Research method for acupuncture point classification and multi-channel electroencephalogram correlation analysis based on convolutional network

Abstract

The invention discloses a research method for acupuncture point classification and multi-channel electroencephalogram signal correlation analysis based on a convolutional network, and belongs to the field of brain-computer interfaces and the field of signal processing. The method is specifically implemented according to the following steps: firstly, acquiring required data by using an electroencephalogram signal acquisition device, and preprocessing the acquired electroencephalogram signal: filtering and segmenting; the correlation between the channels is studied, and visual display is carried out; extracting time-frequency characteristics of the electroencephalogram signals through wavelet transformation; drawing a time-frequency graph; inputting a time-frequency diagram result into an SE-ConvNeXt network for training, and obtaining an acupoint classification result; and analyzing the result, and visually expressing the accuracy by using a confusion matrix. The method disclosed by the invention has important reference significance for deeply studying the regulation mechanism of the acupuncture point acupuncture traditional Chinese medicine therapy on the central nervous system and studying the acupuncture point acupuncture classification according to the electroencephalogram signals.

Description

Technical Field

[0001] The present invention belongs to the field of brain-computer interface and signal processing, and specifically relates to a research method for acupuncture point classification and multi-channel electroencephalogram signal correlation analysis based on convolutional networks. Background Art

[0002] Acupuncture originated from traditional Chinese medicine and has a history of thousands of years. It was originally used to relieve pain and treat various diseases. With the development of modern science, acupuncture has gradually attracted the attention of researchers around the world. In recent years, research has focused on its physiological mechanism, its impact on the nervous system, and its clinical effects in relieving pain, anxiety, insomnia, etc.

[0003] Acupuncture, as an important traditional Chinese medicine therapy, is widely used in the clinic to treat various diseases, especially in relieving pain, improving mood, regulating body functions, etc. However, the mechanism of action of acupuncture has not yet been fully revealed.

[0004] In recent years, electroencephalogram (EEG) signals have been used to study psychological processes such as emotions, attention, and memory, helping scientists gain a deeper understanding of the working mechanism of the brain and the neural basis of various psychological states. With the continuous advancement of technology, the application prospects of EEG research are becoming more and more broad. As a non-invasive and real-time means of reflecting the state of brain activity, EEG provides important technical support for studying the regulation of acupuncture on the central nervous system.

[0005] By combining acupuncture therapy with modern EEG technology, we can further study the regulatory mechanism of acupuncture on the central nervous system. These background technologies provide theoretical and technical support for experimental research on the effects of acupuncture on EEG, and are an important basis for promoting the scientific and modern development of acupuncture therapy. Summary of the invention

[0006] In order to solve the problem that the mechanism of acupuncture is not obvious, the present invention proposes a research method based on convolutional network for acupuncture point classification and multi-channel EEG signal correlation analysis, which uses existing EEG technology to expand the effect of Chinese acupuncture and moxibustion, and shows that acupuncture is a reaction state of EEG, making Chinese acupuncture technology more scientific and rational. This method classifies the acupuncture effects of different acupoints by inputting relevant EEG data to reversely verify the effectiveness of acupuncture, that is, acupuncture does have a certain impact on human electrical signals.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] A research method for acupuncture point classification and multi-channel EEG signal correlation analysis based on convolutional networks, comprising the following steps:

[0009] a. Collect the required data using EEG signal acquisition equipment;

[0010] b. EEG signal preprocessing;

[0011] c. Perform correlation analysis on EEG signals and then visualize them;

[0012] d. Use wavelet transform to extract the time-frequency characteristics of EEG signals and draw time-frequency diagrams;

[0013] e. Put the time-frequency graph results into SE-ConvNeXt network training to obtain the prediction results of acupuncture point effect classification;

[0014] f. Result analysis, use confusion matrix to intuitively express the accuracy.

[0015] In step a, an electroencephalogram (EEG) device is used to collect EEG signal responses during acupuncture of different acupoints as training sets and test sets to train the acupuncture point classification network; the original EEG data is collected by a 32-lead device with a sampling rate of 250 Hz; the EEG is collected before acupuncture, during acupuncture, and after needle removal, and the EEG is continuously collected after needle insertion. The needle retention time is set to 25 minutes, and the needle sensation is adjusted every 5 minutes to ensure the continuity of the acupuncture stimulation. The EEG recording of the entire acupuncture process is a block; an event marker is synchronously recorded in the EEG during needle insertion and needle sensation adjustment for subsequent data collation and analysis.

[0016] In step b, a Butterworth bandpass filter is selected to filter out and sort out the information of the main frequency band of the EEG signal, and only the 30-second duration after stimulation is taken as valid data as the content of the subsequent data set.

[0017] In step c, in order to explore the response rules of acupuncture in different brain regions (channels), that is, whether there are two signals with the same or different amplitudes in any band during stimulation; any two channels with any frequency bands are subjected to cross-correlation analysis, and a correlation image is drawn; the normalized cross-correlation calculation formula in the correlation analysis is:

[0018]

[0019] Among them, x(k) and y(k) are EEG signals of different channels; r(n) is the correlation between EEG signals.

[0020] In step d, the method of extracting the time-frequency characteristics of the EEG signal by wavelet transform is as follows:

[0021] Use wavelet transform to perform time-frequency analysis on the filtered signal to obtain the frequency characteristic information of each channel and each segment. The corresponding wavelet transform is:

[0022]

[0023] Among them, W fi (a,τ) is the different power density of each channel at time t = τ, f i,j (t) is the jth segment EEG signal of channel i, is the complex conjugate of the wavelet basis function, and the wavelet basis function is expressed as:

[0024]

[0025] in, is the normalization factor, a is the scale factor, and b is the translation factor. The wavelet basis function commonly used in EEG analysis is Morlet wavelet, which is expressed as After substituting into the above formula, the specific expression is:

[0026]

[0027] Among them, f 0 is the center frequency and σ is the standard deviation.

[0028] In the step e, the feature map extracted by wavelet transform and the sorted related data are classified, and the training set is input into the SE-ConvNeXt network for training to obtain the prediction classification results of acupuncture points; the SE-ConvNeXt network is improved on the basis of ConvNeXt, and the SE model (Squeeze-and-Excitation Model) is added before the feature is input into the network to improve the representation ability of the network, and the weights of different channels are adaptively allocated and learned by the network itself; each layer of the input feature map X is convolved with a 2D spatial kernel to finally obtain C output feature maps, which constitute the feature maps U, F sq Indicates that the global pooling is reduced to a one-dimensional vector, F ex The gate mechanism consists of two fully connected layers. The first fully connected layer compresses C channels into C / r channels to reduce the amount of calculation, and then passes through a RELU nonlinear activation layer. The second fully connected layer restores the number of channels to C channels, and then uses Sigmoid activation to obtain the weight s. The dimension of the final s is 1×1×C, which is used to characterize the weights of the C feature maps in the feature map U; F scale The attention weights obtained previously are weighted to the features of each channel; the next layer consists of convolutional layer, ConvNeXt Block, Downsample, local pooling and fully connected layer; the activation function used in ConvNeXt Block is GELU; the Downsample layer consists of layer normalization and two-dimensional convolution kernel.

[0029] In step f, the model effect needs to be judged. First, a confusion matrix is ​​constructed for each i acupoint category classification effect, and the TP (true positive), FP (false positive), FN (false negative) and TN (true negative) values ​​of each category are counted. Then, the overall precision (Micro-Precision) can be calculated based on these values. The calculation formula is:

[0030]

[0031] The effectiveness of the model is judged by calculating the overall accuracy. The closer the value is to 1, the more effective the model is in predicting acupoints through EEG features, and vice versa.

[0032] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0033] 1. The method based on EEG correlation analysis of the present invention visualizes the correlation characteristics of EEG signal sequences of different channels, adding a new option for a more intuitive display of the scientific nature of the field of traditional Chinese medicine acupuncture.

[0034] 2. Subdividing EEG signals into signals of specific different bands can also demonstrate the correlation analysis effect of arbitrary frequency bands of different channels and different frequency bands of the same channel during acupuncture.

[0035] 3. The present invention uses a ConvNeXt network to classify the features extracted by wavelet transform. The network increases the size of the convolution kernel and the model width, but through refined design, it maintains a high computational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a flowchart of the program of the present invention.

[0037] Figure 2 This is the flow chart of the EEG acquisition experiment.

[0038] Figure 3 Cross-correlation analysis of channels C3 and Cz.

[0039] Figure 4 This is the network architecture diagram of the classification network SE-ConvNeXt proposed in the present invention. DETAILED DESCRIPTION

[0040] The present invention provides a research method for acupuncture point classification and multi-channel EEG signal correlation analysis based on convolutional networks, which can not only visualize the correlation of each channel between the control and experimental groups, but also serve as a verification of the effectiveness of acupuncture points. Figure 1 , which comprises the following steps:

[0041] a. Use EEG signal acquisition equipment to collect the required data.

[0042] Use EEG equipment to collect relevant data information, specify the experimental plan, connect to the computer, and prepare for the experiment.

[0043] Design an acupuncture EEG acquisition experiment. The purpose of the experiment is to collect EEG signal responses when acupuncture is performed on different acupoints as training and test sets to train the acupuncture point classification network. The original EEG data of the experiment is collected by a 32-lead device with a sampling rate of 250 Hz. The experimental content includes baseline acquisition (5 minutes before acupuncture), during acupuncture (continuous EEG acquisition after needle insertion, setting the needle retention time to 25 minutes, and adjusting the needle sensation every 5 minutes (such as slight twisting) to ensure the continuity of acupuncture stimulation), and after needle removal (5 minutes). The EEG recording of the entire acupuncture process is a block. An event marker (event markers) will be synchronously recorded in the EEG during acupuncture and needle sensation adjustment (see Figure 2 ) for subsequent data analysis.

[0044] According to the purpose of the experiment, acupoints with clear therapeutic effects and research basis are selected. For example: Shenmen (HT7): Located on the ulnar side of the wrist crease, in the radial depression of the ulnar wrist flexor tendon, Shenmen acupoint has the effect of calming the mind and calming the nerves in traditional Chinese medicine. Neiguan (PC6): Located on the inner side of the forearm, it is often used to relieve anxiety and improve sleep, and its regulatory effect on EEG is studied. Jianjing (GB21): Located at the highest point of the shoulder, it is often used to treat pain and stiffness in the shoulder and neck. Hegu (LI4): Located on the back of the hand, it is used for analgesia and mood regulation.

[0045] The operator must have extensive clinical experience to ensure the accuracy and consistency of the technique. The number of acupoints is selected as 1 to avoid too many acupoints interfering with the research results. To ensure that the number of subjects meets statistical requirements, at least 30-50 volunteers are required to ensure the reliability and representativeness of the data. The frequency of acupuncture is selected once a day or once every other day for 2-4 weeks to ensure the accumulation and stability of the therapeutic effect. Acupuncture operation safety records are established to observe the subjects' reactions at any time to ensure the safety and comfort of the subjects. If there is discomfort or abnormal reaction, it should be handled and recorded in a timely manner.

[0046] b. EEG signal preprocessing.

[0047] A Butterworth bandpass filter (0.5-40Hz) was selected to filter out and organize the information of the main frequency band of the EEG signal. Since a single piece of data is too long, it is not conducive to accurate feature extraction and subsequent calculations, so the data is segmented at the marked points and divided into 7 segments. At this time, the length of each time segment is 5 minutes. From the perspective of the effect of acupuncture points, most of the data is still invalid and lengthy, so only the 30 seconds after stimulation of the acupuncture part is taken as valid data as the content of the subsequent data set.

[0048] c. Perform correlation analysis on EEG signals and then display them visually.

[0049] In order to explore the response rules of acupuncture in different brain regions (channels), that is, whether there are two signals with the same or different amplitudes in any band during stimulation. The cross-correlation analysis of any two channels and any frequency bands was performed separately, and the correlation image was drawn. The normalized cross-correlation calculation formula in the correlation analysis is:

[0050]

[0051] Among them, x(k) and y(k) are EEG signals of different channels; r(n) is the correlation between EEG signals.

[0052] The cross-correlation calculated using the Xcorr function is measured by moving a time series vector step by step and repeatedly calculating the correlation between the two signals. If the peak of the correlation is at the center (offset = 0), it means that the two time series are most correlated at this time. However, if one signal is leading the other, the peak of the correlation may be at different coordinates. Using channels C3 and Cz as an example, draw a cross-correlation graph (see Appendix Figure 3 ). As can be seen from the figure, the maximum correlation value exceeds 0.8, and the peak of the correlation is located in the center. Generally speaking, if the correlation coefficient exceeds 0.75, it can be said to have a strong correlation.

[0053] d. Use wavelet transform to extract the time-frequency characteristics of EEG signals and draw the time-frequency diagram.

[0054] Use wavelet transform to perform time-frequency analysis on the filtered signal to obtain the frequency characteristic information of each channel and each segment. The corresponding wavelet transform is:

[0055]

[0056] Among them, W fi (a,τ) is the different power density of each channel at time t = τ, f i,j (t) is the jth EEG signal of channel i, is the complex conjugate of the wavelet basis function, which is generally expressed as:

[0057]

[0058] in, is the normalization factor, a is the scale factor, and b is the translation factor. The wavelet basis function commonly used in EEG analysis is Morlet wavelet, which is expressed as After substituting into the above formula, the specific expression is:

[0059]

[0060] Among them, f 0 is the center frequency and σ is the standard deviation (which controls the width of the time window).

[0061] e. Put the time-frequency graph results into SE-ConvNeXt network training to obtain the prediction results of acupuncture point effect classification.

[0062] The feature map extracted from the time-frequency information collected by wavelet transform in the previous step is input into the SE-ConvNeXt network (see Figure 4 ), predict the classification of acupuncture points.

[0063] ConvNeXt is a pure convolution-based network that has significantly improved classification accuracy compared to traditional convolutional networks. Compared with the 3×3 convolution kernel commonly used in traditional CNN, ConvNeXt uses larger convolution kernels (such as 7×7 and 9×9), which can extract feature information more widely. In addition, ConvNeXt not only optimizes the depth of the network, but also increases the width of the network, that is, it introduces more channels, which improves the model's ability to express complex features. Its network architecture also draws on ResNet's staged downsampling strategy. Each stage reduces the amount of calculation by reducing the resolution of the feature map, while retaining more fine-grained features, thereby achieving higher accuracy in classification and segmentation tasks. In addition, the network has made many adjustments in details, such as using fewer activation functions and normalization layers; using a separate downsampling layer, and so on.

[0064] The SE-ConvNeXt network constructed in this invention is an improvement on ConvNeXt. Before the features are input into the network, SE model (Squeeze-and-Excitation Model) is added to improve the representation ability of the network. The core idea is that the weights of different channels should be adaptively allocated and learned by the network itself. Its main structure is as follows: Figure 4 As shown in the SE Model, F tr It represents the convolution operation. Each layer of the input feature map X undergoes a convolution with a 2D spatial kernel and finally obtains C output feature maps, which constitute feature maps U, F sq Indicates that the global pooling is reduced to a one-dimensional vector, F exThe gate mechanism consists of two fully connected layers. The first fully connected layer compresses C channels into C / r channels to reduce the amount of calculation. Then, through a RELU nonlinear activation layer, the second fully connected layer restores the number of channels to C channels. Then, through Sigmoid activation, the weight s is obtained. The dimension of the final s is 1×1×C, which is used to characterize the weights of the C feature maps in the feature map U. r refers to the compression ratio. Finally, F scale The attention weights obtained above are weighted to the features of each channel. Next, it consists of a convolutional layer, a ConvNeXt Block, a downsampling layer, a local pooling layer, and a fully connected layer; the activation function used in the ConvNeXt Block is GELU; the downsampling layer consists of layer normalization and a two-dimensional convolution kernel.

[0065] f. Result analysis, use confusion matrix to intuitively express the accuracy.

[0066] To judge the model effect, firstly, a confusion matrix is ​​constructed for each i-point category classification effect. Usually, in multi-classification problems, the TP (true positive), FP (false positive), FN (false negative) and TN (true negative) values ​​of each category are concerned, and then some indicators can be calculated based on these values. The overall precision (Micro-Precision) is used in this method, and its calculation formula is:

[0067]

[0068] The effectiveness of the model is judged by calculating the overall accuracy. The closer the value is to 1, the more effective the model is in predicting acupoints through EEG features, and vice versa.

[0069] The above contents are merely examples and explanations of the concept of the present invention. The technicians in this technical field may make various modifications or additions to the specific embodiments described or replace them in a similar manner. As long as they do not deviate from the concept of the invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A research method for acupuncture point classification and multi-channel EEG signal correlation analysis based on convolutional networks, characterized in that: The steps include: a. Collect the required data using EEG signal acquisition equipment; b. EEG signal preprocessing; c. Perform correlation analysis on EEG signals and then visualize them; d. Use wavelet transform to extract the time-frequency characteristics of EEG signals and draw time-frequency diagrams; e. Put the time-frequency graph results into SE-ConvNeXt network training to obtain the prediction results of acupuncture point effect classification; f. Result analysis, use confusion matrix to intuitively express the accuracy.

2. The research method according to claim 1, characterized in that: In step a, an electroencephalogram (EEG) device is used to collect EEG signal responses during acupuncture of different acupoints as training sets and test sets to train the acupuncture point classification network; the original EEG data is collected by a 32-lead device with a sampling rate of 250 Hz; the EEG is collected before acupuncture, during acupuncture, and after needle removal, and the EEG is continuously collected after needle insertion. The needle retention time is set to 25 minutes, and the needle sensation is adjusted every 5 minutes to ensure the continuity of the acupuncture stimulation. The EEG recording of the entire acupuncture process is a block; an event marker is synchronously recorded in the EEG during needle insertion and needle sensation adjustment for subsequent data collation and analysis.

3. The research method according to claim 2, characterized in that: In step b, a Butterworth bandpass filter is selected to filter out and sort out the information of the main frequency band of the EEG signal, and only the 30-second duration after stimulation is taken as valid data as the content of the subsequent data set.

4. The research method according to claim 3, characterized in that: In step c, in order to explore the response rules of acupuncture in different brain regions (channels), that is, whether there are two signals with the same or different amplitudes in any band during stimulation; any two channels with any frequency bands are subjected to cross-correlation analysis, and a correlation image is drawn; the normalized cross-correlation calculation formula in the correlation analysis is: Among them, x(k) and y(k) are EEG signals of different channels; r(n) is the correlation between EEG signals.

5. The research method according to claim 4, characterized in that: In step d, the method of extracting the time-frequency characteristics of the EEG signal by wavelet transform is as follows: Use wavelet transform to perform time-frequency analysis on the filtered signal to obtain the frequency characteristic information of each channel and each segment. The corresponding wavelet transform is: Among them, W fi (a,τ) is the different power density of each channel at time t = τ, f i,j (t) is the jth segment EEG signal of channel i, is the complex conjugate of the wavelet basis function, and the wavelet basis function is expressed as: in, is the normalization factor, a is the scale factor, and b is the translation factor. The wavelet basis function commonly used in EEG analysis is Morlet wavelet, which is expressed as After substituting into the above formula, the specific expression is: Where f0 is the center frequency and σ is the standard deviation.

6. The research method according to claim 5, characterized in that: In the step e, the feature map extracted by wavelet transform and the sorted related data are classified, and the training set is input into the SE-ConvNeXt network for training to obtain the prediction classification results of acupuncture points; the SE-ConvNeXt network is improved on the basis of ConvNeXt, and the SE model (Squeeze-and-Excitation Model) is added before the feature is input into the network to improve the representation ability of the network, and the weights of different channels are adaptively allocated and learned by the network itself; each layer of the input feature map X is convolved with a 2D spatial kernel to finally obtain C output feature maps, which constitute the feature maps U, F sq Indicates that the global pooling is reduced to a one-dimensional vector, F ex The gate mechanism consists of two fully connected layers. The first fully connected layer compresses C channels into C / r channels to reduce the amount of calculation, and then passes through a RELU nonlinear activation layer. The second fully connected layer restores the number of channels to C channels, and then uses Sigmoid activation to obtain the weight s. The dimension of the final s is 1×1×C, which is used to characterize the weights of the C feature maps in the feature map U; F scale The attention weights obtained previously are weighted to the features of each channel; the next layer consists of convolutional layer, ConvNeXt Block, Downsample, local pooling and fully connected layer; the activation function used in ConvNeXt Block is GELU; the Downsample layer consists of layer normalization and two-dimensional convolution kernel.

7. The research method according to claim 6, characterized in that: In step f, the model effect needs to be judged. First, a confusion matrix is ​​constructed for each i acupoint category classification effect, and the TP (true positive), FP (false positive), FN (false negative) and TN (true negative) values ​​of each category are counted. Then, the overall precision (Micro-Precision) can be calculated based on these values. The calculation formula is: The effectiveness of the model is judged by calculating the overall accuracy. The closer the value is to 1, the more effective the model is in predicting acupoints through EEG features, and vice versa.

Images