A dementia recognition method based on brain-computer network space cross-attention fusion

Abstract

This invention discloses a dementia identification method based on spatial cross-attention fusion using brain-computer networks, comprising: acquiring resting-state electroencephalogram (EEG) signals, performing preprocessing and brain network construction, calculating frequency band power ratios, and dividing the dataset into training and testing sets; fusing spatial information through the brain network, and obtaining spatial features via an isovariant graph neural network and a local feature enhancement module. The frequency band power ratio is encoded by a multilayer perceptron to obtain spectral statistical characteristics. ;Will and Input a bidirectional cross-attention module to extract complementary fusion features; The complementary fusion features are mixed and pooled, then concatenated and fed into a Chebyshev Kolmogorov–Arnold network classifier to output classification results for Alzheimer's disease / frontotemporal dementia / healthy controls (AD / FTD / HC) and cognitive scale evaluation scores (MMSE). This invention effectively integrates complementary information from multiple features by jointly modeling spatial graph features and global spectral features, thereby improving the accuracy, stability, and generalization ability of AD and FTD classification in complex brain networks.

Description

A brain-computer network spatial cross-attention fusion method for dementia identification Technical Field

[0001] This invention relates to neurodegenerative disease identification technology, specifically to a dementia identification method based on spatial cross-attention fusion of brain computer networks. This technical field primarily targets the early screening, accurate diagnosis, and disease progression monitoring of neurodegenerative diseases such as Alzheimer's disease (AD) and frontotemporal dementia (FTD). It utilizes neurophysiological signals such as electroencephalography (EEG) to construct quantifiable brain network biomarkers. Through functional connectivity analysis, spectral feature extraction, and graph neural network modeling, it reveals abnormalities in brain region functional interactions and topological structures, providing technical support for the intelligent, non-invasive, and highly sensitive identification of neurodegenerative diseases. Background Technology

[0002] Alzheimer's disease (AD) and frontotemporal dementia (FTD) are two common neurodegenerative diseases. Early and accurate identification of these two diseases is crucial for developing treatment strategies, intervening in the disease course, and improving prognosis. Among numerous diagnostic methods, resting electroencephalography (EEG) is widely used for the detection and assessment of neurodegenerative diseases due to its advantages such as non-invasiveness, low cost, and high temporal resolution. An individual's resting brain activity can reflect functional connectivity patterns between different brain regions, the synchronicity of neuronal populations, and multi-band energy distribution patterns. Through signal processing and feature extraction techniques, multidimensional features reflecting brain functional status can be extracted from EEG, providing biomarkers for disease identification.

[0003] Compared to imaging methods (such as PET and MRI) and cerebrospinal fluid analysis, resting EEG offers advantages such as ease of operation, high reproducibility, and suitability for long-term monitoring. However, despite ongoing research, significant challenges remain in EEG-based differential diagnosis of Alzheimer's disease (AD) and frontotemporal dementia (FTD). While some research has yielded results in using EEG to differentiate between AD and FTD, EEG-based differential diagnosis still faces the following major technical difficulties:

[0004] 1. Insufficient Collaborative Modeling of Multiple EEG Features: Electroencephalogram (EEG) signals simultaneously contain functional connectivity features reflecting brain region interactions (such as coherence) and spectral features characterizing local neural activity states (such as power spectral density (PSD) and band energy ratio). These different features describe brain network function at different levels and exhibit significant nonlinear correlations and interdependencies. However, existing methods typically process different EEG features independently or fuse them only through superficial methods such as simple splicing or static weighting. They lack interactive modeling mechanisms capable of characterizing the dynamic coupling relationships between different features, making it difficult to fully reveal the complementarity between functional connectivity features and spectral features and their biological significance in differentiating between Alzheimer's disease (AD) and frontotemporal dementia (FTD).

[0005] 2. Insufficient Utilization of Functional Connectivity Structure Information and Spatial Topology: Existing EEG-based dementia classification methods typically rely solely on the connection strength reflected by the functional connectivity matrix (such as coherence) when constructing brain networks, treating them as static graph structures. However, these methods often neglect the topological layout of EEG electrodes in three-dimensional space and the geometric constraints they imply, resulting in graph structures based solely on connection weights and lacking the participation of spatial structural information. Furthermore, traditional graph models often simply use spatial coordinates as additional feature inputs, lacking a unified modeling mechanism that can simultaneously utilize the spatial geometry of electrodes and functional connectivity strength. This makes it impossible to maintain consistency in spatial transformations during graph convolution and to effectively represent the true organizational relationships between different brain regions. Therefore, existing methods struggle to accurately characterize the spatial structural differences between AD and FTD brain networks and are insufficient to reflect the damage caused to spatial topology by pathological network lesions, thus limiting the classification performance and generalization ability of the models.

[0006] 3. Insufficient capture of local features and cross-interaction information: Models that rely solely on global functional connectivity or spectral features are unable to fully capture subtle interactive changes in local brain regions and lack modeling of the complex interactive relationship between spatial structure and spectral features. This results in insufficient sensitivity of the model to distinguish between AD and FTD and limits the accuracy and stability of cognitive score prediction.

[0007] Therefore, there is an urgent need to provide a dementia identification method based on spatial cross-attention fusion of brain-computer networks to solve the bottleneck problems of the existing technologies mentioned above, thereby improving the performance of AD and FTD disease classification and the clinical applicability of the model. Summary of the Invention

[0008] This invention addresses the shortcomings of existing technologies by providing a dementia identification method based on spatial cross-attention fusion using brain-computer networks. This method fully integrates electrode 3D space, functional connectivity graph structure, and bidirectional cross-attention to achieve joint modeling of spatial graph features and global spectral features. By combining the nonlinear representation capabilities of deep learning, it effectively integrates complementary information from multiple features, enhancing feature discriminative power and improving the accuracy, stability, and generalization ability of Alzheimer's disease (AD) and frontotemporal dementia (FTD) classification in complex brain networks.

[0009] The present invention adopts the following technical solution to solve the technical problem.

[0010] The present invention discloses a dementia identification method based on spatial cross-attention fusion of brain-computer networks, characterized by comprising the following steps:

[0011] Step 1: EEG signal acquisition and preprocessing steps;

[0012] Step 2: Feature extraction step; obtaining spatial map structural features of EEG signals. Spectral statistical characteristics of EEG signals ;

[0013] Step 3: Determine the structural features of the spatial graph With spectral statistical characteristics A common input bidirectional cross-attention module models cross-feature associations through a multi-head interactive attention mechanism and extracts complementary fusion features;

[0014] Step 4: Structural features of the spatial graph With integrated interactive features After hybrid pooling and concatenation, the data are fed into the Chebyshev Kolmogorov–Arnold network (Chebyshev KAN) classifier. Nonlinear discrimination is achieved using Chebyshev graph kernels, and the classification results of Alzheimer's disease, frontotemporal dementia, and healthy controls, as well as the MMSE cognitive scale evaluation scores, are output.

[0015] The dementia identification method based on spatial cross-attention fusion of brain-computer networks of the present invention is also characterized by:

[0016] Further, step 1 includes the following steps:

[0017] Step 101: Acquisition of EEG signals and channel standardization;

[0018] Step 102: Preprocessing of EEG signals;

[0019] Step 103: Data segmentation and feature extraction;

[0020] Step 104: Construction of the functional connectivity graph and partitioning of the training set.

[0021] Furthermore, in step 102, a Chebyshev bandpass filter is used to filter the original EEG signal during the preprocessing.

[0022] Furthermore, in step 102, independent component analysis is used to eliminate artifact interference from the filtered raw EEG signal.

[0023] Further, step 2 includes the following steps:

[0024] Step 201: Input the brain network functional connectivity map and electrode coordinate features into the equivariant graph neural network EGNN;

[0025] Step 202: Node feature update steps of the Local Feature Enhancement Module (CLFE);

[0026] Step 203: Spectral Statistical Characteristics The extraction steps.

[0027] Furthermore, in step 201, the two-layer equivariant graph neural network EGNN is used for feature modeling.

[0028] Furthermore, in step 202, the local feature enhancement module CLFE uses a neighbor feature weighted mean aggregation method for feature enhancement.

[0029] Furthermore, step 3 includes the following steps:

[0030] Step 301: Cross-feature attention weights A sf Calculation;

[0031] Step 302: Reverse attention weights A fs Calculation;

[0032] Step 303: Feature fusion step.

[0033] Furthermore, step 4 includes the following steps:

[0034] Step 401: Hybrid pooling step;

[0035] Step 402: Disease classification steps.

[0036] Furthermore, in step 402, the Chebyshev KAN classifier is used for disease identification.

[0037] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0038] This invention discloses a dementia identification method based on spatial cross-attention fusion using brain-computer networks, comprising: acquiring resting-state electroencephalogram (EEG) signals, performing preprocessing and brain network construction, calculating frequency band power ratios, and dividing the dataset into training and testing sets; fusing spatial information through the brain network, and obtaining spatial features via an isovariant graph neural network and a local feature enhancement module. The frequency band power ratio is encoded by a multilayer perceptron to obtain spectral statistical characteristics. ;Will and Input a bidirectional cross-attention module to extract complementary fusion features; The complementary fusion features are mixed and pooled, then concatenated and fed into a Chebyshev Kolmogorov–Arnold network classifier to output classification results for Alzheimer's disease / frontotemporal dementia / healthy controls (AD / FTD / HC) and cognitive scale evaluation scores (MMSE). This invention effectively integrates complementary information from multiple features by jointly modeling spatial graph features and global spectral features, thereby improving the accuracy, stability, and generalization ability of AD and FTD classification in complex brain networks.

[0039] This invention addresses the shortcomings of existing technologies by providing a dementia identification method based on spatial cross-attention fusion using brain-computer networks. This method fully integrates electrode 3D space, functional connectivity graph structure, and bidirectional cross-attention to achieve joint modeling of spatial graph features and global spectral features. By combining the nonlinear representation capabilities of deep learning, it effectively integrates complementary information from multiple features, enhancing feature discriminative power and improving the accuracy, stability, and generalization ability of Alzheimer's disease (AD) and frontotemporal dementia (FTD) classification in complex brain networks. Attached Figure Description

[0040] Figure 1 is a flowchart illustrating a brain-computer network spatial cross-attention fusion method for dementia identification according to the present invention.

[0041] Figure 2 is a functional connectivity diagram of the brain network. The blue connecting lines represent the coherence between nodes, and the color bars on the nodes represent node features.

[0042] Figure 3 is a framework diagram of the dementia identification method based on spatial cross-attention fusion of brain-computer networks.

[0043] Figure 4 is a detailed architecture diagram of the local feature enhancement module;

[0044] Figure 5 is a detailed architecture diagram of the bidirectional cross-attention module;

[0045] Figure 6 shows the accuracy and loss curves of the present invention on the ds004504 and BrainLat datasets. In Figure 6(a), the training and testing process curves of the present invention's method on the ds004504 dataset are shown. In Figure 6(b), the training and testing process curves of the present invention's method on the BrainLat dataset are shown.

[0046] Figure 7 is a box plot of the actual MMSE cognitive scores and predicted MMSE cognitive scores on the ds004504 dataset in the three categories of HC, AD and FTD.

[0047] The present invention will be further described below through specific embodiments and in conjunction with the accompanying drawings. Detailed Implementation

[0048] Referring to Figures 1 to 7, the present invention provides a brain-computer network spatial cross-attention fusion method for dementia identification, characterized by comprising the following steps:

[0049] Step 1: EEG signal acquisition and preprocessing steps;

[0050] Preprocessed EEG signals were acquired, and each subject's EEG signal was divided into 10-second segments, making each time segment an independent feature extraction and modeling unit. For each time segment's EEG signal x, coherence and differential entropy (DE) features were extracted. A functional connectivity graph was constructed using electrodes (19 electrodes in the specific implementation) as nodes and coherence as edge weights (DE features were used as node features). Simultaneously, the power spectral density ratio features between each frequency band were calculated and integrated into a vector. A five-fold cross-validation method was used to train a graph neural network model for spatial cross-attention fusion in a brain-computer interface, dividing the dataset into training and testing sets.

[0051] Step 2: Feature extraction step; obtaining spatial map structural features of EEG signals. Spectral statistical characteristics of EEG signals ;

[0052] As shown in Figure 3, the internationally recognized MNI standard space was first used to map the head surface locations of the 19 EEG electrodes to corresponding three-dimensional coordinates. Then, the brain network functional connectivity map (including node features and edge weights) and the node three-dimensional spatial coordinates were input into an Equivariant Graph Neural Network (EGNN) for feature extraction. After passing through a first multilayer perceptron, the node three-dimensional spatial coordinates, encoded by a second multilayer perceptron, were input into a Conditional Local Feature Encoding (CLFE) module, which outputs spatial graph structural features focusing on local brain region interactions. Simultaneously, the global spectral feature vector g PSD The input is encoded by a third-level multilayer perceptron (MLP) to extract nonlinear interaction patterns between frequency bands, thereby obtaining global statistical spectral features. .

[0053] Step 3: Determine the structural features of the spatial graph With spectral statistical characteristics A common input bidirectional cross-attention module models cross-feature associations through a multi-head interactive attention mechanism and extracts complementary fusion features;

[0054] Spatial graph structural features With spectral statistical characteristics The input is a Bidirectional Cross Attention Block (BCAB), as shown in Figure 4. By dynamically mining complementary information from two types of features through scaled dot-product attention and a multi-head mechanism, it alleviates the one-sidedness of single-feature representation and outputs interactive features that fuse cross-feature associations. The BCAB is used to achieve bidirectional interaction and complementary fusion of spatial graph structural features and spectral statistical features, enhancing the model's ability to represent multidimensional coupling relationships between brain regions through a cross-feature attention mechanism.

[0055] Step 4: Structural features of the spatial graph With integrated interactive features After hybrid pooling and concatenation, the data are fed into the Chebyshev Kolmogorov–Arnold network (Chebyshev KAN) classifier; nonlinear discrimination is achieved using Chebyshev graph kernels, and the AD / FTD / HC classification results and MMSE cognitive scale evaluation scores are output.

[0056] As shown in Figure 3, the spatial graph structure features The fused interactive features F output from step 3 fusionSeparate pooling operations are performed, and the results of both are then concatenated directly along the channel dimension to form the final fused feature vector. This final fused feature vector is input into the Chebyshev KAN classifier, which serves as the final discrimination module to improve the modeling ability of complex nonlinear relationships in EEG features. A graph convolution kernel with Chebyshev polynomial approximation is used to adapt to the non-Euclidean structural characteristics of EEG features. Finally, a two-layer classification head outputs three-class classification results for diseases AD / FTD / HC and MMSE (Mini-Mental State Examination) cognitive scale evaluation scores.

[0057] In practice, step 1 includes the following steps:

[0058] Step 101: Acquisition of EEG signals and channel standardization;

[0059] This invention uses the publicly available datasets ds004504 and BrainLat for experimental validation. The ds004504 dataset contains 88 subjects, including 36 patients with Alzheimer's disease (AD), 23 patients with frontotemporal dementia (FTD), and 29 healthy controls (HC). The BrainLat dataset contains 35 patients with Alzheimer's disease (AD), 19 patients with behaviorally variant frontotemporal dementia (FTD), and 32 healthy controls (HC). Electroencephalogram (EEG) signals were acquired using a Nihon Kohden EEG-2100 system (19 channels, 200 Hz sampling rate) and a Biosemi ActiveTwo system (128 channels, 128 Hz sampling rate), respectively. To ensure consistency across datasets, the electrode channels in both datasets were standardized to 19 electrode channels according to the international 10-20 system.

[0060] Step 102: Preprocessing of EEG signals;

[0061] First, the acquired raw EEG signals are represented as a matrix. ;in, This represents the number of electrode channels. The sampling time points are specified. The original signal is resampled to a uniform sampling frequency of 128Hz, and a sixth-order Chebyshev bandpass filter is used to perform bandpass filtering in the range of [1, 45] Hz to remove power frequency noise and low-frequency noise, resulting in a filtered signal. Subsequently, independent component analysis (ICA) algorithm is used to eliminate artifacts such as eye movement and electromyography (EMG) interference, obtaining a clean EEG signal. .

[0062] Step 103: Data segmentation and feature extraction;

[0063] The preprocessed clean EEG signals were divided into 10-second epochs, with each subject's EEG signal containing several epochs. After division, each subject's EEG signal was represented as follows: , Where K is the number of time segments in the subject's EEG, r represents the r-th time segment, 1≤r≤K; since the sampling rate is 128 Hz and the duration of each time segment is 10 s, each channel contains There are 100 sampling points, therefore the size of the EEG signal for a single time segment is 100. .

[0064] For the EEG signal of the r-th time segment This invention extracts the following three types of features: coherence (COH), differential entropy (DE), and inter-band power spectral density ratios (IB-PSDR).

[0065] 1. Coherence Characteristics (COH): Calculates the frequency domain synchronicity between any two channels, quantifying the strength of functional connectivity between brain regions. The formula for calculating COH is shown in the following formula (1);

[0066] (1);

[0067] In formula (1), f is the frequency (unit: Hz); P xx (f) and P yy (f) represents the power spectral density (PSD) of signals x and y, respectively, characterizing the energy distribution of a single signal at frequency f; P xy (f) represents the cross-spectral density of signals x and y, characterizing the phase correlation between signals x and y at frequency f; both signals x and y are 10-second segments of single-channel EEG signals. In a preferred embodiment, the coherence spectrum in the range of 1–45 Hz is calculated, and the coherence values ​​are averaged within this frequency range to obtain a full-band coherence index reflecting the overall functional connectivity level, ultimately yielding the coherence matrix for the r-th time segment: The coherence matrix A r As a weighted adjacency matrix of the functional connectivity map of brain networks, This corresponds to the index of 19 EEG electrode channels.

[0068] 2. Differential Entropy (DE): Used to measure the complexity of local EEG activity, it is an important characteristic describing EEG activity. Assuming that the signals in each frequency band approximately follow a Gaussian distribution, the corresponding differential entropy characteristics can be estimated based on the power spectrum. For each time segment and each electrode channel, this invention divides the frequency domain in a step size of 0.5 Hz within the frequency range of 1 Hz to 45 Hz, obtaining frequency band signals at 89 frequency points, where 1 ≤ k ≤ 89; for the k-th frequency point... Calculate its corresponding differential entropy. The differential entropy sequence of local frequency points is formed, as shown in the following formula (2);

[0069] , (2);

[0070] In formula (2), Frequency point The power at that point; e is the natural constant.

[0071] To further enhance the ability of the brain-computer network spatial cross-attention fusion graph neural network model of this invention to characterize the overall frequency domain dynamics, this invention calculates the comprehensive power P across the entire frequency band from 1Hz to 45Hz based on the aforementioned 89 differential entropy features DE. broad : And based on this, the differential entropy of the entire frequency band is obtained: Finally, the differential entropy eigenvector h of a single channel DE The composition is as follows: Combine the features of all 19 nodes (19 electrodes, each electrode as a node) to form a node feature matrix H: .

[0072] 3. Inter-band Power Spectral Density Ratio (IB-PSDR): This ratio characterizes the relative energy distribution between different EEG frequency bands and is a crucial global feature for depicting the overall spectral structure of the subject. By comparing the power ratios of typical frequency bands, the dynamic changes of EEG signals between different frequency components are reflected. In practice, it is necessary to calculate the power ratios between each frequency band of the EEG signal, including 10 ratios: δ / θ, δ / α, δ / β, δ / γ, θ / α, θ / β, θ / γ, α / β, α / γ, and β / γ, to characterize the dynamic changes between frequency bands. First, the power spectral density function is calculated for a single-channel EEG signal. And integrate within a typical frequency band to obtain the corresponding frequency band power:

[0073] Delta band: 1 ~ 4 Hz, total power ;

[0074] Theta band: 4 ~ 8 Hz, total power ;

[0075] Alpha band: 8 ~ 13 Hz, total power ;

[0076] β band: 13 ~ 30 Hz, total power ;

[0077] γ band: 30 ~ 45 Hz, total power .

[0078] Based on the power in the aforementioned frequency bands, 10 typical ratios PR1 to PR2 are defined. 10 See formula (3) below;

[0079] (3);

[0080] For each time segment of the EEG signal, this invention integrates the above 10 ratios into a global spectral feature vector g. PSD for: .

[0081] Step 104: Construction of the functional connectivity graph and partitioning of the training set.

[0082] A brain network functional connectivity graph was constructed using electrodes as nodes and coherence as edge weights. Differential entropy features were used as node features to form the brain network functional connectivity graph, as shown in Figure 2. The blue connecting lines represent the coherence matrix, and the color bars on the nodes represent node features. Simultaneously, the integrated global spectral feature vector g... PSD This information is used for subsequent spectral statistical modeling. Simultaneously, a five-fold cross-validation method is employed to divide the training and testing sets, thereby enhancing the generalization ability and stability of the brain-computer network spatial cross-attention fusion graph neural network model.

[0083] In specific implementation, in step 102, a Chebyshev bandpass filter is used to filter the original EEG signal during the preprocessing process.

[0084] In specific implementation, in step 102, independent component analysis is used to eliminate artifact interference from the original EEG signal after filtering.

[0085] In practice, step 2 includes the following steps:

[0086] Step 201: Input the brain network functional connectivity map and electrode coordinate features into the equivariant graph neural network EGNN;

[0087] The 3D spatial coordinates of the nodes and the brain network functional connectivity map obtained in step 104 are used as inputs and fed into the existing isovariant graph neural network (EGNN) backbone structure used in this invention. The innovation of this invention lies in the fact that the isovariant graph neural network (EGNN) is used for the first time in the modeling and cross-feature fusion framework of EEG brain network structural features.

[0088] As shown in Figure 3, this invention uses two layers of EGNN for feature modeling. In each EGNN layer, the network input includes a brain network functional connectivity graph and node spatial coordinate information. The brain network functional connectivity graph consists of a node feature matrix. Weight matrix of edges between nodes The node spatial coordinates are composed of the node coordinate matrix. It indicates. Among them, The dimension of the node feature vector. The dimension representing the node's spatial coordinates during training. and Remain unchanged; This represents the feature vector of node i in the l-th layer (l=1,2) of the EGNN. Let be the functional connection weight between node i and node j. Let i represent the three-dimensional spatial coordinates of node i in the l-th layer EGNN, and N be the total number of nodes (in specific implementation, there are 19 electrodes, i.e., 19 nodes); 1≤i≤N, 1≤j≤N.

[0089] EGNN utilizes its existing edge message passing mechanism to comprehensively leverage node features, neighborhood structure relationships, node spatial locations, and functional connection edge weights to calculate neighborhood messages between nodes, which are then used for updates in the following three aspects: (1) Updating node coordinates: Node coordinates in EGNN are used to construct a geometric embedding representation space that integrates functional connection information. During message passing, by introducing functional connection weights between nodes to participate in the calculation of coordinate update terms, the model can adaptively adjust the relative positions of brain regions in the embedding space during training, thereby making brain regions with similar functional connection patterns exhibit higher similarity in geometric representation, providing a more sensitive spatial representation of functional topology for subsequent feature learning; (2) Updating edge weights between nodes: During message passing, the connection weights between nodes are dynamically modeled to characterize the changes in the intensity of functional interaction between brain regions; (3) Updating node features: Each node can integrate the connection information and spatial topology information of adjacent brain regions. It should be emphasized that in the above update process, this invention constrains the output tensor dimension of EGNN, and the dimension of the updated node feature matrix remains unchanged. The tensor dimension of the edge weight matrix between nodes remains unchanged. The dimensions of the node space coordinate matrix remain unchanged. In other words, the input and output tensor sizes of EGNN are completely identical, thus ensuring smooth connection between network layers and maintaining the stability of the overall model structure. Therefore, the output of EGNN can be regarded as a feature representation of brain network functional connectivity graph that explicitly integrates node spatial location information, that is, it obtains brain network functional connectivity graph features that are sensitive to spatial structure, providing a basic representation for subsequent local feature enhancement and cross-feature fusion.

[0090] Step 202: Node feature update steps of the Local Feature Enhancement Module (CLFE);

[0091] As shown in Figure 4, the Local Feature Enhancement (CLFE) module receives graph features from the EGNN and the 3D spatial coordinates of nodes encoded by a multilayer perceptron as input. This invention improves upon the existing CLFE module by adding the participation of node 3D coordinates to the original structure, enabling explicit modeling of spatial location information during inter-layer fusion. Simultaneously, it introduces edge-weighted GraphSAGE mean aggregation, allowing each node to adaptively integrate local feature patterns from its neighboring brain regions, thereby enhancing the ability to capture dynamic interactions within local brain regions. The module achieves cross-layer feature coupling through a conditional feature fusion mechanism (concatenating and normalizing features from the previous and current layers and mapping them via a multilayer perceptron), ultimately outputting spatial graph structure features focusing on local brain region interactions. .

[0092] The encoded 3D coordinate set is set as follows: , This represents the encoded 3D coordinates of the Nth node; the input node feature matrix is: Let l be the l-th layer local feature enhancement module. The input after position enhancement is the node feature enhancement matrix. : .

[0093] In the neighbor feature aggregation stage, functional connection weights between node i and node j are introduced. The neighbor node features are aggregated by weighted average, as shown in the following formula (4).

[0094] , , (4);

[0095] In formula (4), Represents a node The set of neighbors; Represents the normalized coefficients based on functional connection weights; For neighbor node indexes, used to index nodes The weights of all neighboring edges are summed; MLP is the fourth layer perceptron, and Norm is the normalization operation (LayerNorm). Node feature enhancement matrix The node characteristics of the j-th node in the array. This represents the result of weighted average aggregation of the features of neighboring nodes; The aggregation result for all nodes The matrix formed;

[0096] Enhanced Neighbor Information Extraction: Where Concat represents the feature concatenation operation, MLP is the fifth multilayer perceptron, and Dropout represents regularization to prevent overfitting; it was set to 0.3 in the experiment. Feature fusion output: .in, Indicates the first The node feature enhancement terms in the layer local feature enhancement module are driven by neighborhood information. Specifically, By using the location-enhanced node feature matrix The weighted aggregation result of its corresponding neighbor features The data is spliced ​​together and obtained through multilayer perceptron mapping and Dropout regularization. This data is used to characterize the high-order coupling relationship between the node's own attributes and the interaction information of its neighboring brain regions. It reflects the local dynamic interaction features between nodes and their neighboring brain regions under the constraints of functional connectivity weights, and plays a supplementary and enhancing role to the original node representation. Indicates the first The intermediate node representation in the local feature enhancement module after fusing neighbor enhancement information represents the weighted aggregation result of neighbor features. Enhance information with neighbors It is constructed by adding elements one by one, thereby realizing a comprehensive model of the local structural information of the node and the interaction mode of the neighborhood. As the input for subsequent nonlinear transformations and residual connections, it carries the local spatial structure features of the nodes in the current layer and is an important intermediate representation for realizing cross-layer feature transfer and stable training.

[0097] The final node represents the update formula as follows: , where ReLU is a non-linear activation function.

[0098] final = , For spatial graph structure features, including a node feature matrix (dimension is...). The weight matrix of edges between nodes (dimension is...) ).

[0099] Step 203: Spectral Statistical Characteristics The extraction steps.

[0100] The global spectral feature vector g extracted in step 103 PSD Inputting the third-level multilayer perceptron (MLP) network yields spectral statistical features. See formula (5) below;

[0101] (5);

[0102] In formula (5), This represents the global spectral feature vector extracted in step 103; , and , They are respectively Weight matrices and bias terms for each layer; The activation function is nonlinear. The nonlinear transform of the multilayer perceptron is used to extract implicit coupling and dynamic correlation patterns between frequency bands, obtaining statistically significant spectral features. Spectral statistical characteristics It reflects the relationship between the energy distribution ratio and neural rhythm interaction between different frequency bands, supplementing the global spectral information missing in brain functional connectivity.

[0103] In specific implementation, in step 201, the two-layer equivariant graph neural network EGNN is used for feature modeling.

[0104] In specific implementation, in step 202, the local feature enhancement module CLFE uses a neighbor feature weighted mean aggregation method for feature enhancement.

[0105] In practice, step 3 includes the following steps:

[0106] Step 301: Cross-feature attention weights A sf Calculation;

[0107] Compute the attention mapping from space to spectrum: , , .in Indicates the number of electrode nodes. The number of tokens representing spectral statistical features. d represents the attention head dimension. Wherein, These are learnable linear mapping matrices for queries, keys, and values, respectively. To make spatial graph structural features The mapping result to the query subspace. To extract spectral statistical characteristics The mapping result to the key subspace; To extract spectral statistical characteristics The mapping result to the value subspace; D is the node feature dimension.

[0108] To achieve spatial graph structure features With spectral statistical characteristics In attention computation, this invention achieves dimensional alignment between spatial and spectral features through feature mapping and dimensional recombination. It projects and expands spectral statistical features into token representations consistent with the number of electrode nodes, thereby realizing one-to-one alignment between spatial and spectral features in attention computation. and equal.

[0109] Based on spatial graph structural features For querying, based on spectral statistical characteristics Calculate cross-feature attention weights for key / value pairs. See formula (6) below;

[0110] (6);

[0111] In formula (6), T is the matrix transpose.

[0112] Through cross-feature attention weights The enhanced spatial features were calculated. See formula (7) below;

[0113] (7);

[0114] This process enables cross-feature selective attention of spatial features to spectral features, allowing spatial topological information to adaptively absorb supplementary information from spectral patterns.

[0115] Step 302: Reverse attention weights A fs Calculation;

[0116] Calculate the attention mapping from the spectrum to the space: , , .in and Similar to step 302, represents the number of electrode nodes and the number of tokens for spectral statistical features, respectively. The feature mapping and dimension alignment strategy from step 302 are also used to align the features. d represents the attention head dimension. Wherein, It is a learnable linear mapping matrix. Indicates the statistical characteristics of the spectrum Mapped to the query subspace, This indicates the structural features of the spatial graph. Mapped to key space, This indicates the structural features of the spatial graph. Mapped to a value subspace; D is the node feature dimension. Based on spectral statistical features. For querying, based on spatial graph structural features Calculate the inverse attention weights for key / value pairs. See formula (8) below;

[0117] (8);

[0118] Through the weight matrix The enhanced spectral features were calculated. See formula (9) below;

[0119] (9);

[0120] This directional mapping enables spectral features to capture the modulation effects of structural connections between brain regions, thus reflecting spatial dependence at the frequency domain level.

[0121] Step 303: Feature fusion step;

[0122] To enhance cross-feature nonlinear interactions, the enhanced spatial features will be... With enhanced spectral characteristics First, the dimensions are aligned using a multilayer perceptron, then residual connections and layer normalization are performed, and finally, the data is encoded using a multilayer perceptron (MLP), as shown in formulas (10) and (11) below.

[0123] (10);

[0124] (11);

[0125] In formulas (10) and (11), The fused spatial features represent the final output of the spatial graph structure features after cross-feature attention enhancement. The outermost MLP is the sixth layer perceptron. ; The fused spectral characteristics represent the final output of the spectral statistical characteristics modulated from the spatial graph structure information. The outermost MLP is the seventh layer perceptron. .

[0126] Subsequently, spatial features will be integrated. and fusion spectral characteristics These two features are concatenated along the feature dimension and fused using MLP, as shown in the following formula (12);

[0127] (12);

[0128] Formula (12) ultimately outputs the fused interactive features. Where B is BatchSize, which is the number of samples input into the network for each forward propagation. It is the eighth multilayer perceptron. This serves as input for the next stage, the global pooling and classification module (step 4).

[0129] In practice, step 4 includes the following steps:

[0130] Step 401: Hybrid pooling step;

[0131] The spatial graph structural features F obtained in step 2 S and the fused feature F fusion Hybrid pooling is used. Hybrid pooling combines mean pooling and max pooling, achieving a balance between global statistical stability and local saliency, thereby generating a more discriminative graph-level global representation.

[0132] Let the node comprehensive feature matrix F be: , where f i Let N be the feature vector of the i-th node, N be the number of nodes, and D be the feature dimension (distinguishing it from "brain network functional connectivity graph consisting of node feature matrices"). The node comprehensive feature matrix Unlike the node feature matrix used to construct the functional connectivity map of brain networks mentioned earlier. The node comprehensive feature matrix F is a general term for the node feature matrices after processing and feature enhancement by relevant modules, used to uniformly represent the input node representations of subsequent modules. In this invention, F can specifically represent the spatial graph structure feature F. S Or the fused interaction feature F after cross-modal fusion fusion It is used to characterize the comprehensive representation of a node in spatial structure or after the fusion of multiple features.

[0133] eigenvector f i Average pooling results See the following formula (13);

[0134] (13);

[0135] The result of max pooling f max，t See formula (14) below;

[0136] (14);

[0137] In formula (14), f it This represents the value of the i-th node in the t-th feature dimension (1≤t≤D). D is the node feature dimension.

[0138] The final hybrid global representation f of hybrid pooling pool See formula (15) below;

[0139] (15);

[0140] In formula (15), This represents a hybrid global representation that integrates global statistics and local peak responses. This mechanism helps to capture overall trends in brain networks and significant activity patterns in key brain regions.

[0141] Step 402: Disease classification steps.

[0142] The hybrid pooling global representation obtained in step 401 As input, the data is fed into the Chebyshev KAN classifier for disease identification. The classifier utilizes graph convolution kernels in the form of Chebyshev multinomials to adapt to the non-Euclidean structure of the EEG functional connectivity graph and extracts discriminative features from node features and connectivity topology. Subsequently, a fully connected predictor outputs three-class classification results for Alzheimer's disease (AD), frontotemporal dementia (FTD), and healthy controls (HC), along with corresponding MMSE cognitive score predictions. The MMSE score is only introduced as a supervision signal during the training phase, constraining the deviation between the model output and the actual cognitive level through a constructed regression loss function, and is not used as a model input feature in the inference process.

[0143] During the training process of regression loss function constraint (the joint loss function acts on the output prediction stage of the model, specifically reflected in the Chebyshev KAN-based multi-task prediction module at the very end of Figure 3, which achieves joint optimization of model parameters by simultaneously constraining the classification results and MMSE prediction results), the joint loss function is used to simultaneously optimize the classification task and the cognitive rating prediction task. Its loss function is defined as follows: formula (16) and formula (17).

[0144] (16);

[0145] (17);

[0146] In formulas (16) and (17), Cross-entropy loss used for the three-class classification task; The regression loss for the MMSE score is used to guide the model to perceive the degree of cognitive function decline in subjects while ensuring disease classification performance. Specifically, the model sets an independent regression branch to output the predicted MMSE score based on shared feature representations. The corresponding actual MMSE score is recorded as ,in, Indicates the first The training samples are [number]. This regression loss participates in backpropagation during the training phase by minimizing the error between the predicted and true values; it serves only as an auxiliary constraint and does not participate in model input construction. By introducing a cognitive rating regression task, the model can learn discriminative features related to cognitive states, thereby improving the stability and reliability of brain disease diagnosis results within a multi-task joint optimization framework. Parameters The balance coefficient between classification loss and regression loss was set to [value] in the experiment. By employing the aforementioned joint learning strategy, the model can simultaneously optimize its disease classification ability and cognitive score prediction performance, further enhancing the reliability of the overall diagnosis.

[0147] In specific implementation, in step 402, the Chebyshev KAN classifier is used for disease identification.

[0148] As shown in Figure 1, the dementia identification method based on spatial cross-attention fusion of brain-computer networks of the present invention was tested on two public datasets, ds004504 and BrainLat. To quantitatively evaluate the model classification results, sensitivity, specificity, F1 score, and accuracy were used as evaluation metrics. Table 1 below shows the experimental results of different model methods on the ds004504 dataset and the ablation experiment results, demonstrating the comparison between different model methods and the present invention. Figure 6(a) shows the training and testing process curves of the method of the present invention on the ds004504 dataset.

[0149] Table 1. Experimental results using different modeling methods on the ds004504 dataset.

[0150]

[0151] Table 1 shows the experimental results of different model methods and the results of ablation experiments on the ds004504 dataset. The method described in this invention achieves the best performance in all four metrics: sensitivity 96.45%, specificity 97.02%, F1 score 96.41%, and accuracy 96.76%. Compared with the second-best method, SAT (Structure-Aware Transformer), the method of this invention improves accuracy by 7.55%, sensitivity by 7.78%, specificity by 7.88%, and F1 score by 7.69%. Furthermore, the method of this invention outperforms other comparable models such as GNN (Graph Neural Network), GAT (Graph Attention Network), and GPS (General Powerful Scalable Graph Transformer) in all performance metrics, with improvements exceeding 30% in some cases. Ablation experiments showed that after removing the CLFE module (w / o CLFE), the model's sensitivity, specificity, F1 score, and accuracy decreased to 93.84%, 93.47%, 93.21%, and 93.39%, respectively, indicating a significant decline in overall performance. Removing the bidirectional cross-attention module BCAB (w / o BCAB) further reduced model performance, with the four metrics at 94.58%, 94.88%, 94.01%, and 94.05%, respectively, still lower than the complete model. These results demonstrate that the complete method described in this invention achieves the highest performance across all four key metrics, verifying that both the CLFE and BCAB modules contribute to performance improvement, and that their collaborative design is crucial for achieving optimal classification results.

[0152] Figure 6(a) shows the training-test curves of the method of this invention on the ds004504 dataset. The training and testing accuracy increases rapidly with the number of rounds and then tends to stabilize, with the final test accuracy remaining at a high level. The trends of the two are highly consistent. The training and testing loss decreases rapidly with the number of rounds and then tends to stabilize, with the final test loss remaining at an extremely low level. This indicates that the model of this invention does not exhibit overfitting and has good convergence.

[0153] Table 2 below shows the experimental results of different model methods on the BrainLat dataset and the ablation experiment results, which can be seen as a comparison between different model methods and the present invention.

[0154] Table 2. Experimental results using different modeling methods on the BrainLat dataset.

[0155]

[0156] On the BrainLat dataset, this invention achieved a sensitivity of 97.36%, a specificity of 98.22%, an F1 score of 97.50%, and an accuracy of 97.87%. Compared to the suboptimal method SAT, it improved accuracy by 8.06%, sensitivity by 6.95%, specificity by 8.84%, and F1 score by 7.53%. All metrics outperformed comparable methods such as GNN, GAT, and GPS, with improvements exceeding 30% in some cases. Further ablation experiments showed that removing the CLFE module (w / o CLFE) reduced the model's sensitivity, specificity, F1 score, and accuracy to 93.07%, 95.14%, 92.88%, and 94.21%, respectively, indicating a significant decrease in overall performance. This demonstrates that the CLFE module played a crucial role in local feature enhancement. When the BCAB module was removed (w / oBCAB), the model's four metrics were 94.14%, 96.31%, 93.75%, and 95.62%, respectively, outperforming the results when CLFE was removed, but still lower than the complete model. This indicates that BCAB also makes a significant contribution to cross-domain information interaction and feature fusion. In contrast, when using the complete method described in this invention, the model achieved improvements in all four metrics, reaching 97.36%, 98.22%, 97.50%, and 97.87%, respectively. These results fully demonstrate that both CLFE and BCAB modules can effectively improve model performance, and their synergistic effect enabled the model to achieve optimal performance on the BrainLat dataset, thus verifying the scientific validity and necessity of the method design in this invention.

[0157] Figure 6(b) shows the training and testing curves of the method of this invention on the BrainLat dataset. It shows that the training and testing accuracy increases rapidly with each round and then stabilizes, with the final testing accuracy remaining at a high level of nearly 98%, exhibiting a highly consistent trend. The training and testing losses decrease rapidly with each round and then stabilize, with the final testing loss approaching a stable low value range, indicating that the model exhibits no overfitting, good convergence, and excellent generalization ability. Although the sample distribution and feature patterns of this dataset differ from those of ds004504, the method of this invention still maintains a stable lead, confirming its strong generalization ability.

[0158] Experimental results on the ds004504 and BrainLat datasets demonstrate that the key structures and modules of this invention play an irreplaceable role in overall performance: EGNN utilizes the three-dimensional spatial coordinates of electrodes to extract isovariant features of the spatial topology of brain regions, effectively enhancing the spatial structural representation ability of the functional connectivity graph; the CLFE module enhances local brain region interaction information through a neighbor feature mean aggregation and conditional feature fusion mechanism, playing a crucial role in the accurate representation of local features; the BCAB module employs a bidirectional cross-attention mechanism to achieve complementary fusion of spatial graph structural features and global spectral features, enhancing the representation ability of multidimensional coupling relationships between brain regions. The synergistic effect of these modules enables the method of this invention to achieve optimal classification results on both datasets, while also exhibiting good convergence and generalization capabilities, verifying the scientific validity and necessity of the network design. Furthermore, the Chebyshev KAN classifier effectively adapts to the non-Euclidean structure of EEG features through Chebyshev graph convolution kernels, avoiding the loss of topological information; hybrid pooling extracts the global representation of the graph structure and complements the interactive features output by BCAB, enabling the method to robustly handle cross-subject feature differences, thereby achieving accurate identification of AD and FTD.

[0159] To further verify the ability of the method described in this invention to capture cognitive function-related features, regression prediction of Mini-Mental State Examination (MMSE) scores was performed on the ds004504 dataset based on the extracted EEG features. The results are shown in Table 3.

[0160] Table 3. Comparison of actual and predicted MMSE cognitive scores among groups on the ds004504 dataset.

[0161]

[0162] Table 3 shows that the mean true MMSE cognitive score for the healthy control group (HC) was 30.00 (standard deviation 0.00), and the predicted mean was 29.24 (standard deviation 0.31); the mean true MMSE cognitive score for the Alzheimer's disease group (AD) was 17.75 (standard deviation 4.44), and the predicted mean was 18.51 (standard deviation 3.88); the mean true MMSE cognitive score for the frontotemporal dementia group (FTD) was 22.17 (standard deviation 2.58), and the predicted mean was 22.42 (standard deviation 2.14). ANOVA analysis showed highly significant differences between the true and predicted MMSE cognitive scores, with F=119.74 (p<0.001) and F=140.63 (p<0.001), respectively. This indicates that the predicted MMSE cognitive scores largely preserved the cognitive function difference pattern among the three groups (HC>FTD>AD), reflecting the gradient characteristics of cognitive decline.

[0163] Figure 7 shows box plots of the actual and predicted MMSE cognitive scores for each category, clearly reflecting the differences in predicted scores between groups, while also demonstrating the high consistency in distribution between actual and predicted scores within the same group. These results confirm that the method described in this invention can effectively encode EEG features related to cognitive states without using MMSE as an input feature, and possesses good inter-group discrimination ability. Since the BrainLat dataset does not provide cognitive scores, MMSE regression analysis was not performed on this dataset.

[0164] The dementia identification method based on spatial cross-attention fusion of the present invention has the following technical features.

[0165] 1. Cross-feature collaborative modeling framework. This invention addresses the problem that existing methods model only a single feature, making it difficult to characterize the complex nonlinear dependencies between EEG spectral features and functional connectivity structures. It proposes a cross-feature collaborative modeling framework. This framework jointly models the frequency band power and statistical characteristics reflecting local neural oscillations with the functional connectivity graph structure representing macroscopic brain region interactions in a unified feature space. Furthermore, it achieves complementary fusion of different features through a bidirectional cross-attention module (BCAB), thereby improving the ability to identify abnormal network patterns in Alzheimer's disease (AD) and frontotemporal dementia (FTD).

[0166] 2. Spatial-Structure Integrated Graph Modeling Strategy. Addressing the insufficient utilization of functional connectivity and spatial topology in existing EEG dementia classification methods, this invention proposes a graph modeling strategy using the three-dimensional spatial coordinates of electrodes as geometric constraints. By employing an equivariant graph neural network (EGNN), explicit modeling of brain region topology is achieved while maintaining the equivariance of geometric transformations such as rotation and translation. Simultaneously, by combining functional connectivity and spatial distance information in graph construction, adjacency relationships are freed from fixed threshold limitations and can adaptively adjust with training, thereby capturing dynamic interaction patterns across brain regions.

[0167] 3. Deep Synergy Between Local Feature Enhancement and Cross-Feature Fusion. Building upon the fusion of functional connectivity structure and spatial topological information, a Local Feature Enhancement (CLFE) module is introduced to strengthen local interaction features in neighboring brain regions. Simultaneously, a Bidirectional Cross-Attention (BCAB) module is combined to achieve deep interaction between spatial structural features and spectral statistical features. Through this synergistic mechanism, the model can simultaneously capture subtle functional changes in local brain regions and abnormal patterns in the global brain network, improving its sensitivity and discriminative ability to the differences between AD and FTD brain networks. This results in higher accuracy in disease classification and more robust cross-dataset generalization.

[0168] The present invention provides a brain network-based dementia identification method based on spatial cross-attention fusion, which aims to improve the accuracy and stability of EEG-based classification and diagnosis of Alzheimer's disease and frontotemporal dementia. It can achieve dynamic interactive learning of spatial graph structural features and spectral statistical features while maintaining the topological consistency of brain regions, thereby significantly improving the disease classification performance and the clinical applicability of the model.

[0169] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0170] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A data analysis method for spatial cross-attention fusion in brain-computer networks, characterized in that, The process includes the following steps: Step 1: EEG signal acquisition and preprocessing; For each time segment of the EEG signal x, extract its coherence features and differential entropy features, construct a brain network functional connectivity map with electrodes as nodes, coherence as edge weights, and differential entropy features as node features, and calculate the power spectral density ratio features between each frequency band and integrate them into a vector; Step 2: Feature extraction. Acquiring spatial structure features of EEG signals Spectral statistical characteristics of EEG signals The head surface locations of the EEG electrodes are mapped to corresponding three-dimensional coordinates. The brain network functional connectivity map and the three-dimensional spatial coordinates of the nodes are input into the graph neural network (EGNN) for feature extraction. After passing through a first multilayer perceptron, the feature extraction module (CLFE) inputs the node three-dimensional spatial coordinates encoded by a second multilayer perceptron together with the first multilayer perceptron input. The output module focuses on the spatial map structural features of local brain region interactions. Simultaneously, the global spectral feature vector g PSD The input is encoded into a third-level multilayer perceptron (MLP) to extract nonlinear interaction patterns between frequency bands, thus obtaining global statistical spectral features. Step 3: Determine the structural features of the spatial graph With spectral statistical characteristics A common input bidirectional cross-attention module is used to model cross-feature associations through a multi-head interactive attention mechanism, extracting complementary and fused features; Step 4: Spatial graph structure features With integrated interactive features After hybrid pooling and concatenation, the data are fed into a Chebyshev Kolmogorov–Arnold network classifier. Nonlinear discrimination is achieved using Chebyshev graph convolution kernels, and the classification results of Alzheimer's disease, frontotemporal dementia, and healthy controls, as well as the MMSE cognitive scale evaluation scores, are output.

2. The data analysis method for spatial cross-attention fusion in brain-computer networks according to claim 1, characterized in that, Step 1 includes the following steps: Step 101: Acquisition of EEG signals and channel standardization; standardizing the electrode channels to 19 electrode channels under the international 10-20 system; Step 102: Preprocessing of EEG signals; Step 103: Data segmentation and feature extraction; Step 104: Construction of brain network functional connectivity map and division of training set.

3. The data analysis method for spatial cross-attention fusion in brain-computer networks according to claim 2, characterized in that, In step 102, the original EEG signal is filtered using a Chebyshev bandpass filter during the preprocessing.

4. The data analysis method for spatial cross-attention fusion in brain-computer networks according to claim 3, characterized in that, In step 102, independent component analysis is used to eliminate artifact interference from the filtered raw EEG signal.

5. The data analysis method for spatial cross-attention fusion in brain-computer networks according to claim 1, characterized in that, In step 2, the two-layer equivariant graph neural network EGNN is used for feature modeling.

6. The data analysis method for spatial cross-attention fusion in brain-computer networks according to claim 1, characterized in that, In step 2, the local feature enhancement module CLFE uses a neighbor feature weighted mean aggregation method for feature enhancement.

7. The data analysis method for spatial cross-attention fusion in brain-computer networks according to claim 1, characterized in that, Step 3 includes the following steps: Step 301: Cross-feature attention weights The calculation; based on the structural features of the spatial graph. For querying, based on spectral statistical characteristics For the key value, calculate the cross-feature attention weight. ; through cross-feature attention weights The enhanced spatial features were calculated. Step 302: Reverse attention weights Calculation; based on spectral statistical characteristics For querying, based on spatial graph structural features For the key value, calculate the inverse attention weights. ; through reverse attention weights The enhanced spectral features were calculated. Step 303: Feature fusion step.

Images