A rhythm movement decoding method based on electroencephalogram signals

Abstract

The application discloses a rhythm movement decoding method based on electroencephalogram signals, relates to the field of electroencephalogram signal processing and brain-computer interface, and comprises a Task-Discriminant Component Analysis (TDCA) method and a common spatial pattern (CSP) method, which are collectively referred to as TDCA-CSP technology; samples are firstly subjected to the TDCA to obtain predicted frequencies, and then subjected to the CSP to obtain predicted limbs, including left hands and right hands; in the training stage, TDCA spatial filters for frequency classification and CSP spatial filters for limb classification are obtained through training of training data of various categories; in the testing stage, samples are fed into the TDCA-CSP decoding framework to perform frequency and limb prediction. The method can realize simultaneous decoding of the frequency and corresponding limbs of movement based on the TDCA-CSP decoding framework, so that classification of various rhythm movements is realized.

Description

Technical Field

[0001] This invention relates to EEG signal processing and brain-computer interfaces, and more particularly to a method for decoding rhythmic motion based on EEG signals. Background Technology

[0002] Brain-computer interfaces (BCIs) offer an effective approach for the reconstruction and rehabilitation of motor function. Among them, BCIs based on electroencephalograms (EEGs) have attracted widespread attention due to the non-invasive nature and high temporal resolution of EEG signals.

[0003] The foundation for reconstructing motor function lies in decoding movement-related information from electroencephalogram (EEG) signals. Traditional movement decoding methods are based on sensorimotor rhythm (SMR). When a limb, such as the left or right hand, is actually or imagined to move, energy changes in the corresponding area of ​​the sensorimotor cortex in the limb's frequency bands occur. By analyzing the brain regions where energy changes occur, the limb involved in the actual or imagined movement can be decoded. However, SMR-based movement decoding can only decode the limb corresponding to the movement, and cannot decode more refined movement-related information. Furthermore, the spatial resolution of EEG is limited, making it difficult to support the decoding of finer details such as different fingers. Therefore, SMR-based movement decoding can basically only achieve four classifications: left hand / right hand / both feet / tongue.

[0004] Furthermore, current multi-class (more than 4 classes) motor decoding based on EEG signals generally has low accuracy. For example, Schwarz et al. conducted a 7-class motor decoding study on single / double hand extension / grasping tasks and rest, achieving an accuracy of 38.0% ± 6.6%. Jia et al. also decoded 7 motor tasks, including elbow flexion / extension, forearm pronation / pronation, hand closure / opening, and rest, achieving an accuracy of 41.93% ± 7.8%. While these accuracy rates are significantly higher than random levels, they still fall far short of the requirements for practical applications.

[0005] In recent years, some researchers have discovered that when performing movement, motion observation, or motion imagery, if the movement is rhythmic—that is, repeated at a fixed frequency—the corresponding motor cortex of that limb will also produce brain electrical activity at the same frequency and twice that frequency. This brain activity is called steady-state movement-related rhythm (SSMRR). If the characteristics of SSMRR can be fully utilized, it is hoped that both the limb and the frequency of movement can be decoded simultaneously, thereby expanding the number of classifiable movements. However, currently, there is a lack of effective decoding methods to fully utilize the characteristics of SSMRR to decode rhythmic movements of different frequencies.

[0006] Therefore, those skilled in the art are dedicated to developing a rhythmic motion decoding method based on SSMRR in EEG signals, which can simultaneously decode moving limbs and motion frequencies with high accuracy, increase the number of classifiable brain-computer interface systems, and lay a solid foundation for achieving more flexible brain-computer interaction. Summary of the Invention

[0007] In view of the above-mentioned deficiencies of the prior art, the technical problem to be solved by the present invention is how to distinguish motion frequencies.

[0008] To achieve the above objectives, this invention provides a rhythmic motion decoding method based on electroencephalogram (EEG) signals. The method is characterized by including Task-Discriminant Component Analysis (TDCA) and Common Spatial Pattern (CSP), collectively referred to as TDCA-CSP technology. The sample first obtains the predicted frequency through TDCA, and then obtains the predicted limb (left and right hand) through CSP. During the training phase, TDCA obtains an optimal spatial filter using training data from various categories. During the testing phase, the sample is fed into the corresponding CSP for left-hand and right-hand decoding.

[0009] Furthermore, the data in the method requires preprocessing, including bandpass filtering, 50Hz notch filtering to remove power frequency interference, downsampling, and artifact subspace reconstruction, i.e., ASR method to remove electrooculography and motion artifacts; the bandpass filtering refers to using a 4th-order Butterworth filter for bandpass filtering from 0.8 to 40Hz; the downsampling refers to downsampling from 1024Hz to 256Hz.

[0010] Furthermore, when generating the spatial filter, TDCA is required to perform two-stage data augmentation on the data, assuming that the data for each trial is... Nc represents the number of channels, and Nt represents the number of time sampling points; the first stage enhancement involves expanding X to... in This is the enhanced data. The delayed data is obtained by delaying X by l time sampling points; the second stage enhancement is to expand the data to in This is the enhanced data. yes Projection into the orthogonal space formed by the sine-cosine reference signals.

[0011] Furthermore, if the delayed data exceeds the range of trial, the data is padded with zeros; the frequency of the sine-cosine reference signal used includes the frequency of the rhythmic motion task and its second harmonic.

[0012] Furthermore, after the data augmentation, the optimal spatial filter is obtained through two-dimensional linear discriminant analysis (LDA) to maximize inter-class differences and minimize intra-class differences.

[0013] Furthermore, the spatial filter solution process includes constructing a matrix representing the differences between classes. for Construct a matrix representing intra-class differences for Where N trial N represents the total number of trials in the training dataset. class Represents the total number of categories; This represents the data for the i-th trial. The average of the data corresponding to the category of the i-th trial; Represents the average of the i-th class; The average of all trials across all classes; optimal spatial filter The following optimization problem was solved to obtain the following result: Select The first N sub Subspace, obtain

[0014] Furthermore, in the testing phase, the samples used for testing will first undergo the same data augmentation process as the training data; then, they will undergo the following template matching process.

[0015]

[0016]

[0017] Finally, the predicted category is obtained.

[0018] Furthermore, since each task category corresponds to both frequency and limb labels, the predicted category... It also includes labels for both frequency and limb.

[0019] Furthermore, the CSP in the method comprises two parts: CSP feature extraction and Support Vector Machine (SVM) classification, which also require training with training data. Specifically, the CSP at that frequency is trained using left-hand / right-hand data at the same frequency. The frequency range selected for CSP is a 1Hz band centered on the second harmonic of the task frequency. This frequency range must cover the second harmonic of the task frequency and is a relatively narrow band. After the CSP extracts features, the corresponding SVM classifier is trained using these features. A separate CSP-SVM needs to be trained for each left-hand and right-hand data at each frequency.

[0020] Furthermore, after the test sample is processed by the CSP-SVM, the left-hand / right-hand prediction results are obtained, which together with the frequency prediction of TDCA form a complete prediction label, and finally predict which class the sample belongs to.

[0021] The present invention has the following technical effects:

[0022] (1) Based on the TDCA-CSP decoding framework, the frequency of the movement and the corresponding limb can be decoded simultaneously, thereby realizing the classification of multiple rhythmic movements.

[0023] (2) TDCA-CSP can make full use of the time and space information in the SSMRR signal, thereby achieving a higher decoding accuracy.

[0024] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description

[0025] Figure 1 This is a diagram of the TDCA-CSP algorithm according to a preferred embodiment of the present invention;

[0026] Figure 2 This is an electrode channel diagram of a preferred embodiment of the present invention;

[0027] Figure 3 This is a preferred embodiment of the experimental paradigm for verification.

[0028] Figure 4 This is a decoding result confusion matrix of a preferred embodiment of the present invention. Detailed Implementation

[0029] The following description, with reference to the accompanying drawings, illustrates several preferred embodiments of the present invention to make its technical content clearer and easier to understand. The present invention can be embodied in many different forms, and the scope of protection of the present invention is not limited to the embodiments mentioned herein.

[0030] In the accompanying drawings, components with the same structure are indicated by the same numerical designation, and components with similar structures or functions are indicated by similar numerical designations. The dimensions and thicknesses of each component shown in the drawings are arbitrary, and the present invention does not limit the dimensions and thicknesses of each component. To make the illustrations clearer, the thickness of some components has been appropriately exaggerated in the drawings.

[0031] To address the issues that existing motion decoding methods based on EEG signals can only distinguish the movements of different limbs but not the frequency of the movements, and that the accuracy of multi-class (more than 4 classes) motion decoding based on EEG signals is generally low, we propose the TDCA-CSP decoding method for rhythmic motion.

[0032] Our data acquisition uses a 64-channel Biosemi Active Two system with a sampling frequency of 1024Hz. The channel arrangement uses a 10-20 system (e.g., ...). Figure 2 ).

[0033] Data needs to be preprocessed before use. Common preprocessing procedures include bandpass filtering (specifically, a 4th-order Butterworth filter can be used for bandpass filtering from 0.8 to 40 Hz), 50 Hz notch filtering to remove power frequency interference, downsampling (specifically downsampling from 1024 Hz to 256 Hz), and using the artifact subspace reconstruction (ASR) method to remove electrooculography and motion artifacts.

[0034] The TDCA-CSP method comprises two parts: Task-Discriminant Component Analysis (TDCA) and Common Spatial Pattern (CSP). Samples are first analyzed using TDCA to obtain predicted frequencies, and then CSP is used to obtain predicted limbs (left / right hand). Specifically, TDCA-CSP is a trained method. A test dataset must be collected for training before use.

[0035] TDCA obtains an optimal spatial filter using training data from various categories. First, TDCA performs two-stage data augmentation. Assume the data for each trial is... Nc represents the number of channels, and Nt represents the number of time sampling points. The first stage enhancement involves expanding X to... in This is the enhanced data. It is obtained by delaying X by l time sampling points. If the data exceeds the range of trial after the delay, the data is padded with zeros. The second stage enhancement is to expand the data to... in This is the enhanced data. yes The projection onto the orthogonal space formed by the sine-cosine reference signals. The frequencies of the sine-cosine reference signals used include the frequencies of the rhythmic motion task and their second harmonics.

[0036] After data augmentation, the optimal spatial filter is obtained through two-dimensional linear discriminant analysis (LDA), with the principle of maximizing inter-class variance and minimizing intra-class variance. A matrix representing inter-class variance is then constructed. for Construct a matrix representing intra-class differences for Where N trial N represents the total number of trials in the training dataset. class Represents the total number of categories. This represents the data for the i-th trial. This represents the average of the data corresponding to the category of the i-th trial. This represents the average of the i-th class. This represents the average of all trials across all classes. Optimal spatial filter. The following optimization problem was solved to obtain the following result: Generally, select The first N sub Subspace, obtain

[0037] During the testing phase, the samples used for testing undergo the same data augmentation process as the training data. Then, the predicted category is obtained through a template matching process as follows.

[0038]

[0039]

[0040] In particular, since each task category corresponds to both frequency and limb labels, the predicted category... This also corresponds to two parts of labels. In the TDCA-CSP method, we only adopt TDCA's predictions for frequency, and do not adopt predictions for limbs such as left / right hand. Based on the frequency prediction results, the samples are sent to the corresponding CSP for left / right hand decoding.

[0041] CSP (Common Prompt Service) consists of two parts: CSP feature extraction and SVM (Support Vector Machine) classification. It also requires training data. Specifically, the CSP for that frequency is trained using left-hand / right-hand data at the same frequency. The frequency range selected for CSP is a 1Hz band centered on the second harmonic of the task frequency (e.g., if the rhythmic movement task frequency is 1.5Hz, the selected band would be 2.5Hz-3.5Hz. This selection method can be adjusted according to the actual situation, but the principle is to cover the second harmonic of the task frequency, and it should be a relatively narrow band). After CSP extracts features, the corresponding SVM classifier is trained using these features. (For example, if there are 6 types of rhythmic movement tasks: 1Hz-left hand, 1Hz-right hand; 2Hz-left hand, 2Hz-right hand; 3Hz-left hand, 3Hz-right hand, then a separate CSP-SVM needs to be trained for each frequency's left and right hands).

[0042] After the test sample is processed by CSP (CSP-SVM), left-hand / right-hand prediction results are obtained. These results, together with the frequency predictions from TDCA, form a complete prediction label, ultimately predicting which class the sample belongs to. The prediction process for the test sample is as follows: Figure 1 As shown in the figure. (TDCA and CSP in the figure are both pre-trained).

[0043] To verify the effectiveness of the method of this invention in rhythmic motion decoding, data from 15 participants were collected through an experiment. The experimental procedure is described below: Throughout the experiment, participants sat in comfortable chairs in an electric shock shielded room. Task cues were presented via a monitor in front of them. A sound metronome provided beats at specified frequencies. Participants were instructed to perform rhythmic flexion and extension movements using the index and middle fingers of a designated hand. There were three task frequencies: 1.2Hz, 2.0Hz, and 2.8Hz, resulting in six task categories involving the left and right hands and three frequency combinations. The experiment consisted of 10 runs, each containing 30 trials, with each of the six task categories performed five times (in random order). The timing of each trial was as follows: Figure 3 As shown. At the start of the trial, a cross appears on the black screen, and the participant needs to focus their attention on the center of the cross. After 3 seconds, task instructions appear, including an arrow pointing left or right (corresponding to the left or right hand) and a task frequency cue. A sound metronome starts after 1 second and lasts for 5 seconds. During this phase, the participant needs to perform finger flexion and extension movements synchronized with the sound rhythm using the corresponding hand. This is followed by a 3-second rest period.

[0044] Based on the data collected in the above experiments, 10×5-fold cross-validation was performed. The results show that the TDCA-CSP algorithm in this invention can achieve a 6-class motion decoding accuracy of 72.92±14.83% (confusion matrix as shown). Figure 4 This demonstrates the effectiveness of the TDCA-CSP method for decoding rhythmic motion.

[0045] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for decoding rhythmic motion based on electroencephalogram (EEG) signals, characterized in that, This includes Task-Discriminant Component Analysis (TDCA) and Common Spatial Pattern (CSP), collectively known as TDCA-CSP technology. Samples are first analyzed using TDCA to obtain predicted frequencies, and then CSP is used to obtain predicted limbs, including left and right hands. During the training phase, training data from various categories is used to train TDCA spatial filters for frequency classification and CSP spatial filters for limb classification. During the testing phase, samples are fed into the TDCA-CSP decoding framework for frequency and limb prediction. When generating the spatial filter, TDCA is required to perform two-stage data augmentation on the data. Assume the data for each trial is... ， Nc is the number of channels, and Nt is the number of time sampling points; The first phase of enhancement is to... Expand to ,in This is the enhanced data. Depend on Delay l Delay data obtained from each time sampling point; The second phase enhancement expands the data to... ,in This is the enhanced data. yes Projection into the orthogonal space formed by the sine-cosine reference signal; The spatial filter solution process includes constructing a matrix representing the differences between classes. for ; Construct a matrix representing intra-class differences for ;in This represents the total number of trials in the training dataset. Represents the total number of categories; This represents the data for the i-th trial. The average of the data corresponding to the category of the i-th trial; Represents the average of the i-th class; The average of all trials across all classes; optimal spatial filter The following optimization problem was solved to obtain the following result: Select The former Subspace, obtain .

2. The rhythmic motion decoding method based on electroencephalogram (EEG) signals as described in claim 1, characterized in that, The data in the method requires preprocessing, including bandpass filtering, 50Hz notch filtering to remove power frequency interference, downsampling, and artifact subspace reconstruction, i.e., ASR method to remove electrooculography and motion artifacts; the bandpass filtering refers to using a 4th-order Butterworth filter for bandpass filtering from 0.8 to 40Hz; the downsampling refers to downsampling from 1024Hz to 256Hz.

3. The rhythmic motion decoding method based on electroencephalogram (EEG) signals as described in claim 2, characterized in that, If the delayed data exceeds the range of trial, the data is padded with zeros; the frequency of the sine-cosine reference signal used includes the frequency of the rhythmic motion task and its second harmonic.

4. The rhythmic motion decoding method based on electroencephalogram (EEG) signals as described in claim 3, characterized in that, After data augmentation, the optimal spatial filter is obtained through two-dimensional linear discriminant analysis (LDA) to maximize inter-class differences and minimize intra-class differences.

5. The rhythmic motion decoding method based on electroencephalogram (EEG) signals as described in claim 4, characterized in that, During the testing phase, the samples used for testing will first undergo the same data augmentation process as the training data; then, they will be subjected to the following template matching process. Finally, the predicted category is obtained. .

6. The rhythmic motion decoding method based on electroencephalogram (EEG) signals as described in claim 5, characterized in that, Since each task category corresponds to both frequency and limb labels, the predicted category... It also includes labels for both frequency and limb.

7. The rhythmic motion decoding method based on electroencephalogram (EEG) signals as described in claim 6, characterized in that, The CSP in the method comprises two parts: CSP feature extraction and Support Vector Machine (SVM) classification, and also requires training with training data. Specifically, the CSP at that frequency is trained using left-hand / right-hand data at the same frequency. The frequency range selected for CSP is a 1Hz band centered on the second harmonic of the task frequency. This frequency range should cover the second harmonic of the task frequency and be a relatively narrow band. After CSP extracts features, the corresponding SVM classifier is trained using these features. A separate CSP-SVM needs to be trained for each left-hand and right-hand data at each frequency.

8. The rhythmic motion decoding method based on electroencephalogram (EEG) signals as described in claim 7, characterized in that, After the test sample is processed by the CSP-SVM, the left-hand / right-hand prediction results are obtained. Together with the frequency prediction of TDCA, they form a complete prediction label, and finally predict which class the sample belongs to.

Images