Online medical intelligent doctor guiding system based on text graph embedding

Abstract

The application discloses an online medical intelligent doctor guiding system based on text graph embedding and belongs to the technical field of medical doctor guiding. Through fine-grained medical entity recognition and ambiguity resolution on online medical consultation texts, fine-grained entities can be accurately recognized, and the problem of fine semantic granularity can be solved. Through matching of a medical dictionary and context, the ambiguity generated by the entity can be effectively eliminated. Through a dynamic updating mechanism, the latest medical knowledge can be ensured to be incorporated into the graph, and the knowledge can be prevented from being outdated. From text preprocessing to graph construction, a complete link is formed, which can provide a high-quality knowledge base for subsequent graph embedding and doctor guiding reasoning. The entity vector obtained through processing contains graph structure, deep semantics and clinical logic, and can effectively solve the defects of single mode of traditional embedding. Through a constraint loss function, the vector can be ensured to conform to the path rules of authoritative clinical guidelines, clinical logic consistency can be realized, and vectors that violate medical common sense can be avoided.

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 technical implementation in the online medical guidance scenario faces a series of unique problems and defects, such as the same symptom corresponding to multiple diseases, such as headache being a migraine, brain tumor, or high blood pressure; a large number of synonyms, abbreviations, and ICD coding variants in disease names, such as "AMI" = "acute myocardial infarction" = "heart attack"; the relationship between graph nodes has directionality, conditionality, and hierarchy, such as "fever + rash → measles", but "fever → cold", and traditional graph embedding only models undirected adjacency relationships, ignoring clinical reasoning paths; there are defects such as fine semantic granularity and strong ambiguity of medical entities, and general graph embedding is difficult to capture clinical logic. SUMMARY

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

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] The online medical intelligent guidance system based on text graph embedding comprises:

[0007] A fine-grained medical entity processing graph construction module: a medical pre-training large language model is used for fine-grained medical entity recognition and ambiguity resolution of online medical consultation text, and a dynamic medical knowledge graph containing a real-time clinical knowledge update mechanism is constructed;

[0008] 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 the clinical path constraints to design a multi-modal enhanced graph embedding model to generate entity vector representations with deep semantic and clinical logical associations;

[0009] The multi-modal input processing doctor guiding path recommendation module: according to the processed entity vector representation, combined with the multi-modal input of the user consultation, a doctor guiding path recommendation is generated through the medical pre-training large language model chain reasoning module.

[0010] Further, the online medical consultation text input by the user is collected and processed to obtain the consultation processing text;

[0011] When performing fine-grained medical entity recognition, a pre-trained medical pre-training large language model is used as a basic model; and a BIO tag system is used to identify fine-grained entities.

[0012] Further, when performing medical entity ambiguity resolution, it includes candidate entity generation, semantic similarity calculation and result selection; wherein the candidate entity with the highest similarity is selected as the final result.

[0013] 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 ability of the medical rules or the medical pre-training large language model, and stored in the Neo4j database;

[0014] The real-time updating mechanism includes timing synchronization, user feedback driving and threshold updating.

[0015] 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 the graph structure data;

[0016] 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 the model semantic vector data;

[0017] 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.

[0018] 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;

[0019] 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;

[0020] 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;

[0021] The constraint embedding module is used for converting the authoritative clinical pathway rule into a vector feature, injecting clinical logic, and containing constraint rule coding, attention weight calculation and constraint vector generation;

[0022] The fusion module is used for integrating the vectors of the three modalities of structure, semantics and constraint into a final entity vector, and contains modal weight learning, weighted fusion calculation and vector post-processing.

[0023] Further, when performing the weighted fusion calculation, the vectors of the three modalities are weighted and summed, and the expression involved is ; wherein, is the fusion vector; are the structure weight, the semantic weight and the constraint weight, respectively; are the structure vector of the entity, the normalized semantic vector and the constraint vector, respectively.

[0024] Further, the normalized semantic vector involves the expression ; wherein, is the original pre-training large language model vector of the entity v; is the L2 norm.

[0025] Further, when performing the attention weight calculation, for each entity v, all the constraint rules involved in the entity v are collected, the attention weight of each rule to the entity v is calculated, and the expression involved is:

[0026] ; wherein, exp() is a natural exponential function; is the structure vector of the entity v; is the embedding vector of the rule r; sim() is the cosine similarity, and R(v) is the rule set involved in the entity v; represents the embedding vector of all constraint rules involved in the entity v is summed up; represents the embedding vector of any one constraint rule in the rule set R(v).

[0027] Further, when performing the constraint vector generation, the weighted sum of all rule vectors of the entity v is performed to obtain the constraint vector .

[0028] Compared with the prior art, the present application has the following beneficial effects:

[0029] The present application can accurately identify fine-grained entities and solve the problem of fine semantic granularity by performing fine-grained medical entity recognition and ambiguity resolution on online medical consultation text; the ambiguity of the entity can be effectively eliminated by matching the medical dictionary with the context; the latest medical knowledge is ensured to be incorporated into the graph through a dynamic updating mechanism, avoiding outdated knowledge; from text preprocessing to graph construction, a complete link is formed, which can provide a high-quality knowledge base for subsequent graph embedding and doctor reasoning.

[0030] The present application processes the obtained entity vector, which contains graph structure, deep semantic and clinical logic, and can effectively solve the defects of traditional single modal embedding; by constraining the loss function, the vector is ensured to comply with the path rules of the authoritative clinical guidelines, realizing the consistency of clinical logic and avoiding the generation of vectors that violate medical common sense; when the generated vector is used for doctor reasoning, it can effectively improve the accuracy of disease diagnosis and department recommendation; through the above steps, the multi-modal fusion of structure, semantics and constraints is realized, which can provide high-quality entity vector basis for subsequent doctor reasoning tasks and effectively support the core functions of the system. BRIEF DESCRIPTION OF DRAWINGS

[0031] The present application will be further described below with reference to the accompanying drawings.

[0032] Figure 1 The present application is a flowchart of the operation of the online medical intelligent doctor system based on text graph embedding. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0034] As shown in Figure 1 The present application is an online medical intelligent doctor system based on text graph embedding, which includes a fine-grained medical entity processing and graph construction module, a graph embedding model construction and entity vector representation module, and a multi-modal input processing and doctor path recommendation module.

[0035] Fine-grained medical entity processing and 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 real-time clinical knowledge updating mechanism is constructed; the specific steps include:

[0036] An online medical consultation text input by a user is collected and standardized to obtain a consultation processing text; wherein the standardization processing includes but is not limited to noise cleaning, format unification, word segmentation and sentence segmentation, which are all existing conventional technical solutions, and the specific implementation steps are not described here;

[0037] When performing fine-grained medical entity recognition, a pre-trained medical pre-training large language model is used as a base model, such as BioBERT-v2;

[0038] When performing task setting, the BIO tagging system is used to identify fine-grained entities; the BIO tagging system (BIO Tagging Scheme) is a labeling specification widely used in sequence labeling tasks (such as named entity recognition, NER) in natural language processing, and the core idea is to identify the position and role of a certain "named entity" by labeling each word or subword unit with a specific label. Specifically, for example, B-symptom, I-symptom, B-disease, I-disease, B-drug, etc.; B represents that the current word is the beginning of a certain named entity; I represents that the current word belongs to a certain named entity, but is not the beginning, i.e. inside the entity; O represents that the current word does not belong to any named entity;

[0039] The key parameters for training include but are not limited to hidden layer dimension, learning rate, and training batch size;

[0040] The hidden layer dimension is specifically d=768;

[0041] The learning rate is specifically lr=2e-5;

[0042] The training batch size is specifically batch size =16; wherein the training of the base model based on the key parameters is an existing conventional technical solution, and the specific implementation steps are not described here;

[0043] The preprocessed consultation processing text is input into the trained model as input, and an annotated entity list is output; for example: [“fever 38.5℃” (symptom), “cough with yellow sputum” (symptom), “cold balm” (drug)];

[0044] When performing medical entity ambiguity resolution, candidate entity generation, semantic similarity calculation and result selection are included;

[0045] When implementing candidate entity generation, the identified entity is obtained from a medical terminology dictionary to obtain a candidate set; wherein the medical terminology dictionary is, for example, UMLS; when obtaining the candidate set, for example, the entity “cold” corresponds to the candidate: {common cold (ICD-10: J00), influenza (ICD-10: J10-J11)};

[0046] When implementing semantic similarity calculation, cosine similarity is used to measure the matching degree of the candidate entity and the context, and the expression involved is:

[0047] ; wherein, is the similarity; e is the embedding vector of the candidate medical entity, that is, the numerical vector representation corresponding to a certain candidate entity extracted from the medical terminology dictionary; c is the embedding vector of the context where the candidate entity is located, that is, the numerical vector representation of the user consultation text segment containing the entity; is the vector dot product; is the L2 norm;

[0048] When implementing result selection, the candidate entity with the highest similarity is selected as the final result;

[0049] When constructing a dynamic medical knowledge graph and updating it in real time, the graph structure uses an attribute graph model, and the node types include symptoms, diseases, drugs, and departments; the edge types include symptom-associated disease, disease-recommended drug, and disease-corresponding department.

[0050] When constructing, the resolved entity is used as a node, and the edge connection is established through medical rules or the relationship extraction capability of a medical pre-trained large language model, and is stored in a Neo4j database; medical rules, for example, symptom X is commonly associated with disease Y.

[0051] The real-time updating mechanism includes timing synchronization, user feedback-driven, and threshold updating.

[0052] Specifically, when timing synchronization is performed, the latest literature abstracts from PubMed Central are connected daily, new entities and relationships are extracted through a medical pre-trained large language model, and the graph is updated; new entities such as new drugs; PubMed Central (PMC for short) is a free, full-text, open digital archive operated by the National Center for Biotechnology Information under the National Library of Medicine.

[0053] When user feedback-driven is performed, when the user feedback indicates that the doctor's result is incorrect, the corresponding entity relationship is re-verified and corrected.

[0054] When threshold updating is performed, when the number of new entities is ≥100 or the number of new relationships is ≥500, full-index updating is performed to ensure query efficiency.

[0055] In the embodiment of the present application, by performing fine-grained medical entity recognition and ambiguity resolution on online medical consultation text, fine-grained entities can be accurately identified, and the problem of fine semantic granularity can be solved; by matching the medical dictionary with the context, the ambiguity of the entity can be effectively eliminated; by the dynamic updating mechanism, the latest medical knowledge is ensured to be incorporated into the graph, and the knowledge is prevented from being outdated; from text preprocessing to graph construction, a complete link is formed, which can provide a high-quality knowledge base for subsequent graph embedding and doctor reasoning.

[0056] The 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 and the clinical path constraints are fused, a multi-modal enhanced graph embedding model is designed, and entity vector representation with deep semantic and clinical logical association is generated; the specific steps include:

[0057] Extract the entity set and edge set from the constructed dynamic medical knowledge graph, and convert them into an adjacency matrix format to obtain graph structure data;

[0058] The entity set includes but is not limited to symptoms, diseases, drugs, departments, etc.

[0059] The edge set includes but is not limited to symptoms→associated diseases, drugs→treatment of diseases, etc.

[0060] Call the medical pre-training large language model to generate a 768-dimensional semantic vector for each entity, save it as a vector matrix, and obtain model semantic vector data; wherein the entity is, for example, acute bacterial pneumonia.

[0061] Collect the path rules in authoritative clinical guidelines and convert them into logical constraint triples or weighted rules to obtain clinical path constraint data; wherein the authoritative clinical guidelines are, for example, China Clinical Diagnosis and Treatment Guidelines; in addition, the triples are, for example, (acute pneumonia, must do, chest X-ray examination); the weighted rules are, for example, acute pneumonia→first choice antibiotic treatment, weight=0.9.

[0062] Input: constructed medical knowledge graph, medical pre-training large language model interface, and clinical guideline text.

[0063] Output: adjacency matrix, semantic vector matrix, and constraint rule set, corresponding to structure, semantics, and clinical logic, respectively.

[0064] When designing the multi-modal enhanced graph embedding model, the structure embedding module, the semantic embedding module, the constraint embedding module, and the fusion module are included.

[0065] The structure embedding module is used to extract topological correlation features between entities from the constructed medical knowledge graph, is realized by using a Graph Convolutional Network (GCN), and includes adjacency matrix preprocessing, GCN layer construction, and dimension alignment.

[0066] In the adjacency matrix preprocessing, the original adjacency matrix of the medical knowledge graph is symmetrically normalized, the degree matrix D of each node is calculated first, and then the normalized adjacency matrix is generated, and the expression involved is: ; wherein, is the normalized adjacency matrix; A is the original adjacency matrix.

[0067] In the GCN layer construction, a 2-layer or 3-layer GCN network is designed, and each layer includes the following operations:

[0068] The node vector of the current layer is input, which is initially randomly initialized or entity ID embedding;

[0069] is multiplied by the normalized adjacency matrix , to aggregate the features of the neighbor nodes;

[0070] Nonlinearity is introduced through a ReLU activation function, and a node structure vector of the next layer is output;

[0071] In the dimension alignment, the structure vector output by the GCN layer is adjusted in dimension to be consistent with the subsequent semantic vector, to ensure the compatibility of the modal fusion;

[0072] The semantic embedding module is used to introduce entity semantic features of a medical pre-trained large language model, to supplement the deep semantic information missing in the structure embedding, and includes semantic vector acquisition, vector standardization, and semantic preservation.

[0073] In the semantic vector acquisition, the medical pre-trained large language model is called to encode the text description of each entity, to generate an original semantic vector of 768 dimensions; the text description of each entity, for example, "acute bacterial pneumonia: lung inflammation caused by bacterial infection";

[0074] In the vector standardization, L2 normalization is performed on the original semantic vector, and the formula is ; wherein, is the normalized semantic vector. is the original pre-training large language model vector for entity v; the normalized semantic vector is used to eliminate the influence of vector length difference on fusion;

[0075] When performing semantic preservation, the normalized semantic vector is directly taken as the module output without additional modification, ensuring the preservation of medical professional semantics learned by the pre-training large language model;

[0076] The constraint embedding module is used to convert authoritative clinical pathway rules into vector features and inject clinical logic, including constraint rule coding, attention weight calculation, and constraint vector generation;

[0077] When performing constraint rule coding, the pathway rules in the clinical guidelines are converted into a calculable vector form; the pathway rules are, for example, “acute pneumonia must undergo chest X-ray examination”; specifically, the text description of each rule is encoded using a medical pre-training large language model to generate a rule embedding vector;

[0078] When performing attention weight calculation, for each entity v, all constraint rules related to it are collected, and the attention weight of each rule on entity v is calculated, and the expression involved is:

[0079] ; wherein exp() is the natural exponential function; is the structure vector of entity v; is the embedding vector of rule r; sim() is the cosine similarity, and R(v) is the rule set related to entity v; represents all constraint rules related to entity v performing summation; represents the embedding vector of any constraint rule in the rule set R(v), which is generated in the same way as ;

[0080] When generating the constraint vector, the weighted sum of all rule vectors of entity v is performed to obtain the constraint vector , and the clinical logic features are injected;

[0081] The fusion module is used to integrate the vectors of the three modalities of structure, semantics, and constraints into the final entity vector, including modal weight learning, weighted fusion calculation, and vector post-processing;

[0082] When performing modal weight learning, a learnable weight is assigned to each modality, the initial value is set to 0.3 for the structure weight, 0.4 for the semantic weight, and 0.3 for the constraint weight, and the weight is dynamically adjusted through model training to adapt to the requirements of downstream tasks;

[0083] When performing weighted fusion calculation, the vectors of the three modalities are weighted and summed, and the expression involved is ; wherein, is a fusion vector; are structure weight, semantic weight, constraint weight, respectively;

[0084] When performing vector post-processing, L2 normalization is performed on the fusion vector to ensure the stability and comparability of the vector, and finally a 768-dimensional multi-modal embedding vector of the entity v is output;

[0085] And the model is optimized by a multi-objective loss function, and the expression involved is:

[0086] ; wherein, are structure loss, semantic loss, constraint loss, respectively;

[0087] Among them, the structure loss is based on the output of the GCN layer, and adopts cross-entropy loss or link prediction loss. The cross-entropy loss compares the entity pair connection probability output by the GCN layer with the adjacency matrix of the original graph, and calculates the penalty for prediction errors. The link prediction loss is through negative sampling, and the similarity of the real existing edge is higher than that of the negative sample;

[0088] The semantic loss adopts cosine similarity loss or mean square error loss. The cosine similarity loss calculates the cosine similarity between the fusion vector and the semantic vector ; the mean square error loss calculates the mean square error between the fusion vector and the semantic vector ;

[0089] The constraint loss is based on the design of clinical constraint rules, such as logical constraint loss. For entity pairs that meet the rules, the vector similarity is required to be higher than the threshold τ, otherwise a loss is generated. The expression involved is: ; wherein, is the vector similarity of entity A and entity B; are multi-modal embedding vectors of entity A and entity B, respectively; max() represents the maximum value function;

[0090] The constructed model is trained to obtain a multi-modal enhanced graph embedding model; wherein, training the constructed model is an existing conventional technical solution, and the specific implementation steps will not be repeated here;

[0091] All entities in the entity set are input into the graph embedding model to generate a final 768-dimensional entity vector for each entity.

[0092] In the embodiment of the present application, the obtained entity vector processing contains atlas structure, deep semantic and clinical logic, which can effectively solve the defects of traditional single mode embedding; by constraining the loss function, it ensures that the vector conforms to the path rules of the authoritative clinical guidelines, realizes the consistency of clinical logic, and avoids generating vectors that violate medical common sense; when the generated vector is used for doctor reasoning, it can effectively improve the accuracy of disease diagnosis and department recommendation; through the above steps, the multi-modal fusion of structure, semantics and constraints is realized, which can provide high-quality entity vector basis for subsequent doctor reasoning tasks and effectively support the core functions of the system.

[0093] Multi-modal input processing doctor path recommendation module: according to the processed entity vector representation, combined with the multi-modal input of user consultation, the doctor path recommendation is generated through the medical pre-training large language model chain reasoning module. The specific steps include:

[0094] When analyzing and extracting features of the multi-modal input of user consultation, the medical pre-training large language model is used to perform semantic encoding on the user text to generate a text feature vector T, which contains information such as symptoms, disease course, and body temperature;

[0095] In addition, a medical image model such as CheXNet is used to extract image features from X-ray films to generate a picture feature vector I, which contains image information such as lung texture and inflammation area;

[0096] The text feature vector T and the picture feature vector I are fused to obtain a fused feature vector F, and the expression involved is: ; Wherein a, b, c are weight coefficients, the default values are 0.4, 0.3, 0.3, which are used to control the contribution of text, image and entity vectors; The average value of the entity vector related to the user input is calculated;

[0097] The fused feature vector F and the user text are analyzed through the medical pre-training large language model chain reasoning module to output an intent label;

[0098] The intent label, the fused feature vector F and the constraint rule set are analyzed through the medical pre-training large language model chain reasoning module to output a constraint set; the constraint rule set is obtained based on existing clinical constraint rules;

[0099] The constraint set, the intent label and the entity vector set are analyzed through the medical pre-training large language model chain reasoning module to generate a doctor path recommendation; the entity vector set is obtained based on all entity vectors;

[0100] It should be noted that the medical pre-training large language model chain reasoning module is prior art, and the reasoning analysis of the intent label, the constraint set and the doctor path recommendation can be realized through the existing technical solutions, and the specific implementation steps are not described here.

[0101] In the embodiment of the application, multi-modal coverage is realized by analyzing and extracting features from the multi-modal input of user consultation, and a full-process closed loop from user input to personalized doctor path is realized through multi-modal fusion and chain reasoning, which guarantees the rigor of medical logic and takes into account individual needs of users.

[0102] In several embodiments provided by the present application, it should be understood that the disclosed system can be implemented in other ways. For example, the above-described embodiments of the application are merely illustrative, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner.

[0103] The modules illustrated as separate components can or can not be physically separated, and the components illustrated as modules can or can not be physical modules, which can be located in one place or distributed to multiple network modules. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment of the present application.

[0104] In addition, the functional modules in each embodiment of the present application can be integrated in one processing module, or each module can exist physically, or two or more modules can be integrated in one module. The integrated module can be realized in the form of hardware or hardware plus software function module.

[0105] For those skilled in the art, it is obvious that the present application is not limited to the details of the above exemplary embodiments, and the present application can be realized in other specific forms without departing from the essential characteristics of the present application.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. An online medical intelligent doctor guiding system based on text graph embedding, characterized in that, Comprise: Fine-grained medical entity processing graph construction module: adopt medical pre-training large language model to conduct fine-grained medical entity recognition and ambiguity resolution on online medical consultation text, and construct dynamic medical knowledge graph containing real-time clinical knowledge update mechanism; Graph embedding model construction and entity vector representation module: based on the constructed dynamic medical knowledge graph, the entity semantic vector generated by the medical pre-training large language model is fused with the clinical pathway constraints, and a multi-modal enhanced graph embedding model is designed to generate entity vector representation with deep semantic and clinical logical association; including: Extract the entity set and edge set from the constructed dynamic medical knowledge graph, and convert it into an adjacency matrix format to obtain graph structure data; Call the medical pre-training large language model to generate a 768-dimensional semantic vector for each entity, save it as a vector matrix, and obtain model semantic vector data; Collect the path rules in authoritative clinical guidelines and convert them into logical constraint triples or weighted rules to obtain clinical pathway constraint data; 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 deep semantic information missing in the structure embedding, including semantic vector acquisition, vector standardization, and semantic preservation. The constraint embedding module is used to convert authoritative clinical pathway rules into vector features and inject clinical logic, including constraint rule encoding, attention weight calculation, and constraint vector generation. When calculating the attention weight, for each entity v, collect all the constraint rules related to it, calculate the attention weight of each rule to the entity v, and the expression involved is: ; wherein exp() is the natural exponential function; is the structure vector of the entity v; is the embedding vector of the rule r; sim() is the cosine similarity, and R(v) is the rule set related to the entity v; indicates that all constraint rules related to the entity v are summed up; indicates the embedding vector of any one constraint rule in the rule set R(v). When generating the constraint vector, all rule vectors for the entity v are summed up with weights ; 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. When the weighted fusion calculation is performed, the vectors of the three modalities are weighted and summed, and the expression involved is ; wherein, is a fusion vector; are a structure weight, a semantic weight, and a constraint weight, respectively; are a structure vector, a normalized semantic vector, and a constraint vector of an entity, respectively. Multi-modal input processing doctor guidance path recommendation module: according to the entity vector representation obtained by processing, combined with the multi-modal input of user consultation, generate doctor guidance path recommendation through medical pre-training large language model chain reasoning module.

2. The text graph embedding based online medical intelligent doctor guiding system according to claim 1, wherein, Collect online medical consultation text input by users and standardize it to obtain consultation processing text; When performing fine-grained medical entity recognition, use the pre-trained medical pre-training large language model as the base model; and use the BIO tag system to identify fine-grained entities.

3. The text graph embedding based online medical intelligent doctor guiding system according to claim 2, wherein, When performing medical entity ambiguity resolution, it includes candidate entity generation, semantic similarity calculation, and result selection; among them, the candidate entity with the highest similarity is selected as the final result.

4. The text graph embedding based online medical intelligent doctor guiding system according to claim 3, characterized in that, When constructing a dynamic medical knowledge graph and updating it in real time, the graph structure adopts the property graph model, the resolved entities are used as nodes, the edges are connected through the relationship extraction ability of medical rules or medical pre-training large language model, and are stored in the Neo4j database; The real-time update mechanism includes timed synchronization, user feedback driving, and threshold update.

5. The text graph embedding based online medical intelligent doctor guiding system according to claim 1, wherein, Normalized semantic vector The expression involved is ; wherein, is the original pre-trained large language model vector for entity v; is the L2 norm.

Images