Neuropsychiatric disease electroencephalogram diagnosis method based on iterative polar coordinate attention

Abstract

The invention belongs to the field of electroencephalogram signal classification detection, and particularly relates to a neuropsychiatric disease electroencephalogram diagnosis method based on iteration polar coordinate attention, which comprises the following steps: acquiring a multi-channel EEG (electroencephalogram) signal and preprocessing the multi-channel EEG signal; inputting the EEG preprocessing signal into a pre-trained improved LaBraM model to obtain a plurality of node features; calculating a Pearson's correlation coefficient matrix and a cosine similarity matrix according to the EEG preprocessing signal, and then constructing a brain function fusion connection matrix as an adjacent matrix through an iteration polar coordinate attention mechanism; according to the method, the large-scale pre-trained EEG model and polar coordinate attention are used for brain graph structure construction for the first time, and collaborative optimization of time feature generalization and space structure modeling capacity is achieved.

Description

Technical Field

[0001] This invention belongs to the field of electroencephalogram (EEG) signal classification and detection, specifically relating to an EEG diagnostic method for neuropsychiatric disorders based on iterative polar coordinate attention. Background Technology

[0002] Neuropsychiatric disorders are a group of complex illnesses involving the nervous system and mental state, including attention deficit hyperactivity disorder (ADHD), Alzheimer's disease, and schizophrenia. These disorders severely affect patients' cognitive, emotional, and behavioral abilities, imposing a heavy burden on individuals, families, and society. Early and accurate diagnosis is crucial for the intervention and treatment of neuropsychiatric disorders. Electroencephalography (EEG), as a non-invasive, high-temporal-resolution, and low-cost brain signal acquisition technique, has broad application prospects in the diagnosis of neuropsychiatric disorders.

[0003] In recent years, deep learning models such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Transformers have made significant progress in EEG analysis. However, these methods treat multi-channel EEG signals as two-dimensional Euclidean data (such as images or sequences), ignoring the non-Euclidean topological characteristics of brain functional networks. Graph Convolutional Networks (GCNs), due to their ability to directly process graph-structured data, have become a key technology for addressing the non-Euclidean characteristics of EEG. However, current intelligent diagnosis of neuropsychiatric diseases using GCN analysis of EEG still suffers from three major shortcomings:

[0004] 1) Insufficient generalization ability of GCN node features: Existing GCN-based methods mostly use original EEG signals, statistical features (such as mean and standard deviation) or features extracted by task-specific models. However, these features are significantly affected by task differences and individual differences, making it difficult to apply them across subjects.

[0005] 2) GCN's brain functional connectivity modeling is incomplete: Using indicators such as Pearson correlation coefficient (PCC) and phase lock value (PLV) alone cannot fully reflect neural interactions (e.g., PCC cannot capture nonlinear correlations, and PLV is not sensitive to low-amplitude signals). Existing fusion methods (e.g., element-wise addition) do not consider indicator dependence and are difficult to effectively integrate multi-dimensional connectivity information.

[0006] 3) Insufficient generalization ability across subjects: There are significant differences in EEG signals among individuals, but existing models mostly use mixed subject validation methods, which may cause the model to learn individual pattern differences rather than disease characteristics. Summary of the Invention

[0007] To address the above problems, this invention provides a neuroelectroencephalogram (EEG) diagnostic method for neuropsychiatric disorders based on iterative polar coordinate attention, characterized by the following steps:

[0008] S1. Acquire multi-channel EEG signals and perform preprocessing to obtain preprocessed EEG signals;

[0009] S2. Input the preprocessed EEG signal into the pre-trained improved LaBraM model to obtain multiple node features;

[0010] S3. Calculate the Pearson correlation coefficient matrix and cosine similarity matrix based on the EEG preprocessed signal, and then obtain the brain function fusion connectivity matrix as the adjacency matrix through the iterative polar coordinate attention mechanism;

[0011] S4. Construct a brain graph structure using node features and adjacency matrix, and input it into a graph convolutional network and classifier to obtain the prediction results for the auxiliary diagnosis of mental illness.

[0012] The beneficial effects of this invention are:

[0013] Improving cross-subject generalization ability: By using the improved LaBraM model to extract topic-invariant node features, the inter-individual differences in EEG signals are effectively alleviated, and the model's cross-subject generalization ability is significantly improved.

[0014] Accurately characterizing brain functional connectivity: An iterative polar coordinate attention mechanism is proposed to model brain functional connectivity in polar coordinate space. The Pearson correlation coefficient and phase difference are used as the radial and angular coordinates of the polar coordinates, respectively. By fusing the complementary information of the two through a dual-path polar coordinate iterative cross-attention mechanism, the functional connectivity between brain regions can be more comprehensively and accurately characterized.

[0015] Enhanced feature representation capability: By combining pre-trained models and graph convolutional networks, the temporal dynamic features and spatial topological features of EEG signals are fully explored, thereby improving the feature representation capability.

[0016] This invention is the first to apply a large-scale pre-trained EEG model and polar coordinate attention to brain map structure construction, achieving synergistic optimization of temporal feature generalization and spatial structure modeling capabilities. It proposes an Iterative Polar Coordinate Attention (IPCA) mechanism, unifying PCC and phase difference into polar coordinate orthogonal components. Through iterative cross-attention, complementary information is dynamically enhanced, overcoming the representational bottleneck of traditional linear fusion. Attached Figure Description

[0017] Figure 1 This is an overall framework diagram of the EEG diagnostic method for neuropsychiatric diseases based on iterative polar coordinate attention according to the present invention.

[0018] Figure 2 This is a model diagram of an EEG diagnostic method for neuropsychiatric diseases based on iterative polar coordinate attention, as proposed in this invention.

[0019] Figure 3 This is the i-th update of the iterative polar coordinate attention module. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Figure 1 This is a flowchart illustrating an electroencephalogram (EEG) diagnostic method for neuropsychiatric disorders based on iterative polar coordinate attention, according to some embodiments of the present invention. Figure 2 This is a model framework diagram of an EEG diagnostic method for neuropsychiatric diseases based on iterative polar coordinate attention, as shown in some embodiments of the present invention.

[0022] Some embodiments of the present invention provide an electroencephalogram (EEG) diagnostic method for neuropsychiatric disorders based on iterative polar coordinate attention, such as... Figure 1-2 As shown, it includes the following steps:

[0023] S1. Acquire multi-channel EEG signals and perform preprocessing to obtain EEG preprocessed signals.

[0024] In some embodiments, preprocessing of multi-channel EEG signals includes bandpass filtering, denoising and baseline correction, independent component analysis, and other operations to remove noise and artifacts and improve signal quality.

[0025] As an example, a 0.5-32Hz bandpass filter was applied to the multi-channel EEG signal. An average reference and ICLabel tool were used to remove artifacts such as electrooculography (EOG) and electromyography (EMG), and abnormal data with amplitudes exceeding 100μV were discarded. Furthermore, the multi-channel EEG signal after the above operations was segmented into 2-second intervals.

[0026] S2. Input the preprocessed EEG signal into the pre-trained improved LaBraM model to obtain multiple node features.

[0027] In some embodiments, to extract subject-invariant node features, the present invention employs a pre-trained improved LaBraM model. The improved LaBraM model removes spatial embeddings from the original LaBraM model to avoid introducing predefined spatial priors, allowing it to focus on capturing the temporal dynamic features of EEG signals.

[0028] Step S2 inputs the pre-processed EEG signal into the pre-trained improved LaBraM model to obtain node features including:

[0029] S21. Divide the preprocessed EEG signal into multiple segments;

[0030] S22. Input each segment into a temporal encoder to obtain temporal encoded features; the temporal encoder includes 4 temporal convolutional blocks, each temporal convolutional block including a one-dimensional convolutional layer, a Gaussian error linear unit activation function (GELU) and a group normalization (GN);

[0031] S23. Perform temporal embedding on each temporal coding feature to incorporate temporal sequence information, thereby obtaining temporal embedded features;

[0032] S24. Input each temporal embedded feature into the Transformer encoder, learn long-distance temporal dependencies through the self-attention mechanism, and obtain the corresponding node features.

[0033] S3. Calculate the PCC matrix and cosine similarity matrix based on the EEG preprocessed signal, and then obtain the brain function fusion connectivity matrix as the adjacency matrix through the iterative polar coordinate attention mechanism.

[0034] In some embodiments, step S3 includes:

[0035] S31. Calculate the Pearson correlation coefficient (PCC) between the EEG signals of each channel in the preprocessed EEG signal, and construct the Pearson correlation coefficient matrix.

[0036] Specifically, the Pearson correlation coefficient can be used to measure the linear dependence between channels, reflecting the strength of the connection. The calculation formula is as follows:

[0037]

[0038] Where, x m x n These refer to the EEG signals of the m-th and n-th channels in the preprocessed EEG signal, respectively; cov() represents the covariance. Let A and B represent the variances of the m-th and n-th channels in the EEG preprocessed signal, respectively; and construct the PCC matrix A accordingly. PCC as follows:

[0039]

[0040] N represents the number of channels.

[0041] S32. Calculate the cosine similarity between the EEG signals of each channel in the preprocessed EEG signal, and construct the cosine similarity matrix.

[0042] Specifically, cosine similarity is used to measure the directional consistency between EEG signals in each channel, reflecting phase synchronization. The calculation formula is as follows:

[0043]

[0044] Where ||·|| represents the norm of the signal; thus, the cosine similarity matrix A is constructed. cos as follows:

[0045]

[0046] S33. Apply the inverse cosine function to the cosine similarity matrix to obtain the phase difference matrix.

[0047] Specifically, the phase difference matrix A θ as follows:

[0048]

[0049] Where, θ m,n This represents the phase difference between the m-th and n-th channels in the preprocessed EEG signal.

[0050] S34. The complex matrix is ​​obtained by fusing the Pearson correlation coefficient matrix and the phase difference matrix using an iterative polar coordinate attention mechanism.

[0051] In some embodiments, step S34 includes:

[0052] S341. Perform maximum and minimum normalization on the Pearson correlation coefficient matrix and phase difference matrix respectively to ensure numerical stability, and obtain the normalized Pearson correlation coefficient matrix and normalized phase difference matrix.

[0053] The normalization formula is:

[0054]

[0055] Represents the normalized PCC matrix The element in the m-th row and n-th column, A PCC (m,n) represents the PCC matrix A. PCC The element in the m-th row and n-th column, max[] represents the maximum value and min[] represents the minimum value; Represents the normalized phase difference matrix The element in the m-th row and n-th column, A θ (m,n) represents the phase difference matrix A. θ The element in the m-th row and n-th column.

[0056] S342. The normalized Pearson correlation coefficient matrix and the normalized phase difference matrix are respectively passed through a CNN block of a shared architecture to enhance the discriminative ability of the features, resulting in an enhanced Pearson correlation coefficient matrix and an enhanced phase difference matrix; the CNN block includes a convolutional layer, a batch normalization layer, an activation function layer, and a Dropout layer.

[0057] The enhancement process is represented as:

[0058]

[0059] in, The matrix represents the enhancement matrix, Dropout() is the Dropout layer, σ(·) is the activation function, BN() is the batch normalization function, and Conv() is the convolutional layer.

[0060] S343. Take the absolute value of all elements in the enhanced Pearson correlation coefficient matrix to obtain the absolute value matrix of the Pearson correlation coefficient. Use the absolute value matrix of the Pearson correlation coefficient as the radial coordinate of polar coordinates to represent the connection strength; use the enhanced phase difference matrix as the angular coordinate of polar coordinates to reflect the phase synchronization mode, as shown below:

[0061]

[0062] Where, r 0 Represents the initial radial coordinate, θ 0 This represents the initial angular coordinates.

[0063] A dual-path polar coordinate iterative cross-attention mechanism is used to dynamically update the radial and angular coordinates, resulting in the K1-th updated radial and angular coordinates. In some embodiments, step S343 uses a dual-path polar coordinate iterative cross-attention mechanism to dynamically update the radial and angular coordinates, wherein the process of the i-th update is as follows: Figure 3 As shown, it includes:

[0064] A1. Radial coordinate i-th update: The angular coordinate θ obtained from the (i-1)-th update. (i-1) As a query, the radial coordinate r obtained from the (i-1)th update is... (i-1) As keys and values, updates are performed via multi-head cross-attention, represented as follows:

[0065]

[0066] Where, r (i) MHCA represents the radial coordinates obtained from the i-th update. r () indicates multi-head cross-attention, and concat() indicates concatenation. This indicates that the h-th attention head (h = 1, 2, ..., H) is used in the i-th update of the radial coordinates, where H represents the number of attention heads. Let d represent the learnable projection matrix and d represent the hidden layer dimension, then d = h × d h , |·| represents taking the absolute value; attention head The formula for calculation is:

[0067]

[0068] in All are learnable matrices; T Re represents the conjugate transpose, and Re() represents extracting the real part of the complex dot product; These represent attention heads respectively. The query matrix, key matrix, and value matrix; d h Let h represent the hidden dimension of the h-th attention head;

[0069] A2. Angular coordinate update for the i-th time: The radial coordinate r obtained from the (i-1)-th update is... (i-1) As a query, the angle coordinate θ obtained from the (i-1)th update is... (i-1) As keys and values, updates are performed via multi-head cross-attention, represented as follows:

[0070]

[0071] Where, θ (i) MHCA represents the angle coordinates obtained in the i-th update. θ () indicates multi-headed cross-attention. This indicates that the h-th attention head (h = 1, 2, ..., H) is used in the i-th update of the angle coordinates. Represents the learnable projection matrix; attention head The formula for calculation is:

[0072]

[0073] in Both are learnable matrices, and arg() represents the phase angle for extracting the complex dot product; These represent attention heads respectively. The query matrix, key matrix, and value matrix.

[0074] S344. Update the radial coordinates after the K1th iteration. and angular coordinates They are merged into a complex matrix.

[0075] S35. Extract the real part of the complex matrix and perform top-k% sparsification to retain the strongest connections, reduce noise and overfitting risk, and obtain the brain function fusion connectivity matrix A, denoted as:

[0076]

[0077] Specifically, in this embodiment of the invention, the radial coordinates and angular coordinates after the K1=3th update are fused into a complex matrix, and top-70% sparsification is applied.

[0078] S4. Construct a brain graph structure using node features and adjacency matrix, and input it into a graph convolutional network and classifier to obtain the prediction results for the auxiliary diagnosis of mental illness.

[0079] In some embodiments, the constructed brain graph structure (node ​​features and multifunctional adjacency matrix) is input into a graph convolutional network (GCN) for classification. The graph convolutional network includes K2 graph convolutional layers, and the propagation rule of the i = 1, 2, ..., K2 layers is as follows:

[0080]

[0081] Among them, X i This represents the node features input to the i-th graph convolutional layer. W represents the normalized adjacency matrix. i denoted by , D represents the diagonal matrix of the adjacency matrix, σ() represents the nonlinear activation function; A represents the brain function fusion connectivity matrix, i.e., the adjacency matrix of the brain map structure.

[0082] The output of the last graph convolutional layer is passed through a fully connected layer (i.e., a classifier) ​​and then processed by a softmax function to obtain the prediction result for the auxiliary diagnosis of mental illness.

[0083] For example only, a graph convolutional network with three stacked graph convolutional layers is used to propagate node features. The first layer has an input dimension of 400 and an output dimension of 256, the second layer has an input dimension of 256 and an output dimension of 1024, and the third layer has an input dimension of 1024 and an output dimension of 2. Then, the feature vector is input into a fully connected layer to output the classification result.

[0084] In some embodiments, the present invention employs a cross-subject validation method to train and evaluate the model, including hold-out validation and leave-one-subject-out cross-validation (LOSOCV). The Adam optimizer is used to adjust the learning rate according to different datasets, with the classification cross-entropy loss function as the optimization objective, to update the model parameters and calculate metrics such as accuracy (ACC), sensitivity (SEN), specificity (SPE), F1 score, and area under the receiver operating characteristic curve (AUC) to evaluate the diagnostic performance of the model.

[0085] Experimental results on six public EEG datasets demonstrate that the proposed method achieves excellent diagnostic performance in cross-subject validation settings, with leave-one cross-validation accuracy ranging from 86.77% to 95.69% and leave-one cross-validation accuracy ranging from 88.81% to 97.49%, outperforming many existing state-of-the-art methods. This method can assist physicians in diagnosing various neuropsychiatric disorders, including Attention Deficit Hyperactivity Disorder (ADHD), Alzheimer's Disease (AD), Frontotemporal Dementia (FTD), Mild Cognitive Impairment (MCI), and Schizophrenia (SCZ), and has broad clinical application prospects.

[0086] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "setting," "connection," "fixing," "rotation," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction between two components. Unless otherwise explicitly limited, those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0087] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A neuroelectroencephalogram (EEG) diagnostic method for neuropsychiatric disorders based on iterative polar coordinate attention, characterized in that, Includes the following steps: S1. Acquire multi-channel EEG signals and perform preprocessing to obtain preprocessed EEG signals; S2. Input the preprocessed EEG signal into the pre-trained improved LaBraM model to obtain multiple node features; S3. Calculate the Pearson correlation coefficient matrix and cosine similarity matrix based on the EEG preprocessed signal, and then obtain the brain function fusion connectivity matrix as the adjacency matrix through the iterative polar coordinate attention mechanism; S4. Construct a brain graph structure using node features and adjacency matrix, and input it into a graph convolutional network and classifier to obtain the prediction results for the auxiliary diagnosis of mental illness.

2. The EEG diagnostic method for neuropsychiatric disorders based on iterative polar coordinate attention according to claim 1, characterized in that, The improved LaBraM model removes spatial embeddings from the original LaBraM model; Step S2 inputs the pre-processed EEG signal into the pre-trained improved LaBraM model to obtain node features including: S21. Divide the preprocessed EEG signal into multiple segments; S22. Input each segment into a temporal encoder to obtain temporal encoded features; the temporal encoder includes 4 temporal convolutional blocks, each temporal convolutional block including a one-dimensional convolutional layer, a Gaussian error linear unit activation function and group normalization; S23. Perform temporal embedding on each temporal encoded feature to obtain temporal embedded features; S24. Input each temporal embedding feature into the Transformer encoder to obtain the corresponding node features.

3. The EEG diagnostic method for neuropsychiatric disorders based on iterative polar coordinate attention according to claim 1, characterized in that, Step S3 includes: S31. Calculate the Pearson correlation coefficient between the EEG signals of each channel in the preprocessed EEG signal, and construct the Pearson correlation coefficient matrix; S32. Calculate the cosine similarity between the EEG signals of each channel in the preprocessed EEG signal, and construct the cosine similarity matrix; S33. Apply the inverse cosine function to the cosine similarity matrix to obtain the phase difference matrix; S34. A complex matrix is ​​obtained by fusing the Pearson correlation coefficient matrix and the phase difference matrix using an iterative polar coordinate attention mechanism. S35. Extract the real part of the complex matrix and perform top-k% sparsification to obtain the brain function fusion connectivity matrix.

4. The EEG diagnostic method for neuropsychiatric disorders based on iterative polar coordinate attention according to claim 3, characterized in that, Step S34 includes: S341. Perform maximum and minimum normalization on the Pearson correlation coefficient matrix and the phase difference matrix respectively to obtain the normalized Pearson correlation coefficient matrix and the normalized phase difference matrix; S342. The normalized Pearson correlation coefficient matrix and the normalized phase difference matrix are respectively passed through a CNN block to obtain the enhanced Pearson correlation coefficient matrix and the enhanced phase difference matrix; the CNN block includes a convolutional layer, a batch normalization layer, an activation function layer, and a Dropout layer; S343. Take the absolute value of all elements in the enhanced Pearson correlation coefficient matrix to obtain the absolute value matrix of the Pearson correlation coefficient. Use the absolute value matrix of the Pearson correlation coefficient as the radial coordinate of the polar coordinates and the enhanced phase difference matrix as the angular coordinate of the polar coordinates. Use a dual-path polar coordinate iterative cross-attention mechanism to dynamically update the radial coordinates and angular coordinates to obtain the radial coordinates and angular coordinates after the K1th update. S344. Merge the radial and angular coordinates after the K1th update into a complex matrix.

5. The EEG diagnostic method for neuropsychiatric disorders based on iterative polar coordinate attention according to claim 4, characterized in that, Step S343 uses a dual-path polar coordinate iterative cross-attention mechanism to dynamically update the radial and angular coordinates, wherein the i-th update process includes: A1. Radial coordinate i-th update: The angular coordinate θ obtained from the (i-1)-th update. (i-1) As a query, the radial coordinate r obtained from the (i-1)th update is... (i-1) As keys and values, updates are performed via multi-head cross-attention, represented as follows: Where, r (i) MHCA represents the radial coordinates obtained from the i-th update. r () indicates multi-head cross-attention, and concat() indicates concatenation. This indicates that the h-th attention head (h = 1, 2, ..., H) is used in the i-th update of the radial coordinates, where H represents the number of attention heads. The learnable projection matrix is ​​represented by |·|, which represents taking the absolute value; attention head. The formula for calculation is: in All are learnable matrices, () T Re represents the conjugate transpose, and Re() represents extracting the real part of the complex dot product; These represent attention heads respectively. The query matrix, key matrix, and value matrix; d h Let h represent the hidden dimension of the h-th attention head; A2. Angular coordinate update for the i-th time: The radial coordinate r obtained from the (i-1)-th update is... (i-1) As a query, the angle coordinate θ obtained from the (i-1)th update is... (i-1) As keys and values, updates are performed via multi-head cross-attention, represented as follows: Where, θ (i) MHCA represents the angle coordinates obtained in the i-th update. θ () indicates multi-headed cross-attention. This indicates that the h-th attention head (h = 1, 2, ..., H) is used in the i-th update of the angle coordinates. Represents the learnable projection matrix; attention head The formula for calculation is: in Both are learnable matrices, and arg() represents the phase angle for extracting the complex dot product; These represent attention heads respectively. The query matrix, key matrix, and value matrix.

6. The EEG diagnostic method for neuropsychiatric disorders based on iterative polar coordinate attention according to claim 1, characterized in that, The graph convolutional network consists of K2 graph convolutional layers. The propagation rule for the i = 1, 2, ..., K2-th layer is as follows: Among them, X i This represents the node features input to the i-th graph convolutional layer. W represents the normalized adjacency matrix. i Let represent the trainable weight parameters, D represent the diagonal matrix of the adjacency matrix, σ() represent the nonlinear activation function, and A represent the adjacency matrix. The output of the last graph convolutional layer is passed through a fully connected layer and processed by the softmax function to obtain the prediction result for the auxiliary diagnosis of mental illness.

Images