on autonomic motor imagery electroencephalogram | |
| CN115346676A (en) | Movement function reconstruction dynamic model construction method based on cortical muscle network | |
| CN105809124B (en) | The feature extracting method of Mental imagery EEG signals based on DWT and Parametric t-SNE | |
| CN110367967B (en) | Portable lightweight human brain state detection method based on data fusion | |
| CN114533086B (en) | Motor imagery brain electrolysis code method based on airspace characteristic time-frequency transformation | |
| CN112244873A (en) | Electroencephalogram time-space feature learning and emotion classification method based on hybrid neural network | |
| CN111067514B (en) | Multi-channel electroencephalogram coupling analysis method based on multi-scale multivariable transfer entropy | |
| CN104771163A (en) | Electroencephalogram feature extraction method based on CSP and R-CSP algorithms | |
| CN113191225B (en) | Emotion electroencephalogram recognition method and system based on graph attention network | |
| Chatterjee et al. | Effects of wavelets on quality of features in motor-imagery EEG signal classification | |
| CN106618561A (en) | High-robustness analysis method for electroencephalogram (ECG) signal of disabled people in motion state | |
| CN111544855A (en) | Pure idea control intelligent rehabilitation method based on distillation learning and deep learning and application | |
| CN106503733B (en) | The useful signal recognition methods clustered based on NA-MEMD and GMM | |
| CN114298216A (en) | Electroencephalogram vision classification method based on time-frequency domain fusion Transformer | |
| Morabito et al. | Deep learning approaches to electrophysiological multivariate time-series analysis | |
| Janapati et al. | Towards a more theory-driven BCI using source reconstructed dynamics of EEG time-series | |
| Wang et al. | Multiband decomposition and spectral discriminative analysis for motor imagery BCI via deep neural network | |
| Fauzi et al. | Energy extraction method for EEG channel selection | |
| Ai et al. | Flexible coding scheme for robotic arm control driven by motor imagery decoding | |
| CN113128384B (en) | Brain-computer interface software key technical method of cerebral apoplexy rehabilitation system based on deep learning | |
| CN112259228B (en) | Depression screening method by dynamic attention network non-negative matrix factorization |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |