Robust discovery method and system for medical entity association relationship based on heterogeneous graph

Abstract

This invention discloses a robust method and system for discovering medical entity relationships based on heterogeneous graphs. The method includes acquiring medical heterogeneous graph data and defining a meta-path space based on the heterogeneous graph data and learning objectives; transforming input features into corresponding target meta-path feature spaces based on the meta-path space to obtain a set of embedding representations on different meta-paths; aggregating the embedding representations on different meta-paths according to the weight coefficients of each neighbor connected by the meta-paths; optimizing the GNN information transmission process on each meta-path using a multi-path semantic capture optimization function calculated using the Hilbert independence criterion to capture multi-path semantic information; and obtaining prediction results based on the multi-path semantic information. This invention can improve the robustness of heterogeneous graph models in the medical field.

Description

Technical Field

[0001] This invention relates to the field of biomedical technology, specifically to a robust method and system for discovering medical entity associations based on heterogeneous graphs. Background Technology

[0002] Discovering relationships between medical entities is one of the most fundamental and core tasks in the biomedical field. It can be used to uncover potential relationships between medical entities, helping researchers develop drugs and assisting doctors in diagnosis. For example, discovering drug-drug interactions helps explore the correlations between drugs and develop new drugs; discovering drug-disease interactions helps improve the efficiency of disease treatment. In particular, traditional methods for discovering relationships between medical entities rely heavily on extensive experiments, a process that is quite complex, time-consuming, and costly. Therefore, designing and developing computational methods is essential.

[0003] Existing methods primarily rely on medical heterogeneous graphs for entity relationship discovery, with data often derived from large drug or medical databases or preprocessed using text mining techniques. However, these methods typically assume that the observed medical entities and their interactions are reliable. In reality, due to experimental errors and omissions, the observed medical data may not be entirely accurate and reliable, potentially threatening the robustness of the discovered medical entity relationships.

[0004] Intelligent diagnostic and treatment systems are medical aids based on artificial intelligence technology, used to assist doctors and other medical professionals in disease diagnosis and treatment. By integrating and analyzing patients' medical data and medical history, combined with advanced algorithms and models, the system generates precise diagnostic and treatment plans tailored to individual patients, thereby improving medical efficiency and quality and reducing misdiagnosis and missed diagnosis rates. Intelligent diagnostic and treatment systems can be applied to various disease areas, such as cardiovascular and cerebrovascular diseases, oncology, and neurological diseases. They also support the processing and analysis of various medical images and signals, such as CT, MRI, ultrasound, and electronic physiological signals. Furthermore, the system can provide telemedicine services, offering convenient medical consultation and treatment services to doctors and patients.

[0005] Due to the special nature of the medical field, the diagnostic and treatment system must be able to output correct results stably. In related technologies, the patent application document with publication number CN114883001A proposes to expand the current electronic medical record of the patient to be predicted, and construct an electronic medical record heterogeneous graph based on the expanded data; on the electronic medical record heterogeneous graph, the electronic medical record is embedded based on the learning representation of the meta-path, and the meta-path neighbor nodes are aggregated through the attention mechanism; finally, the disease prediction of the patient to be predicted is realized. However, in actual application scenarios, the optimization of the intelligent diagnostic and treatment model based on heterogeneous graphs still faces huge challenges: (1) During the graph propagation process, the model's task performance is very easily affected by the subtle perturbations on the nodes or edges. If the data provider provides low-quality or incorrect original data or model parameters, it will lead to a decrease in model quality; (2) For heterogeneous graph neural networks, the complexity of heterogeneous graphs makes it possible for adjacent nodes to contain noise information that is unrelated to the target node. Through the propagation aggregation mechanism, this unrelated information can be easily propagated to adjacent and higher-order nodes, thereby interfering with the embedding quality of the heterogeneous graph.

[0006] The paper "A Hybrid Compression Method for Neural Networks Based on Information Bottleneck, Computer Application Research, Excellence, etc." proposes a hybrid compression scheme for neural networks based on the information bottleneck theory. This compression scheme based on the information bottleneck theory (IB) reduces model parameters by using a derived loss function, mainly relying on mutual information calculation based on the estimated probability density. Its goal is to find the model pruning position through the loss function based on IB, that is, to compress the model size and reduce the computational and storage pressure in engineering applications. Summary of the Invention

[0007] The technical problem to be solved by this invention is how to address the system security issues caused by the low robustness of heterogeneous graph models in the medical field.

[0008] The present invention solves the above-mentioned technical problems through the following technical means:

[0009] A robust method for discovering medical entity relationships based on heterogeneous graphs is proposed, the method comprising:

[0010] Acquire medical heterogeneous graph data, and define a meta-path space based on the heterogeneous graph data and the learning objective;

[0011] Based on the meta-path space, the input features are transformed to the corresponding target meta-path feature space to obtain a set of embedding representations on different meta-paths;

[0012] Based on the weight coefficient of each neighbor connected by the meta-path, the embedding representations on different meta-paths are aggregated;

[0013] The multi-path semantic capture optimization function, calculated using the Hilbert independence criterion, optimizes the GNN information transmission process on each meta-path to capture multi-path semantic information.

[0014] The prediction result is obtained based on the multi-path semantic information.

[0015] Furthermore, the step of acquiring medical heterogeneous graph data and defining a meta-path space based on the heterogeneous graph data and the learning objective includes:

[0016] Acquire medical heterogeneous graph data and establish an adjacency matrix based on the relationships between entities in the heterogeneous graph data. The entities include four types: patients, diseases, symptoms, and drugs.

[0017] A multimodal medical heterogeneous graph network is constructed based on the adjacency matrix. , V A set of nodes representing different types of entities. E A collection of relationships between different types of entities;

[0018] Based on the characteristics of the medical heterogeneous graph data and the model learning objective, multiple meta-path aggregation nodes are selected, and based on the selected multiple meta-paths... Define metapath space .

[0019] Further, based on the meta-path space, the input features are transformed to the corresponding target meta-path feature space to obtain a set of embedding representations on different meta-paths, including:

[0020] The input features are transformed to the corresponding target metapath feature space to obtain a set of embedding representations on different metapaths, as expressed by the formula:

[0021]

[0022] In the formula: X Indicates input features, This represents the features after transformation to the corresponding target metapath feature space. Metapath The transformation matrix in the diagram.

[0023] Further, the aggregation of embedding representations on different meta-paths based on the weight coefficients of each neighbor connected by the meta-path includes:

[0024] Self-attention is used to learn the weights of each neighbor connected by the meta-path, and the weight coefficients of each neighbor are obtained through normalized attention values.

[0025] The attention mechanism between nodes is used K times repeatedly. Based on the weight coefficient of each neighbor connected by the meta-path, the embedding representations on different meta-paths are aggregated, as shown in the formula:

[0026]

[0027] In the formula, || denotes the cascading operation of embedding. Metapath The weight, This represents a set containing all metapaths. The metapath learned by the self-attention mechanism between nodes. Embedded features on This represents the metapath obtained by repeating the inter-node attention mechanism K times and concatenating its features. Embedded features on.

[0028] Furthermore, the formula for the multi-path semantic capture optimization function is expressed as follows:

[0029]

[0030]

[0031]

[0032]

[0033] In the formula: The semantic capture loss function represents the relationship between input features, ground truth, and hidden features, and is used to impose semantic embeddings and constraints between the input and ground truth for each meta-path; Indicates the number of hidden layers and the predicted label. Consider it the last hidden layer , Metapath The first j Features of each hidden layer This indicates the HSIC dependency between input features and hidden features. This represents the HSIC dependency between the true value and the hidden features. Indicates the trade-off coefficient. X Indicates input features, Y Represents the actual value; This represents the semantic capture loss function between hidden layers, which is used to add additional constraints between hidden layers on each meta-path to capture the relationships between hidden layers. The tradeoff coefficients representing the loss function. The HSIC value represents the degree of dependence between adjacent hidden layer features. Metapath The first j+ Features of a hidden layer.

[0034] Furthermore, for a given joint probability of variables U and V ,sampling The formula for calculating the Hilbert independence criterion is as follows:

[0035]

[0036] In the formula: This represents the degree of dependency between two variables U and V in different spaces. tr () denotes the trace of the matrix. N Let represent the row number of the matrix, where . It is the identity matrix. A column vector containing all elements equal to 1. Denotes the kernel matrix, and its given... i lines and j The elements of the column are ,in This represents the kernel function bandwidth, primarily serving a smoothing function. and The definition is similar.

[0037] Furthermore, this invention also proposes a robust discovery system for medical entity associations based on heterogeneous graphs, the system comprising:

[0038] The acquisition module is used to acquire medical heterogeneous graph data and define a meta-path space based on the heterogeneous graph data and the learning objective;

[0039] The feature transformation module is used to transform the input features to the corresponding target meta-path feature space based on the meta-path space, so as to obtain a set of embedding representations on different meta-paths;

[0040] The feature aggregation module is used to aggregate the embedding representations on different meta-paths based on the weight coefficients of each neighbor connected by the meta-path.

[0041] The semantic capture module is used to optimize the GNN information transmission process on each meta-path by using the multi-path semantic capture optimization function calculated using the Hilbert independence criterion, thereby capturing multi-path semantic information.

[0042] The prediction module is used to obtain prediction results based on the multi-path semantic information.

[0043] Furthermore, the feature aggregation module includes:

[0044] The weight coefficient calculation unit is used to learn the weight of each neighbor connected by the meta-path using self-attention, and obtain the weight coefficient of each neighbor through the normalized attention value.

[0045] The aggregation unit is used to repeatedly apply the attention mechanism between nodes K times. Based on the weight coefficient of each neighbor connected by the meta-path, it aggregates the embedding representations on different meta-paths, as expressed by the formula:

[0046]

[0047] In the formula, || denotes the cascading operation of embedding. Metapath The weight, This represents a set containing all metapaths. The metapath learned by the self-attention mechanism between nodes. Embedded features on This represents the metapath obtained by repeating the inter-node attention mechanism K times and concatenating its features. Embedded features on.

[0048] Furthermore, the formula for the multi-path semantic capture optimization function is expressed as follows:

[0049]

[0050]

[0051]

[0052]

[0053] In the formula: The semantic capture loss function represents the relationship between input features, ground truth, and hidden features, and is used to impose semantic embeddings and constraints between the input and ground truth for each meta-path; Indicates the number of hidden layers and the predicted label. Consider it the last hidden layer , Metapath The first j Features of each hidden layer This indicates the HSIC dependency between input features and hidden features. This represents the HSIC dependency between the true value and the hidden features. Indicates the trade-off coefficient. X Indicates input features, Y Represents the actual value; This represents the semantic capture loss function between hidden layers, which is used to add additional constraints between hidden layers on each meta-path to capture the relationships between hidden layers. The tradeoff coefficients representing the loss function. The HSIC value represents the degree of dependence between adjacent hidden layer features. Metapath The first j+ Features of a hidden layer.

[0054] Furthermore, the formula for calculating the Hilbert independence criterion is as follows:

[0055]

[0056]

[0057] In the formula: This represents the degree of dependency between two variables U and V in different spaces. tr () denotes the trace of the matrix. N Let represent the row number of the matrix, where . It is the identity matrix. A column vector containing all elements equal to 1. and Denotes the kernel matrix, and its given... i lines and j The elements of the column are ,in This represents the kernel function bandwidth.

[0058] The advantages of this invention are:

[0059] (1) In order to extract useful information to embed nodes to the maximum extent, this invention uses the Hilbert-Schmidt Independence Criterion (HSIC) to constrain the aggregated information during feature aggregation based on the information bottleneck in the metapath, thereby limiting the dependency between different propagation layers, thus guiding the propagation of information that has a positive impact on the target node, and using it to measure the dependency between different metapath semantic spaces, thereby integrating accurate and comprehensive node information, thus effectively improving the embedding quality of heterogeneous graphs and improving the robustness of the model in practical applications.

[0060] By using the Hilbert–Schmidt Independence Criterion (HSIC) to calculate the dependence between two features during graph neural network training, and by using a loss function to guide the model's learning for different meta-paths and semantic information of each layer of the network, the robustness enhancement effect of removing useless information and capturing key features is achieved. Compared with mutual information, HSIC does not require estimation of probability density, making it more accurate in calculation and simpler to implement.

[0061] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating a robust discovery method for medical entity associations based on heterogeneous graphs according to an embodiment of the present invention.

[0063] Figure 2 This is a block diagram illustrating the principle of enhancing the robustness of intelligent diagnosis and treatment based on medical heterogeneity graphs in this invention embodiment.

[0064] Figure 3 This is a schematic diagram of the structure of a robust discovery system for medical entity associations based on heterogeneous graphs in an embodiment of the present invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0066] like Figure 1 As shown, the first embodiment of the present invention proposes a robust method for discovering medical entity associations based on heterogeneous graphs, the method comprising the following steps:

[0067] S10. Obtain medical heterogeneous graph data, and define a meta-path space based on the heterogeneous graph data and the learning objective;

[0068] S20. Based on the meta-path space, transform the input features to the corresponding target meta-path feature space to obtain a set of embedding representations on different meta-paths;

[0069] It should be noted that the input features here refer to the original features of the nodes in the heterogeneous graph network.

[0070] S30. Aggregate the embedding representations on different meta-paths based on the weight coefficients of each neighbor connected by the meta-path;

[0071] S40. The multi-path semantic capture optimization function calculated using the Hilbert independence criterion optimizes the GNN information transmission process on each meta-path to capture multi-path semantic information.

[0072] S50. Based on the multi-path semantic information, the prediction result is obtained.

[0073] This embodiment considers the effectiveness of aggregating information between adjacent nodes in heterogeneous graphs under real-world conditions, as well as the information loss during semantic fusion between meta-paths. In this context, to maximize the extraction of useful information embedded in nodes, the Hilbert-Schmidt independence criterion is used within meta-paths to limit dependencies between different propagation layers based on information bottlenecks, thereby guiding the propagation of information that positively impacts the target node. Furthermore, the Hilbert-Schmidt independence criterion is used between meta-paths to measure the dependencies between the semantic spaces of different meta-paths, thus integrating accurate and comprehensive node information. Through these two parts, the heterogeneous graph model can effectively improve the embedding quality of heterogeneous graphs and enhance the model's robustness in practical applications.

[0074] In one embodiment, step S10, which involves acquiring medical heterogeneous graph data and defining a meta-path space based on the heterogeneous graph data and the learning objective, specifically includes the following steps:

[0075] S11. Obtain medical heterogeneous graph data and establish an adjacency matrix based on the relationships between entities in the heterogeneous graph data. The entities include four types: patients, diseases, symptoms, and drugs.

[0076] It should be noted that after obtaining the medical heterogeneous graph data in this embodiment, the format of medical data with different metadata formats in different fields is unified, and then an adjacency matrix is ​​established based on the relationship between the entities.

[0077] S12. Construct a multimodal medical heterogeneous graph network based on the adjacency matrix. , V A set of nodes representing different types of entities. E A collection of relationships between different types of entities;

[0078] It should be noted that this embodiment is based on four types of entities: patient, disease, symptom, and drug. By relying on the mapping of entities and their relationships, a multimodal medical heterogeneous graph can be constructed. ,in This represents a set of nodes (patient set, disease set, symptom set, drug set). It is a set of relationships (patient-disease, disease-symptom, symptom-drug, drug-drug, ...).

[0079] S13. Based on the characteristics of the medical heterogeneous graph data and the model learning objective, select multiple meta-path aggregation nodes, and based on the selected multiple meta-paths... Define metapath space .

[0080] In one embodiment, since heterogeneous nodes have diverse original features in different feature spaces, it is necessary to perform node feature space transformation based on metapaths. Step S20: Based on the metapath space, the input features are transformed to the corresponding target metapath feature space to obtain a set of embedding representations on different metapaths, specifically including:

[0081] The input features are transformed to the corresponding target metapath feature space to obtain a set of embedding representations on different metapaths, as expressed by the formula:

[0082]

[0083] In the formula: X Indicates input features, This represents the features after transformation to the corresponding target metapath feature space. Metapath The transformation matrix in the diagram.

[0084] In one embodiment, step S30: aggregating the embedding representations on different meta-paths based on the weight coefficients of each neighbor connected by the meta-path, specifically includes the following steps:

[0085] S31. Use self-attention to learn the weights of each neighbor connected by the meta-path, and obtain the weight coefficients of each neighbor through normalized attention values.

[0086] Specifically, by transforming the feature space, a set of specific embeddings on different meta-paths are obtained. Then, self-attention is used to learn the weights of each neighbor connected by its meta-path, and the weight coefficients are obtained through normalized attention values. For a pair of nodes on a meta-path ( m , n The attention coefficient is calculated as follows:

[0087]

[0088] In the formula: LR It is the activation function LeakyReLU, where exp() represents the power of e. Metapath Attention vectors of all nodes. Indicates meta-path The adjacent nodes, Represents node pairs ( m , n Attention coefficient, This indicates the transformation of node m to the metapath. Features after feature space This represents the transformation of node n to the metapath. Features after feature space This indicates the transformation of node j to the metapath. Features after feature space Metapath The weight matrix on, i Indicates the number of paths. j Represents nodes on the path. It represents a real number space of 1×2D dimensions.

[0089] S32. Apply multi-head attention, repeat the attention mechanism between nodes K times, and aggregate the embedding representations on different meta-paths based on the weight coefficients of each neighbor connected by the meta-path. The formula is as follows:

[0090]

[0091] In the formula, || denotes the cascading operation of embedding. Metapath The weight, This represents a set containing all metapaths. The metapath learned by the self-attention mechanism between nodes. Embedded features on This represents the metapath obtained by repeating the inter-node attention mechanism K times and concatenating its features. Embedded features on.

[0092] In one embodiment, such as Figure 2 As shown, in step S40, the formula for the multi-path semantic capture optimization function is expressed as follows:

[0093]

[0094]

[0095]

[0096]

[0097] In the formula: The semantic capture loss function represents the relationship between input features, ground truth, and hidden features, and is used to impose semantic embeddings and constraints between the input and ground truth for each meta-path; Indicates the number of hidden layers and the predicted label. Consider it the last hidden layer , Metapath The first j Features of each hidden layer This indicates the HSIC dependency between input features and hidden features. This represents the HSIC dependency between the true value and the hidden features. Indicates the trade-off coefficient. X Indicates input features, Y Represents the actual value; This represents the semantic capture loss function between hidden layers, which is used to add additional constraints between hidden layers on each meta-path to capture the relationships between hidden layers. The tradeoff coefficients representing the loss function. The HSIC value represents the degree of dependence between adjacent hidden layer features. Metapath The first j+ Features of a hidden layer.

[0098] In one embodiment, the Hilbert independence criterion is calculated as follows:

[0099]

[0100] In the formula: This represents the degree of dependency between two variables U and V in different spaces. tr () denotes the trace of the matrix. N Let represent the row number of the matrix, where . It is the identity matrix. A column vector containing all elements equal to 1. Denotes the kernel matrix, and its given... i lines and j The elements of the column are ,in This represents the kernel function bandwidth, primarily serving a smoothing function. and These represent sampled values ​​of a set of variables U.

[0101] This embodiment takes into account the problem of information loss during the information propagation process in multi-layer networks. The model may be interfered with by bad nodes or edges during the learning process, deviating from the goal of supervised learning in the wrong direction. Therefore, based on the information bottleneck theory and considering the interrelationship between multiple meta-paths, the objective functions on each path are fused to obtain a multi-path semantic capture method. During graph propagation, the Hilbert independence criterion is used as the loss function to capture the dependencies between the layers of the neural network and compress the noise information, effectively constraining the purity of information aggregation.

[0102] In addition, such as Figure 3 As shown, the second embodiment of the present invention proposes a robust discovery system for medical entity associations based on heterogeneous graphs, the system comprising:

[0103] The acquisition module 10 is used to acquire medical heterogeneous graph data and define a meta-path space based on the heterogeneous graph data and the learning target;

[0104] The feature transformation module 20 is used to transform the input features to the corresponding target meta-path feature space based on the meta-path space, so as to obtain a set of embedding representations on different meta-paths;

[0105] The feature aggregation module 30 is used to aggregate the embedding representations on different meta-paths based on the weight coefficients of each neighbor connected by the meta-path.

[0106] The semantic capture module 40 is used to optimize the GNN information transmission process on each meta-path using the multi-path semantic capture optimization function calculated by the Hilbert independence criterion, thereby capturing multi-path semantic information.

[0107] The prediction module 50 is used to obtain the prediction result based on the multi-path semantic information.

[0108] In this embodiment, based on the information bottleneck, the Hilbert-Schmidt independence criterion is used in the metapath to limit the dependencies between different propagation layers, thereby guiding the propagation of information that has a positive impact on the target node. In addition, the dependencies between the semantic spaces of different metapaths are measured, thereby integrating accurate and comprehensive node information. This allows the heterogeneous graph model to effectively improve the embedding quality of heterogeneous graphs and enhance the robustness of the model in practical applications.

[0109] In one embodiment, the acquisition module 10 includes:

[0110] The data acquisition unit is used to acquire medical heterogeneous graph data and establish an adjacency matrix based on the relationships between entities in the heterogeneous graph data. The entities include four types: patients, diseases, symptoms, and drugs.

[0111] The graph network construction unit is used to construct a multimodal medical heterogeneous graph network based on the adjacency matrix. , V A set of nodes representing different types of entities. E A collection of relationships between different types of entities;

[0112] The meta-path space definition unit is used to select multiple meta-path aggregation nodes based on the characteristics of the medical heterogeneity graph data and the model learning objective, and to define the selected meta-paths. Define metapath space .

[0113] In one embodiment, the feature transformation module 20 is specifically used to: transform the input features to the corresponding target metapath feature space to obtain a set of embedding representations on different metapaths, expressed by the formula:

[0114]

[0115] In the formula: X Indicates input features, This represents the features after transformation to the corresponding target metapath feature space. Metapath The transformation matrix in the diagram.

[0116] In one embodiment, the feature aggregation module 30 includes:

[0117] The weight coefficient calculation unit is used to learn the weight of each neighbor connected by the meta-path using self-attention, and obtain the weight coefficient of each neighbor through the normalized attention value.

[0118] The aggregation unit is used to repeatedly apply the attention mechanism between nodes K times. Based on the weight coefficient of each neighbor connected by the meta-path, it aggregates the embedding representations on different meta-paths, as expressed by the formula:

[0119]

[0120] In the formula, || denotes the cascading operation of embedding. Metapath The weight, This represents a set containing all metapaths. The metapath learned by the self-attention mechanism between nodes. Embedded features on This represents the metapath obtained by repeating the inter-node attention mechanism K times and concatenating its features. Embedded features on.

[0121] In one embodiment, the formula for the multi-path semantic capture optimization function used by the semantic capture module 40 is expressed as follows:

[0122]

[0123]

[0124]

[0125]

[0126] In the formula: The semantic capture loss function represents the relationship between input features, ground truth, and hidden features, and is used to impose semantic embeddings and constraints between the input and ground truth for each meta-path; Indicates the number of hidden layers and the predicted label. Consider it the last hidden layer , Metapath The first j Features of each hidden layer This indicates the HSIC dependency between input features and hidden features. This represents the HSIC dependency between the true value and the hidden features. Indicates the trade-off coefficient. X Indicates input features, Y Represents the actual value; This represents the semantic capture loss function between hidden layers, which is used to add additional constraints between hidden layers on each meta-path to capture the relationships between hidden layers. The tradeoff coefficients representing the loss function. The HSIC value represents the degree of dependence between adjacent hidden layer features. Metapath The first j+ Features of a hidden layer.

[0127] It should be noted that other embodiments or implementation methods of the robust discovery system for medical entity associations based on heterogeneous graphs described in this invention can refer to the above-described method embodiments, and will not be repeated here.

[0128] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0129] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0130] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A robust method for discovering medical entity relationships based on heterogeneous graphs, characterized in that, The method includes: Acquiring medical heterogeneous graph data and defining a meta-path space based on the heterogeneous graph data and learning objectives includes acquiring medical heterogeneous graph data and establishing an adjacency matrix based on the relationships between entities in the heterogeneous graph data. The entities include four types: patients, diseases, symptoms, and drugs. A multimodal medical heterogeneous graph network is constructed based on the adjacency matrix. , V A set of nodes representing different types of entities. E This represents a set of relationships between different types of entities; based on the characteristics of the medical heterogeneous graph data and the model learning objective, multiple meta-path aggregation nodes are selected, and based on the selected multiple meta-paths... Define metapath space ; Based on the meta-path space, the input features are transformed to the corresponding target meta-path feature space to obtain a set of embedding representations on different meta-paths; Based on the weight coefficient of each neighbor connected by the meta-path, the embedding representations on different meta-paths are aggregated; The multi-path semantic capture optimization function, calculated using the Hilbert independence criterion, optimizes the GNN information transmission process on each meta-path to capture multi-path semantic information. Based on the multi-path semantic information, the prediction result is obtained; The formula for the multi-path semantic capture optimization function is as follows: In the formula: The semantic capture loss function represents the semantic embedding between input features, ground truth, and hidden layer features, and is used to impose semantic embeddings and constraints between the input and ground truth for each meta-path; Indicates the number of hidden layers and the predicted label. Consider it the last hidden layer , Metapath The first j Features of each hidden layer This indicates the HSIC dependency between input features and hidden layer features. This represents the HSIC dependency between the true value and the hidden layer features. Indicates the trade-off coefficient. X Indicates input features, Y Represents the actual value; This represents the semantic capture loss function between hidden layers, which is used to add additional constraints between hidden layers on each meta-path to capture the relationships between hidden layers. The weighting coefficients of the loss function are represented by the following: The HSIC value represents the degree of dependence between features of adjacent hidden layers. Metapath The first j+ Features of a hidden layer.

2. The robust discovery method for medical entity associations based on heterogeneous graphs as described in claim 1, characterized in that, The process of transforming the input features to the corresponding target meta-path feature space based on the meta-path space yields a set of embedding representations on different meta-paths, including: The input features are transformed to the corresponding target metapath feature space to obtain a set of embedding representations on different metapaths, as expressed by the formula: In the formula: X Indicates input features, This represents the features after transformation to the corresponding target metapath feature space. Metapath The transformation matrix in the diagram.

3. The robust discovery method for medical entity associations based on heterogeneous graphs as described in claim 1, characterized in that, The step of aggregating the embedding representations on different meta-paths based on the weight coefficients of each neighbor connected by the meta-path includes: Self-attention is used to learn the weights of each neighbor connected by the meta-path, and the weight coefficients of each neighbor are obtained through normalized attention values. The attention mechanism between nodes is used K times repeatedly. Based on the weight coefficient of each neighbor connected by the meta-path, the embedding representations on different meta-paths are aggregated, as shown in the formula: In the formula, || denotes the cascading operation of embedding. Metapath The weight, This represents a set containing all metapaths. The metapath learned by the self-attention mechanism between nodes. Embedded features on This represents the metapath obtained by repeating the inter-node attention mechanism K times and concatenating its features. Embedded features on.

4. The robust discovery method for medical entity associations based on heterogeneous graphs as described in claim 1, characterized in that, The formula for calculating the Hilbert independence criterion is as follows: In the formula: This represents the degree of dependence between two variables U and V in different spaces. tr () denotes the trace of the matrix. N Indicates the number of rows in the matrix. , It is the identity matrix. A column vector containing all elements equal to 1. and Represents the kernel matrix.

5. A robust system for discovering medical entity relationships based on heterogeneous graphs, characterized in that, The system includes: An acquisition module is used to acquire medical heterogeneity graph data and define a meta-path space based on the heterogeneity graph data and the learning objective. The acquisition module includes: The data acquisition unit is used to acquire medical heterogeneous graph data and establish an adjacency matrix based on the relationships between entities in the heterogeneous graph data. The entities include four types: patients, diseases, symptoms, and drugs. The graph network construction unit is used to construct a multimodal medical heterogeneous graph network based on the adjacency matrix. , V A set of nodes representing different types of entities. E A collection of relationships between different types of entities; The meta-path space definition unit is used to select multiple meta-path aggregation nodes based on the characteristics of the medical heterogeneity graph data and the model learning objective, and to define the selected meta-paths. Define metapath space ; The feature transformation module is used to transform the input features to the corresponding target meta-path feature space based on the meta-path space, so as to obtain a set of embedding representations on different meta-paths; The feature aggregation module is used to aggregate the embedding representations on different meta-paths based on the weight coefficients of each neighbor connected by the meta-path. The semantic capture module is used to optimize the GNN information transmission process on each meta-path by using the multi-path semantic capture optimization function calculated using the Hilbert independence criterion, thereby capturing multi-path semantic information. The prediction module is used to obtain the prediction result based on the multi-path semantic information; The formula for the multi-path semantic capture optimization function is as follows: In the formula: The semantic capture loss function represents the semantic embedding between input features, ground truth, and hidden layer features, and is used to impose semantic embeddings and constraints between the input and ground truth for each meta-path; Indicates the number of hidden layers and the predicted label. Consider it the last hidden layer , Metapath The first j Features of each hidden layer This indicates the HSIC dependency between input features and hidden layer features. This represents the HSIC dependency between the true value and the hidden layer features. Indicates the trade-off coefficient. X Indicates input features, Y Represents the actual value; This represents the semantic capture loss function between hidden layers, which is used to add additional constraints between hidden layers on each meta-path to capture the relationships between hidden layers. The weighting coefficients of the loss function are represented by the following: The HSIC value represents the degree of dependence between features of adjacent hidden layers. Metapath The first j+ Features of a hidden layer.

6. The robust discovery system for medical entity associations based on heterogeneous graphs as described in claim 5, characterized in that, The feature aggregation module includes: The weight coefficient calculation unit is used to learn the weight of each neighbor connected by the meta-path using self-attention, and obtain the weight coefficient of each neighbor through the normalized attention value. The aggregation unit is used to repeatedly apply the attention mechanism between nodes K times. Based on the weight coefficient of each neighbor connected by the meta-path, it aggregates the embedding representations on different meta-paths, as expressed by the formula: In the formula, || denotes the cascading operation of embedding. Metapath The weight, This represents a set containing all metapaths. The metapath learned by the self-attention mechanism between nodes. Embedded features on This represents the metapath obtained by repeating the inter-node attention mechanism K times and concatenating its features. Embedded features on.

7. The robust discovery system for medical entity associations based on heterogeneous graphs as described in claim 5, characterized in that, The formula for calculating the Hilbert independence criterion is as follows: In the formula: This represents the degree of dependence between two variables U and V in different spaces. tr () denotes the trace of the matrix. N Indicates the number of rows in the matrix. , It is the identity matrix. A column vector containing all elements equal to 1. and Represents the kernel matrix.

Images