Online medical intelligent medical guide system based on text graph embedding

Abstract

The invention discloses an online medical intelligent medical guide system based on text graph embedding, and belongs to the technical field of medical guide. Fine-grained medical entity recognition and ambiguity resolution are performed on an online medical consultation text, so that a fine-grained entity can be accurately recognized, and the problem of fine semantic granularity is solved; the medical dictionary is matched with the context, so that ambiguity generated by the entity can be effectively eliminated; through a dynamic updating mechanism, newest medical knowledge is guaranteed to be included in the atlas, and outdated knowledge is avoided; a complete link is formed from text preprocessing to graph construction, and a high-quality knowledge basis can be provided for subsequent graph embedding and medical guidance reasoning; the entity vector obtained by processing simultaneously comprises a map structure, deep semantics and clinical logic, so that the defect of traditional embedding of a single mode can be effectively overcome; by constraining the loss function, it is ensured that the vector conforms to a path rule of an authoritative clinical guide, clinical logic consistency is achieved, and the vector violating medical common knowledge is prevented from being generated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical guidance, in particular to an online medical intelligent guidance system based on text graph embedding. BACKGROUND

[0002] The existing online medical guidance solution based on text graph embedding usually constructs medical entities such as symptoms, diseases, departments, and drugs into a knowledge graph, uses graph embedding technologies such as Node2Vec, TransE, and GNN, and combines the natural language description input by the patient to achieve intelligent recommendation from symptoms to diseases to departments.

[0003] Although graph embedding has been widely used in social networks, recommendation systems, and financial risk control, its implementation in the online medical guidance scenario faces a series of unique problems and defects, such as multiple diseases corresponding to the same symptom, e.g., headache may be migraine, brain tumor, or hypertension; a large number of synonyms, abbreviations, and ICD coding variants in disease names, e.g., “AMI” = “acute myocardial infarction” = “heart attack”; the relationship between graph nodes has directionality, conditionality, and hierarchy, e.g., “fever + rash → measles”, but “fever → cold”, while traditional graph embedding only models undirected adjacency relationships and ignores clinical reasoning paths; and the semantic granularity of medical entities is fine and highly ambiguous, and general graph embedding cannot capture clinical logic. SUMMARY

[0004] The present application aims to provide an online medical intelligent guidance system based on text graph embedding to solve the technical problems of fine semantic granularity and high ambiguity of medical entities in existing technical solutions, and the difficulty of general graph embedding in capturing clinical logic.

[0005] The purpose of the present application can be achieved by the following technical solutions: The online medical intelligent guidance system based on text graph embedding comprises: A fine-grained medical entity processing graph construction module: a medical pre-training large language model is used to perform fine-grained medical entity recognition and ambiguity resolution on online medical consultation text, and a dynamic medical knowledge graph containing a real-time clinical knowledge update mechanism is constructed; A graph embedding model construction and entity vector representation module: based on the constructed dynamic medical knowledge graph, the entity semantic vectors generated by the medical pre-training large language model are fused with clinical path constraints to design a multi-modal enhanced graph embedding model, and entity vector representations with deep semantic and clinical logical associations are generated; A multi-modal input processing guidance path recommendation module: according to the processed entity vector representations, combined with the multi-modal input of user consultation, a guidance path recommendation is generated through a medical pre-training large language model chain reasoning module.

[0006] Further, the online medical consultation text input by the user is collected and standardized to obtain a consultation processing text. When performing fine-grained medical entity recognition, a pre-trained medical pre-training large language model is used as a base model; and a BIO tag system is used to identify fine-grained entities.

[0007] Further, when performing medical entity ambiguity resolution, candidate entity generation, semantic similarity calculation, and result selection are included; the candidate entity with the highest similarity is selected as the final result.

[0008] Further, when constructing a dynamic medical knowledge graph and updating it in real time, the graph structure uses an attribute graph model, the resolved entities are used as nodes, and the edges are connected through the relationship extraction capability of the medical rules or the medical pre-training large language model, and are stored in a Neo4j database. The real-time update mechanism includes timed synchronization, user feedback-driven, and threshold update.

[0009] Further, the entity set and edge set are extracted from the constructed dynamic medical knowledge graph and converted into an adjacency matrix format to obtain graph structure data. The medical pre-training large language model is called to generate a 768-dimensional semantic vector for each entity, which is saved as a vector matrix to obtain model semantic vector data. The path rules in the authoritative clinical guidelines are collected and converted into logical constraint triples or weighted rules to obtain clinical path constraint data.

[0010] Further, when designing a multi-modal enhanced graph embedding model, it includes a structure embedding module, a semantic embedding module, a constraint embedding module, and a fusion module. The structure embedding module is used to extract the topological association features between entities from the constructed medical knowledge graph, including adjacency matrix preprocessing, GCN layer construction, and dimension alignment. The semantic embedding module is used to introduce the entity semantic features of the medical pre-training large language model to supplement the missing deep semantic information of the structure embedding, including semantic vector acquisition, vector standardization, and semantic preservation. The constraint embedding module is used to convert the authoritative clinical path rules into vector features and inject clinical logic, including constraint rule coding, attention weight calculation, and constraint vector generation. The fusion module is used to integrate the vectors of the structure, semantic, and constraint three modalities into the final entity vector, including modal weight learning, weighted fusion calculation, and vector post-processing.

[0011] Further, when performing weighted fusion calculation, the vectors of the three modalities are weighted and summed, involving the expression ; wherein, is the fusion vector. These are structural weights, semantic weights, and constraint weights, respectively. These are the entity's structural vector, normalized semantic vector, and constraint vector, respectively.

[0012] Furthermore, the normalized semantic vector The expressions involved are ;in, The original pre-trained large language model vector for entity v; It is an L2 norm.

[0013] Furthermore, when calculating attention weights, for each entity v, all the constraint rules involved are collected, and the attention weight of each rule for entity v is calculated. The relevant expression is: Where exp() is the natural exponential function; The structure vector of entity v; Let be the embedding vector of rule r; sim() is the cosine similarity, and R(v) is the set of rules involved in entity v; This represents all constraint rules applicable to entity v. Perform summation; Represents any constraint rule in the rule set R(v). The embedding vector.

[0014] Furthermore, when generating the constraint vector, a weighted sum is performed on all the rule vectors of entity v to obtain the constraint vector. .

[0015] Compared to existing solutions, the beneficial effects achieved by this invention are: This invention achieves precise identification of fine-grained entities and resolves ambiguities in online medical consultation texts, thus addressing the issue of fine semantic granularity. By matching medical dictionaries with the context, it effectively eliminates ambiguities arising from entities. A dynamic update mechanism ensures that the graph incorporates the latest medical knowledge, preventing knowledge from becoming outdated. From text preprocessing to graph construction, a complete chain is formed, providing a high-quality knowledge foundation for subsequent graph embedding and medical guidance reasoning.

[0016] The entity vectors obtained by this invention simultaneously contain graph structure, deep semantics, and clinical logic, effectively addressing the shortcomings of traditional single-modal embedding. By constraining the loss function, it ensures that the vectors conform to the path rules of authoritative clinical guidelines, achieving consistency in clinical logic and avoiding the generation of vectors that violate medical common sense. When the generated vectors are used for medical guidance reasoning, they can effectively improve the accuracy of disease diagnosis and department recommendation. Through the protocol of the above steps, multimodal fusion of structure, semantics, and constraints is achieved, providing a high-quality entity vector foundation for subsequent medical guidance reasoning tasks and effectively supporting the core functions of the system. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart illustrating the operation of the online medical intelligent guidance system based on text graph embedding of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.

[0020] like Figure 1 As shown, the present invention is an online medical intelligent guidance system based on text graph embedding, including a fine-grained medical entity processing and graph construction module, a graph embedding model construction and entity vector representation module, and a multimodal input processing and guidance path recommendation module; Fine-grained medical entity processing and graph construction module: This module uses a pre-trained medical language model to perform fine-grained medical entity recognition and disambiguation on online medical consultation texts, constructing a dynamic medical knowledge graph that includes a real-time clinical knowledge update mechanism. Specific steps include: The system collects and standardizes online medical consultation text input by users to obtain the consultation processing text. The standardization process includes, but is not limited to, noise removal, format unification, word segmentation and sentence segmentation, all of which are existing conventional technical solutions. The specific implementation steps will not be elaborated here. When performing fine-grained medical entity recognition, a pre-trained medical pre-trained large language model is used as the base model, such as BioBERT-v2. When setting up the task, the BIO tagging scheme is used to identify fine-grained entities. The BIO tagging scheme is a tagging specification widely used in natural language processing for sequence labeling tasks (such as named entity recognition, NER). Its core idea is to identify the position and role of each word or sub-word unit within a named entity by assigning a specific tag. Specifically, examples include B-symptom, I-symptom, B-disease, I-disease, B-drug, etc.; B indicates that the current word is the beginning of a named entity; I indicates that the current word belongs to a named entity but is not the beginning, i.e., it is inside the entity; O indicates that the current word does not belong to any named entity. Key parameters for training include, but are not limited to, hidden layer dimension, learning rate, and training batch size; Specifically, the dimension of the hidden layer is d=768; The learning rate is specifically lr = 2e-5; The specific training batch size is batch. size =16; Among them, training the basic model based on key parameters is an existing conventional technical solution, and the specific implementation steps will not be elaborated here; The preprocessed consultation text is used as input to the trained model, and the labeled entity list is output; for example: ["fever 38.5℃" (symptom), "cough with yellow phlegm" (symptom), "cold medicine" (medication)]; When performing medical entity ambiguity resolution, it includes candidate entity generation, semantic similarity calculation, and result selection; When generating candidate entities, a candidate set is obtained from a medical terminology dictionary for the identified entities; the medical terminology dictionary is, for example, UMLS; when obtaining the candidate set, for example, the entity "cold" corresponds to the following candidates: {common cold (ICD-10:J00), influenza (ICD-10:J10-J11)}; When performing semantic similarity calculation, cosine similarity is used to measure the matching degree between candidate entities and their context. The relevant expression is as follows: ;in, denoted as similarity; e is the embedding vector of the candidate medical entity, which is the numerical vector representation of a candidate entity extracted from the medical terminology dictionary; c is the embedding vector of the context in which the candidate entity is located, which is the numerical vector representation of the user consultation text fragment containing the entity. It is the dot product of vectors; It is an L2 norm; When selecting the results, the candidate entity with the highest similarity will be used as the final result; When constructing and updating a dynamic medical knowledge graph in real time, the graph structure adopts an attribute graph model. Node types include symptoms, diseases, drugs, and departments; edge types include symptom-associated disease, disease-recommended drug, and disease-corresponding department. During construction, the resolved entities are used as nodes, and edge connections are established through medical rules or the relation extraction capabilities of medical pre-trained large language models, and stored in the Neo4j database; medical rules, such as symptom X is common in disease Y; The real-time update mechanism includes scheduled synchronization, user feedback-driven updates, and threshold updates; Specifically, during scheduled synchronization, the system connects daily to the latest literature abstracts from PubMed Central, extracts new entities and relationships using a medical pre-trained large language model, and updates the atlas; new entities include, for example, novel drugs; PubMed Central (PMC) is a free, full-text, open digital archive operated by the National Center for Biotechnology Information, a division of the U.S. National Library of Medicine.

[0021] When user feedback is used to drive the process, if a user reports an error in the medical guidance result, the corresponding entity relationship will be re-verified and corrected. When performing threshold updates, if ≥100 new entities or ≥500 new relationships are added, a full index update is performed to ensure query efficiency.

[0022] In this embodiment of the invention, by performing fine-grained medical entity recognition and ambiguity resolution on online medical consultation text, fine-grained entities can be accurately identified, solving the problem of fine semantic granularity; by matching the medical dictionary with the context, ambiguity generated by entities can be effectively eliminated; a dynamic update mechanism ensures that the graph incorporates the latest medical knowledge, avoiding knowledge obsolescence; and a complete link is formed from text preprocessing to graph construction, providing a high-quality knowledge foundation for subsequent graph embedding and medical guidance reasoning.

[0023] Graph Embedding Model Construction and Entity Vector Representation Module: Based on the constructed dynamic medical knowledge graph, this module integrates entity semantic vectors generated by a pre-trained medical language model with clinical pathway constraints to design a multimodal enhanced graph embedding model, generating entity vector representations that possess both deep semantic and clinical logical connections. Specific steps include: Entity sets and edge sets are extracted from the constructed dynamic medical knowledge graph and converted into adjacency matrix format to obtain graph structure data; The entity set includes, but is not limited to, symptoms, diseases, drugs, departments, etc. Edge sets include, but are not limited to, symptoms → associated diseases, drugs → treatments for diseases, etc. The medical pre-trained large language model is invoked to generate a 768-dimensional semantic vector for each entity, which is then saved as a vector matrix to obtain the model semantic vector data; among which, entities include, for example, acute bacterial pneumonia; Pathway rules from authoritative clinical guidelines are collected and transformed into logical constraint triples or weighted rules to obtain clinical pathway constraint data. Authoritative clinical guidelines include, for example, the "Chinese Clinical Practice Guidelines". Triples include, for example, (acute pneumonia, mandatory, chest X-ray); weighted rules include, for example, acute pneumonia → antibiotic treatment first, weight = 0.9. Inputs: Constructed medical knowledge graph, medical pre-trained large language model interface, clinical guideline text; Outputs: adjacency matrix, semantic vector matrix, and constraint rule set, corresponding to structure, semantics, and clinical logic, respectively; When designing a multimodal enhanced graph embedding model, it includes building a structural embedding module, a semantic embedding module, a constraint embedding module, and a fusion module; The structural embedding module is used to extract topological relationship features between entities from the constructed medical knowledge graph. It is implemented using Graph Convolutional Network (GCN) and includes adjacency matrix preprocessing, GCN layer construction, and dimension alignment. Graph Convolutional Network (GCN) is a deep learning model specifically designed for processing graph structure data. It extends the operations of traditional convolutional neural networks on regular grids (such as images) to non-Euclidean spaces, such as irregular graph structures like social networks, knowledge graphs, and molecular structures, thereby enabling representation learning and prediction of nodes, edges, or the entire graph. In the adjacency matrix preprocessing, the original adjacency matrix of the medical knowledge graph is symmetrically normalized. First, the degree matrix D of each node is calculated, and then the normalized adjacency matrix is ​​generated. The relevant expressions are as follows: ;in, A is the normalized adjacency matrix; A is the original adjacency matrix. When building the GCN layer, design a 2- or 3-layer GCN network, with each layer containing the following operations: Input the node vector of the current layer, which is initially either randomly initialized or embedded with entity IDs; With normalized adjacency matrix Multiplication aggregates the features of neighboring nodes; Nonlinearity is introduced by using the ReLU activation function to output the node structure vector of the next layer; When performing dimension alignment, the dimensions of the structure vector output by the GCN layer are adjusted to be consistent with the subsequent semantic vectors to ensure the compatibility of modality fusion; The semantic embedding module is used to introduce entity semantic features into the medical pre-trained large language model to supplement the deep semantic information missing in the structural embedding, including semantic vector acquisition, vector standardization and semantic preservation. When acquiring semantic vectors, a medical pre-trained large language model is called to encode the text description of each entity, generating a 768-dimensional original semantic vector; the text description of each entity is, for example, "acute bacterial pneumonia: lung inflammation caused by bacterial infection"; When performing vector standardization, L2 normalization is applied to the original semantic vector, using the following formula: ;in, A normalized semantic vector; The original pre-trained large language model vectors for entity v; normalized semantic vectors are used to eliminate the impact of vector length differences on fusion; When preserving semantics, the standardized semantic vector is directly used as the module output without additional modification, ensuring that the medical semantics learned by the pre-trained large language model are preserved. The constraint embedding module is used to transform authoritative clinical pathway rules into vector features and inject clinical logic, including constraint rule encoding, attention weight calculation, and constraint vector generation. When encoding constraint rules, the path rules in clinical guidelines are transformed into computable vector forms; a path rule is, for example, "acute pneumonia must be examined by chest X-ray"; specifically, the text description of each rule is encoded using a medical pre-trained large language model to generate rule embedding vectors. When calculating attention weights, for each entity v, all the constraint rules involved are collected, and the attention weight of each rule for entity v is calculated. The relevant expression is as follows: Where exp() is the natural exponential function; The structure vector of entity v; Let be the embedding vector of rule r; sim() is the cosine similarity, and R(v) is the set of rules involved in entity v; This represents all constraint rules applicable to entity v. Perform summation; Represents any constraint rule in the rule set R(v). The embedding vector, and The generation methods are consistent; When generating constraint vectors, a weighted sum is performed on all rule vectors of entity v to obtain the constraint vector. And inject clinical logic features; The fusion module is used to integrate the vectors of the three modalities of structure, semantics, and constraints into the final entity vector, which includes modal weight learning, weighted fusion calculation and vector post-processing; When learning modal weights, learnable weights are assigned to each modality. The initial values ​​are set as follows: structural weight 0.3, semantic weight 0.4, and constraint weight 0.3. The weights are dynamically adjusted through model training to adapt to the requirements of downstream tasks. When performing weighted fusion calculations, the vectors of the three modes are summed using weighted methods, and the relevant expression is as follows: ;in, This is the fusion vector; These are structural weights, semantic weights, and constraint weights, respectively. During vector post-processing, L2 normalization is performed on the fused vector to ensure the stability and comparability of the vector, and finally the 768-dimensional multimodal embedding vector of entity v is output. Furthermore, the model is optimized using a multi-objective loss function, and the relevant expression is: ;in, These are structural loss, semantic loss, and constraint loss, respectively. Among them, the structural loss is based on the output of the GCN layer and adopts either cross-entropy loss or link prediction loss. Cross-entropy loss calculates the penalty for prediction errors by comparing the connection probability of entity pairs output by the GCN layer with the adjacency matrix of the original graph. Link prediction loss uses negative sampling to make the similarity of real edges higher than that of negative samples. Semantic loss can be achieved using cosine similarity loss or mean squared error loss. Cosine similarity loss is calculated by merging the vector. With semantic vectors Cosine similarity; mean squared error loss is calculated by merging the vectors. With semantic vectors The mean of the squared differences; The constraint loss is designed based on clinical constraint rules, such as logical constraint loss. For entity pairs that conform to the rules, a loss is incurred if their vector similarity is higher than a threshold τ. The relevant expression is: ;in, The vector similarity between entity A and entity B; These are the multimodal embedding vectors for entities A and B, respectively; max() represents the maximum value function. The constructed model is trained to obtain a multimodal augmented graph embedding model; the training of the constructed model is a conventional technical solution, and the specific implementation steps are not described here. Embed the input graph of all entities in the entity set into the model to generate the final 768-dimensional entity vector for each entity.

[0024] In this embodiment of the invention, the processed entity vectors simultaneously contain graph structure, deep semantics, and clinical logic, effectively addressing the shortcomings of traditional single-modal embedding. By constraining the loss function, the vectors are ensured to conform to the path rules of authoritative clinical guidelines, achieving consistency in clinical logic and avoiding the generation of vectors that violate medical common sense. When the generated vectors are used for medical guidance reasoning, they can effectively improve the accuracy of disease diagnosis and department recommendation. Through the protocol of the above steps, multimodal fusion of structure, semantics, and constraints is achieved, providing a high-quality entity vector foundation for subsequent medical guidance reasoning tasks and effectively supporting the core functions of the system.

[0025] Multimodal input processing and medical route recommendation module: Based on the processed entity vector representations and combined with the multimodal input from user inquiries, a medical route recommendation is generated through a chain-based reasoning module of a pre-trained medical language model. Specific steps include: When parsing and extracting features from the multimodal user consultation input, the user text is semantically encoded using a medical pre-trained large language model to generate a text feature vector T, which contains information such as symptoms, disease course, and body temperature. In addition, medical imaging models, such as CheXNet, are used to extract image features from X-ray films and generate image feature vector I, which contains image information such as lung texture and inflammatory areas. The text feature vector T is fused with the image feature vector I to obtain the fused feature vector F. The relevant expression is: Where a, b, and c are all weighting coefficients, with default values ​​of 0.4, 0.3, and 0.3, respectively, used to control the contribution of text vectors, images, and entity vectors; Input the average value of the relevant entity vectors for the user; The fused feature vector F and user text are used for inference analysis through the chain inference module of the medical pre-trained large language model to output intent labels; The intent label, fused feature vector F, and constraint rule set are analyzed through the chain reasoning module of the medical pre-trained large language model to output the constraint set; the constraint rule set is obtained by combining existing clinical constraint rules. The constraint set, intent label, and entity vector set are used for reasoning analysis through the chain reasoning module of the medical pre-trained large language model to generate medical guidance route recommendations; the entity vector set is obtained based on the combination of all entity vectors. It should be noted that the chain reasoning module of the medical pre-trained large language model is an existing technology. The reasoning analysis of intent labels, constraint sets, and medical guidance path recommendations can all be achieved through existing technical solutions. The specific implementation steps will not be elaborated here.

[0026] In this embodiment of the invention, multimodal coverage is achieved by parsing and extracting features from the multimodal input of user consultations; through multimodal fusion and chain reasoning, a closed loop of the entire process from user input to personalized medical guidance path is realized, which not only ensures the rigor of medical logic, but also takes into account the individual needs of users.

[0027] In the several embodiments provided by this invention, it should be understood that the disclosed system can be implemented in other ways. For example, the embodiments of the invention described above are merely illustrative; for example, the division of modules is only a logical functional division, and there may be other division methods in actual implementation.

[0028] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0029] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or in the form of hardware plus software functional modules.

[0030] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the essential characteristics of the present invention.

[0031] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An online medical intelligent guidance system based on text-graph embedding, characterized in that, include: Fine-grained medical entity processing and graph construction module: Using a pre-trained medical language model, fine-grained medical entity recognition and ambiguity resolution are performed on online medical consultation texts to construct a dynamic medical knowledge graph that includes a real-time clinical knowledge update mechanism; Graph Embedding Model Construction and Entity Vector Representation Module: Based on the constructed dynamic medical knowledge graph, this module integrates entity semantic vectors generated by a pre-trained medical language model with clinical pathway constraints, designs a multimodal enhanced graph embedding model, and generates entity vector representations that combine deep semantics and clinical logical connections. Multimodal input processing and medical route recommendation module: Based on the processed entity vector representations and combined with the multimodal input of user consultation, the module generates medical route recommendations through the chain reasoning module of the medical pre-trained large language model.

2. The online medical intelligent guidance system based on text graph embedding according to claim 1, characterized in that, Collect and standardize the online medical consultation text input by users to obtain the consultation processing text; When performing fine-grained medical entity recognition, a pre-trained medical pre-trained large language model is used as the base model; and the BIO tag system is used to identify fine-grained entities.

3. The online medical intelligent guidance system based on text graph embedding according to claim 2, characterized in that, When performing medical entity ambiguity resolution, the process includes candidate entity generation, semantic similarity calculation, and result selection; among these, the candidate entity with the highest similarity is selected as the final result.

4. The online medical intelligent guidance system based on text graph embedding according to claim 3, characterized in that, When constructing and updating a dynamic medical knowledge graph in real time, the graph structure adopts an attribute graph model, with the resolved entities as nodes. Edge connections are established through medical rules or the relation extraction capabilities of a pre-trained medical language model, and the graph is stored in the Neo4j database. The real-time update mechanism includes timed synchronization, user feedback-driven updates, and threshold updates.

5. The online medical intelligent guidance system based on text graph embedding according to claim 4, characterized in that, Entity sets and edge sets are extracted from the constructed dynamic medical knowledge graph and converted into adjacency matrix format to obtain graph structure data; The medical pre-trained large language model is called to generate a 768-dimensional semantic vector for each entity, which is then saved as a vector matrix to obtain the model semantic vector data. Pathway rules from authoritative clinical guidelines are collected and transformed into logical constraint triples or weighted rules to obtain clinical pathway constraint data.

6. The online medical intelligent guidance system based on text graph embedding according to claim 5, characterized in that, When designing a multimodal enhanced graph embedding model, it includes building a structural embedding module, a semantic embedding module, a constraint embedding module, and a fusion module; The structural embedding module is used to extract topological association features between entities from the constructed medical knowledge graph, including adjacency matrix preprocessing, GCN layer construction, and dimension alignment. The semantic embedding module is used to introduce entity semantic features into the medical pre-trained large language model to supplement the deep semantic information missing in the structural embedding, including semantic vector acquisition, vector standardization and semantic preservation. The constraint embedding module is used to transform authoritative clinical pathway rules into vector features and inject clinical logic, including constraint rule encoding, attention weight calculation, and constraint vector generation. The fusion module is used to integrate the vectors of the three modalities of structure, semantics, and constraint into the final entity vector, including modality weight learning, weighted fusion calculation, and vector post-processing.

7. The online medical intelligent guidance system based on text graph embedding according to claim 6, characterized in that, When performing weighted fusion calculations, the vectors of the three modes are summed using weighted methods, and the relevant expression is as follows: ;in, This is the fusion vector; These are structural weights, semantic weights, and constraint weights, respectively. These are the entity's structural vector, normalized semantic vector, and constraint vector, respectively.

8. The online medical intelligent guidance system based on text graph embedding according to claim 7, characterized in that, Normalized semantic vectors The expressions involved are ;in, The original pre-trained large language model vector for entity v; It is an L2 norm.

9. The online medical intelligent guidance system based on text graph embedding according to claim 6, characterized in that, When calculating attention weights, for each entity v, all the constraint rules involved are collected, and the attention weight of each rule for entity v is calculated. The relevant expression is as follows: Where exp() is the natural exponential function; The structure vector of entity v; Let be the embedding vector of rule r; sim() is the cosine similarity, and R(v) is the set of rules involved in entity v; This represents all constraint rules applicable to entity v. Perform summation; Represents any constraint rule in the rule set R(v). The embedding vector.

10. The online medical intelligent guidance system based on text graph embedding according to claim 9, characterized in that, When generating constraint vectors, a weighted sum is performed on all rule vectors of entity v to obtain the constraint vector. .

Images