Hierarchical functional alignment hypergraph learning method and system for auxiliary diagnosis of brain diseases

Abstract

The application discloses a hierarchical function alignment hypergraph learning method and system for auxiliary diagnosis of brain diseases. In view of the influence of high-order interaction level heterogeneity of changes caused by brain diseases and presented among different anatomical levels, a hierarchical high-order knowledge capturing mechanism is adopted to jointly construct pair graphs and hypergraph relationships on multiple anatomical levels, so that the model can cooperatively learn and effectively capture high-order interaction knowledge of a specific level, and the inappropriate modeling caused by multi-level rough learning is avoided. In view of the influence of functional role level heterogeneity of the same brain element presented under different observation levels, the functional role of a specific level is evaluated through hierarchical functional role alignment combined with signal variability and structural characteristics, and the functional role of the brain element is aligned among multiple anatomical levels, so as to solve the knowledge conflict problem caused by inconsistent functional roles, thereby improving the generalization ability of the graph neural network across levels and realizing more accurate diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of artificial intelligence and medical health, and relates to a hierarchical functional alignment hypergraph learning method and system for auxiliary diagnosis of brain diseases. Specifically, it relates to a hierarchical functional alignment hypergraph learning method and system for auxiliary diagnosis of brain diseases based on graph neural networks (GNN) and hypergraph learning of electroencephalogram (EEG) or functional near-infrared spectroscopy (fNIRS) signals. Background Technology

[0002] Brain diseases, including mental disorders and neurodegenerative diseases, have become a major public health challenge worldwide, urgently requiring effective auxiliary diagnostic solutions. Modern neuroscience research regards these diseases as abnormal disorders of brain network function. Therefore, constructing brain functional networks and performing intelligent analysis using multi-channel neurophysiological signals such as electroencephalogram (EEG) and functional near-infrared spectroscopy (fNIRS) has become an important means of assisting medical diagnosis ([Reference 1]).

[0003] Graph neural networks (GNNs) have shown great potential in the field of brain network modeling due to their ability to capture complex dependencies and non-Euclidean structure data. Existing brain map learning methods typically treat brain elements such as sensor channels as graph nodes and learn pathological features by aggregating neighborhood information ([Reference 2][Reference 3]).

[0004] However, GNNs have the following problems when learning from brain networks: 1. Insufficient Capture of Higher-Order Interaction Knowledge: Neural changes caused by brain diseases exhibit significant heterogeneity across different anatomical levels. Existing methods often model only at a single level or employ simple multi-level feature stacking. This coarse learning approach fails to effectively capture the higher-order interaction knowledge specific to particular anatomical levels, resulting in inaccurate modeling of pathological features and difficulty in addressing complex brain functional changes.

[0005] 2. Cross-level functional role conflict: In neurophysiology, the same brain element often plays different functional roles under different observation perspectives and anatomical levels. Existing models, when learning in multi-level parallel processing, typically ignore this hierarchical heterogeneity of functional roles. Due to the lack of alignment mechanisms, the feature representations of the same brain element at different levels may exhibit knowledge conflicts, severely impairing the cross-level generalization ability of graph neural networks and limiting diagnostic accuracy.

[0006] Therefore, how to construct a graph learning architecture that can collaboratively capture multi-level high-order interactive knowledge and effectively align the functional roles of brain elements at different levels to improve the accuracy and reliability of brain disease diagnosis is a technical problem that urgently needs to be solved in the field of brain network intelligent assisted diagnosis.

[0007] [Literature 1]Gao Y, Fu [Literature 2] Li G, Duda M, Zhang [Literature 3] Roy Y, Banville H, Albuquerque I, et al. Deep learning-basedelectroencephalography analysis: a systematic review[J]. Journal of neural engineering, 2019, 16(5): 051001. Summary of the Invention To address the negative impacts of the high-order interaction hierarchy heterogeneity of changes caused by brain diseases across different anatomical levels, and the hierarchical heterogeneity of functional roles exhibited by the same brain element at different observation levels, this invention provides a hierarchical functional alignment hypergraph learning method and system for auxiliary diagnosis of brain diseases.

[0008] The technical solution adopted by the method of the present invention is: a hierarchical functional alignment hypergraph learning method for auxiliary diagnosis of brain diseases, which adopts a hierarchical high-order knowledge capture mechanism, jointly constructs pair graphs and hypergraph relationships at multiple anatomical levels, so that the graph neural network model can learn collaboratively and effectively capture high-order interactive knowledge at specific levels; By aligning functional roles in a hierarchical manner, combining signal variability and structural features to assess functional roles at specific levels, and aligning the functional roles of brain elements across multiple anatomical levels, the generalization ability of graph neural network models across levels is improved, thereby achieving more accurate assisted diagnosis.

[0009] Preferably, a hierarchical high-order knowledge capture mechanism is adopted, which jointly constructs pairwise graphs and hypergraphs at multiple anatomical levels, enabling the graph neural network model to learn collaboratively and effectively capture high-order interaction knowledge at specific levels; the specific implementation includes the following sub-steps: Step A1: Extract the instantaneous phase of the multi-channel neurophysiological signals, calculate the phase lock value (PLV) between channels, and construct the PLV similarity matrix S; Step A2: Jointly construct graph structures at three anatomical levels, including: a channel-level paired graph based on direct functional interactions, a channel-level hypergraph that captures higher-order interactions, and a brain region-level hypergraph that aggregates channels within anatomical regions and models cross-regional interactions; Step A3: Perform parallel message passing on each level view to learn the node representation of a specific level, and finally retain and splice the representations from different views, which are then processed by a classifier for the final brain disease auxiliary diagnosis.

[0010] Preferably, in step A1, the multi-channel neurophysiological signal matrix is ​​extracted using Hilbert transform. Instantaneous phase of each channel ,in for Analytical form, Represents the Hilbert transform. and For the channel, Indicates the time index of channel i The signal It represents the real number field. It is the number of time samples. It is the number of channels; calculate the number of channels. and Phase lock value between:

[0011] Aggregating all channel pairs yields the PLV similarity matrix. ,in .

[0012] Preferably, in step A2, the normalized propagation operator of the channel-level paired graph is obtained by symmetric normalization of the similarity matrix S. ,in It is a diagonal degree matrix; By analyzing each channel node Select the one with the highest PLV similarity. Each neighbor defines a superedge. Export the corresponding correlation matrix Thus, the normalized propagation operator of the channel-level hypergraph is obtained. ; ; By integrating specific brain regions Create internal hyperedges for all channels. Calculate the region-level similarity matrix And based on this choice The most similar neighboring brain regions are used to construct pairs of cross-regional hyperborders. Merge all region-level hyperedges to derive the correlation matrix. The normalized propagation operator of the brain region-level hypergraph is obtained. ;in, Indicates a specific brain region All internal channels are combined; ;in and These are the node degree and hyperedge degree matrices, respectively. It is the hyperedge weight matrix.

[0013] Preferably, in step A3, an initial node feature matrix is ​​given. In each view The above uses a graph neural network to perform parallel message passing to obtain the node embedding matrix. ,in For weight matrices shared across views, For activation function, For view exist The node embedding matrix of the last layer; the node embedding matrix of the last layer. Concatenate and flatten the data along the node dimension to obtain a unified vector. Predicting class probabilities using a multilayer perceptron classifier And utilize cross-entropy loss Optimization was carried out, including It is the one-hot encoding of the real label, and C represents the total number of categories.

[0014] As a preliminary step, the hierarchical functional role alignment, which combines signal variability and structural features, is used to evaluate functional roles at specific levels and aligns the functional roles of brain elements across multiple anatomical levels. The specific implementation includes the following sub-steps: Step B1: Calculate the graph / hypergraph Laplacian matrix for each level view and construct a fractional Laplacian operator to emphasize long-range dependencies; Step B2: Combine the learned node representations to calculate the Fisher-inspired diffusion sensitivity score, which reflects the degree of energy variation of node features under fractional-order diffusion. Step B3: Combining discrete geometry knowledge, use Forman-Ricci curvature quantification to determine the dispersion and concentration characteristics of the local composite structure of the neighborhood of nodes in each view; Step B4: Combine diffusion sensitivity with Forman-Ricci curvature to calculate the role index of each node, characterizing the functional role of the node in a specific view; Step B5: Achieve consistent alignment across levels by calculating the absolute difference in node role indices between view pairs to minimize functional role differences between levels.

[0015] As a preliminary step, in step B1, for each view The symmetric normalized Laplace matrix is And perform spectral decomposition on it. ;in, For view The normalization propagation operator; In step B2, node diffusion sensitivity score ;in It is a fractional exponent; Embed the matrix for the nodes of the last layer; In step B3, node In view The Forman-Ricci curvature in the value is the mean curvature of the associated hyperedge at that node. ;in, Indicates in view In and nodes Associative hyperedge set, Represents a view Middle node and super edge Node-hyperedge curvature between; In step B4, node In view Role Index ;in This is the curvature scaling factor.

[0016] Preferably, the overall objective function for training the graph neural network model is: , To control the balance hyperparameter of alignment regularization intensity; Among them, cross-entropy loss , It is the one-hot encoding of the real label, where C represents the total number of categories; Alignment loss ,in Represents all different view index pairs The set, where K represents the total number of channels. Represents a view Middle node Role index.

[0017] The technical solution adopted by the system of this invention is: a hierarchical functional alignment hypergraph learning system 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 hierarchical functional alignment hypergraph learning method for brain disease-assisted diagnosis.

[0018] The present invention also provides a hierarchical functional alignment hypergraph learning product for auxiliary diagnosis of brain diseases, including computer program instructions, which, when run on a computer, cause the computer to execute the hierarchical functional alignment hypergraph learning method for auxiliary diagnosis of brain diseases.

[0019] Compared with the prior art, the beneficial effects of the present invention include: (1) This invention reveals and solves for the first time the challenge of heterogeneity in high-order interaction levels among different anatomical levels of pathological changes in brain diseases. By designing a hierarchical high-order knowledge capture mechanism, a unified source of functional connectivity is mapped into three structures: channel-level pairwise graphs, channel-level hypergraphs, and brain region-level hypergraphs. This design enables the model to collaboratively and precisely learn multi-level pairwise interactions and high-order interaction features, overcoming the shortcomings of previous methods that folded multi-level information into a single view or modeled too coarsely, and greatly enhancing the ability to capture hierarchical specific interaction patterns.

[0020] (2) This invention creatively proposes a hierarchical functional role alignment mechanism to address the hierarchical heterogeneity of functional roles exhibited by the same brain element in interactions at different levels. Instead of directly and crudely aligning node embedding features, this method accurately assesses the connectivity functional role of nodes by combining the signal diffusion sensitivity obtained from fractional-order Laplace operations with the Forman-Ricci curvature reflecting geometric properties. Furthermore, the alignment mechanism forcibly reduces differences in role interpretation between levels, alleviating knowledge conflicts and endowing the network with strong cross-level generalization consistency.

[0021] (3) The method of the present invention is scientifically designed and easy to integrate, and can provide effective collaborative learning supervision when capturing brain pathological connectivity. Experiments based on multiple real electroencephalogram (EEG) and functional near-infrared spectroscopy (fNIRS) datasets show that the present invention is significantly more accurate than existing graph- and hypergraph-based baseline methods when handling brain disease auxiliary diagnosis tasks, and has significant application potential and practical value for medical auxiliary diagnosis. Attached Figure Description

[0022] The present invention provides a detailed description of the technical solution through the following specific embodiments and their implementation methods. For ease of understanding, relevant illustrations are provided for illustrative purposes. Those skilled in the art, upon understanding the design concept of the present invention, can derive other structural diagrams or implementation schemes from these illustrations without any creative effort.

[0023] Figure 1 A schematic diagram illustrating the method principle of an embodiment of the present invention is shown. Detailed Implementation

[0024] 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.

[0025] Please see Figure 1 This embodiment provides a hierarchical functional alignment hypergraph learning method suitable for brain disease diagnosis. Addressing the impact of high-order interaction heterogeneity in brain disease-induced changes across different anatomical levels, a hierarchical high-order knowledge capture mechanism is employed. Pair graphs and hypergraph relationships are jointly constructed across multiple anatomical levels, enabling the graph neural network model to collaboratively learn and effectively capture high-order interaction knowledge at specific levels, avoiding unsuitable modeling caused by coarse multi-level learning. Furthermore, addressing the impact of functional role heterogeneity of the same brain element at different observation levels, hierarchical functional role alignment is used, combining signal variability and structural features to evaluate functional roles at specific levels. The functional roles of brain elements are aligned across multiple anatomical levels, resolving knowledge conflicts caused by inconsistent functional roles, thereby improving the cross-level generalization ability of the graph neural network model and achieving more accurate diagnosis.

[0026] In one implementation, a hierarchical high-order knowledge capture mechanism is employed to jointly construct pairwise graph and hypergraph relationships at multiple anatomical levels, enabling the graph neural network model to collaboratively learn and effectively capture high-order interaction knowledge at specific levels; the specific implementation includes the following sub-steps: Step A1: Signal preprocessing and phase-based connectivity matrix construction Consider treating the multichannel neurophysiological recordings from an experiment as a matrix. ,in It is the number of time samples. This refers to the number of channels. First, the instantaneous phase of each channel is extracted using the analytic signal obtained through the Hilbert transform. Let... Indicates the time index of channel i The signal, whose analytical form is: ,in This represents the Hilbert transform. The instantaneous phase is... Subsequently, the calculation channel... and Phase lock value (PLV) between: ; This equation measures how well the phase difference remains consistent over time, producing a The similarity score is calculated within the specified range. All channel pairs are then combined to obtain the PLV similarity matrix. Its elements This matrix serves as the source of connectivity for the subsequent three levels.

[0027] Step A2: Based on the PLV matrix, jointly construct three anatomical hierarchy maps and a hypergraph; (1) Channel-level Paired Graph: A global channel-level propagation operator is constructed based on PLV similarity to capture paired functional relationships. Symmetric normalization is applied to the PLV similarity matrix: ; in For the angle matrix, It is a normalized propagation operator.

[0028] (2) Channel-level hypergraph: Constructing a channel-level hypergraph that forms higher-order relationships. For each node... Use PLV to select its Find the nearest neighbors and define a superedge:

[0029] This generates a hyperedge set. And obtain the normalized channel-level hypergraph propagation operator. .

[0030] (3) Brain region-level hypermap: Forces cohesion within regions by creating a hyperedge for each region. By calculating the average PLV of all channel pairs across regions, channel-level PLV is elevated to region-level similarity. Identify each region based on region-level similarity. The most similar neighborhoods are used to construct pairs of cross-regional superedges. Finally, a normalized brain region-level hypergraph propagation operator is obtained. .in, Indicates a specific brain region All internal channels are combined; The above view The corresponding normalized propagation operator formula is uniformly expressed as: ; in and These are the node degree and hyperedge degree matrices, respectively. It is the hyperedge weight matrix.

[0031] Step A3: Parallel message passing and classification; Before applying the graph neural network model, the PLV matrix was directly used as the node feature source, and the initial node feature matrix was... Message passing is performed in parallel on all views. exist The node embedding matrix of the layer is updated as follows: ; Where L represents the total number of layers; Embed the nodes of the last layer into the matrix Concatenate and flatten the data along the node dimension to obtain a unified vector. ; Predicting class probabilities using a multilayer perceptron classifier And utilize cross-entropy loss Optimization was carried out, including It is a unique hot encoding of the real label.

[0032] In one implementation, the hierarchical functional role alignment module assesses functional roles at specific levels by combining signal variability and structural features, and aligns the functional roles of brain elements across multiple anatomical levels; specifically, this includes the following sub-steps: Step B1: Calculate the graph / hypergraph Laplacian matrix for each level view and construct a fractional Laplacian operator to emphasize long-range dependencies; For each view Calculate the Laplacian matrix of the symmetric normalized hypergraph. A fractional-order Laplace matrix is ​​constructed by performing a full-spectrum decomposition on the Laplace matrix. This emphasizes long-range dependence.

[0033] Step B2: Combine the learned node representations to calculate the Fisher-inspired diffusion sensitivity score, which reflects the degree of energy variation of node features under fractional-order diffusion. For each node Calculate the diffusion sensitivity score based on Fisher's inspiration: ; This score measures the energy of a node representation under fractional-order diffusion, reflecting the degree of change in node features relative to nonlocal structure operators.

[0034] Step B3: Combining discrete geometry knowledge, use Forman-Ricci curvature quantification to determine the dispersion and concentration characteristics of the local composite structure of the neighborhood of nodes in each view; Incorporating Forman-Ricci curvature into a discrete geometric perspective. Computing node-hyperedge curvature. And calculate the average curvature scalar of the incident hyperedge at that node. ;in, Indicates in view In and nodes The set of associated hyperedges; Step B4: Combine diffusion sensitivity with geometric curvature Forman-Ricci curvature to calculate the role index of each node, characterizing the functional role of the node in a specific view; Define nodes by combining sensitivity and geometric properties. In view Functional Role Index: ; in This is the curvature scaling factor; This index characterizes the functional role of a node in the view by modulating Fisher diffusion sensitivity through curvature-based exponential decay.

[0035] Step B5: Achieve consistent alignment across levels by calculating the absolute difference in node role indices between view pairs to minimize functional role differences between levels.

[0036] Cross-hierarchical functional role alignment: To align roles across views, pairwise difference loss is calculated: ; in A collection of all distinct view index pairs. Indicates in view In and nodes The set of associated hyperedges. The overall training objective combines classification loss and alignment loss: ,in To balance the hyperparameters.

[0037] To verify the effectiveness of this invention, this embodiment evaluated four brain time-series datasets under a binary classification experimental setup, including three publicly available electroencephalogram (EEG) datasets: ADFTD, APAVA, and TDBrain, and a proprietary functional near-infrared spectroscopy (fNIRS) dataset, M-fNIRS. These datasets differ in imaging modalities, number of channels, and number of subjects, providing a comprehensive testing platform for evaluating the effectiveness of the method across different brain signal features. Regarding evaluation metrics, this experiment employed two threshold-free metrics: Area Under the Receiver Operating Characteristic (AUROC) and Area Under the Precision-Recall (AUPRC). This experiment compared the method of this invention with traditional machine learning methods KNN and Random Forest, as well as state-of-the-art baseline methods for brain time-series signal learning and brain mapping, including: GroupINN, IGS, EvoBrain, Medformer, DMSGL, FSTA-EC, TarDiff, LaBraM, and EEGPT.

[0038] Table 1

[0039] This experiment demonstrates the performance of various graph learning and time series learning methods in the diagnosis of brain diseases on four different datasets. The experimental results are shown in Table 1, indicating that the method of this invention outperforms all other state-of-the-art baseline methods in terms of AUROC and AUPRC metrics in the vast majority of experimental settings.

[0040] This invention innovates primarily in two aspects: Addressing the diagnostic challenges posed by the heterogeneity of hierarchical anatomical structures, it achieves system synergistic optimization by integrating a hierarchical high-order knowledge capture mechanism with a functional role alignment module. First, by jointly constructing channel-level paired graphs, channel-level hypergraphs, and brain region-level hypergraphs, this invention can synergistically capture multi-granularity high-order interaction features at different anatomical levels, overcoming the limitation of a single view being insufficient for modeling complex brain network changes. Secondly, by utilizing fractional-order diffusion sensitivity and discrete geometric curvature features to quantify the functional roles of nodes at each level and implementing cross-level alignment, it effectively alleviates knowledge conflicts caused by inconsistent functional roles of the same brain element at different observation levels. These two mechanisms complement each other; the former extracts diverse structural patterns, while the latter ensures consistency and generalization performance across multiple views, jointly significantly improving the model's accuracy in modeling and diagnosing complex pathological features of brain diseases.

[0041] 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.

[0042] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered 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, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A hierarchical functional alignment hypergraph learning method for auxiliary diagnosis of brain diseases, characterized in that: A hierarchical high-order knowledge capture mechanism is adopted to jointly construct pair graph and hypergraph relationships at multiple anatomical levels, enabling graph neural network models to learn collaboratively and effectively capture high-order interaction knowledge at specific levels; By aligning functional roles in a hierarchical manner, combining signal variability and structural features to assess functional roles at specific levels, and aligning the functional roles of brain elements across multiple anatomical levels, the generalization ability of graph neural network models across levels is improved, thereby achieving more accurate assisted diagnosis.

2. The hierarchical functional alignment hypergraph learning method for auxiliary diagnosis of brain diseases according to claim 1, characterized in that: The method employs a hierarchical high-order knowledge capture mechanism, jointly constructing pairwise graphs and hypergraph relationships at multiple anatomical levels, enabling the graph neural network model to collaboratively learn and effectively capture high-order interaction knowledge at specific levels; the specific implementation includes the following sub-steps: Step A1: Extract the instantaneous phase of the multi-channel neurophysiological signals, calculate the phase lock value (PLV) between channels, and construct the PLV similarity matrix S; Step A2: Jointly construct graph structures at three anatomical levels, including: a channel-level paired graph based on direct functional interactions, a channel-level hypergraph that captures higher-order interactions, and a brain region-level hypergraph that aggregates channels within anatomical regions and models cross-regional interactions; Step A3: Perform parallel message passing on each level view to learn the node representation of a specific level, and finally retain and splice the representations from different views, which are then processed by a classifier for the final brain disease auxiliary diagnosis.

3. The hierarchical functional alignment hypergraph learning method for auxiliary diagnosis of brain diseases according to claim 2, characterized in that: In step A1, the multi-channel neurophysiological signal matrix is ​​extracted using Hilbert transform. Instantaneous phase of each channel ,in for Analytical form, Represents the Hilbert transform. and For the channel, Indicates the time index of channel i The signal It represents the real number field. It is the number of time samples. It is the number of channels; calculate the number of channels. and Phase lock value between: Aggregating all channel pairs yields the PLV similarity matrix. ,in .

4. The hierarchical functional alignment hypergraph learning method for auxiliary diagnosis of brain diseases according to claim 3, characterized in that: In step A2, the normalized propagation operator of the channel-level paired graph is obtained by symmetric normalization of the similarity matrix S. ,in It is a diagonal degree matrix; By analyzing each channel node Select the one with the highest PLV similarity. Each neighbor defines a superedge. Export the corresponding correlation matrix Thus, the normalized propagation operator of the channel-level hypergraph is obtained. ; ; By integrating specific brain regions Create internal hyperedges for all channels. Calculate the region-level similarity matrix and based on this choice The most similar neighboring brain regions are used to construct pairs of cross-regional hyperborders. Merge all region-level hyperedges to derive the correlation matrix. The normalized propagation operator of the brain region-level hypergraph is obtained. ;in, Indicates a specific brain region All internal channels are combined; , , and These are the node degree and hyperedge degree matrices, respectively. It is the hyperedge weight matrix.

5. The hierarchical functional alignment hypergraph learning method for auxiliary diagnosis of brain diseases according to claim 4, characterized in that: In step A3, given the initial node feature matrix In each view The above uses a graph neural network to perform parallel message passing to obtain the node embedding matrix. ,in For weight matrices shared across views, For activation function, For view exist The node embedding matrix of the last layer; the node embedding matrix of the last layer. Concatenate and flatten the data along the node dimension to obtain a unified vector. Predicting class probabilities using a multilayer perceptron classifier And utilize cross-entropy loss Optimization was carried out, including It is the one-hot encoding of the real label, and C represents the total number of categories.

6. The hierarchical functional alignment hypergraph learning method for auxiliary diagnosis of brain diseases according to any one of claims 1-5, characterized in that: The method involves hierarchical functional role alignment, combining signal variability and structural features to assess functional roles at specific levels, and aligning the functional roles of brain elements across multiple anatomical levels. The specific implementation includes the following sub-steps: Step B1: Calculate the graph / hypergraph Laplacian matrix for each level view and construct a fractional Laplacian operator to emphasize long-range dependencies; Step B2: Combine the learned node representations to calculate the Fisher-inspired diffusion sensitivity score, which reflects the degree of energy variation of node features under fractional-order diffusion. Step B3: Combining discrete geometry knowledge, use Forman-Ricci curvature quantification to determine the dispersion and concentration characteristics of the local composite structure of the neighborhood of nodes in each view; Step B4: Combine diffusion sensitivity with Forman-Ricci curvature to calculate the role index of each node, characterizing the functional role of the node in a specific view; Step B5: Achieve consistent alignment across levels by calculating the absolute difference in node role indices between view pairs to minimize functional role differences between levels.

7. The hierarchical functional alignment hypergraph learning method for auxiliary diagnosis of brain diseases according to claim 6, characterized in that: In step B1, for each view The symmetric normalized Laplace matrix is And perform spectral decomposition on it. ;in, For view The normalization propagation operator; In step B2, node diffusion sensitivity score ;in It is a fractional exponent; Embed the matrix for the nodes of the last layer; In step B3, node In view The Forman-Ricci curvature in the value is the mean curvature of the associated hyperedge at that node. ;in, Indicates in view In and nodes The set of associated hyperedges Represents a view Middle node and super edge Node-hyperedge curvature between; In step B4, node In view The character index in the middle is ;in This is the curvature scaling factor.

8. The hierarchical functional alignment hypergraph learning method for auxiliary diagnosis of brain diseases according to claim 6, characterized in that: The overall objective function for training the graph neural network model is: , To control the balance hyperparameter of alignment regularization intensity; Among them, cross-entropy loss , It is the one-hot encoding of the real label, where C represents the total number of categories; Alignment loss ,in Represents all different view index pairs The set, where K represents the total number of channels. Represents a view Middle node Role index.

9. A hierarchical functional alignment hypergraph learning system 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 hierarchical functional alignment hypergraph learning method for assisted diagnosis of brain diseases as described in any one of claims 1 to 8.

10. A hierarchical functional alignment hypergraph learning product for auxiliary diagnosis of brain diseases, comprising computer program instructions, characterized in that: When the computer program instructions are executed on a computer, the computer performs the hierarchical functional alignment hypergraph learning method for assisted diagnosis of brain diseases as described in any one of claims 1 to 8.

Images