Traditional Chinese medicine intelligent ai-assisted diagnosis and treatment ecosystem and method

Abstract

The application provides a traditional Chinese medicine intelligent AI assisted diagnosis and treatment ecological system and method, rare subgraph mining is performed on each multi-source data set by determining the base sequence node set of each base sequence mode to obtain a global rare subgraph set, a multi-structure fusion supergraph is constructed based on the global rare subgraph set, an initial graph structure and all base sequence node sets; the cross-correlation matrix of each supergraph neural network model is determined, all cross-correlation matrices are subjected to de-correlation constraints to obtain a plurality of collaborative distillation losses; the all cross-correlation matrices are subjected to an antagonistic disturbance to generate a plurality of antagonistic collaborative losses, and all supergraph neural network models are subjected to federal optimization according to all antagonistic collaborative losses and all collaborative distillation losses; the real-time supergraph is input into the corresponding supergraph neural network model to obtain the confidence of the rare syndrome of the target patient. According to the scheme of the application, rare feature patterns can be mined according to the high-order structure correlation information of multi-source heterogeneous data to accurately identify the rare syndromes of traditional Chinese medicine.

Description

Technical Field

[0001] This application relates to the field of auxiliary diagnosis and treatment technology, and more specifically, to the ecosystem and methods of TCM digital AI-assisted diagnosis and treatment. Background Technology

[0002] Traditional Chinese medicine (TCM) diagnostic data is characterized by its multi-source heterogeneity, involving various modalities such as tongue appearance, pulse diagnosis, and patient consultation texts. Furthermore, the data distribution varies significantly across different medical institutions, forming a typical data silo phenomenon. With the development of artificial intelligence technology, machine learning-based assisted diagnostic methods have been introduced into the field of TCM.

[0003] Existing TCM auxiliary diagnostic methods mostly employ single graph structure models or knowledge distillation frameworks, which struggle to effectively capture hidden high-order structural associations within the data. They fail to extract unique rare feature patterns specific to rare syndromes from limited samples, and require high-quality and high-quantity labeled data, exhibiting limited recognition capabilities when dealing with unevenly distributed rare syndromes. Furthermore, existing methods face challenges in cross-institutional collaborative scenarios, such as difficulties in aligning heterogeneous data distributions and stringent privacy protection requirements. Therefore, how to mine rare feature patterns based on high-order structural associations from multi-source heterogeneous data to accurately identify rare TCM syndromes has become a challenging problem for the industry. Summary of the Invention

[0004] This application provides a digital AI-assisted diagnosis and treatment ecosystem and method for traditional Chinese medicine, which can mine rare feature patterns based on high-order structural correlation information of multi-source heterogeneous data to accurately identify rare syndromes in traditional Chinese medicine.

[0005] Firstly, this application provides a method for identifying rare syndrome types in multi-source heterogeneous data, applicable to the TCM digital AI-assisted diagnosis and treatment ecosystem. This method includes the following steps: Obtain multi-source datasets from various medical institutions and extract the basal patterns from each multi-source dataset; The set of motif nodes for each motif pattern is determined, and rare subgraph mining is performed on each multi-source dataset to obtain a global rare subgraph set. Based on the global rare subgraph set, the initial graph structure of each multi-source dataset, and the set of all motif nodes, a multi-structure fusion hypergraph of each multi-source dataset is constructed. Each multi-structure fusion hypergraph is converted into a corresponding hypergraph neural network model. The cross-correlation matrix of each hypergraph neural network model in the public unlabeled TCM dataset is determined. Then, the cross-correlation matrix is ​​de-correlation-constrained to obtain multiple collaborative distillation losses when co-aligning heterogeneous data. Adversarial perturbations are applied to all cross-correlation matrices to generate multiple adversarial cooperative losses. Federated optimization is then performed on all hypergraph neural network models based on all adversarial cooperative losses and all cooperative distillation losses. Diagnostic data of the target patient is collected and a real-time hypergraph is constructed. The real-time hypergraph is then input into the corresponding hypergraph neural network model after federated optimization to obtain the confidence level of the rare syndrome type of the target patient.

[0006] In some embodiments, obtaining multi-source datasets from various medical institutions and extracting the basalt patterns from each multi-source dataset specifically includes: Obtain multi-source datasets and preset sequence templates from various medical institutions; Each multi-source dataset is transformed into an initial graph structure containing multiple nodes and edges; Based on the aforementioned sequence template, subgraph isomorphism detection is performed on all initial graph structures to obtain the sequence patterns of each multi-source dataset.

[0007] In some embodiments, determining the set of motif nodes for each motif pattern specifically includes: Deduplication of nodes is performed on each motif pattern, and nodes that appear repeatedly in the same motif pattern are merged into a single node. Verify the topology of the merged node set in each basal sequence pattern; The set of nodes that pass the verification in each motif pattern is taken as the motif node set of the corresponding motif pattern.

[0008] In some embodiments, performing rare subgraph mining on each multi-source dataset to obtain a global rare subgraph set specifically includes: Rare subgraphs are enumerated for the initial graph structure of each multi-source dataset to generate multiple candidate subgraphs; Determine the support of each candidate subgraph in the corresponding initial graph structure; All candidate subgraphs are filtered based on all support scores; Structural similarity clustering is performed on all candidate subgraphs after screening to obtain a global rare subgraph set.

[0009] In some embodiments, constructing a multi-structure fusion hypergraph for each multi-source dataset based on the global rare subgraph set, the initial graph structure of each multi-source dataset, and the set of all motif nodes specifically includes: All nodes in the initial graph structure of each multi-source dataset are used as vertices of the hypergraph; Construct multiple rare subgraph superedges based on the set of nodes contained in each subgraph in the global rare subgraph set; Construct corresponding ordering hyperedges based on each ordering node set; Map each edge in the initial graph structure of each multi-source dataset to a pair of superedges containing two nodes; Determine the set of neighborhood nodes that each hypergraph vertex in the initial graph structure of each multi-source dataset can reach within a set step size, and then construct the hop count neighborhood hyperedge corresponding to that vertex. By fusing all rare subgraph hyperedges, radix hyperedges, paired edge hyperedges, and hop-count neighborhood hyperedges, a multi-structure fused hypergraph is generated for each multi-source dataset.

[0010] In some embodiments, determining the cross-correlation matrix of feature associations for each hypergraph neural network model in a public unlabeled TCM dataset specifically includes: Obtain public unlabeled TCM datasets; The public unlabeled TCM dataset is input into each hypergraph neural network model, and the corresponding output feature vectors are extracted. The average eigenvectors of all features are averaged to obtain the global average eigenvectors. Determine the correlation coefficients of each feature vector with the global average feature vector in each dimension, and construct the cross-correlation matrix of the feature associations of the corresponding hypergraph neural network model.

[0011] In some embodiments, applying decorrelation constraints to all cross-correlation matrices to obtain multiple cooperative distillation losses for cooperatively aligning heterogeneous data specifically includes: Extract the diagonal elements of each cross-correlation matrix to form the corresponding homogeneous vector; Extract the off-diagonal elements of each cross-correlation matrix to construct the corresponding heterogeneous decorrelation matrix; For any cross-correlation matrix, construct a similar feature constraint term based on the corresponding similar consistent vector, and construct a dissimilar feature constraint term based on the corresponding dissimilar decorrelation matrix; The cooperative distillation loss for co-aligning heterogeneous data is determined based on the heterogeneous feature constraint term and the homogeneous feature constraint term, and the corresponding cross-correlation matrix is ​​used to determine the cooperative distillation loss.

[0012] Secondly, this application provides a TCM digital AI-assisted diagnosis and treatment ecosystem, which includes a rare syndrome identification unit, comprising: The acquisition module is used to acquire multi-source datasets from various medical institutions and extract the basal sequence patterns from each multi-source dataset. The processing module is used to determine the set of motif nodes for each motif pattern, perform rare subgraph mining on each multi-source dataset to obtain a global rare subgraph set, and construct a multi-structure fusion hypergraph for each multi-source dataset based on the global rare subgraph set, the initial graph structure of each multi-source dataset, and all motif node sets. The processing module is also used to convert each multi-structure fusion hypergraph into a corresponding hypergraph neural network model, determine the cross-correlation matrix of the feature association of each hypergraph neural network model in the public unlabeled TCM dataset, and then perform decorrelation constraints on all cross-correlation matrices to obtain multiple collaborative distillation losses when collaboratively aligning heterogeneous data. The processing module is also used to apply adversarial perturbations to all cross-correlation matrices, generate multiple adversarial cooperative losses, and perform federated optimization on all hypergraph neural network models based on all adversarial cooperative losses and all cooperative distillation losses. The execution module is used to collect diagnostic data of the target patient and construct a real-time hypergraph. The real-time hypergraph is then input into the corresponding hypergraph neural network model after federated optimization to obtain the confidence level of the rare syndrome type of the target patient.

[0013] Thirdly, this application provides a computer device, the computer device including a memory and a processor, the memory storing code, the processor being configured to acquire the code and execute the above-described method for identifying rare patterns in heterogeneous multi-source data.

[0014] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for identifying rare certificate types oriented towards heterogeneous multi-source data.

[0015] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects: In the process of identifying rare syndrome types in heterogeneous multi-source data, this application first determines the set of motif nodes for each motif pattern, performs rare subgraph mining on each multi-source dataset to obtain a global rare subgraph set, and constructs a multi-structure fusion hypergraph for each multi-source dataset based on the global rare subgraph set, the initial graph structure of each multi-source dataset, and all motif node sets; wherein, the multi-structure fusion hypergraph is a complete hypergraph structure that comprehensively represents the multi-level high-order associations of the multi-source dataset; secondly, each multi-structure fusion hypergraph is converted into a corresponding hypergraph neural network model, and the characteristics of each hypergraph neural network model in public unlabeled TCM datasets are determined. The method involves identifying the cross-correlation matrix of the data, then applying disassociation constraints to all cross-correlation matrices to obtain multiple collaborative distillation losses for aligning heterogeneous data. The collaborative distillation loss is a comprehensive loss value used to constrain the alignment of similar features and the decoupling of dissimilar features. Finally, adversarial perturbations are applied to all cross-correlation matrices to generate multiple adversarial collaborative losses. Federated optimization is then performed on all hypergraph neural network models based on all adversarial collaborative losses and all collaborative distillation losses. Diagnostic data of the target patient is collected and a real-time hypergraph is constructed. This real-time hypergraph is then input into the corresponding federated optimized hypergraph neural network model to obtain the confidence level of the rare syndrome type of the target patient. This approach can mine rare feature patterns based on the high-order structural correlation information of multi-source heterogeneous data to accurately identify rare syndrome types in Traditional Chinese Medicine. Attached Figure Description

[0016] Figure 1 This is an exemplary flowchart of a rare disease type identification method for heterogeneous multi-source data, as shown in some embodiments of this application. Figure 2 This is an exemplary flowchart illustrating the determination of a set of motif nodes according to some embodiments of this application; Figure 3 This is an exemplary flowchart illustrating the determination of resistance to cooperation loss according to some embodiments of this application; Figure 4 This is a schematic diagram of the structure of a rare disease type identification method according to some embodiments of this application; Figure 5 This is a schematic diagram of the structure of a computer device for implementing a rare disease type identification method for heterogeneous multi-source data, according to some embodiments of this application. Detailed Implementation

[0017] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0018] refer to Figure 1The figure is an exemplary flowchart of a rare disease identification method for heterogeneous multi-source data, according to some embodiments of this application. The rare disease identification method for heterogeneous multi-source data mainly includes the following steps: In step 101, multi-source datasets from various medical institutions are obtained, and the motif patterns of each multi-source dataset are extracted.

[0019] In specific implementation, acquiring multi-source datasets from various medical institutions and extracting the motif patterns from each dataset can be achieved in the following way: First, collect multi-source datasets containing various types of data, such as images, text, and numerical data, from various medical institutions. Simultaneously, predefine a set of motif templates for identifying local structural patterns. Each motif template describes a specific node connection pattern, such as a triangle, star, or chain structure. Second, perform structuring processing on each multi-source dataset: map each data sample in the dataset to a node, and establish edges between nodes based on the relationships between data, thereby generating an initial graph structure composed of nodes and edges for each dataset. This initial graph structure reflects the local connectivity within the data and serves as the base graph representation for subsequent motif matching and rare subgraph mining. Finally, for each initial graph structure, use a subgraph isomorphism detection algorithm to traverse all node combinations in the graph and find subgraph instances that completely match the motif template structure. For each successfully matched instance, record its set of nodes and corresponding template type, and use the set of all matching instances as the motif pattern for that dataset. Other implementation methods can also be used in other embodiments, which are not limited here.

[0020] It should be noted that the multi-source dataset in this application is a heterogeneous dataset used to integrate various TCM diagnosis and treatment data from medical institutions. It is used to provide TCM diagnosis and treatment information in all dimensions for subsequent structural feature extraction and model training, so as to lay a data foundation for comprehensively mining the structural features and correlation patterns in the data and avoid the identification bias of rare syndromes due to the lack of data dimensions. The initial graph structure is a basic graph representation structure used to transform the multi-source dataset into a node and edge association form. The order pattern is a set of subgraph instances used to represent the multi-source dataset and the preset order pattern template that are successfully matched.

[0021] In step 102, the set of motif nodes for each motif pattern is determined, rare subgraph mining is performed on each multi-source dataset to obtain a global rare subgraph set, and a multi-structure fusion hypergraph for each multi-source dataset is constructed based on the global rare subgraph set, the initial graph structure of each multi-source dataset, and all motif node sets.

[0022] In some embodiments, reference Figure 2 As shown, this figure is an exemplary flowchart of determining the set of motif nodes in some embodiments of this application. In this embodiment, determining the set of motif nodes for each motif pattern can be achieved by the following steps: First, in step 1021, node deduplication is performed on each motif pattern, and nodes that appear repeatedly in the same motif pattern are merged into a single node. Secondly, in step 1022, the topological structure of the merged node set in each basal sequence pattern is verified; Finally, in step 1023, the set of nodes that pass the verification in each motif pattern is taken as the motif node set of the corresponding motif pattern.

[0023] In practical implementation, node deduplication is performed on each motif pattern, merging duplicate nodes within the same motif pattern into a single node. This can be achieved as follows: For each motif pattern containing multiple motif instances, count the nodes appearing in all instances. If the same node appears in multiple instances, retain only that node once, thus eliminating redundancy and obtaining the set of unique nodes corresponding to that motif pattern. The topology of the merged node sets in each motif pattern can be verified as follows: For each motif pattern's unique node set, check if these nodes still constitute the corresponding motif in the initial graph structure. If the topological relationship is consistent with the template, that is, the edge connection relationship between nodes is the same as the template, then the verification is deemed to pass, to ensure that the deduplication operation does not destroy the original structural semantics; the above topological verification result is the basis for judging the selection of valid basal sequence nodes; the set of nodes that pass the verification in each basal sequence pattern can be used as the basal sequence node set of the corresponding basal sequence pattern in the following way: take the set of non-repeating nodes corresponding to each basal sequence pattern that passes the topological verification as the basal sequence node set of that basal sequence pattern; if the verification fails, discard the set to ensure the structural accuracy of the basal sequence nodes; other methods can also be used in other embodiments, which are not limited here.

[0024] In some embodiments, rare subgraph mining of each multi-source dataset to obtain a global rare subgraph set can be achieved by the following steps: Rare subgraphs are enumerated for the initial graph structure of each multi-source dataset to generate multiple candidate subgraphs; Determine the support of each candidate subgraph in the corresponding initial graph structure; All candidate subgraphs are filtered based on all support scores; Structural similarity clustering is performed on all candidate subgraphs after screening to obtain a global rare subgraph set.

[0025] In specific implementation, the rare subgraph enumeration of the initial graph structure of each multi-source dataset to generate multiple candidate subgraphs can be achieved in the following way: For the initial graph structure of each multi-source dataset, a graph mining algorithm, such as the depth-first search-based gSpan algorithm, is used to traverse all possible connected subgraph structures in the initial graph structure, generating a set of subgraphs of various sizes and shapes as multiple candidate subgraphs for the corresponding multi-source dataset; where the above candidate subgraphs are the original subgraph instances used for support statistics; the support of each candidate subgraph appearing in the corresponding initial graph structure can be determined in the following way: For any candidate subgraph of each multi-source dataset, the number of times it appears as a subgraph isomorphic instance in the initial graph structure of the corresponding multi-source dataset is counted, and this number is defined as the support of the candidate subgraph; where support is a parameter value reflecting the frequency of the subgraph structure appearing in the corresponding multi-source dataset, and is a quantitative indicator used to screen rare subgraphs; the remaining support is then determined; other implementation methods can also be used in other embodiments, which are not limited here.

[0026] In specific implementation, filtering all candidate subgraphs based on all support can be achieved in the following way: set a support threshold, for example, the number of occurrences is less than or equal to 5, remove all candidate subgraphs with support higher than the threshold, and retain candidate subgraphs with support lower than or equal to the threshold; the filtered subgraphs here are the set of rare subgraphs used for clustering; performing structural similarity clustering on all filtered candidate subgraphs to obtain the global rare subgraph set can be achieved in the following way: perform isomorphism detection on all selected candidate subgraphs, group candidate subgraphs with the same structure, that is, completely identical node connection relationships, into the same class, select a representative subgraph for each class, and the set of all representative subgraphs constitutes the global rare subgraph set; where the global rare subgraph set is the unique structural template set used to construct the superedges of rare subgraphs; other embodiments can also be implemented in other ways, which are not limited here.

[0027] In some embodiments, constructing a multi-structure fusion hypergraph for each multi-source dataset based on the global rare subgraph set, the initial graph structure of each multi-source dataset, and the set of all motif nodes can be achieved using the following steps: All nodes in the initial graph structure of each multi-source dataset are used as vertices of the hypergraph; Construct multiple rare subgraph superedges based on the set of nodes contained in each subgraph in the global rare subgraph set; Construct corresponding ordering hyperedges based on each ordering node set; Map each edge in the initial graph structure of each multi-source dataset to a pair of superedges containing two nodes; Determine the set of neighborhood nodes that each hypergraph vertex in the initial graph structure of each multi-source dataset can reach within a set step size, and then construct the hop count neighborhood hyperedge corresponding to that vertex. By fusing all rare subgraph hyperedges, radix hyperedges, paired edge hyperedges, and hop count neighborhood hyperedges, a multi-structure fusion hypergraph for each multi-source dataset is generated. For example, an initial weight of 0.4 is assigned to rare subgraph hyperedges (to prioritize capturing rare-type related structures), and an initial weight of 0.2 is assigned to radix hyperedges, paired edge hyperedges, and hop count neighborhood hyperedges. All weighted hyperedges are then fused to generate a multi-structure fusion hypergraph for each multi-source dataset. During training, the weights of each hyperedge can be adaptively adjusted according to the model performance.

[0028] In specific implementation, using all nodes in the initial graph structure of each multi-source dataset as hypergraph vertices can be achieved as follows: For the initial graph structure of each multi-source dataset, all its contained nodes are directly used as vertices of the corresponding hypergraph of the multi-source dataset, ensuring a one-to-one correspondence between the hypergraph vertex set and the entities in the original data; constructing multiple rare subgraph hyperedges based on the node set contained in each subgraph of the global rare subgraph set can be achieved as follows: Traverse each subgraph in the global rare subgraph set, combine all the nodes contained in each subgraph to form a hyperedge, obtaining the rare subgraph hyperedge, and collectively form the rare subgraph hyperedge set; constructing corresponding radix hyperedges based on each radix node set can be achieved as follows: For each radix node set, the nodes in the radix node set are used as vertices of the hypergraph of the radix node set. All nodes are combined to form a superedge, called a basalt superedge; each set of basalt nodes corresponds to a basalt superedge, and all basalt superedges constitute a basalt superedge set; the above-mentioned basalt superedges are used to capture higher-order associations of a preset semantic structure; mapping each edge in the initial graph structure of each multi-source dataset to a pair of edge superedges containing two nodes can be implemented in the following way: for each edge in the initial graph structure of each multi-source dataset, its two endpoint nodes are combined to form a binary node set as a pair of edge superedges, and the pair of edge superedges of the remaining initial graph structure are determined, and the pair of edge superedges of the remaining multi-source datasets are determined; the superedges corresponding to all edges constitute a pair of edge superedges set; the above-mentioned pair of edge superedges are the basic units used to preserve the pairwise connection relationships in the original graph; other embodiments can also be implemented in other ways, which are not limited here.

[0029] In specific implementation, determining the set of neighboring nodes reachable by each hypergraph vertex within a set step size in the initial graph structure of each multi-source dataset, and thus constructing the hop count neighborhood hyperedge corresponding to that vertex, can be achieved in the following way: For each vertex in the initial graph structure of each multi-source dataset, starting from that vertex, calculate all nodes reachable within a preset number of steps (including the starting point itself) in the corresponding initial graph structure using breadth-first search. Form a set of these nodes, assign a unique hyperedge identifier to this set, and establish a record of the inclusion relationship between the hyperedge identifier and all nodes in the set, thereby completing the transformation from the node set to the hyperedge, i.e., obtaining the unique hop count neighborhood hyperedge corresponding to the vertex. Perform the above operation independently on each vertex in each initial graph structure to generate a hop count neighborhood hyperedge uniquely corresponding to the corresponding vertex. Each vertex corresponds to one such hyperedge, and the hop count neighborhood hyperedges of all vertices constitute the hop count neighborhood hyperedge set. Here, the hop count neighborhood hyperedge is used to capture higher-order associations of local extended structures. Other implementation methods can also be used in other embodiments, which are not limited here.

[0030] In specific implementation, the multi-structure fusion hypergraph of each multi-source dataset can be generated by fusing all rare subgraph hyperedges, radix hyperedges, paired edge hyperedges, and hop count neighborhood hyperedges. This can be achieved by merging all rare subgraph hyperedges, radix hyperedges, paired edge hyperedges, and hop count neighborhood hyperedges together with all radix node sets to form a complete hypergraph structure. Specifically, each hyperedge is assigned a unique hyperedge identifier, and a bidirectional index relationship is established between the hyperedge identifier and all vertices it contains. That is, all vertices contained in the hyperedge can be queried through the hyperedge identifier, and all hyperedges containing that vertex can be queried through the vertex. After all hyperedges are added, the multi-structure fusion hypergraph of the corresponding medical institution, which contains a set of vertices and four types of hyperedge sets, is completed. This multi-structure fusion hypergraph structure contains high-order association information extracted from different perspectives. Other implementation methods can also be used in other embodiments, which are not limited here.

[0031] It should be noted that the set of basal nodes captures local semantic patterns in the data that conform to the preset template, rare subgraph mining discovers low-frequency implicit structural patterns in the data, paired edges preserve basic pairwise associations, and hop count neighborhood expands the local context. The fusion of these four types of information in the hypergraph structure realizes multi-level high-order association modeling from micro to macro and from explicit to implicit. The beneficial effects of the above steps are reflected in the fact that through this multi-source structural fusion, the hypergraph can more comprehensively reflect the complex multi-dimensional interaction relationships within the data, providing semantically rich structured feature inputs for subsequent model learning, significantly enhancing the ability to perceive potential patterns in the data, and laying a rich structural feature foundation, especially for identifying rare syndromes with few samples.

[0032] It should be noted that the support in this application is a numerical indicator used to quantify the frequency of candidate subgraphs appearing in the initial graph structure. The global rare subgraph set is a set of connected subgraphs that appear infrequently and have unique structures in the initial graph structure of each multi-source dataset. It is used to mine implicit rare common structural patterns in multi-source heterogeneous data of traditional Chinese medicine, and to provide a unique structural template for the rare subgraph hyperedges in the construction of multi-structure fusion hypergraphs, capturing the rare structural features that are prevalent in the data. The hypergraph vertices are the set of nodes used as the basic building blocks of the multi-structure fusion hypergraph, corresponding one-to-one with the nodes in the initial graph structure. The rare subgraph hyperedges are the hyperedge forms used to represent the higher-order associations between nodes in the global rare subgraph set. The basalt hyperedges are the preset semantic templates used to represent the relationships between basalt node sets. The hyperedges are categorized into four types: paired hyperedges, which preserve the basic relationships between nodes in the initial graph structure; hop count neighborhood hyperedges, which represent higher-order relationships between nodes in the neighborhood of a hypergraph vertex within a set step size; and multi-structure fusion hypergraphs, which are complete hypergraph structures used to fuse four types of hyperedges and comprehensively represent multi-level higher-order relationships in multi-source datasets. These structures integrate the multi-dimensional relationships of motifs, rare subgraphs, paired edges, and hop count neighborhoods into a unified hypergraph structure, providing semantically rich structured feature inputs for hypergraph neural network models. This solves the problem that traditional graph structures cannot effectively represent the complex and multi-dimensional interactive relationships in multi-source data of traditional Chinese medicine. Through multi-dimensional hyperedge fusion, the structural relationships in the data are comprehensively mined, building a rich foundation of structured features for rare syndrome identification.

[0033] In addition, this application integrates all rare subgraph hyperedges, radix hyperedges, paired edge hyperedges, and hop count neighborhood hyperedges to generate multi-structure fusion hypergraphs for each multi-source dataset. For example, rare subgraph hyperedges are assigned an initial weight of 0.4 (prioritizing the capture of rare-type related structures), while radix hyperedges, paired edge hyperedges, and hop count neighborhood hyperedges are each assigned an initial weight of 0.2. All weighted hyperedges are integrated to generate multi-structure fusion hypergraphs for each multi-source dataset. During training, the weights of each hyperedge can be adaptively adjusted according to the model performance. This is only an example and is not intended to limit the scope of this application.

[0034] In step 103, each multi-structure fusion hypergraph is converted into a corresponding hypergraph neural network model, the cross-correlation matrix of the feature association of each hypergraph neural network model in the public unlabeled TCM dataset is determined, and then the cross-correlation constraint is applied to all cross-correlation matrices to obtain multiple collaborative distillation losses when collaboratively aligning heterogeneous data.

[0035] In practical implementation, converting each multi-structure fusion hypergraph into a corresponding hypergraph neural network model can be achieved as follows: design at least one hypergraph convolutional layer for feature learning of higher-order associations in the hypergraph structure; specifically, for each hypergraph convolutional layer, firstly, construct an association matrix based on the inclusion relationship between hyperedges and nodes; then, aggregate and update node features through hypergraph convolution operations. This operation consists of two stages: the first stage is aggregation from nodes to hyperedges, that is, for each hyperedge, aggregate the features of all nodes it contains to obtain the feature representation of the hyperedge; the second stage is aggregation from hyperedges to nodes, that is, for each node, aggregate the features of all hyperedges containing that node and update the hyperedge feature representation. The node features are represented; through the aggregation operations in the two stages mentioned above, feature propagation of one layer of hypergraph convolution is completed; after each hypergraph convolution layer, a non-linear activation function is applied to the updated node features, and a batch normalization layer is optionally added to stabilize the training process; finally, after the last hypergraph convolution layer, a fully connected layer is connected as a classification output layer, which maps the node features extracted by multiple hypergraph convolutions to a preset number of categories, and finally outputs the initial prediction score of each node (i.e., each data sample) belonging to each category; the hypergraph neural network model here is the basic model structure for subsequent federated collaborative optimization and local distillation training; other embodiments can also be implemented in other ways, which are not limited here.

[0036] In some embodiments, determining the cross-correlation matrix of feature associations for each hypergraph neural network model in a public unlabeled TCM dataset can be achieved using the following steps: Obtain public unlabeled TCM datasets; The public unlabeled TCM dataset is input into each hypergraph neural network model, and the corresponding output feature vectors are extracted. The average eigenvectors of all features are averaged to obtain the global average eigenvectors. Determine the correlation coefficients of each feature vector with the global average feature vector in each dimension, and construct the cross-correlation matrix of the feature associations of the corresponding hypergraph neural network model.

[0037] In specific implementation, the public unlabeled TCM dataset is input into each hypergraph neural network model, and the corresponding output feature vectors are extracted. This can be achieved as follows: Each sample in the public unlabeled TCM dataset is sequentially input into the hypergraph neural network model already trained by each medical institution. Through the forward propagation of the model, the high-dimensional vector output by the last hidden layer of the model is obtained as the feature vector of that sample. Each model outputs a feature vector for each sample, and finally each model obtains a set of sample feature vectors. These feature vectors are the basic numerical representation used to construct the cross-correlation matrix. The correlation coefficients between each feature vector and the global average feature vector in each dimension are determined. The cross-correlation matrix of the feature associations of the corresponding hypergraph neural network model can be constructed in the following way: For each hypergraph neural network model, the feature vector output by the hypergraph neural network model for each sample is compared with the global average feature vector, and the Pearson correlation coefficient of the two in each dimension is calculated. The correlation coefficients in all dimensions are arranged in rows and columns to form a square matrix, which is the cross-correlation matrix corresponding to the hypergraph neural network model. This matrix quantifies the degree of association between the hypergraph neural network model and global knowledge in each dimension. Continue to determine the cross-correlation matrix of the feature associations of the remaining hypergraph neural network models; other methods can also be used in other embodiments, which are not limited here.

[0038] In some embodiments, the following steps can be used to obtain multiple cooperative distillation losses for cooperatively aligning heterogeneous data by applying decorrelation constraints to all cross-correlation matrices: Extract the diagonal elements of each cross-correlation matrix to form the corresponding homogeneous vector; Extract the off-diagonal elements of each cross-correlation matrix to construct the corresponding heterogeneous decorrelation matrix; For any cross-correlation matrix, construct a similar feature constraint term based on the corresponding similar consistent vector, and construct a dissimilar feature constraint term based on the corresponding dissimilar decorrelation matrix; The cooperative distillation loss for co-aligning heterogeneous data is determined based on the heterogeneous feature constraint term and the homogeneous feature constraint term, and the corresponding cross-correlation matrix is ​​used to determine the cooperative distillation loss.

[0039] In practical implementation, extracting the diagonal elements of each cross-correlation matrix to construct the corresponding homogeneous consistent vector can be achieved as follows: For each cross-correlation matrix, extract all elements on its diagonal (i.e., elements with the same row and column numbers), and arrange these elements in order to form a vector called the homogeneous consistent vector; this homogeneous consistent vector reflects the consistency between the model output and global knowledge in the same dimension. Extracting the off-diagonal elements of each cross-correlation matrix to construct the corresponding heterogeneous decorrelation matrix can be achieved as follows: For each cross-correlation matrix, extract all off-diagonal elements (i.e., elements with different row and column numbers), and arrange these elements according to the row and column positions of the original matrix to form a matrix of the same size as the original matrix but with zero diagonal elements; this is called the heterogeneous decorrelation matrix; this heterogeneous decorrelation matrix reflects the cross-correlation between different dimensions. The above heterogeneous decorrelation... The correlation matrix is ​​the basis for constructing heterogeneous feature constraint terms. For any cross-correlation matrix, homogeneous feature constraint terms are constructed based on the corresponding homogeneous consistent vectors. Heterogeneous feature constraint terms are constructed based on the corresponding heterogeneous decorrelation matrix in the following manner: For each cross-correlation matrix, firstly, homogeneous feature constraint terms are calculated based on its homogeneous consistent vectors: the difference between each element in the homogeneous consistent vector and the value 1 is squared and summed to obtain homogeneous feature constraint terms; secondly, heterogeneous feature constraint terms are calculated based on its heterogeneous decorrelation matrix: the difference between each element in the heterogeneous decorrelation matrix and the value 0 is squared and summed, or it is made to approach -1 as needed to obtain heterogeneous feature constraint terms; wherein, the above homogeneous feature constraint terms and heterogeneous feature constraint terms are loss components used to jointly constrain the cross-correlation matrix; other methods can also be used in other embodiments, which are not limited here.

[0040] In specific implementation, the collaborative distillation loss for coordinating heterogeneous data by determining the corresponding cross-correlation matrix based on the heterogeneous feature constraint term and the homogeneous feature constraint term can be implemented in the following way: weighted combination of homogeneous feature constraint term and heterogeneous feature constraint term, for example, adding the two together, to obtain the collaborative distillation loss corresponding to the cross-correlation matrix; this loss is used to force the model output features of each client to align on the same dimension and decouple on different dimensions in subsequent optimization, and is the core supervision signal for federated optimization; other embodiments can also be implemented in other ways, which are not limited here.

[0041] It should be noted that the collaborative distillation loss is obtained by extracting the diagonal elements of the cross-correlation matrix to construct a homogeneous consistency vector and extracting the off-diagonal elements to construct a heterogeneous decorrelation matrix. Homogeneous and heterogeneous feature constraint terms are then constructed based on these two types of constraint terms, and finally, the two types of constraint terms are weighted and combined to obtain the collaborative distillation loss. The working principle is as follows: the diagonal elements reflect the consistency between the output of each client model and the global consensus on the same feature dimension. The diagonal constraints force all clients to converge towards the global consensus on the same dimension, achieving collaborative alignment of homogeneous features. The off-diagonal elements characterize the cross-correlation between different feature dimensions. The off-diagonal constraints promote the correlation between dimensions to approach zero or negative values, achieving decoupling and independence of heterogeneous features. This dual constraint mechanism allows the feature distributions of different client models to maintain their respective structural characteristics while achieving alignment in the shared semantic space, effectively alleviating the feature distribution shift problem caused by heterogeneous data. Simultaneously, the decoupling of feature dimensions enhances the semantic independence of each dimension, improving the model's ability to distinguish fine-grained categories (especially rare types), and providing high-quality supervision signals for knowledge transfer and robustness enhancement in the subsequent federated optimization process.

[0042] It should be noted that the cross-correlation matrix of feature association in this application is a square matrix used to quantify the degree of correlation between the feature vectors of each model and the global average feature vector in each dimension. It is used to characterize the difference in feature dimension-level correlation between each local model and global knowledge, and serves as the numerical basis for subsequent decorrelation constraint and collaborative distillation loss calculation. It provides a quantifiable basis for model feature distribution alignment, solves the problem that the model feature distribution offset caused by heterogeneous data cannot be accurately quantified. By realizing dimension-level feature association analysis through the cross-correlation matrix, the deviation between each model and global knowledge can be accurately located, providing an operable quantitative supervision basis for subsequent collaborative alignment of heterogeneous data.

[0043] The federated optimization process adopts a collaborative model where data does not leave the local area. Each medical institution only transmits the cross-correlation matrix and loss value of the model, while the original diagnosis and treatment data is always kept locally. This can effectively protect patient privacy and medical institution data security. In specific implementation, in step 104, adversarial perturbations are applied to all cross-correlation matrices to generate multiple adversarial collaborative losses. Based on all adversarial collaborative losses and all collaborative distillation losses, federated optimization is performed on all hypergraph neural network models.

[0044] In some embodiments, reference Figure 3 As shown, this figure is an exemplary flowchart for determining adversarial cooperation losses in some embodiments of this application. In this embodiment, adversarial perturbations are applied to all cross-correlation matrices to generate multiple adversarial cooperation losses, which can be achieved by the following steps: First, in step 1041, the gradient direction of each cooperative distillation loss in the corresponding cross-correlation matrix is ​​determined; Secondly, in step 1042, for any cross-correlation matrix, a sign perturbation of a preset strength is added along the corresponding gradient direction to generate an adversarial perturbation matrix; Finally, in step 1043, the adversarial cooperation loss of the corresponding cross-correlation matrix is ​​determined based on the adversarial perturbation matrix.

[0045] In specific implementation, determining the gradient direction of each cooperative distillation loss in the corresponding cross-correlation matrix can be achieved as follows: For each cross-correlation matrix, the partial derivative of its cooperative distillation loss with respect to each element in the matrix is ​​calculated using automatic differentiation techniques to obtain the gradient matrix; this gradient matrix indicates the sensitive direction of the loss function to changes in matrix elements; this gradient matrix is ​​used as the directional basis for generating adversarial perturbations; for any cross-correlation matrix, adding a sign perturbation of a preset strength along the corresponding gradient direction to generate the adversarial perturbation matrix can be achieved as follows: For each cross-correlation matrix, the sign function of its gradient matrix is ​​taken, i.e., each element is +1 or -1 to represent the gradient direction, and then multiplied by a factor of 1. A preset small constant (e.g., 0.01) is used as the perturbation intensity. The resulting perturbation matrix is ​​added to the original cross-correlation matrix to generate an adversarial perturbation matrix. This matrix simulates the change of the cross-correlation matrix in the worst case. The adversarial cooperative loss corresponding to the cross-correlation matrix can be determined based on the adversarial perturbation matrix in the following way: substitute the adversarial perturbation matrix into the cooperative distillation loss calculation formula, recalculate the loss value, and obtain the adversarial cooperative loss corresponding to the cross-correlation matrix. Here, the loss is used to enhance the robustness of the model to the input perturbation, and the adversarial cooperative loss is a robust supervision signal used to fuse with the original cooperative distillation loss. Other embodiments may also use other methods, which are not limited here.

[0046] In some embodiments, federated optimization of all hypergraph neural network models based on all adversarial cooperative losses and all cooperative distillation losses can be achieved by the following steps: We weight and fuse all adversarial cooperative losses and all cooperative distillation losses to obtain the robust cooperative loss of each hypergraph neural network model. Backpropagation is performed on all hypergraph neural network models based on all robust cooperative losses, and local model parameters are updated. Iterative optimization continues until all hypergraph neural network models converge.

[0047] In practice, all adversarial cooperation losses and all cooperative distillation losses are weighted and fused to obtain the robust cooperation loss for each hypergraph neural network model. This can be achieved as follows: For each medical institution hypergraph neural network model, its corresponding cooperative distillation loss and adversarial cooperation loss are multiplied by preset weight coefficients and then added together. For example, the cooperative distillation loss has a weight of 0.7 and the adversarial loss has a weight of 0.3, resulting in the robust cooperation loss specific to that hypergraph neural network model. This loss integrates the original alignment target and the adversarial enhancement target. Here, the robust collaboration loss is the final loss function used to update the model parameters. Backpropagation is performed on all hypergraph neural network models based on all robust collaboration losses, and the local model parameters are updated. Iterative optimization until all hypergraph neural network models converge can be achieved as follows: using the robust collaboration loss of each hypergraph neural network model, the gradient of the loss with respect to the model parameters is calculated through backpropagation, and the local parameters of each model are updated using an optimizer (such as stochastic gradient descent). The robust collaboration loss is repeatedly calculated and executed multiple times until the robust collaboration loss of all hypergraph neural network models tends to stabilize or reaches the preset number of iterations. At this point, the feature distributions of each hypergraph neural network model have been aligned and possess robustness. Other implementation methods can also be used in other embodiments, which are not limited here.

[0048] It should be noted that the collaborative distillation loss in this application is a comprehensive loss value used to constrain the alignment of similar features and the decoupling of dissimilar features in the model. It is used to achieve collaborative alignment of feature distributions under heterogeneous data, which facilitates the alignment and decoupling of features through dual constraints, providing a high-quality feature foundation for rare disease identification and improving the classification accuracy of the model.

[0049] In step 105, diagnostic data of the target patient is collected and a real-time hypergraph is constructed. The real-time hypergraph is then input into the corresponding hypergraph neural network model after federated optimization to obtain the confidence level of the rare syndrome type of the target patient.

[0050] In some embodiments, the confidence level of the rare syndrome type of the target patient can be obtained by inputting the real-time hypergraph into the federated optimized corresponding hypergraph neural network model through the following steps: Construct a real-time supermap of the target patient based on the diagnostic data; The real-time hypergraph is input into the corresponding hypergraph neural network model after federated optimization for feature aggregation, thereby obtaining the rare syndrome confidence of the target patient.

[0051] In specific implementation, constructing a real-time hypergraph for the target patient based on the diagnostic data can be achieved in the following way: Real-time multi-source diagnostic data of the target patient is collected from the patient's medical database, such as tongue images, pulse waveforms, and medical history text, and converted into an initial graph structure. Then, a multi-structure fusion hypergraph specific to the patient is constructed through steps such as motif matching and rare subgraph mining; this is the real-time hypergraph for the target patient. Here, the real-time hypergraph is a specific graph structure for the target patient used as input to the model for inference. The real-time hypergraph is then input into the federated optimized corresponding hypergraph neural network model for feature aggregation, thereby obtaining the rare syndrome confidence score of the target patient. The following method is used to achieve this: the constructed real-time hypergraph is input into a federated optimization hypergraph neural network model belonging to the target medical institution. The model aggregates and propagates the node features within the hyperedge through hypergraph convolutional layers, and finally outputs the probability value of each syndrome category. At the same time, by analyzing the stability of the model's response to rare categories, such as the variance of the output of multiple forward propagations or the consistency between the output of the teacher model and the student model, an additional confidence score is calculated. This score represents the degree of certainty of the corresponding hypergraph neural network model in judging the rare syndrome of the target patient. Other embodiments may also use other methods, which are not limited here.

[0052] Furthermore, in another aspect of this application, in some embodiments, this application provides a TCM digital AI-assisted diagnosis and treatment ecosystem, which includes a rare syndrome type identification unit, as referenced. Figure 4 The figure is a schematic diagram of the structure of a rare disease identification unit according to some embodiments of this application. The rare disease identification unit includes: an acquisition module 401, a processing module 402, and an execution module 403, which are described below: The acquisition module 401 in this application is mainly used to acquire multi-source datasets from various medical institutions and extract the basal sequence patterns from each multi-source dataset. Processing module 402, in this application, is used to determine the set of motif nodes for each motif pattern, perform rare subgraph mining on each multi-source dataset to obtain a global rare subgraph set, and construct a multi-structure fusion hypergraph for each multi-source dataset based on the global rare subgraph set, the initial graph structure of each multi-source dataset, and all motif node sets. It should be noted that the processing module 402 in this application is also used to convert each multi-structure fusion hypergraph into a corresponding hypergraph neural network model, determine the cross-correlation matrix of the feature association of each hypergraph neural network model in the public unlabeled TCM dataset, and then perform decorrelation constraints on all cross-correlation matrices to obtain multiple collaborative distillation losses when collaboratively aligning heterogeneous data. Additionally, it should be noted that the processing module 402 in this application is also used to apply adversarial perturbations to all cross-correlation matrices, generate multiple adversarial cooperative losses, and perform federated optimization on all hypergraph neural network models based on all adversarial cooperative losses and all cooperative distillation losses. The execution module 403 in this application is mainly used to collect the diagnostic data of the target patient and construct a real-time hypergraph. The real-time hypergraph is then input into the corresponding hypergraph neural network model after federated optimization to obtain the confidence level of the rare syndrome type of the target patient.

[0053] In addition, this application also provides a computer device, the computer device including a memory and a processor, the memory storing code, the processor being configured to acquire the code and execute the above-described method for identifying rare patterns in heterogeneous multi-source data.

[0054] In some embodiments, reference Figure 5 The figure is a schematic diagram of the structure of a computer device implementing a rare disease type identification method for heterogeneous multi-source data, according to some embodiments of this application. The rare disease type identification method for heterogeneous multi-source data in the above embodiments can be implemented through... Figure 5 The computer device shown is used to implement this, and the computer device includes at least one processor 501, a communication bus 502, a memory 503, and at least one communication interface 504.

[0055] Processor 501 can be a general-purpose central processing unit (CPU) or an application-specific integrated circuit (ASIC).

[0056] The communication bus 502 can be used to transmit information between the aforementioned components.

[0057] Memory 503 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CDROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital versatile optical discs, Blu-ray discs, etc.), magnetic disks or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. Memory 503 may exist independently and be connected to processor 501 via communication bus 502. Memory 503 may also be integrated with processor 501.

[0058] The memory 503 stores program code for executing the scheme of this application, and its execution is controlled by the processor 501. The processor 501 executes the program code stored in the memory 503. The program code may include one or more software modules. The method used in the above embodiments can be implemented by the processor 501 and one or more software modules in the program code in the memory 503.

[0059] Communication interface 504 uses any transceiver-like device to communicate with other devices or communication networks, such as Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc.

[0060] In a specific implementation, as one example, a computer device may include multiple processors, each of which may be a single-core (single CPU) processor or a multi-core (multi CPU) processor. Here, a processor may refer to one or more devices, circuits, and / or processing cores used to process data (e.g., computer program instructions).

[0061] The aforementioned computer device can be a general-purpose computer device or a special-purpose computer device. In specific implementations, the computer device can be a desktop computer, a portable computer, a network server, a handheld digital assistant (PDA), a mobile phone, a tablet computer, a wireless terminal device, a communication device, or an embedded device. This application does not limit the type of computer device.

[0062] In addition, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for identifying rare certificate types in heterogeneous multi-source data.

[0063] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0064] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for identifying rare syndrome types in multi-source heterogeneous data, applied to a TCM digital AI-assisted diagnosis and treatment ecosystem, characterized in that... The method includes the following steps: Obtain multi-source datasets from various medical institutions and extract the basal patterns from each multi-source dataset; The set of motif nodes for each motif pattern is determined, and rare subgraph mining is performed on each multi-source dataset to obtain a global rare subgraph set. Based on the global rare subgraph set, the initial graph structure of each multi-source dataset, and the set of all motif nodes, a multi-structure fusion hypergraph of each multi-source dataset is constructed. Each multi-structure fusion hypergraph is converted into a corresponding hypergraph neural network model. The cross-correlation matrix of each hypergraph neural network model in the public unlabeled TCM dataset is determined. Then, the cross-correlation matrix is ​​de-correlation-constrained to obtain multiple collaborative distillation losses when co-aligning heterogeneous data. Adversarial perturbations are applied to all cross-correlation matrices to generate multiple adversarial cooperative losses. Federated optimization is then performed on all hypergraph neural network models based on all adversarial cooperative losses and all cooperative distillation losses. Diagnostic data of the target patient is collected and a real-time hypergraph is constructed. The real-time hypergraph is then input into the corresponding hypergraph neural network model after federated optimization to obtain the confidence level of the rare syndrome type of the target patient.

2. The method as described in claim 1, characterized in that, Obtaining multi-source datasets from various medical institutions and extracting the basalt patterns from each dataset specifically includes: Obtain multi-source datasets and preset sequence templates from various medical institutions; Each multi-source dataset is transformed into an initial graph structure containing multiple nodes and edges; Based on the aforementioned sequence template, subgraph isomorphism detection is performed on all initial graph structures to obtain the sequence patterns of each multi-source dataset.

3. The method as described in claim 1, characterized in that, Determining the set of motif nodes for each motif pattern specifically includes: Deduplication of nodes is performed on each motif pattern, and nodes that appear repeatedly in the same motif pattern are merged into a single node. Verify the topology of the merged node set in each basal sequence pattern; The set of nodes that pass the verification in each motif pattern is taken as the motif node set of the corresponding motif pattern.

4. The method as described in claim 1, characterized in that, Rare subgraph mining is performed on various multi-source datasets to obtain a global rare subgraph set, specifically including: Rare subgraphs are enumerated for the initial graph structure of each multi-source dataset to generate multiple candidate subgraphs; Determine the support of each candidate subgraph in the corresponding initial graph structure; All candidate subgraphs are filtered based on all support scores; Structural similarity clustering is performed on all candidate subgraphs after screening to obtain a global rare subgraph set.

5. The method as described in claim 1, characterized in that, The construction of a multi-structure fusion hypergraph for each multi-source dataset based on the aforementioned global rare subgraph set, the initial graph structure of each multi-source dataset, and the set of all motif nodes specifically includes: All nodes in the initial graph structure of each multi-source dataset are used as vertices of the hypergraph; Construct multiple rare subgraph superedges based on the set of nodes contained in each subgraph in the global rare subgraph set; Construct corresponding ordering hyperedges based on each ordering node set; Map each edge in the initial graph structure of each multi-source dataset to a pair of superedges containing two nodes; Determine the set of neighborhood nodes that each hypergraph vertex in the initial graph structure of each multi-source dataset can reach within a set step size, and then construct the hop count neighborhood hyperedge corresponding to that vertex. By fusing all rare subgraph hyperedges, radix hyperedges, paired edge hyperedges, and hop-count neighborhood hyperedges, a multi-structure fused hypergraph is generated for each multi-source dataset.

6. The method as described in claim 1, characterized in that, Determining the cross-correlation matrix of feature associations for each hypergraph neural network model in a public unlabeled TCM dataset specifically includes: Obtain public unlabeled TCM datasets; The public unlabeled TCM dataset is input into each hypergraph neural network model, and the corresponding output feature vectors are extracted. The average eigenvectors of all features are averaged to obtain the global average eigenvectors. Determine the correlation coefficients of each feature vector with the global average feature vector in each dimension, and construct the cross-correlation matrix of the feature associations of the corresponding hypergraph neural network model.

7. The method as described in claim 1, characterized in that, By applying discorrelation constraints to all cross-correlation matrices, the specific collaborative distillation losses for collaboratively aligning heterogeneous data include: Extract the diagonal elements of each cross-correlation matrix to form the corresponding homogeneous vector; Extract the off-diagonal elements of each cross-correlation matrix to construct the corresponding heterogeneous decorrelation matrix; For any cross-correlation matrix, construct a similar feature constraint term based on the corresponding similar consistent vector, and construct a dissimilar feature constraint term based on the corresponding dissimilar decorrelation matrix; The cooperative distillation loss for co-aligning heterogeneous data is determined based on the heterogeneous feature constraint term and the homogeneous feature constraint term, and the corresponding cross-correlation matrix is ​​used to determine the cooperative distillation loss.

8. A TCM digital AI-assisted diagnosis and treatment ecosystem, the ecosystem comprising a rare syndrome type recognition unit, characterized in that, The rare disease identification unit includes: The acquisition module is used to acquire multi-source datasets from various medical institutions and extract the basal sequence patterns from each multi-source dataset. The processing module is used to determine the set of motif nodes for each motif pattern, perform rare subgraph mining on each multi-source dataset to obtain a global rare subgraph set, and construct a multi-structure fusion hypergraph for each multi-source dataset based on the global rare subgraph set, the initial graph structure of each multi-source dataset, and all motif node sets. The processing module is also used to convert each multi-structure fusion hypergraph into a corresponding hypergraph neural network model, determine the cross-correlation matrix of the feature association of each hypergraph neural network model in the public unlabeled TCM dataset, and then perform decorrelation constraints on all cross-correlation matrices to obtain multiple collaborative distillation losses when collaboratively aligning heterogeneous data. The processing module is also used to apply adversarial perturbations to all cross-correlation matrices, generate multiple adversarial cooperative losses, and perform federated optimization on all hypergraph neural network models based on all adversarial cooperative losses and all cooperative distillation losses. The execution module is used to collect diagnostic data of the target patient and construct a real-time hypergraph. The real-time hypergraph is then input into the corresponding hypergraph neural network model after federated optimization to obtain the confidence level of the rare syndrome type of the target patient.

9. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing code, and the processor being configured to retrieve the code and execute the rare certificate identification method for heterogeneous multi-source data as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the rare certificate identification method for heterogeneous multi-source data as described in any one of claims 1 to 7.

Images