Tinnitus grading method and system based on EEG (electroencephalogram) analysis and deep learning

Abstract

The invention discloses a tinnitus grading method and system based on EEG electroencephalogram analysis and deep learning, and relates to the technical field of tinnitus grading, and the method comprises the steps: in a pre-training stage, converting EEG signals of tinnitus patients with different tinnitus severity degrees into EEG images, inputting the EEG images into an EEG self-encoder, extracting potential features in the EEG signals, and carrying out the recognition of the potential features in the EEG signals; and inputting into a five-classification linear classifier in a downstream fine tuning stage to obtain the probability of each tinnitus severity level, and predicting the tinnitus severity according to the probability. Through training, a deep neural network model containing an EEG auto-encoder used for EEG feature extraction and a linear classifier used for generating the tinnitus severity of a patient is obtained. The probability that the tinnitus severity of the to-be-graded tinnitus patient belongs to each severity level in the five levels is obtained by using the deep neural network model, so that the tinnitus severity is accurately predicted, and medical personnel are helped to formulate a personalized rehabilitation scheme for the patient.

Description

Technical Field

[0001] The present invention relates to the technical field of tinnitus classification, and more specifically, to a tinnitus classification method and system based on EEG analysis and deep learning. Background Art

[0002] Tinnitus is a common health problem that affects about 10% of the world's population and is clinically considered a common but challenging symptom. Traditional tinnitus classification methods mainly rely on patients' subjective scale tests, which are relatively general and difficult to accurately describe the heterogeneity of tinnitus. They have problems such as strong subjectivity and lack of standardization. Although new classification advances have been explored from different perspectives such as tinnitus perception, causal risk factors, tinnitus-related pain, and response to treatment, the neural mechanisms of tinnitus have not been deeply explored. Therefore, a method that can objectively grade tinnitus more accurately and sensitively is needed.

[0003] EEG signals are non-invasive and can provide real-time objective reflection of brain activity, so they can be used as objective markers for tinnitus severity classification tasks. The EEG graph constructed using EEG signals can simultaneously represent temporal and spatial information, and can extract high-level features that are difficult to capture with traditional signal processing methods, which helps to analyze the dynamic changes of brain activity and network structure. The GCN network is an extension of the convolutional neural network combined with spectral theory. The GCN network can be used to extract discriminant features of EEG graphs from the time domain and the spatial domain. Therefore, how to use the GCN framework to extract EEG signal features and achieve accurate tinnitus classification is a problem that needs to be solved. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a tinnitus grading method and system based on EEG analysis and deep learning, which can achieve accurate prediction of the severity of tinnitus in patients and help medical personnel to develop personalized rehabilitation plans for patients.

[0005] The first aspect of the present invention provides a tinnitus grading method based on EEG analysis and deep learning, comprising the following steps:

[0006] Collect and preprocess the resting EEG signals of tinnitus patients with different tinnitus severity, and convert the preprocessed EEG signals into EEG graphs;

[0007] In the pre-training stage, the EEG image is input into an EEG autoencoder based on a GCN framework for pre-training to obtain an EEG autoencoder for EEG feature extraction, and in the downstream fine-tuning stage, a five-class linear classifier is trained to obtain a linear classifier for generating the severity of the patient's tinnitus, and a deep neural network model is constructed using the trained EEG autoencoder and linear classifier;

[0008] The preprocessed EEG signal of the patient with tinnitus to be classified is obtained, converted into an EEG graph and input into the deep neural network model, and the probability that the severity of tinnitus of the patient to be classified belongs to the five levels of severity is output to predict the severity of tinnitus.

[0009] In this scheme, resting EEG signals of tinnitus patients with different tinnitus severity are collected and preprocessed, specifically:

[0010] Collect resting EEG signals of tinnitus patients with different tinnitus severity, clean the collected EEG signals, remove redundant and erroneous data, and eliminate noise and interference of EEG signals through mean filtering;

[0011] The filtered EEG signal is segmented, the segmented EEG signal is baseline corrected to eliminate the baseline offset, and then the data is re-referenced to align the EEG signal in a standard coordinate system, and the missing data is estimated and replaced by an interpolation method during the processing;

[0012] Independent component analysis is used to extract independent components from the aligned EEG signals, and damaged or noisy independent components are screened out from the independent components to obtain the processed EEG signals.

[0013] In this scheme, the preprocessed EEG signal is converted into an EEG graph, specifically:

[0014] Get the preprocessed EEG signal X∈R of tinnitus patients with different tinnitus severity C×T , where C is the number of electrodes and T is the number of sampling points of each electrode;

[0015] The preprocessed EEG signal is represented by a graph, each electrode is regarded as a node, and an edge structure is constructed according to the correlation between the electrodes. The EEG graph obtained by the graph representation is denoted as G=(X, A), where X represents the attribute of the node, and A represents the weighted adjacency matrix of the edge structure;

[0016] For the weighted adjacency matrix A, the matrix element a ij Represents node X i and node X j The delayed coherence between the EEG signals is expressed as:

[0017] a ij (τ) = E[X i (t)·X j (t+τ)]

[0018] Where X i (t) represents node X i The EEG signal at time t, Xj (t+τ) represents X j For the signal after a delay of τ time units, E[·] represents the expectation operator of the mean value of the random variable.

[0019] In this scheme, in the pre-training stage, the EEG image is input into the EEG autoencoder based on the GCN framework for pre-training to obtain the EEG autoencoder for EEG feature extraction, specifically:

[0020] The encoder and decoder of the EEG autoencoder are constructed based on the GCN framework. The encoder consists of two GCN layers. In the pre-training stage, the encoder is used to encode the EEG image to obtain the low-dimensional feature representation of the EEG image, which is expressed as:

[0021]

[0022] Among them, H (l) represents the input feature matrix of the lth layer, Add the adjacency matrix of self-loop I, express The corresponding degree matrix, represents the learnable weight matrix, σ represents the ReLU activation function;

[0023] The decoder includes an attribute decoder and a structure decoder. The attribute decoder is used to reconstruct node attributes from the encoded low-dimensional feature representation, and a series of deconvolutions are applied to expand the data back to the original feature space, which is expressed as:

[0024]

[0025] in, Represents the reconstructed node feature matrix;

[0026] The structural decoder is used to obtain the relationship between nodes in the EEG image according to the encoded low-dimensional feature representation to reconstruct the adjacency matrix and restore the connectivity relationship in the EEG image. The structural decoder predicts whether there is connectivity between each pair of nodes, which is expressed as:

[0027]

[0028] Among them, z i ,z j Represents node X i and node X j The potential feature representation of P represents the probability of connectivity between nodes. Represents the reconstructed node X i and node X j connectivity relationship;

[0029] In the decoder of the EEG autoencoder, the connection prediction layer is trained according to the output of the attribute decoder Z, which is expressed as:

[0030]

[0031] in, Represents the reconstructed adjacency matrix.

[0032] In this scheme, in the EEG autoencoder, the original node feature matrix X and the reconstructed node feature matrix The L2 norm between them balances the reconstruction validity of the attribute, where the reconstruction validity of the attribute R A It is expressed as:

[0033]

[0034] Through the original adjacency matrix A and the reconstructed adjacency matrix The L2 norm between them balances the reconstruction validity of the structure, where the reconstruction validity of the structure R S It is expressed as:

[0035]

[0036] The reconstruction is comprehensively evaluated from the aspects of attributes and structure, and the loss function of the EEG autoencoder is defined. The EEG autoencoder is iteratively trained until the loss function converges. The loss function L is expressed as:

[0037] L=αR A +(1-α)R S

[0038] Where α represents a hyperparameter, which is used to balance the importance of attribute reconstruction and structure reconstruction.

[0039] In this solution, in the downstream fine-tuning stage, a five-category linear classifier is trained to obtain a linear classifier for generating the severity of the patient's tinnitus, specifically:

[0040] Freeze the encoder part of the pre-trained EEG autoencoder, use the encoder part to generate the latent feature representation of the EEG graph, calculate the average value of the latent feature representation, and obtain the m-dimensional vector Input the m-dimensional vector into a three-layer multi-layer perceptron to train a linear classifier for five categories;

[0041] The structure of the linear classifier is: (l) =σ(W (l) x (l) +b (l) ), where y (l) represents the output of the lth layer, W (l)represents the weight matrix of the lth layer, x (l) represents the m-dimensional input vector of the lth layer, b (l) represents the bias vector of the lth layer, σ represents the activation function;

[0042] The probability of each tinnitus severity level output by the classifier is obtained through iterative training, and the classification performance of the linear classifier is verified using the verification data. When the classification performance meets the preset standard, a linear classifier for generating the patient's tinnitus severity is obtained.

[0043] In this solution, a deep neural network model is constructed based on the pre-trained encoder and linear classifier;

[0044] Obtaining a preprocessed EEG signal from a tinnitus patient to be classified, and converting the EEG signal into an EEG graph as an input of a deep neural network model;

[0045] The deep neural network model is used to calculate and output the probability that the severity of tinnitus of the patient to be classified belongs to one of five severity levels, and obtain a prediction result of the severity of tinnitus of the patient to be classified.

[0046] The second aspect of the present invention provides a tinnitus grading system based on EEG analysis and deep learning, the system comprising a data acquisition unit, a graph processing unit, a tinnitus grading unit and an output unit;

[0047] The data acquisition unit is responsible for collecting EEG signals of the tinnitus patient to be classified and preprocessing the collected EEG signals;

[0048] The graph processing unit is responsible for converting the preprocessed EEG signal into an EEG graph;

[0049] The tinnitus grading unit is responsible for constructing an EEG autoencoder based on a GCN framework, training the EEG autoencoder through pre-training, and training a five-class linear classifier in a downstream fine-tuning stage, using the trained EEG autoencoder and linear classifier to construct a deep neural network model, and using the trained deep neural network model to calculate the probability that the tinnitus severity of the tinnitus patient to be graded belongs to one of the five severity levels;

[0050] The output unit is responsible for determining the corresponding tinnitus severity using the probability that the tinnitus patient to be classified belongs to the five severity levels, and outputting the tinnitus classification result.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The present invention generates a high-quality, large-scale EEG data set from resting-state EEG data of tinnitus patients with different tinnitus severities. In the pre-training stage, the converted EEG graph is input into the EEG autoencoder to extract the potential features in the EEG data and generate a simplified and accurate representation, and is input into the linear classifier of the five categories in the downstream fine-tuning stage to obtain the probability of each tinnitus severity level, and the severity of tinnitus is predicted based on the probability. After training, a deep neural network model including an EEG autoencoder for EEG feature extraction and a linear classifier for generating the severity of the patient's tinnitus is obtained. When the user inputs the EEG signal of the tinnitus patient after preprocessing, the EEG signal is first converted into an EEG graph, and then used as the input of the above-mentioned deep neural network model, and finally the probability of the patient's tinnitus severity belonging to each of the five levels is output, thereby predicting the severity of tinnitus, which helps medical personnel to formulate personalized rehabilitation plans for patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the drawings required for use in the embodiments or exemplary descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to the drawings without paying creative work.

[0054] Figure 1 A flowchart of a tinnitus grading method based on EEG analysis and deep learning is shown;

[0055] Figure 2 A flow chart showing tinnitus severity grading using EEG signals;

[0056] Figure 3 Schematic diagram showing the reconstruction loss of the EEG autoencoder;

[0057] Figure 4 A flowchart of a tinnitus grading system based on EEG analysis and deep learning is shown. DETAILED DESCRIPTION

[0058] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0059] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0060] like Figure 1 As shown, in the first embodiment of the present invention, a tinnitus grading method based on EEG analysis and deep learning is provided, including:

[0061] S102, collecting and preprocessing resting EEG signals of tinnitus patients with different tinnitus severity levels, and converting the preprocessed EEG signals into EEG graphs;

[0062] S104, in the pre-training stage, the EEG image is input into an EEG autoencoder based on a GCN framework for pre-training, an EEG autoencoder for EEG feature extraction is obtained, in the downstream fine-tuning stage, a five-category linear classifier is trained to obtain a linear classifier for generating the severity of the patient's tinnitus, and a deep neural network model is constructed using the trained EEG autoencoder and linear classifier;

[0063] S106, obtaining the pre-processed EEG signal of the patient with tinnitus to be classified, converting it into an EEG graph and inputting it into a deep neural network model, outputting the probability that the severity of tinnitus of the patient with tinnitus to be classified belongs to one of the five severity levels, and predicting the severity of tinnitus.

[0064] It should be noted that the resting state EEG signals of tinnitus patients with different tinnitus severity are collected, and the collected EEG signals are cleaned to remove redundant and erroneous data, and the noise and interference of the EEG signals are eliminated by mean filtering, so as to extract useful signal features; the filtered EEG signals are segmented, and the sampling frequency is preferably set to 1278Hz, each segment contains two seconds of data, and each segment generates 256 power values. The segmented EEG signals are baseline corrected to eliminate baseline offset and enhance the reliability of the data. Then the data is re-referenced to align the EEG signals in the standard coordinate system for easy comparison and integration, and the interpolation method is used to estimate and replace the missing data during the processing; independent component analysis is used to extract independent components from the aligned EEG signals to further reveal the underlying structure, and damaged or noisy independent components are screened and eliminated from the independent components to obtain and process the EEG signals to ensure the purity of the data set.

[0065] It should be noted that the preprocessed EEG signals X∈R of tinnitus patients with different tinnitus severity are obtained. C×T , where R C×T represents a real number space of size C×T, C is the number of electrodes, and T is the number of sampling points of each electrode; the preprocessed EEG signal is represented by a graph, each electrode is regarded as a node, and an edge structure is constructed according to the correlation between the electrodes. The EEG graph obtained by the graph representation is denoted as G=(X,A), wherein X represents the attribute of the node, and A represents the weighted adjacency matrix of the edge structure.

[0066] Due to the large number of original sampling time points, it is impossible to directly use the original data as the time feature of each node. Therefore, data compression is performed, and the average value of the data in each segment is calculated and used as a dimension of the node feature vector. Data compression provides a time feature that describes EEG fluctuations over a longer period of time, while omitting short-term detail changes. For the weighted adjacency matrix A, the matrix element a ij Represents node X i and node X j The delayed coherence between the EEG signals is expressed as:

[0067] a ij (τ) = E[X i (t)·X j (t+τ)]

[0068] Where X i (t) represents node X i The EEG signal at time t, X j (t+τ) represents X j For the signal after a delay of τ time units, E[·] represents the expectation operator of the mean value of the random variable.

[0069] like Figure 2 As shown, the EEG autoencoder consists of an encoder and a decoder, both of which are built on the GCN framework. In the pre-training stage, the EEG image is input into the EEG autoencoder based on the GCN framework for pre-training to obtain the EEG autoencoder for EEG feature extraction, specifically:

[0070] The encoder and decoder of the EEG autoencoder are constructed based on the GCN framework. The encoder consists of two GCN layers, which can effectively model the properties of a single node and the interactions between them. The application of GCN in the encoder effectively reduces the dimension of the EEG data while retaining the basic features. In the pre-training stage, the encoder is used to encode the EEG image to obtain the low-dimensional feature representation of the EEG image, which is expressed as:

[0071]

[0072] Among them, H (l) represents the input feature matrix of the lth layer, Add the adjacency matrix of self-loop I, express The corresponding degree matrix, represents the learnable weight matrix, σ represents the ReLU activation function;

[0073] The decoder decodes the low-dimensional feature representation, trying to restore the data to a state as close as possible to the pre-encoded data. If the restored data is very similar to the pre-encoded data, it can be considered that the encoder has extracted the key information of the original data well, and the decoder has good recovery ability. The decoder includes an attribute decoder and a structure decoder. The attribute decoder is used to reconstruct the node attributes from the encoded low-dimensional feature representation, and a series of deconvolution layers are applied to expand the data back to the original feature space, which is expressed as:

[0074]

[0075] in, Represents the reconstructed node feature matrix;

[0076] The structural decoder is used to obtain the relationship between nodes in the EEG image according to the encoded low-dimensional feature representation to reconstruct the adjacency matrix and restore the connectivity relationship in the EEG image. The structural decoder predicts whether there is connectivity between each pair of nodes, which is expressed as:

[0077]

[0078] Among them, z i ,z j Represents node X i and node X j The potential feature representation of P represents the probability of connectivity between nodes. Represents the reconstructed node X i and node X j connectivity relationship;

[0079] In the decoder of the EEG autoencoder, the connection prediction layer is trained according to the output of the attribute decoder Z, which is expressed as:

[0080]

[0081] in, Represents the reconstructed adjacency matrix.

[0082] like Figure 3 As shown, in the EEG autoencoder, the original node feature matrix X and the reconstructed node feature matrix The L2 norm between them balances the reconstruction validity of the attribute, where the reconstruction validity of the attribute R A It is expressed as:

[0083]

[0084] Through the original adjacency matrix A and the reconstructed adjacency matrix The L2 norm between them balances the reconstruction validity of the structure, where the reconstruction validity of the structure RS It is expressed as:

[0085]

[0086] The reconstruction is comprehensively evaluated from the aspects of attributes and structure, and the loss function of the EEG autoencoder is defined. The EEG autoencoder is iteratively trained until the loss function converges. The loss function L is expressed as:

[0087] L=αR A +(1-α)R S

[0088] Where α represents a hyperparameter, which is used to balance the importance of attribute reconstruction and structure reconstruction when evaluating the overall reconstruction quality.

[0089] In the downstream fine-tuning stage, the encoder part of the pre-trained EEG autoencoder is frozen and used to generate the latent feature representation X∈R of the EEG map. C×M , where R C×M represents a real number space of size C×M, where C is the number of nodes and M is the dimension of the encoded feature. To classify the latent feature representation, the average value of the latent feature representation is calculated along the first dimension to obtain an m-dimensional vector X i represents the potential feature representation of the i-th node.

[0090] The m-dimensional vector is input into a three-layer multi-layer perceptron to train a linear classifier for five categories. The structure of the linear classifier is: (l) =σ(W (l) x (l) +b (l) ), where y (l) represents the output of the lth layer, W (l) represents the weight matrix of the lth layer, x (l) represents the m-dimensional input vector of the lth layer, b (l) represents the bias vector of the lth layer, σ represents the ReLU activation function; the probability of each tinnitus severity level output by the classifier is obtained through iterative training, the tinnitus severity level with the highest probability is selected as the final diagnosis result, and the classification performance of the linear classifier is verified using the verification data. When the classification performance meets the preset standard, the linear classifier used to generate the patient's tinnitus severity is obtained.

[0091] According to an embodiment of the present invention, a preprocessed EEG signal of a patient with tinnitus to be classified is obtained, and the EEG signal is converted into an EEG graph as an input of a deep neural network model; the deep neural network model is used to calculate and output the probability that the severity of tinnitus in the patient to be classified belongs to one of five severity levels as a severity score, and the severity category corresponding to the maximum severity score is obtained as a tinnitus severity prediction result for the patient to be classified.

[0092] Figure 4 A flowchart of a tinnitus grading system based on EEG analysis and deep learning is shown.

[0093] The second embodiment of the present invention provides a tinnitus grading system 4 based on EEG analysis and deep learning, the system comprising a data acquisition unit 401, a graph processing unit 402, a tinnitus grading unit 403 and an output unit 404;

[0094] The data acquisition unit is responsible for collecting EEG signals of the tinnitus patient to be classified and preprocessing the collected EEG signals;

[0095] The graph processing unit is responsible for converting the preprocessed EEG signal into an EEG graph;

[0096] The tinnitus grading unit is responsible for constructing an EEG autoencoder based on a GCN framework, training the EEG autoencoder through pre-training, and training a five-class linear classifier in a downstream fine-tuning stage, using the trained EEG autoencoder and linear classifier to construct a deep neural network model, and using the trained deep neural network model to calculate the probability that the tinnitus severity of the tinnitus patient to be graded belongs to one of the five severity levels;

[0097] The output unit is responsible for determining the corresponding tinnitus severity using the probability that the tinnitus patient to be classified belongs to the five severity levels, and outputting the tinnitus classification result.

[0098] A third embodiment of the present invention provides a computer-readable storage medium, which includes a tinnitus grading method program based on EEG electroencephalogram analysis and deep learning. When the tinnitus grading method program based on EEG electroencephalogram analysis and deep learning is executed by a processor, the steps of the tinnitus grading method based on EEG electroencephalogram analysis and deep learning are implemented.

[0099] In the several embodiments provided in the present application, it should be understood that the disclosed method can be implemented in other ways. The embodiments described above are merely illustrative. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored, or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0100] Those skilled in the art can understand that: all or part of the steps of implementing the above method embodiments can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiments; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, and other media that can store program codes.

[0101] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software functional unit and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0102] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A tinnitus grading method based on EEG analysis and deep learning, characterized in that: The following steps are involved: Collect and preprocess the resting EEG signals of tinnitus patients with different tinnitus severity, and convert the preprocessed EEG signals into EEG graphs; In the pre-training stage, the EEG image is input into an EEG autoencoder based on a GCN framework for pre-training to obtain an EEG autoencoder for EEG feature extraction, and in the downstream fine-tuning stage, a five-class linear classifier is trained to obtain a linear classifier for generating the severity of the patient's tinnitus, and a deep neural network model is constructed using the trained EEG autoencoder and linear classifier; The preprocessed EEG signal of the patient with tinnitus to be classified is obtained, converted into an EEG graph and input into the deep neural network model, and the probability that the severity of tinnitus of the patient to be classified belongs to the five levels of severity is output to predict the severity of tinnitus.

2. The tinnitus grading method based on EEG analysis and deep learning according to claim 1, characterized in that: The resting EEG signals of tinnitus patients with different tinnitus severity were collected and preprocessed, specifically: Collect resting EEG signals of tinnitus patients with different tinnitus severity, clean the collected EEG signals, remove redundant and erroneous data, and eliminate noise and interference of EEG signals through mean filtering; The filtered EEG signal is segmented, the segmented EEG signal is baseline corrected to eliminate the baseline offset, and then the data is re-referenced to align the EEG signal in a standard coordinate system, and the missing data is estimated and replaced by an interpolation method during the processing; Independent component analysis is used to extract independent components from the aligned EEG signals, and damaged or noisy independent components are screened out from the independent components to obtain the processed EEG signals.

3. The tinnitus grading method based on EEG analysis and deep learning according to claim 1, characterized in that: The preprocessed EEG signal is converted into an EEG graph, specifically: Get the preprocessed EEG signal X∈R of tinnitus patients with different tinnitus severity C×T , where C is the number of electrodes and T is the number of sampling points of each electrode; The preprocessed EEG signal is represented by a graph, each electrode is regarded as a node, and an edge structure is constructed according to the correlation between the electrodes. The EEG graph obtained by the graph representation is denoted as G=(X, A), where X represents the attribute of the node, and A represents the weighted adjacency matrix of the edge structure; For the weighted adjacency matrix A, the matrix element a ij Represents node X i and node X j The delayed coherence between the EEG signals is expressed as: a ij (τ)=E[X i (t)·X j (t+τ)] Where X i (t) represents node X i The EEG signal at time t, X j (t+τ) represents X j For the signal after a delay of τ time units, E[·] represents the expectation operator of the mean value of the random variable.

4. The tinnitus grading method based on EEG analysis and deep learning according to claim 1, characterized in that: In the pre-training stage, the EEG image is input into the EEG autoencoder based on the GCN framework for pre-training to obtain the EEG autoencoder for EEG feature extraction, specifically: The encoder and decoder of the EEG autoencoder are constructed based on the GCN framework. The encoder consists of two GCN layers. In the pre-training stage, the encoder is used to encode the EEG image to obtain the low-dimensional feature representation of the EEG image, which is expressed as: Among them, H (l) represents the input feature matrix of the lth layer, Add the adjacency matrix of self-loop I, express The corresponding degree matrix, represents the learnable weight matrix, σ represents the ReLU activation function; The decoder includes an attribute decoder and a structure decoder. The attribute decoder is used to reconstruct node attributes from the encoded low-dimensional feature representation, and a series of deconvolutions are applied to expand the data back to the original feature space, which is expressed as: in, Represents the reconstructed node feature matrix; The structural decoder is used to obtain the relationship between nodes in the EEG image according to the encoded low-dimensional feature representation to reconstruct the adjacency matrix and restore the connectivity relationship in the EEG image. The structural decoder predicts whether there is connectivity between each pair of nodes, which is expressed as: Among them, z i ,z j Represents node X i and node X j The potential feature representation of P represents the probability of connectivity between nodes. Represents the reconstructed node X i and node X j connectivity relationship; In the decoder of the EEG autoencoder, the connection prediction layer is trained according to the output of the attribute decoder Z, which is expressed as: in, Represents the reconstructed adjacency matrix.

5. The tinnitus grading method based on EEG analysis and deep learning according to claim 4, characterized in that: In the EEG autoencoder, the original node feature matrix X and the reconstructed node feature matrix The L2 norm between them balances the reconstruction validity of the attribute, where the reconstruction validity of the attribute R A It is expressed as: Through the original adjacency matrix A and the reconstructed adjacency matrix The L2 norm between them balances the reconstruction validity of the structure, where the reconstruction validity of the structure R S It is expressed as: The reconstruction is comprehensively evaluated from the aspects of attributes and structure, and the loss function of the EEG autoencoder is defined. The EEG autoencoder is iteratively trained until the loss function converges. The loss function L is expressed as: L=αR A +(1-α)R S Where α represents a hyperparameter, which is used to balance the importance of attribute reconstruction and structure reconstruction.

6. The tinnitus grading method based on EEG analysis and deep learning according to claim 1, characterized in that: In the downstream fine-tuning stage, a five-category linear classifier is trained to obtain a linear classifier for generating the severity of the patient's tinnitus, specifically: Freeze the encoder part of the pre-trained EEG autoencoder, use the encoder part to generate the latent feature representation of the EEG graph, calculate the average value of the latent feature representation, and obtain the m-dimensional vector Input the m-dimensional vector into a three-layer multi-layer perceptron to train a linear classifier for five categories; The structure of the linear classifier is: (l) =σ(W (l) x (l) +b (l) ), where y (l) represents the output of the lth layer, W (l) represents the weight matrix of the lth layer, x (l) represents the m-dimensional input vector of the lth layer, b (l) represents the bias vector of the lth layer, σ represents the activation function; The probability of each tinnitus severity level output by the classifier is obtained through iterative training, and the classification performance of the linear classifier is verified using the verification data. When the classification performance meets the preset standard, a linear classifier for generating the patient's tinnitus severity is obtained.

7. The tinnitus grading method based on EEG analysis and deep learning according to claim 1, characterized in that: Build a deep neural network model based on the pre-trained encoder and linear classifier; Obtaining a preprocessed EEG signal from a tinnitus patient to be classified, and converting the EEG signal into an EEG graph as an input of a deep neural network model; The deep neural network model is used to calculate and output the probability that the severity of tinnitus of the patient to be classified belongs to one of five severity levels, and obtain a prediction result of the severity of tinnitus of the patient to be classified.

8. A tinnitus grading system based on EEG analysis and deep learning, characterized in that: Implementing the tinnitus grading method based on EEG analysis and deep learning as described in any one of claims 1 to 7, the system comprises a data acquisition unit, a graph processing unit, a tinnitus grading unit and an output unit; The data acquisition unit is responsible for collecting EEG signals of the tinnitus patient to be classified and preprocessing the collected EEG signals; The graph processing unit is responsible for converting the preprocessed EEG signal into an EEG graph; The tinnitus grading unit is responsible for constructing an EEG autoencoder based on a GCN framework, training the EEG autoencoder through pre-training, and training a five-class linear classifier in a downstream fine-tuning stage, using the trained EEG autoencoder and linear classifier to construct a deep neural network model, and using the trained deep neural network model to calculate the probability that the tinnitus severity of the tinnitus patient to be graded belongs to one of the five severity levels; The output unit is responsible for determining the corresponding tinnitus severity using the probability that the tinnitus patient to be classified belongs to the five severity levels, and outputting the tinnitus classification result.

Images