Path-sensitive and semantic-unified graph learning method suitable for auxiliary diagnosis of brain diseases

Abstract

The application discloses a path-sensitive and semantic unified graph learning method and system suitable for auxiliary diagnosis of brain diseases. In view of the problems that the existing graph neural network is path-agnostic and cannot capture the propagation flow of neural signals when analyzing multi-channel neural signals, causal path signal coding is adopted, the step-by-step diffusion process of neural signals is simulated through a predefined propagation operator and a learnable propagation operator, and path-aware node features with directional propagation characteristics are extracted. In view of the problem that the layered representation semantics cannot be effectively unified due to the message transmission conflict easily caused in multi-scale brain network modeling, hierarchical brain semantic unification is adopted, the most relevant semantic dimension features between the regional graph, the global graph and the fusion representation thereof are identified and aligned to effectively resolve the message transmission conflict, and finally the deep unification of cross-scale brain activity characteristics is realized while the integrity of the brain network topology is reserved.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of artificial intelligence and medical health, and relates to a brain network analysis and disease auxiliary diagnosis method and system, specifically a path-sensitive and semantic unified graph learning method and system applicable to auxiliary diagnosis of brain diseases. Background Technology

[0002] Brain diseases such as Alzheimer's and Parkinson's are characterized by cognitive impairment, and early and accurate identification and intervention are crucial for patient recovery. Currently, the clinical diagnosis of brain diseases mainly relies on standardized assessment scales and the professional experience of clinicians. These methods depend on the subjective cooperation of the subjects, are easily affected by environmental interference, and carry the risk of misdiagnosis. The emergence of objective neural signals such as electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS) has made it possible to develop auxiliary diagnostic tools. However, neural signals are highly complex and environmentally sensitive; manual analysis is time-consuming and easily influenced by subjective biases.

[0003] Since the spatial arrangement of electrodes naturally defines a graph, graph neural networks (GNNs) have shown great potential in modeling brain network data and diagnosing brain diseases [Reference 1]. Although existing graph neural network-based brain network analysis methods have achieved some success, they still face two major challenges in processing complex multichannel neural signals: 1. Existing models often suffer from the limitation of "path agnosticness." From a neurological perspective, brain activity involves the directional flow of neural signals between different regions and channels. Relying solely on static graph structures and statistical correlations is insufficient to capture this dynamic flow of neural signal propagation [Reference 2]. The neglect of the brain's directional signal transmission mechanisms leads to the inability of existing GNN methods to distinguish between direct neural interactions and indirect information transmission, and the extracted features lack a profound description of the evolution of the brain's internal state.

[0004] 2. The "message passing conflict" problem exists in multi-scale brain network modeling. Brain neural activity can be analyzed from different scale views, such as local brain regions and global whole-brain networks [Reference 3]. Integrating these multi-scale views in graph neural networks can lead to message passing conflicts, hindering the effective unification of hierarchical representations across brain regions and the whole-brain network. This results in the model being unable to accurately capture multi-dimensional neural representations while preserving the topological integrity of the brain network.

[0005] [Literature 1] Yan Y, Zhu J, Duda M, et al. Groupinn: Grouping-basedinterpretable neural network for classification of limited, noisy brain data[C] / / Proceedings of the 25th ACM SIGKDD international conference on knowledgediscovery&data mining. 2019: 772-782. [Literature 2]Avvaru S, Peled N, Provenza NR, et al. Region-level functional and effective network analysis of human brain during cognitivetask engagement[J]. IEEE Transactions on Neural Systems and RehabilitationEngineering, 2021, 29: 1651-1660. [Literature 3]Kong W, Zhou Z, Jiang B, et al. Assessment of driving fatiguebased on intra / inter-region phase synchronization[J]. Neurocomputing, 2017,219: 474-482. Summary of the Invention To address the problems of path agnosticness in multi-channel neural signal analysis and message passing conflicts that are easily triggered in multi-scale modeling of existing graph neural networks, this invention provides a path-sensitive and semantically unified graph learning method and system suitable for auxiliary diagnosis of brain diseases.

[0006] The technical solution adopted by the method of the present invention is: a path-sensitive and semantic unified graph learning method suitable for auxiliary diagnosis of brain diseases, which adopts causal path signal encoding, simulates the gradual diffusion process of neural signals through predefined propagation operators and learnable propagation operators, thereby extracting path-aware node features with directional propagation characteristics; By employing hierarchical brain semantic unification, we effectively resolve message passing conflicts by identifying and aligning the most relevant semantic dimension features among region maps, global maps, and their fused representations. Ultimately, this achieves deep unification of cross-scale brain activity features while preserving the integrity of the brain network topology.

[0007] Preferably, the method employs causal path signal coding to extract path-aware node features with directional propagation characteristics. This is specifically implemented through the following steps: Step 1: Construct regional and global map representations based on phase consistency between EEG channels to quantify cortical functional connectivity and spatial topology; Step 2: Construct a propagation matrix, including a learnable propagation matrix and a fixed propagation matrix, responsible for the global map and regional map respectively, to capture complex propagation paths across brain regions and internal pathways; Step 3: Initialize the random feature matrix and perform an iterative diffusion process through the propagation matrix to generate the path-aware feature matrix, and integrate it with the original node attributes to distinguish between direct neural interaction and indirect information transmission.

[0008] Preferably, in step 1, the instantaneous phase of the signal is extracted by Hilbert transform, and the phase lock value is calculated as an indicator for evaluating synchronization.

[0009] in, Representing the The first electrode and the first The weights between the electrodes To observe the total number of sampling points within the window, as well as For the first The first electrode and the first Each electrode corresponds to time. Extracted instantaneous phase; Through the phase consistency matrix Building a global graph Its adjacency matrix is ,in, Total channel count; Regional mapping focusing on local neural features is constructed by eliminating connections between electrodes in different brain regions. Its adjacency matrix is ​​calculated as follows ,in It is Hadamaji. It is a binary mask matrix, if the... The and the first If the channels belong to the same functional area, then equal Otherwise, it equals 0.

[0010] Preferably, in step 2, the features of the global graph are transformed through graph convolutional network layers to extract high-level representations, which are then processed by a multi-head attention mechanism to evaluate dependencies across multiple representation subspaces; the average attention scores of all heads are calculated to obtain the learnable propagation matrix. ;in, This represents the total number of heads of attention. and The first Query and key projection of size, This represents the scaling factor determined by the dimension of the key.

[0011] Preferably, in step 2, the fixed propagation matrix is ​​derived directly from the adjacency matrix of the region graph using a normalized Laplace transform. : ; in, Represents the identity matrix. The adjacency matrix represents the region's graph. It is the angle matrix, defined as , Adjacency matrix The Middle line, number The elements of the column.

[0012] Preferably, in step 3, the random features of the propagation are generated:

[0013]

[0014] in, It is a random matrix sampled from the standard normal distribution. Represents the propagation matrix. Indicates the propagation step index, It is the total number of propagation iterations. For orthogonalization interval, This represents the QR decomposition that extracts orthogonal components to maintain numerical stability; This is the total number of channels. For the initial random matrix The number of columns; A path-aware feature matrix is ​​constructed by concatenating features from all propagation steps. The concatenated features are transformed through a learnable embedding layer, mapping multi-view information to a high-level latent space. This process is represented as follows:

[0015] in, This represents the final node representation. It is the original feature matrix of the nodes, and the operators. This represents the concatenation of eigenvectors along the channel axis. and These are the trainable projected weight matrix and the bias vector, respectively. This represents a non-linear activation function.

[0016] Preferably, the hierarchical brain semantic unification is adopted, which first uses a graph attention network to extract high-order semantic information from the regional graph and the global graph to obtain the updated node representation. and Then, the multi-view features are integrated, fusion weights are calculated, and weighted aggregation is performed to obtain the fused graph representation. ;

[0017] The fusion coefficient matrix pass It is derived that; This represents the fusion coefficient of the regional map. This represents the fusion coefficient of the global graph. and This represents two learnable projection matrices; It is a Hadama pile; Fusion diagram representation Subsequently, a topology-aware pooling layer is used to calculate the selection score of each node by combining a local voting mechanism based on neighborhood topology and a global voting mechanism based on feature semantics, thereby aggregating the results to obtain a graph-level embedding. .

[0018] Preferably, the local voting mechanism based on neighborhood topology is configured such that the structural feature matrix is... Where N is the number of channels and S is the dimension of the structural features; the structural features are... Other characteristics Includes the first with the highest gradient magnitude Indexes for each feature dimension; Local score of each node The calculation formula is:

[0019]

[0020]

[0021] in, This is a learnable similarity weight matrix. It represents the Hadamah accumulation. This represents a vector of length N consisting entirely of 1s. Given an adjacency matrix with self-loops, The corresponding degree matrix is; The global voting mechanism based on semantic features is defined as follows: Let the semantic feature matrix be... Where N is the number of channels; C is the number of semantic features, i.e. The feature indexes included are k; firstly, the semantic features are locally smoothed using graph topology, and then the global semantic score is calculated. :

[0022]

[0023] Where P is a learnable global projection vector. ; The final selection score for each node is:

[0024] Where λ is the penalty hyperparameter, d∈R N The degree of each node; And the Softmax function is used to map them into attention weights. :

[0025] The final pooling yields : .

[0027] As a preferred approach, to align semantic consistency and resolve conflicts between local and global information flows, a semantic feature selection mechanism is implemented, utilizing the focus loss function L. fc In supervised classification tasks, gradient sensitivity analysis is used to identify the most discriminative feature dimensions.

[0028]

[0029] in, Includes the first with the highest gradient magnitude Index of each feature dimension Indicates fusion graph embedding, This represents the category balancing hyperparameter. This indicates the focusing hyperparameter. This represents the predicted probability of category m. Indicates the number of categories; Calculate the neighbor jump weight matrix to quantify the structural influence of second-order neighbors:

[0030] in, This represents the neighbor hop weight matrix, which reflects the probability of signal spread in the global graph. Represents the adjacency matrix with self-loops in the global graph. Indicates and The corresponding angle matrix, This indicates row normalization.

[0031] Based on Jensen-Shannon divergence The semantic alignment loss is divided into two parts, and the calculation formula is as follows:

[0032]

[0033] in, Indicates the number of channels. The graph signal diffusion probability weights between node i and node j are represented. This represents the semantic feature vector of node i in the region graph. This represents the projected feature vector of node j in the global graph along the semantic dimension. This represents the projected feature vector of node i in the semantic dimension in the fusion graph; The final objective function is constructed as a weighted combination of classification loss and a dual semantic consistency term: In order to use hyperparameters and The importance of balancing multi-view alignment with the main binary classification task.

[0034] This invention also provides a path-sensitive and semantic unified graph learning system suitable for auxiliary diagnosis of brain diseases, comprising: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the path-sensitive and semantic unified graph learning method applicable to brain disease auxiliary diagnosis.

[0035] The present invention also provides a computer program product, including computer program instructions, which, when run on a computer, cause the computer to execute the path-sensitive and semantic unified graph learning method applicable to brain disease auxiliary diagnosis.

[0036] Compared with the prior art, the beneficial effects of the present invention include: (1) This invention proposes causal path signal encoding, overcoming the limitation of path agnosticness in graph neural networks. By initializing a random feature matrix to simulate the potential diversity of neural activity and combining fixed and learnable propagation operators, the gradual diffusion and evolution of signal flow in multi-hop brain network topology is successfully simulated. This mechanism extracts the features of path-aware nodes and fuses them with the original attributes, effectively distinguishing between direct neural interactions and indirect information transmission within the brain network, thus providing fundamental biomarkers for analyzing the internal state of the brain.

[0037] (2) This invention designs a hierarchical brain semantic unification mechanism, which effectively solves the message passing conflict problem caused by multi-scale feature fusion. Gradient sensitivity analysis identifies the most discriminative key feature dimensions, and a dual semantic consistency alignment based on Jensen-Shannon divergence is introduced between the three different views: the region map, the global map, and the fused map. This joint optimization strategy enables the model to effectively alleviate the conflict between local and global information flows while preserving the integrity of the cross-scale topological structure, thereby achieving deep unification of brain activity features across views and accurately capturing multi-dimensional neural representations. Attached Figure Description

[0038] The technical solutions of the present invention will be further illustrated below using embodiments and specific implementation methods. In addition, some accompanying drawings are used in the description of the technical solutions. Those skilled in the art can obtain other drawings and the intent of the present invention from these drawings without any creative effort.

[0039] Figure 1 This is a diagram illustrating the logical architecture of a method according to an embodiment of the present invention. Detailed Implementation

[0040] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0041] Please see Figure 1This embodiment provides a path-sensitive and semantically unified graph learning method suitable for brain disease diagnosis. Addressing the problem that existing graph neural networks cannot capture the propagation flow of neural signals when analyzing multi-channel neural signals, this embodiment employs causal path signal encoding. It simulates the gradual diffusion process of neural signals using predefined and learnable propagation operators to extract path-aware node features with directional propagation characteristics. Furthermore, addressing the issue that message passing conflicts are easily triggered in multi-scale brain network modeling, leading to ineffective unification of hierarchical representation semantics, this embodiment employs hierarchical brain semantic unification. It effectively resolves message passing conflicts by identifying and aligning the most relevant semantic dimension features among region graphs, global graphs, and their fused representations. Ultimately, it achieves deep unification of cross-scale brain activity features while preserving the topological integrity of the brain network.

[0042] In one implementation, the causal path signal coding is specifically implemented by the following steps: Step A1: Construct regional and global maps of the brain network based on phase synchronization to quantify cortical functional connectivity and spatial topology. Phase lock-in values ​​are used as the primary indicator to assess neural signal synchronicity, calculated by analyzing the instantaneous phase difference of neural signals.

[0043] in, express Connection weights between them This represents the total number of sampling points within the observation window. as well as Corresponding to time The instantaneous phase of the signal is extracted using the Hilbert transform.

[0044] This leads to the construction of a global graph. Its adjacency matrix is ,in This represents the total number of channels. Simultaneously, an area map is constructed. By excluding connections between electrodes in different anatomical regions to focus on local neural features, its adjacency matrix is ​​calculated as follows: ,in For Hadama product, It is a binary mask matrix, when the... The and the first When multiple channels belong to the same functional area, then equal .

[0045] Step A2: Construct a propagation matrix, including a learnable propagation matrix and a fixed propagation matrix, responsible for the global map and the intra-regional map, respectively. Use learnable propagation operators and fixed propagation operators to capture complex signal propagation paths across brain regions and internal channels.

[0046] Learnable propagation matrix The goal is to adaptively identify dynamic functional interactions between nodes. First, global graph features are extracted using high-order features through graph convolutional network layers:

[0047] in, For the input feature matrix, These are the node feature matrices of the input and output. It is an adjacency matrix with self-loops added. Let it be its corresponding degree matrix. It is a learnable weight transformation matrix, and It is a non-linear activation function.

[0048] The learnable propagation matrix is ​​then calculated using a multi-head attention mechanism:

[0049] in, The total number of heads that need attention. and For the first The query and key projection matrix, where $d_k$ represents the key dimension scaling factor.

[0050] At the same time, a fixed propagation matrix is ​​used. The inherent static structure of the graph within the region is represented by a normalized Laplace transform, which is directly derived from the adjacency matrix:

[0051] in, Represents the identity matrix. The adjacency matrix represents the region's graph. It is the angle matrix, defined as , Adjacency matrix The Middle line, number The elements of the column.

[0052] Step A3: Generate physiological signal flow embeddings. The spontaneous diversity of neural activity is represented by an initialized random feature matrix, and the directional signal flow in a multi-hop network is simulated through the iterative diffusion process of the propagation matrix.

[0053] The diffusion generation process is defined as follows:

[0054]

[0055] in, It is a random matrix sampled from the standard normal distribution. Represents the propagation matrix. To spread the number of steps, It is the total number of propagation iterations. For orthogonalization interval, This indicates the QR decomposition operation used to maintain numerical stability. This is the total number of channels. For the initial random matrix The number of columns; Concatenate the features of all propagation steps to construct a path-aware feature matrix:

[0056] Finally, the multi-hop signal diffusion characteristics will be... With original node attributes Concatenation, mapped to a higher-order latent space through learnable embedding layers:

[0057] in, This represents the final node representation, which contains rich propagation characteristics. and This is the trainable projective weight matrix and bias vector.

[0058] In one implementation, the hierarchical brain semantic unification is specifically implemented as follows: Step B1: Cross-scale adaptive feature fusion. A graph attention network (GAT) is used to extract high-order semantic information from the regional graph and the global graph and update the node representations:

[0059] Represents a non-linear activation function. Represents the neighborhood set of node i. The attention coefficient represents the importance of node i to its neighbor node j. The learnable weight matrix for linear transformation. Indicates the first Layer node characteristics; Refined representation of each view and Then, the fusion coefficient matrix is ​​calculated using a nonlinear gated network. The multi-view features are weighted and aggregated to obtain a fused graph representation. :

[0060] Among them, the fusion coefficient matrix pass It is deduced that, This represents the fusion coefficient of the regional map. This represents the fusion coefficient of the global graph. and This represents two learnable projection matrices; It is a Hadama pile; Fusion diagram representation Subsequently, a topology-aware pooling layer is used to calculate the selection score of each node by combining a local voting mechanism based on neighborhood topology and a global voting mechanism based on feature semantics, thereby aggregating the results to obtain a graph-level embedding. .

[0061] The local voting mechanism based on neighborhood topology is as follows: Let the structural feature matrix be Where N is the number of channels and S is the structural feature dimension, which will be discussed in the following text. In addition to the features, the learnable similarity weight matrix is: The adjacency matrix with self-loops is The corresponding degree matrix is Local score for each node The calculation formula is:

[0062]

[0063]

[0064] in It represents the Hadamah accumulation. This represents a vector of length N consisting entirely of 1s.

[0065] The global voting mechanism based on feature semantics is as follows: Let the semantic feature matrix be Where N is the number of channels and C is the number of semantic features, as described below. The number of feature indices contained is k, and the learnable global projection vector is First, local smoothing of semantic features is performed using graph topology, followed by calculation of the global semantic score. :

[0066]

[0067] The final selection score for each node is:

[0068] Where λ is the penalty hyperparameter, d∈R NThe degree of each node; Node selection score Finally, the data is aggregated into graph-level embeddings using a multilayer perceptron (MLP) and the predicted probabilities are output.

[0069] in, Indicates the first The fused feature vector of each node; Step B2: Semantic Consistency Alignment. To resolve message passing conflicts between local and global information flows, a semantic feature selection mechanism based on focus loss is implemented for gradient sensitivity analysis.

[0070]

[0071] in, It includes the most discriminative top performers in classification tasks. A feature index serving as a semantic dimension Indicates fusion graph embedding, This represents the category balancing hyperparameter. This indicates the focusing hyperparameter. This represents the predicted probability of category m. Indicates the number of categories; Then the neighbor jump weight matrix is ​​calculated. To quantify the structural influence of second-order neighbors in the graph and the probability of signal spread:

[0072] in, Represents the adjacency matrix with self-loops in the global graph. Indicates and The corresponding angle matrix, This indicates row normalization.

[0073] To narrow the feature distance of multi-scale neural representations in the latent space, Jensen-Shannon divergence is used. Constructing dual semantic alignment loss and .in Semantic features used to zoom in on region graphs and global graphs Force alignment of the two original single-view features with the fused representation:

[0074]

[0075] in, Indicates the number of channels. The graph signal diffusion probability weights between node i and node j are represented. This represents the semantic feature vector of node i in the region graph. This represents the projected feature vector of node j in the global graph along the semantic dimension. This represents the projected feature vector of node i in the semantic dimension in the fusion graph.

[0076] Therefore, a weighted optimization overall objective function is constructed. With joint training networks:

[0077] in and To balance the hyperparameters.

[0078] The invention will be further illustrated by specific experiments below.

[0079] To verify the effectiveness of this invention, this embodiment used three datasets from two different imaging modalities: electroencephalography (EEG) and functional near-infrared spectroscopy (fNIRS). These included two publicly available EEG datasets, ADFTD and APAVA, and a proprietary fNIRS dataset, M-fNIRS. The ADFTD dataset contained 19 channels of EEG data from 88 participants, including Alzheimer's disease patients, frontotemporal dementia patients, and healthy controls. The APAVA dataset contained 16 channels of EEG data from 23 participants, including Alzheimer's disease patients and healthy controls. The M-fNIRS dataset contained 53 channels of fNIRS data from 1090 participants, covering healthy controls and patients with four mental illnesses, including schizophrenia and bipolar disorder. In this embodiment, a subject-independent binary classification task was used on all datasets, and the data was randomly divided into training, validation, and test sets in an 8:1:1 ratio to ensure that samples from the same subject remained in the same partition, preventing data leakage.

[0080] This embodiment compares the method of the present invention with several state-of-the-art baseline methods for brain time series and graph learning: (1) KNN and Random Forest, two traditional machine learning models; (2) STAGIN, a dynamic graph representation learning framework using timestamp-based spatiotemporal attention; (3) MDGL, a multi-scale Transformer-based graph neural network for dynamic brain connectome learning; (4) ALTER, a long-range perceptual brain graph Transformer using adaptive bias random walks; (5) LaBraM, a large model for learning EEG representations through vector quantization neural networks; (6) EEGPT, a self-supervised Transformer for general representations of EEG signals; (7) Medformer, a multi-granularity Patching Transformer for medical time series classification; (8) DMSGL, a spatiotemporal multi-view graph learning model for EEG emotion recognition; (9) EvoBrain, a dynamic graph neural network for modeling evolutionary brain networks; and (10) TarDiff, a goal-oriented time series generation diffusion model. The experimental evaluation metrics used were the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPRC). The experimental results are shown in Table 1 below.

[0081] Table 1

[0082] This experiment demonstrates the classification and diagnostic performance of various methods across three different neural signal modalities and multiple brain disease diagnostic scenarios, including cross-disease classification experiments and robustness validation for specific modalities (EEG and fNIRS). The experimental results are shown in the table above, summarizing the final average performance of each method on the AUROC and AUPRC metrics. The results show that the method of this invention significantly outperforms all baseline methods on all three core datasets. Notably, this invention exhibits extremely strong robustness in dealing with the dynamics and structural heterogeneity of multi-channel neural signals, surpassing other cutting-edge models, fully validating the practical value and clinical application potential of this invention in the objective diagnosis of brain diseases.

[0083] This invention innovates in two aspects: accurately capturing the causal path of neural signal flow and achieving semantic unification across multiple scales. Regarding causal path capture, inspired by the neuroscience mechanisms of brain networks, this invention employs causal path signal encoding, utilizing predefined and learnable propagation operators to simulate the gradual diffusion process of neural activity. This facilitates a precise characterization of the directed information transmission mechanism in the cerebral cortex and successfully extracts path-aware node features with directional propagation characteristics, overcoming the path-awareness limitations of existing models. In terms of cross-scale semantic unification, this invention designs a hierarchical brain semantic unification module for modeling multi-scale brain activity. It successfully aligns and unifies the most relevant features between the region map and the global map using a combination of adaptive feature fusion and semantic dimension alignment, effectively resolving message passing conflicts caused by multi-scale structural fusion. Furthermore, this invention performs well in various neural signal acquisition modalities such as EEG and functional near-infrared spectroscopy (fNIRS), as well as in recognition tasks across various mental illnesses such as Alzheimer's disease. It can effectively handle various types of clinical scenarios, enhancing the robustness and generalization ability of the model in practical applications. This invention is easily integrated into existing graph learning frameworks, providing an effective method to address the limitations of objective automated diagnosis of brain diseases caused by unknown neural signal transmission paths and multi-scale network message passing conflicts.

[0084] It should be understood that the embodiments described above are only some, not all, of the embodiments of the present invention. Furthermore, the technical features of the various embodiments or individual embodiments provided by the present invention can be arbitrarily combined to form feasible technical solutions. Such combinations are not constrained by the order of steps and / or structural composition patterns, but must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.

[0085] It should be understood that the above description of the preferred embodiments is quite detailed and should not be construed as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, all of which fall within the scope of protection of this invention.

Claims

1. A path-sensitive and semantically unified graph learning method suitable for auxiliary diagnosis of brain diseases, characterized in that: Causal path signal encoding is employed, and the gradual diffusion process of neural signals is simulated through predefined propagation operators and learnable propagation operators, thereby extracting path-aware node features with directional propagation characteristics. By employing hierarchical brain semantic unification, we effectively resolve message passing conflicts by identifying and aligning the most relevant semantic dimension features among region maps, global maps, and their fused representations. Ultimately, this achieves deep unification of cross-scale brain activity features while preserving the integrity of the brain network topology.

2. The path-sensitive and semantically unified graph learning method for auxiliary diagnosis of brain diseases according to claim 1, characterized in that: The method employs causal path signal coding to extract path-aware node features with directional propagation characteristics. The specific implementation includes the following steps: Step 1: Construct regional and global map representations based on phase consistency between EEG channels to quantify cortical functional connectivity and spatial topology; Step 2: Construct a propagation matrix, including a learnable propagation matrix and a fixed propagation matrix, responsible for the global map and regional map respectively, to capture complex propagation paths across brain regions and internal pathways; Step 3: Initialize the random feature matrix and perform an iterative diffusion process through the propagation matrix to generate the path-aware feature matrix, and integrate it with the original node attributes to distinguish between direct neural interaction and indirect information transmission.

3. The path-sensitive and semantic unified graph learning method for auxiliary diagnosis of brain diseases according to claim 2, characterized in that: In step 1, the instantaneous phase of the signal is extracted by Hilbert transform, and the phase lock value is calculated as an indicator for evaluating synchronization. in, Representing the The first electrode and the first The weights between the electrodes To observe the total number of sampling points within the window, as well as For the first The first electrode and the first Each electrode corresponds to time. Extracted instantaneous phase; Through the phase consistency matrix Building a global graph Its adjacency matrix is ,in, Total channel count; Regional mapping focusing on local neural features is constructed by eliminating connections between electrodes in different brain regions. Its adjacency matrix is ​​calculated as follows ,in It is Hadamaji. It is a binary mask matrix, if the... The and the first If the channels belong to the same functional area, then equal Otherwise, it equals 0.

4. The path-sensitive and semantically unified graph learning method for auxiliary diagnosis of brain diseases according to claim 2, characterized in that: In step 2, the features of the global graph are transformed through graph convolutional network layers to extract high-level representations, which are then processed by a multi-head attention mechanism to evaluate dependencies across multiple representation subspaces; the average attention scores of all heads are calculated to obtain the learnable propagation matrix. ;in, This indicates the total number of heads of attention. and The first Query and key projection of size, This represents the scaling factor determined by the dimension of the key.

5. The path-sensitive and semantically unified graph learning method for auxiliary diagnosis of brain diseases according to claim 2, characterized in that: In step 2, the fixed propagation matrix is ​​directly derived from the adjacency matrix of the region graph using a normalized Laplace transform. : in, Represents the identity matrix. The adjacency matrix represents the region's graph. It is the angle matrix, defined as ,in Adjacency matrix The Middle line, number The elements of the column.

6. The path-sensitive and semantically unified graph learning method for auxiliary diagnosis of brain diseases according to claim 2, characterized in that: In step 3, the random features of the propagation are generated: ； ； in, It is a random matrix sampled from the standard normal distribution. Represents the propagation matrix. Indicates the propagation step index, It is the total number of propagation iterations. For orthogonalization interval, This represents the QR decomposition that extracts orthogonal components to maintain numerical stability; This is the total number of channels. For the initial random matrix The number of columns; A path-aware feature matrix is ​​constructed by concatenating features from all propagation steps. The concatenated features are transformed through a learnable embedding layer, mapping multi-view information to a high-level latent space. This process is represented as follows: ;in, This represents the final node representation. It is the original feature matrix of the nodes, and the operators. This represents the concatenation of eigenvectors along the channel axis. and These are the trainable projected weight matrix and the bias vector, respectively. This represents a non-linear activation function.

7. The path-sensitive and semantically unified graph learning method for auxiliary diagnosis of brain diseases according to claim 1, characterized in that: The proposed hierarchical brain semantic unification first utilizes a graph attention network. Extract high-order semantic information from the regional and global graphs to obtain updated node representations. and ;in, Represents a non-linear activation function. Represents the neighborhood set of node i. The attention coefficient represents the importance of node i to its neighbor node j. The learnable weight matrix for linear transformation. Indicates the first Layer node features are identified; then, multi-view features are integrated, fusion weights are calculated, and weighted aggregation is performed to obtain a fused graph representation. ; ； Among them, the fusion coefficient matrix pass It is deduced that, This represents the fusion coefficient of the regional map. This represents the fusion coefficient of the global graph. and This represents two learnable projection matrices; It is a Hadama pile; Fusion diagram representation Subsequently, a topology-aware pooling layer is used to calculate the selection score of each node by combining a local voting mechanism based on neighborhood topology and a global voting mechanism based on feature semantics, thereby aggregating the results to obtain a graph-level embedding. .

8. The path-sensitive and semantically unified graph learning method for auxiliary diagnosis of brain diseases according to claim 7, characterized in that: The local voting mechanism based on neighborhood topology is defined by the structural feature matrix as follows: , where N is the number of channels and S is the structural feature dimension; Structural features are Other characteristics, Includes the first with the highest gradient magnitude Indexes for each feature dimension; Local score of each node The calculation formula is: in, This is a learnable similarity weight matrix. It represents the Hadamah accumulation. This represents a vector of length N consisting entirely of 1s. Given an adjacency matrix with self-loops, The corresponding degree matrix is; The global voting mechanism based on semantic features is defined as follows: Let the semantic feature matrix be... Where N is the number of channels; C is the number of semantic features, i.e. The feature indexes included are k; firstly, the semantic features are locally smoothed using graph topology, and then the global semantic score is calculated. : Where P is a learnable global projection vector. ; The final selection score for each node is: Where λ is the penalty hyperparameter, d∈R N The degree of each node; And the Softmax function is used to map them into attention weights. : The final pooling yields : 。 9. The path-sensitive and semantically unified graph learning method for auxiliary diagnosis of brain diseases according to claim 7, characterized in that: To align semantic consistency and resolve conflicts between local and global information flows, a semantic feature selection mechanism is implemented, utilizing a focus loss function. In supervised classification tasks, gradient sensitivity analysis is used to identify the most discriminative feature dimensions. in, Includes the first with the highest gradient magnitude Index of each feature dimension Indicates fusion graph embedding, This represents the category balancing hyperparameter. This indicates the focusing hyperparameter. This represents the predicted probability of category m. Indicates the number of categories; Calculate the neighbor jump weight matrix to quantify the structural influence of second-order neighbors: ； in, This represents the neighbor hop weight matrix, which reflects the probability of signal spread in the global graph. Represents the adjacency matrix with self-loops in the global graph. Indicates and The corresponding angle matrix, Indicates row normalization; Based on Jensen-Shannon divergence Semantic alignment loss is divided into two parts. and ; in, Indicates the number of channels. The graph signal diffusion probability weights between node i and node j are represented. This represents the semantic projection feature vector of node i in the graph within the region. This represents the projected feature vector of node j in the global graph along the semantic dimension. This represents the projected feature vector of node i in the semantic dimension in the fusion graph; The final objective function is a weighted combination of classification loss and the dual semantic consistency term. In order to use hyperparameters and The importance of balancing multi-view alignment with the main binary classification task.

10. A path-sensitive and semantically unified graph learning system suitable for auxiliary diagnosis of brain diseases, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the path-sensitive and semantic unified graph learning method for brain disease auxiliary diagnosis as described in any one of claims 1 to 9.

Images