NATURAL MOVEMENT EEG RECOGNITION METHOD BASED ON SOURCE LOCALIZATION AND BRAIN NETWORKS

Abstract

Disclosed is a natural movement electroencephalogram (EEG) recognition method based on source localization and a brain network, which includes the following steps: (1) performing multi-channel EEG measurement for natural movements; (2) preprocessing acquired EEG signals, and extracting the movement-related cortical potential (MRCP), and θ, α, β, and γ rhythms; (3) determining a lead field matrix of the signals, calculating initial solutions of sources by means of L1 regularization constraint, and then performing iteration by means of successive over-relaxation to obtain a source localization result; (4) by using the sources as nodes, calculating PLV between each pair of sources at each time point by means of short-time sliding window, and establishing brain networks; and (5) calculating a network adjacency matrix at each time point and five brain network indicators, introducing these features into a classifier for training and testing, and conducting a statistical test for the brain network indicators. The present disclosure makes improvements to the conventional source localization method by using the T-wMNE algorithm in combination with successive over-relaxation, and establishes brain networks by using the sources as nodes, thus improving the EEG decoding accuracy for natural movements and revealing the neural mechanism of the human body.

Claims

1. A natural movement electroencephalogram (EEG) recognition method based on source localization and a brain network, comprising the following steps: (1) performing multi-channel EEG measurement for natural movements; (2) preprocessing acquired EEG signals, removing artefacts, and extracting the movement-related cortical potential (MRCP) , θ rhythm, α rhythm, β rhythm, and γ rhythm; (3) determining a lead field matrix of the signals, and calculating initial solutions of sources by means of L1 regularization constraint; and then performing iteration for the initial solutions by means of successive over-relaxation, and using the latest solution vector as a final estimation result of source localization after iteration completion; (4) by using the sources as nodes, calculating PLV between each pair of sources at each time point by means of short-time sliding window; and when the PLV is greater than a set threshold, constructing an edge between the two sources, and using a standardized value of PLV as the weight of the edge; and (5) calculating the characteristic path length, clustering coefficient, average node strength, average betweenness, efficiency, and network adjacency matrix at each time point; introducing these features into a classifier for training and testing; and conducting a statistical test for the first 5 features, to analyze differences in time or frequency of these features corresponding to different movements.

2. The natural movement EEG recognition method based on source localization and a brain network according to claim 1, wherein step (2) comprises the following sub-steps: (a1) performing pre-filtering for the acquired EEG signals; (a2) eliminating the data channel with abnormal kurtosis and performing spherical interpolation, which uses an average value of four channels closest to the interpolated channel as a value of this channel; (a3) identifying and removing EOG and EMG components from the EEG by means of a blind source separation algorithm; (a4) extracting epochs and correcting baseline for the EEG; (a5) removing trials with absolute value of amplitude greater than 200 μV, abnormal joint probability or abnormal kurtosis, wherein the thresholds of the latter two are 5 times the standard deviation of their statistic; (a6) performing common average reference (CAR) for the EEG; and (a7) performing zero-phase Butterworth bandpass filtering at 0.3 Hz to 3 Hz, 4 Hz to 8 Hz, 8 Hz to 13 Hz, 13 Hz to 30 Hz, and 30 Hz to 45 Hz separately for the EEG after re-reference, and extracting the MRCP and the θ, α, β, and γ rhythms.

3. The natural movement EEG recognition method based on source localization and a brain network according to claim 1, wherein step (3) comprises the following sub-steps: (b1) selecting a head model; (b2) solving a forward problem, to obtain the lead field matrix L; (b3) determining a time point to be analyzed, and setting an iteration error ε and the maximum number K of iterations; (b4) calculating an initial solution of a source vector by means of the T-wMNE algorithm:
s.sub.t.sup.(0)=min∥v.sub.t−Ls.sub.t∥.sub.2+λ∥Ws.sub.t∥.sub.1 wherein W=diag(∥l.sub.1∥,∥.sub.2∥, . . . ,∥l.sub.N∥) is a weighted matrix; N is the number of the sources, which is equal to the number of electrodes herein; s.sub.t denotes the source vector at the time point t; v.sub.t denotes the electrode potential at the time point t; and λ is the regularization coefficient; (b5) performing iteration for the initial solution obtained in sub-step (b4) by means of successive over-relaxation: $\{\begin{array}{c}\begin{array}{c}\begin{array}{c}{s}_{1,t}^{\left(k+1\right)}=\left(1-\omega \right)s_{1,t}^{\left(k\right)}+w\left(v_{1,t}-l_{1,2}{s}_{2,t}^{\left(k\right)}-l_{1,3}{s}_{3,t}^{\left(k\right)}-.Math.-l_{1,N}{s}_{N,t}^{\left(k\right)}\right)/l_{1,1}\\ s_{2,t}^{\left(k+1\right)}=\left(1-\omega \right)s_{2,t}^{\left(k\right)}+w\left(v_{2,t}-l_{2,1}{s}_{1,t}^{\left(k+1\right)}-l_{2,3}{s}_{3,t}^{\left(k\right)}-.Math.-l_{2,N}{s}_{N,t}^{\left(k\right)}\right)/l_{2,2}\end{array}\\ .Math.\end{array}\\ \begin{array}{c}{s}_{N,t}^{\left(k+1\right)}=\left(1-\omega \right)s_{N,t}^{\left(k\right)}+\\ w\left(v_{N,t}-l_{N,1}{s}_{1,t}^{\left(k+1\right)}-l_{N,2}{s}_{2,t}^{\left(k+1\right)}-.Math.-l_{N,N-1}{s}_{N-1,t}^{\left(k+1\right)}\right)/l_{N,N}\end{array}\end{array}$ wherein s.sub.i,t denotes a value of the ith source at the time point t, and i=1,2, . . . , N; v.sub.j,t denotes the potential of the jth electrode at the time point t, and j=1, 2, . . . , N; ω is a relaxation factor; and k denotes the number of iterations; and (b6) when ∥s.sub.t.sup.(k+1)−s.sub.t.sup.(k)∥≤ε or k>K, the iteration ends, and the latest solution vector is used as the final estimation result of source localization; otherwise, continuing the iteration.

4. The natural movement EEG recognition method based on source localization and a brain network according to claim 1, wherein in step (4), Hilbert transform is performed on the source vector of a single trial at each time point, to obtain the phase of the source vector at each time point; and then the PLV of each pair of sources at each time point is calculated: $\widehat{s_l}\left(t\right)=H\left(s_i\left(t\right)\right)={\int }_{-\infty }^{+\infty }\frac{s_i\left(\tau \right)}{\pi \left(t-\tau \right)}d\tau$ $\varphi \left(t\right)=\arctan \left(\frac{\widehat{s_l}\left(t\right)}{s_i\left(t\right)}\right)$ ${\mathrm{\Delta \varphi }}_{\mathrm{ij}}={\varphi }_i-{\varphi }_j$ ${\mathrm{PLV}}_{\mathrm{ij}}=\frac{1}{M}.Math.\underset{m=1}{\overset{M}{.Math.}}{e}^{j\Delta {\varphi }_{\mathrm{ij}}}.Math.$ wherein m=1, 2, . . . , M, which denotes the mth trial; and when PLV.sub.ij is greater than the threshold, an edge between sources i and j is constructed; and the PLV is subjected to standardization processing and the standardized value is used as the weight of the edge: $w_{\mathrm{ij}}=\frac{PL{V}_{\mathrm{ij}}-\min \left(\mathrm{PLV}\right)}{\max \left(PLV\right)-\min \left(PLV\right)}$

5. The natural movement EEG recognition method based on source localization and a brain network according to claim 1, wherein in step (5), the statistical test method is t-test; and the classifier is the sLDA classifier, which is trained and tested by performing five-fold cross-validation for ten times.

6. The natural movement EEG recognition method based on source localization and a brain network according to claim 3, wherein in sub-step (b5), with 0.01 step size between (1, 2), ω that minimizes ∥v.sub.t−Ls.sub.t∥ after 10 iterations is selected as the optimal relaxation factor.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0016] FIG. 1 is a flowchart of a natural movement EEG recognition method based on source localization and brain networks of the present disclosure;

[0017] FIG. 2 is a flowchart of EEG signal preprocessing in the natural movement EEG recognition method based on source localization and brain networks of the present disclosure; and

[0018] FIG. 3 is a flowchart of EEG source localization in the natural movement EEG recognition method based on source localization and brain networks of the present disclosure.

DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] The present disclosure is further elaborated below with reference to the accompanying drawings and a specific embodiment. It should be noted that the following specific implementation is merely used to describe the present disclosure rather than limiting the scope of the present disclosure.

[0020] The present disclosure designs a natural movement EEG recognition method based on source localization and brain networks, which, as shown in FIG. 1, includes the following steps:

(1) performing multi-channel EEG measurement for natural movements;
(2) preprocessing acquired EEG signals, removing artefacts, and extracting the MRCP, θ rhythm, α rhythm, β rhythm, and γ rhythm ;
(3) determining a lead field matrix of the signals, and calculating initial solutions of sources by means of L1 regularization constraint; and then performing iteration for the initial solutions by means of successive over-relaxation, and using the latest solution vector as a final estimation result of source localization after iteration completion;
(4) by using the sources as nodes, calculating PLV between each pair of sources at each time point by means of short-time sliding window; and when the PLV is greater than a set threshold, constructing an edge between the two sources, and using a standardized value of PLV as the weight of the edge; and
(5) calculating the characteristic path length, clustering coefficient, average node strength, average betweenness, efficiency, and network adjacency matrix at each time point; introducing these features into a classifier for training and testing; and conducting a statistical test for the first 5 features, to analyze differences in time or frequency of these features corresponding to different movements.

[0021] As shown in FIG. 2, step (2) includes the following sub-steps:

(a1) performing pre-filtering for the acquired EEG signals;
(a2) eliminating the data channel with abnormal kurtosis and performing spherical interpolation, which uses an average value of four channels closest to the interpolated channel as a value of this channel;
(a3) identifying and removing EOG and EMG components from the EEG by means of a blind source separation algorithm;
(a4) extracting epochs and correcting baseline for the EEG;
(a5) removing trials with absolute value of amplitude greater than 200 μV, abnormal joint probability or abnormal kurtosis, where the thresholds of the latter two are 5 times the standard deviation of their statistic;
(a6) performing CAR for the EEG; and
(a7) performing zero-phase Butterworth bandpass filtering at 0.3 Hz to 3 Hz, 4 Hz to 8 Hz, 8 Hz to 13 Hz, 13 Hz to 30 Hz, and 30 Hz to 45 Hz separately for the EEG after re-reference, and extracting the MRCP, θ rhythm, α rhythm, β rhythm, and γ rhythm.

[0022] As shown in FIG. 3, step (3) includes the following sub-steps:

(b1) selecting a head model;
(b2) dividing the scalp into N small 3D grids; placing three current dipoles with the dipole moment along X, Y, and Z axes in each grid, where their vector sum is equivalent to a possible current dipole; and determining the lead field matrix L according to the following equation:

[00004]$V=\left[\begin{array}{ccc}v\left(r_1,1\right)& .Math.& v\left(r_1,T\right)\\ .Math.& \ddots & .Math.\\ v\left(r_N,1\right)& .Math.& v\left(r_N,T\right)\end{array}\right]$$=\left[\begin{array}{ccc}l\left(r_1,r_1^{*}\right)e_1& .Math.& l\left(r_1,r_N^{*}\right)e_N\\ .Math.& \ddots & .Math.\\ l\left(r_N,r_1^{*}\right)e_1& .Math.& l\left(r_N,r_N^{*}\right)e_N\end{array}\right]\left[\begin{array}{ccc}{s}_{1,1}& .Math.& s_{1,T}\\ .Math.& \ddots & .Math.\\ s_{N,1}& .Math.& s_{N,T}\end{array}\right]$$=\mathrm{LS}$

where the ith column in L denotes a potential distribution generated by the ith current dipole source at each electrode position; r.sub.i* denotes a position vector of the current dipole; r.sub.j denotes a position vector of the measured scalp electrode; s=se.sub.i denotes the dipole moment (s is the size and e.sub.i is the direction) of the current dipole; v(r.sub.j, T) denotes the potential of the jth electrode at the time point t; i=1, . . . , N, which denotes that there are N current dipoles; and j=1, . . . , N, which denotes that there are N measurement electrodes;
(b3) determining a time point to be analyzed, and setting an iteration error ε and the maximum number K of iterations;
(b4) calculating an initial solution of a source vector by means of the T-wMNE algorithm:

s.sub.t.sup.(0)=min ∥v.sub.t−Ls.sub.t∥.sub.2+λ∥Ws.sub.t∥.sub.1

where W=diag(∥l.sub.1∥,∥l.sub.2∥, . . . ,∥l.sub.N∥) is a weighted matrix; N is the number of the sources, which is equal to the number of electrodes herein; s.sub.t denotes the source vector at the time point t; v.sub.t denotes the electrode potential at the time point t; and λ is the regularization coefficient;
(b5) performing iteration for the initial solution obtained in sub-step (b4) by means of successive over-relaxation:

[00005]$\{\begin{array}{c}\begin{array}{c}\begin{array}{c}{s}_{1,t}^{\left(k+1\right)}=\left(1-\omega \right)s_{1,t}^{\left(k\right)}+w\left(v_{1,t}-l_{1,2}{s}_{2,t}^{\left(k\right)}-l_{1,3}{s}_{3,t}^{\left(k\right)}-.Math.-l_{1,N}{s}_{N,t}^{\left(k\right)}\right)/l_{1,1}\\ s_{2,t}^{\left(k+1\right)}=\left(1-\omega \right)s_{2,t}^{\left(k\right)}+w\left(v_{2,t}-l_{2,1}{s}_{1,t}^{\left(k+1\right)}-l_{2,3}{s}_{3,t}^{\left(k\right)}-.Math.-l_{2,N}{s}_{N,t}^{\left(k\right)}\right)/l_{2,2}\end{array}\\ .Math.\end{array}\\ \begin{array}{c}{s}_{N,t}^{\left(k+1\right)}=\left(1-\omega \right)s_{N,t}^{\left(k\right)}+\\ w\left(v_{N,t}-l_{N,1}{s}_{1,t}^{\left(k+1\right)}-l_{N,2}{s}_{2,t}^{\left(k+1\right)}-.Math.-l_{N,N-1}{s}_{N-1,t}^{\left(k+1\right)}\right)/l_{N,N}\end{array}\end{array}$

where ω∈(1,2) is a relaxation factor, and ω that minimizes ∥v.sub.t−Ls.sub.t∥ after 10 iterations is selected as the optimal relaxation factor with 0.01 step size between (1, 2); s.sub.i,t denotes a value of the ith source at the time point t, and i=1,2, . . . , N; v.sub.j,t denotes the potential of the jth electrode at the time point t, and j=1, 2, . . . , N; and k denotes the number of iterations; and
(b6) when ∥s.sub.t.sup.(k+1)−s.sub.t.sup.(k)∥≤ε or k>K, the iteration ends, and the latest solution vector is used as the final localization estimation result of the source; otherwise, continuing the iteration.

[0023] In step (4), Hilbert transform is performed on the source vector of a single trial at each time point, to obtain the phase of the source vector at each time point; and then the PLV of each pair of sources at each time point is calculated:

[00006]$\widehat{s_l}\left(t\right)=H\left(s_i\left(t\right)\right)={\int }_{-\infty }^{+\infty }\frac{s_i\left(\tau \right)}{\pi \left(t-\tau \right)}d\tau$$\varphi \left(t\right)=\arctan \left(\frac{\widehat{s_l}\left(t\right)}{s_i\left(t\right)}\right)$${\mathrm{\Delta \varphi }}_{\mathrm{ij}}={\varphi }_i-{\varphi }_j$${\mathrm{PLV}}_{\mathrm{ij}}=\frac{1}{M}.Math.\underset{m=1}{\overset{M}{.Math.}}{e}^{j\Delta {\varphi }_{\mathrm{ij}}}.Math.$

where m=1, 2, . . . , M, which denotes the mth trial.

[0024] When PLV.sub.ij is greater than the set threshold, an edge between sources i and j is constructed; and the PLV is subjected to standardization processing and the standardized value is used as the weight of the edge:

[00007]$w_{\mathrm{ij}}=\frac{{\mathrm{PLV}}_{\mathrm{ij}}-\min \left(\mathrm{PLV}\right)}{\max \left(\mathrm{PLV}\right)-\min \left(\mathrm{PLV}\right)}$

[0025] In step (5), the statistical test method is t-test; and the classifier is the sLDA classifier, which is trained and tested by performing five-fold cross-validation for ten times.