Electroencephalogram anomaly detection method of multi-feature graph convolution adaptive brain network

Abstract

The invention belongs to the field of electroencephalogram signal anomaly detection, and relates to an electroencephalogram anomaly detection method for a multi-feature map convolutional adaptive brain network, which comprises the following steps: acquiring an original electroencephalogram of a user, and performing short-time Fourier transform on the original electroencephalogram to obtain a time-frequency map; using bandwidth integral to apply power spectral density to represent spectrum characteristics of the time-frequency graph; extracting a difference entropy feature and a Shannon entropy feature of the time-frequency graph; respectively inputting the frequency spectrum features, the difference entropy features and the Shannon entropy features into a trained adaptive graph learning GNN model to obtain a detection result; wherein the adaptive graph learning GNN model is composed of an adaptive brain network learning module and a hybrid-jump GNN module; according to the algorithm, function connection corresponding to electroencephalogram signal fragments does not need to be calculated in advance before model learning, corresponding hyper-parameters are dynamically adjusted in the model fitting process, and therefore the brain network structure of each sample is completely learned in a data-driven mode.

Description

Technical Field

[0001] The present invention belongs to the field of electroencephalogram (EEG) signal anomaly detection, and particularly relates to a method for detecting EEG anomalies of a multi-feature map convolutional adaptive brain network. Background Art

[0002] Many studies have attempted to find EEG biomarkers for depression. A large amount of work is based on the emerging deep learning. Although deep neural network models can automatically extract features of EEG signals and build an end-to-end framework for identifying patients with depression. Wu et al. developed an end-to-end machine learning algorithm for the latent space of Sparse EEG Latent Space Regression (SELSER), which uses the EEG signals of patients with resting-state depression to predict their treatment outcomes. The SELSER algorithm optimizes a latent space model that maps resting-state EEG data to the treatment outcomes of patients by minimizing the prediction error, and this mapping is constrained by the dimension of the latent signal. Taking the power of four typical EEG frequency bands as features, a set of spatial filters is optimized to convert multi-channel linear EEG signals in sensor space into low-dimensional latent signals. Then a linear regression model is established to correlate the band power features of the latent signals with the treatment outcomes. The physiological significance of the parameters of the entire end-to-end model is known, and it has a high degree of interpretability. This model can reliably predict the treatment outcomes of patients with depression and distinguish the responses to sertraline and placebo at the individual patient level, providing guidance for machine learning-driven personalized treatment methods for depression.

[0003] With the rise of Graph neural Networks (GNN) in the fields of deep learning and artificial intelligence, due to the powerful analysis ability of GNN for graph-structured data, many studies have introduced GNN into the identification of depression, especially in the research on brain network connection features. Some studies have proposed a multi-granularity graph convolutional neural network that can simultaneously extract the granularity information of functional connection networks with different thresholds, fully retaining valuable weak connections while removing noise. Some studies have proposed a soft-label self-attention graph pooling model to identify depression in order to solve the problems of insufficient exploration of the topological structure of electroencephalograms, information loss caused by high-dimensional data compression, and neglect of intra-class differences and inter-class similarities in EEG data. Recently, in order to solve the problem of data scarcity in model-driven methods, some scholars have proposed a semi-supervised domain adaptation model based on graph neural networks to identify depression, effectively alleviating the limitations of noise pseudo-labels and class imbalance problems while expanding the data to multiple domain adaptation examples. Although the graph neural networks in the above studies have strong classification capabilities, these studies often need to construct a brain network in advance as input before model learning, and the relationship between brain network lesions and depression has not been fully explained.

[0004] In order to be able to autonomously learn or gradually adjust the construction of the brain network during model training, some research has proposed a multi-view sparse dynamic graph convolutional regional attention fusion network for identifying depression through EEG signals. First, a multi-view feature extractor was designed to simultaneously represent electroencephalogram signals from the time, frequency, and time-frequency views. Subsequently, a sparse dynamic graph convolutional network was introduced to autonomously construct a sparse brain network, avoiding the limitations of over-smoothing and redundant edges existing in traditional graph neural networks. Finally, a regional attention feature fusion network was proposed to effectively fuse multi-view features. This framework effectively improved the recognition accuracy of patients with depression. There is also research that proposed a multi-scale adaptive spatio-temporal graph convolutional network for mining the potential topological structure between EEG channels and capturing discriminative spatio-temporal features. This research integrated adaptive graph convolution, combining the inherent graph construction method with the data-driven graph reconstruction method. This model uses an attention mechanism to learn the adaptive topological structure and semantic information from different layers and classes. Secondly, a multi-scale time convolutional layer was proposed to capture the long-term dependencies of EEG data. Although the above research realizes the autonomous learning of brain network topology construction in the model training of EEG depression recognition, there are still the following problems. First, these methods cannot construct an exclusive network structure for each sample in the model, but use a holistic network topology to represent the entire depression recognition process. This not only does not conform to the principle of the specificity of the subject's brain network, but also makes it difficult to explain the relationship between functional connection abnormalities and depression by batch analyzing the network. Second, most of these studies construct network features based on end-to-end EEG time signal points, and do not specifically construct a specific brain network for the time-domain, frequency-domain, and spatial-domain features of EEG signals, which also increases the difficulty of further brain network analysis of the pathological mechanism of depression. Summary of the Invention

[0005] To solve the above problems existing in the prior art, the present invention proposes an electroencephalogram abnormality detection method based on a multi-feature graph convolutional adaptive brain network, including: collecting the original electroencephalogram of a user, performing short-time Fourier transform on the original electroencephalogram to obtain a time-frequency graph; using bandwidth integration to apply power spectral density to represent the spectral features of the time-frequency graph; extracting the differential entropy feature and Shannon entropy feature of the time-frequency graph; respectively inputting the spectral features, differential entropy features, and Shannon entropy features into a trained adaptive graph learning GNN model to obtain a detection result; wherein the adaptive graph learning GNN model consists of an adaptive brain network learning module and a hybrid-hop GNN module, where the adaptive brain network learning module is used to capture the hidden topological relationship between different EEG channel features, and the hybrid-hop GNN module is used to fuse feature information.

[0006] Advantages of the present invention:

[0007] Through the hybrid jump graph convolutional network with adaptive graph structure learning, the present invention constructs a specific functional connection brain network for multi-class EEG features (power spectral density, differential entropy, Shannon entropy) of each sample in a data-driven manner on the premise of accurately identifying depression patients. Compared with previous depression studies using graph neural networks based on functional connections, the algorithm of the present invention does not need to pre-calculate the functional connections corresponding to EEG signal segments before model learning, but dynamically adjusts the corresponding hyperparameters during the model fitting process, so as to completely learn the brain network structure of each sample in a data-driven manner. The algorithm of the present invention combines the construction of the brain network with the downstream EEG classification task and verifies its effectiveness in depression EEG data. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 is the overall flow chart of the present invention;

[0009] Figure 2 is the block diagram of the multi-feature adaptive graph learning GNN model of the present invention;

[0010] Figure 3 is the structural diagram of the multi-feature adaptive graph learning GNN model of the present invention;

[0011] Figure 4 is the structural diagram of the hybrid-jump GNN module of the present invention;

[0012] Figure 5 is the visualization diagram of the significant brain network connections of MODMA depression patients of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0013] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0014] An EEG abnormality detection method for a multi-feature graph convolutional adaptive brain network, as Figure 1As shown, the method includes: collecting the user's original electroencephalogram (EEG), performing short-time Fourier transform on the original EEG to obtain a time-frequency map; representing the spectral characteristics of the time-frequency map using bandwidth integral applied power spectral density; extracting the differential entropy feature and Shannon entropy feature of the time-frequency map; respectively inputting the spectral feature, differential entropy feature, and Shannon entropy feature into the trained adaptive graph learning GNN model to obtain a detection result; where the adaptive graph learning GNN model consists of an adaptive brain network learning module and a hybrid-hop GNN module, and the adaptive brain network learning module is used to capture the hidden topological relationship between different EEG channel features, and the hybrid-hop GNN module is used to fuse feature information.

[0015] The original EEG record is defined as E = {(Xi i , yi i )|i = 1, 2,..., k}, where Xi i ∈R C×S represents that the i-th subject has C channels and S samples, K is the total number of depressive patients and normal control subjects, and yi i is the corresponding class label. Each sample is transformed into a time-frequency map of the original EEG signal through short-time Fourier transform (STFT), directly generating unique features in different frequency bands, namely δ[0–4Hz], θ[4–8Hz], α[8–12Hz], β[12–30Hz], and γ[30–50Hz]. From the spectral perspective, the power spectral density (PSD) is used to represent the spectral characteristics of the EEG signal through bandwidth integral, and its calculation method can be achieved through the following formula:

[0016]

[0017] where X is the time-frequency map of the EEG signal after fast Fourier transform, c is the current channel, t is the time window, v is the frequency value, and f is the frequency band range to be calculated. Generally speaking, the PSD features extracted from each EEG sample can be expressed as where D PSD represents the dimension of the power spectral density feature, which is equal to the number of corresponding frequency bands.

[0018] The PSD feature can represent the frequency domain characteristics of the EEG signal. To fully analyze the video domain characteristics of the EEG signal, differential entropy (DE) and Shannon entropy (SE) are introduced here to capture deeper manual features of the EEG signal, which have been proven effective in previous EEG detection tasks. The differential entropy feature can be calculated by the following formula:

[0019]

[0020] where X is the Gaussian distribution N(μ, σ 2The cycle of the electroencephalogram (EEG) signal, where x is a variable, e is a constant, σ is the variance, and μ is the mean.

[0021] The Shannon entropy feature can be calculated by the following formula:

[0022]

[0023] Where X is the EEG signal cycle of the Gaussian distribution N(μ,σ 2 ), x represents the variable, and π and e are constants respectively. p(x i ) is the distribution probability of x i among the amplitudes of n total slices. The above calculation operations are performed on all five frequency bands of the EEG signal in sequence to obtain the feature representations of differential entropy DE and Shannon entropy SE: Where D F is the number of frequency bands.

[0024] In this embodiment, as Figure 2 and Figure 3 shown, the adaptive brain network learning module processes the input feature map including: performing feature extraction on the input feature map based on the parameters in the adaptive brain network learning module and the tanh function to obtain the hidden relationships of different EEG channel features; calculating the brain network adjacency matrix according to the hidden relationships of different EEG channel features; and performing coefficient thresholding processing on the brain network adjacency matrix.

[0025] Specifically, after extracting the PSD, DE, and SE features corresponding to the EEG signal, they are respectively input into the adaptive graph learning GNN model, which is divided into an adaptive brain network learning module and a hybrid-hop GNN module. The adaptive brain network learning module realizes data-driven learning of the adjacency matrix of the graph to capture the hidden topological relationships among different EEG channel features. Generally speaking, the brain network adjacency matrix A is adaptively learned through the following formula:

[0026] M1 = tanh(αEθ1)

[0027] M2 = tanh(αEθ2)

[0028]

[0029] Where Relu is the activation function, tanh is the hyperbolic tangent function, α is the hyperparameter controlling the saturation of the activation function, M1 and M2 are the hidden relationships of different EEG channel features, T is the transpose, θ1 and θ2 are two parameter matrices to be learned, E is the feature embedding of graph learning, and the above-mentioned x PSD , x DE , x SE ∈R C×D features. θ1, θ2 ∈ R D×Care two parameter matrices to be learned. α is a hyperparameter that controls the saturation of the activation function. A large number of studies have shown that the functional connections of the human brain follow a sparse connection pattern. In order to make the learned brain network matrix conform to the sparse connection, the following formula is used to perform coefficient thresholding on the adjacency matrix:

[0030] for i = 1, 2, …, N

[0031] idx = argtopk(A[i, :])

[0032] A[i, -idx] = 0

[0033] where the function argtopk(·) returns the indices of the top k maximum values of the vector. For each node, its top k closest nodes are selected as its neighbors. While retaining the weights of the connected nodes, the weights of the non-connected nodes are set to zero.

[0034] In this embodiment, as Figure 4 shown, the processing of the input features by the hybrid-hop GNN module includes: taking the electroencephalogram differential entropy, Shannon entropy, and power spectral density as feature embeddings respectively and parallelly inputting them into a residual graph convolutional neural network with multiple layers of hybrid-hop connections for feature transformation and mapping to obtain three deep features; splicing the three deep features, and inputting the spliced feature map into the classification layer to obtain the recognition result.

[0035] Specifically, the hybrid-hop GNN module is composed of multiple layers of graph convolutional neural networks through progressive skip-residual connections layer by layer, and finally the multiple layers of graph convolutional layers are composed into an output through a fully connected layer. The graph convolutional layer aims to fuse the information of a node with the information of its neighbors to process the spatial dependence in the graph. The graph convolutional module is composed of two hybrid-hop propagation layers, which respectively process the inflowing and outflowing information passing through each node. The net inflow information is obtained by adding the outputs of the two hybrid-hop propagation layers. Given the graph adjacency matrix, a hybrid-hop propagation layer is proposed here to process the information flow on the spatially related electrodes of the brain network. The proposed hybrid-hop propagation layer consists of two steps: information propagation and information selection. The information propagation step is defined as follows:

[0036]

[0037] where β is a hyperparameter that controls the proportion of retaining the original state of the root node; H in is the feature embedding, is the normalized brain network adjacency matrix, and H (k-1) is the feature output by the k - 1 layer residual graph convolutional neural network.

[0038] The information selection step is defined as follows:

[0039]

[0040] where K is the propagation depth, H in represents the input hidden state of the output of the previous layer, H out represents the output hidden state of the current layer.

[0041] The feature outputs of the three parallel graph neural networks are finally concatenated into the output of the entire model

[0042] H (k) = H in

[0043]

[0044] In the information propagation step and the information selection step of the hybrid skip propagation layer, information is first propagated horizontally and then selected vertically. The information propagation step recursively propagates node information and the given graph structure.

[0045] A serious limitation of the graph convolutional network is that when the number of graph convolutional layers tends to infinity, the hidden state of the nodes converges to a single point. This is because a graph convolutional network with multiple layers will reach the limiting distribution of random walks regardless of the initial node state. To solve this problem, this study retains the original state of some nodes during the propagation process, so that the propagated node state can retain locality and explore deep neighborhoods. However, if only the information propagation formula is applied, some node information will be lost. In the extreme case without spatial dependence information, aggregating neighborhood information will only add useless noise to each node. Therefore, an information selection step is introduced to filter out the important information generated by each layer of jumps. According to the information selection step, the parameter matrix W (k) acts as a feature selector. When the given graph structure does not contain spatial dependence information, the above formula can still retain the original node self-information by setting all W (k) with k > 0 to 0.

[0046] The hybrid-jump connection pattern has been frequently used and proven effective in the fields of image and natural language processing. Some studies concatenate information from different jumps. Other studies have proposed an attention mechanism to weight the information between different jumps. They all apply graph convolutional neural networks for information propagation. However, due to the over-smoothing problem faced by graph convolutional layers, information from higher jumps may not contribute positively to the overall performance. To avoid this, the method in this study maintains a balance between local and neighborhood information. In addition, related research in the field of traffic road prediction shows that a model with two hybrid-jump layers can represent the difference between two consecutive jump layers. The method in this study can achieve the same effect with only one hybrid-jump propagation layer.

[0047] Assume that k = 2, W (0) = 0, W (1) = -1, and W (2) = 1, then we can get:

[0048] H out = Δ(H (2) , H (1) ) = H 2 - H 1

[0049] From this perspective, compared with the concatenation method, using summation can more effectively represent the linear interaction of all valid information in different jump layers.

[0050] In the research of electroencephalogram (EEG) depression classification, it is very common that the number of depressive patients is inconsistent with that of normal controls. To solve the problem of unbalanced model training samples caused by the above situation, this study uses the focal loss function to solve this problem. The focal loss is a loss function used to solve the problem of class imbalance. By reducing the attention to easy-to-classify samples and enhancing the learning of difficult-to-classify samples, the performance of the model can be improved. In binary classification problems, the ordinary cross-entropy loss function is defined as:

[0051]

[0052] When the number of samples in a certain class is severely imbalanced, the traditional cross-entropy loss will be dominated by a large number of samples in this class in this case, resulting in insufficient learning of other samples by the model. To solve the problem of sample imbalance of the traditional cross-entropy loss, some studies have proposed the weighted cross-entropy loss on this basis:

[0053]

[0054] Although the weighted cross-entropy loss solves the problem of sample imbalance to a certain extent, it requires fixing the weight parameters in advance to indicate how the model should adjust the learning weights of each class, but still cannot dynamically adjust the impact of easy-to-classify samples. The focal loss introduces a modulation factor on the basis of the weighted cross-entropy to further reduce the loss contribution of easy-to-classify samples. The specific formula is:

[0055]

[0056] where p is the predicted probability of the model for the positive class. y is the true label, taking values of 1 (positive class) or 0 (negative class). α is the weight hyperparameter for positive and negative samples, used to balance the class imbalance problem. γ is the modulation factor, used to control the degree of loss attenuation of easy-to-classify samples, usually taking a value of 2. Through the modulation factor (1 - p) γ and p γ , the focal loss dynamically reduces the loss contribution of easy-to-classify samples, making the model more focused on difficult-to-classify samples. The main advantage of the focal loss compared to the weighted cross-entropy loss is that it dynamically adjusts the loss contribution of samples by introducing a modulation factor, thus effectively solving the class imbalance problem and the impact of easy-to-classify samples. The weighted cross-entropy only balances the losses of positive and negative samples by fixing weights, while the focal loss further exponentially attenuates the losses of easy-to-classify samples through the modulation factor, making the model more focused on the learning of difficult-to-classify samples. This mechanism not only strengthens the sample detection ability for small targets or in complex backgrounds, but also improves the robustness and overall performance of the model.

[0057] In this embodiment, the EEG signal samples are first transformed into time-frequency maps in five frequency bands [0–4Hz], [4–8Hz], [8–12Hz], [12–30Hz], and [30–50Hz] through fast Fourier transform. Based on this time-frequency map, the corresponding EEG signal features are calculated: power spectral density PSD, differential entropy DE, and Shannon entropy SE. Subsequently, the features of each class are used as graph embedding features and input into the adaptive graph learning hybrid jump graph neural network module to construct a corresponding graph structure for each class of features as the corresponding brain network. The subsequently constructed graph structure is used to guide the inflow and outflow information of each node (electrode) in the feature learning in the hybrid jump GNN module. Finally, all the features of the last layer are concatenated and input into the focal loss to realize the EEG depression recognition of multi-feature brain network self-learning.

[0058] The multi-feature adaptive graph learning GNN model is composed of a hybrid-jump GNN module and an adaptive brain network learning module. The adaptive brain network learning module constructs a functional connection network for each class of features, and the hybrid-jump GNN module propagates and aggregates the feature information under the brain network graph structure through graph convolution.

[0059] The multi-feature adaptive graph learning GNN model can adaptively construct the graph structure of each sample under this feature as its corresponding brain network according to the graph embedding features of each layer by controlling the learning of hyperparameters θ1 and θ2.

[0060] The hybrid-jump GNN module inputs the feature map H each time. in It is learned through a residual GCN network with multiple-layer skip connections, and the outputs of each layer obtained are H0, H1, H2…H. k Finally, it is fused through a fully connected network to obtain H. out 。

[0061] In this embodiment, in order to study the model performance of our algorithm, experiments were carried out under the MODMA public dataset. This dataset comes from the electroencephalogram signal data of clinical depression patients and matched normal controls. All patients were carefully diagnosed and screened by professional psychiatrists in the hospital. All participants had normal or corrected-to-normal vision. The depression patients were recruited from inpatients and outpatients of the Second Hospital of Lanzhou University in Gansu, China, and were diagnosed and recommended by at least one clinical psychiatrist. The normal control group was recruited through posters. The study was approved by the Ethics Committee of the Second Affiliated Hospital of Lanzhou University, and all subjects gave written informed consent before the experiment started. For MDD patients, the inclusion criteria included meeting the depression criteria in the MINI diagnosis, the patient's PHQ-9 score being greater than or equal to 5, and no psychotropic drug treatment in the past two weeks. The exclusion criteria included having other mental disorders or brain organic lesions, severe physical diseases, and severe suicidal tendencies. For NC, the exclusion criteria included a personal or family history of mental disorders. The EEG dataset used data collected by a traditional 128-electrode elastic cap. The 128-electrode EEG signals of 53 subjects were recorded in the resting state, and 5 minutes of closed-eye resting-state electroencephalogram was recorded. The subjects were required to stay awake and still, without any body movements, including head or legs, and any unnecessary eye movements, saccades, and blinks. It consisted of 24 depression patients (female / male = 11 / 13, age 30.88 ± 10.37 years old) and 29 healthy controls (female / male = 9 / 20, age 31.45 ± 9.15 years old). The 128-channel EEG signals were recorded at a sampling rate of 250Hz using a HydroCel Geodesic collector.

[0062] First, perform a certain degree of preprocessing on the original EEG data: ① Use a notch filter of 60HZ to remove the working frequency. ② Use a band-pass filter from 0.5Hz to 100Hz to filter out irrelevant frequency bands. ③ Perform average re-reference. After preprocessing, all EEGs are sliced using a non-overlapping window of 2s, and each slice is used as a sample for model training. During the model training process, the data is divided into a training set and a test set in a ratio of 9:1. The training set is further divided into 10 equal parts for ten-fold cross-validation. Each time, 9 folds are used as the training set and 1 fold is used as the validation set. Finally, the test results of the optimal model on the 10 validation sets are saved as the final performance evaluation index. The overall experimental results are shown in the following table:

[0063] 10-fold cross-validation results of the MODMA dataset

[0064]

[0065] The recognition ability of the algorithm of the present invention for depression on this dataset is generally better than the above several comparison algorithms. Next, the distribution of the functional connections constructed by the algorithm of the present invention will be shown. Here, the first 6 most significant functional connections among the depression patients in the MODMA dataset are visualized, and the results are as Figure 5 shown.

[0066] The above embodiments further elaborate on the purpose, technical solutions, and advantages of the present invention. It should be understood that the above embodiments are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made to the present invention within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for detecting electroencephalogram abnormalities in a multi-feature map convolutional adaptive brain network, characterized in that Including: Collect the original electroencephalogram (EEG) of the user, perform short-time Fourier transform on the original EEG to obtain a time-frequency diagram; use bandwidth integration to apply the power spectral density to represent the spectral characteristics of the time-frequency diagram; Extract the differential entropy feature and Shannon entropy feature of the time-frequency diagram; input the spectral feature, differential entropy feature and Shannon entropy feature into the trained adaptive graph learning GNN model respectively to obtain the detection result; Where The adaptive graph learning GNN model consists of an adaptive brain network learning module and a hybrid-hop GNN module. The adaptive brain network learning module is used to capture the hidden topological relationship between different EEG channel features, and the hybrid-hop GNN module is used to fuse feature information.

2. The EEG abnormality detection method for a multi-feature map convolutional adaptive brain network according to claim 1, wherein Using bandwidth integration to apply the power spectral density to represent the spectral characteristics of the time-frequency diagram includes: Where X is the time-frequency diagram after the EEG signal undergoes fast Fourier transform, c is the current channel, t is the time window, v is the frequency value, and f is the frequency band range to be calculated.

3. The EEG abnormality detection method of a multi-feature graph convolutional adaptive brain network according to claim 1, characterized in that, The differential entropy feature is: where X is the EEG signal period of the Gaussian distribution N(μ,σ 2 ), x is a variable, e is a constant, σ is the standard deviation, and μ is the mean value.

4. The electroencephalogram abnormality detection method of a multi-feature map convolutional adaptive brain network according to claim 1, characterized in that The Shannon entropy feature is: Among them, X is the EEG signal period of the Gaussian distribution N(μ,σ 2 ), n is the number of slices, and p(x i ) is the distribution probability of x i in the total amplitudes of n slices.

5. The EEG abnormality detection method of a multi-feature map convolution adaptive brain network according to claim 1, characterized in that The adaptive brain network learning module processes the input feature map including: performing feature extraction on the input feature map based on the parameters in the adaptive brain network learning module and the tanh function to obtain the hidden relationship of different EEG channel features; calculating the brain network adjacency matrix according to the hidden relationship of different EEG channel features; performing coefficient thresholding processing on the brain network adjacency matrix.

6. The EEG abnormality detection method of a multi-feature map convolutional adaptive brain network according to claim 5, wherein, The formula for calculating the brain network adjacency matrix is: M1 = tanh(αEθ1) M2 = tanh(αEθ2) Where Relu is the activation function, tanh is the hyperbolic tangent function, α is the hyperparameter that controls the saturation of the activation function, M1 and M2 are the hidden relationships of different EEG channel features, T is the transpose, θ1 and θ2 are two parameter matrices to be learned, and E is the feature embedding of graph learning.

7. The EEG abnormality detection method of a multi-feature graph convolutional adaptive brain network according to claim 1, characterized in that, The processing of the input features by the hybrid-hop GNN module includes: using the electroencephalogram differential entropy, Shannon entropy, and power spectral density as feature embeddings respectively Input them in parallel into a residual graph convolutional neural network with multi-layer hybrid-hop connections for feature transformation and mapping to obtain three deep features; concatenate the three deep features, and input the concatenated feature map into the classification layer to obtain the recognition result.

8. The EEG abnormality detection method of a multi-feature map convolutional adaptive brain network according to claim 7, characterized in that, The k-th layer residual graph convolutional neural network processes data including: Among them, β is a hyperparameter, and H in is the feature embedding, is the normalized brain network adjacency matrix, and H (k-1) is the feature output by the k-1 layer residual graph convolutional neural network.

9. The EEG abnormality detection method of a multi-feature map convolution adaptive brain network according to claim 1, characterized in that The loss function for training the adaptive graph learning GNN model is: Where p = sigmod(Y) is the prediction probability of the model for the positive class, y is the true label, α is the positive and negative sample weight hyperparameter, and γ is the modulation factor.

Images